Eye guidance in reading and scene perception

EYE GUIDANCE IN READING AND SCENE PERCEPTION

This page intentionally left blank

Eye Guidance in Reading and Scene Perception

Edited by

Geoffrey Underwood Department of Psychology University of Nottingham Nottingham NG7 2RD England

1998 ELSEVIER Amsterdam - Lausanne - New York - Oxford - Shannon - Singapore - Tokyo

ELSEVffiR SCIENCE Ltd The Boulevard, Langford Lane Kidlington, Oxford OX5 1GB

Library of Congress Cataloglng-ln-PublIcatlon Data Eye guidance In reading, driving and scene perception / edited by Geoffrey Underwood. — 1st ed. p. en. Includes Index. ISBN 0-08-043361-8 1. Eye—Movements. (Geoffrey D. M.) QP477.5.E916 1998 152.14—dc21

2. Visual perception.

I. Underwood. Geoffrey

98-7314 CIP

British Library of Cataloguing in Publication Data A catalogue record from the Library of Congress has been applied for. First edition 1998 ISBN: 008 043 3618

© 1998 Elsevier Science Ltd All rights reserved. No part of this publication may be reproduced, stored in a retrieval system or transmitted in any form or by any means: electronic, electrostatic, magnetic tape, mechanical photocopying, recording or otherwise, without permission in writing from the publishers. © The paper used in this publication meets the requirements of ANSI/NISO Z39.48-1992 (Permanence of Paper). Printed in The Netherlands.

vPreface The distinguished contributors to this volume have been set the problem of describing how we know where to move our eyes next. Are they under the control of the visual information available in words and in scenes, or under the control of the meanings of those items, or is there little or no control at all? The origins of this volume are in discussions held at a meeting of the European Society for Cognitive Psychology (ESCOP) that was held in Wiirzburg in September 1996. Informally and formally, a number of us with interests in eye movements were arguing about the landing position effect in reading — an effect that, if substantiated, would provide evidence of the use of parafoveal information in eye guidance. The argument was not resolved in Wiirzburg, and so we decided to continue talking about eye guidance at a small meeting that was held in Chamonix in February 1997. Many of the contributors to this volume were present at that meeting, but the argument was not resolved in Chamonix either. We decided that the way forward was to extend the debate to encourage contributions from psychologists who had been unable to attend the meetings, but who were known to have views on the matter. And this volume is the product. The argument is still unresolved, but the problem is now clearer. ESCOP not only provided a forum for our initial arguments in Wiirzburg, but also generously supported the workshop in Chamonix. We are very grateful to the Society for giving us the opportunity to sit around a table talking about eye movements to fellow enthusiasts. Not all of the arguments in Chamonix centred upon questions of eye guidance of course, and academic rivalries saw expression in a number of competitions during our meetings. The most memorable must be the tall-story competition, which was won by an explanation of the events leading to one colleague arriving at the airport with minutes to spare before departure. This winning story involved a highly improbable overnight vigil and an illegal journey down a motorway. Chamonix also witnessed eye-movement researchers engaged in a glove-throwing competition (the best throw won by several hundred metres) and a self-mutilation-while-snow-boarding competition (the winning entry involved a party of French schoolchildren, blood spilt on the snow and, appropriately enough, an accident researcher). While our meetings have been fun, the most enjoyable aspect of the debate has been in helping this volume to completion. The contributors have responded to the original question with enthusiasm. They have developed a set of arguments, often supported with new data, that help us to understand how our eyes move in

characteristic ways. Each of the chapters has been reviewed by at least two other contributors to the volume, and authors required to respond to reviewer's comments. The authors and reviewers have kept to a strict schedule. For their readiness in making the volume a coherent series of discussions around a single question, I am pleased to acknowledge their contributions as authors and as reviewers. vi Geoffrey Underwood University of Nottingham

vii

Contents Preface Contributors

v ix

Chapter 1. Eye Guidance and Visual Information Processing: Reading, Visual Search, Picture Perception and Driving Geoffrey Underwood and Ralph Radach

1

Chapter 2. Definition and Computation of Oculomotor Measures in the Study of Cognitive Processes Albrecht Werner Inhoffand Ralph Radach Chapter 3. Eye Movements and Measures of Reading Time Simon P. Liversedge, Kevin B. Paterson and Martin J. Pickering

29

55

Chapter 4. Determinants of Fixation Positions in Words During Reading . . . 77 Ralph Radach and George W. McConkie Chapter 5. About Regressive Saccades in Reading and Their Relation to Word Identification Fran^oise Vitu, George W. McConkie and David Zola

101

Chapter 6. Word Skipping: Implications for Theories of Eye Movement Control in Reading Marc Brysbaert and Frangoise Vitu

125

Chapter 7. The Influence of Parafoveal Words on Foveal Inspection Time: Evidence for a Processing Trade-Off Alan Kennedy

149

Chapter 8. Parafoveal Pragmatics Wayne S. Murray Chapter 9. Foveal Processing Load and Landing Position Effects in Reading Simon P. Liversedge and Geoffrey Underwood

181

201

Chapter 10. Individual Differences in Reading and Eye Movement Control. . 223 viii Everatt, Mark F. Bradshaw and Paul B. Hibbard John Chapter 11. Eye Movement Control in Reading: An Overview and Model . . 243 Keith Rayner, Erik D. Reichle and Alexander Pollatsek Chapter 12. Eye Movements During Scene Viewing: An Overview John M. Henderson and Andrew Hollingworth

269

Chapter 13. Eye Guidance and Visual Search John M. Findlay and lain D. Gilchrist

295

Chapter 14. Prefixational Object Perception in Scenes: Objects Popping OutofSchemas Peter De Graef

313

Chapter 15. Functional Division of the Visual Field: Moving Masks and Moving Windows Paul M.J. van Diepen, Martien Wampers and Gery d'Ydewalle

337

Chapter 16. Film Perception: The Processing of Film Cuts Gery d'Ydewalle, Geert Desmet and Johan Van Rensbergen

357

Chapter 17. Visual Search of Dynamic Scenes: Event Types and the Role of Experience in Viewing Driving Situations Peter R. Chapman and Geoffrey Underwood

369

Chapter 18. How Much Do Novice Drivers See? The Effects of Demand on Visual Search Strategies in Novice and Experienced Drivers David E. Crundall, Geoffrey Underwood and Peter R. Chapman

395

Chapter 19. The Development of the Eye Movement Strategies of Learner Drivers Damion C. Dishart and Michael F. Land

419

Chapter 20. What the Driver's Eye Tells the Car's Brain Andrew Liu

431

Author Index Subject Index

453 461

ix

Contributors Mark F. Bradshaw Department of Psychology, University of Surrey, Guildford GU2 5XH, England Marc Brysbaert Department of Experimental Psychology, University of Ghent, Henri Dunantlaan 2, 9000 Ghent, Belgium Peter R. Chapman Department of Psychology, University of Nottingham, Nottingham NG7 2RD, England David E. Crundall Department of Psychology, University of Nottingham, Nottingham NG7 2RD, England Peter De Graef Laboratory of Experimental Psychology, University of Leuven, Tiensestraat 102, B-3000 Leuven, Belgium Geert Desmet Laboratory of Experimental Psychology, University of Leuven, Tiensestraat 102, B-3000 Leuven, Belgium Damion C. Dishart School of Biological Sciences, University of Sussex, Brighton BN1 9QG, England Gery d'Ydewalle Laboratory of Experimental Psychology, University of Leuven, Tiensestraat 102, B-3000 Leuven, Belgium John Everatt Department of Psychology, University of Surrey, Guildford GU2 5XH, England John M. Findlay Department of Psychology, University of Durham, Durham DH1 3LE, England Iain D. Gilchrist Department of Psychology, University of Durham, Durham DH1 3LE, England

xJohn M. Henderson Department of Psychology, Michigan State University, East Lansing, MI 48824-1117, US A Paul B. Hibbard Department of Psychology, University of Surrey, Guildford GU2 5XH, England Andrew Hollingworth Department of Psychology, Michigan State University, East Lansing, MI 48824-1117, USA Albrecht W. Inhoff Department of Psychology, State University of New York, Binghamton, NY 13902-6000, US A Alan Kennedy Department of Psychology, University of Dundee, Dundee DD1 4HN, Scotland Michael F. Land School of Biological Sciences, University of Sussex, Brighton BN1 9QG, England Andrew Liu Nissan Cambridge Basic Research, 4 Cambridge Center, Cambridge MA 02142, USA Simon P. Liversedge Department of Psychology, University of Durham, Durham DH1 3LE, England George W. McConkie Beckman Institute, 405 North Mathews Avenue, University of Illinois at Urbana-Champaign, IL 61801, USA Wayne S. Murray Department of Psychology, University of Dundee, Dundee DD1 4HN, Scotland Kevin B. Paterson Department of Psychology, University of Nottingham, Nottingham NG7 2RD, England Martin J. Pickering Department of Psychology, University of Glasgow, Glasgow G12 8QQ, Scotland Alexander Pollatsek Department of Psychology, University of Masachusetts, Amherst, MA 01003, USA

xi

Ralph Radach Institute of Psychology, Technical University of Aachan, Jaegerstrasse 17, 52064 Aachen, Germany Keith Rayner Department of Psychology, University of Masachusetts, Amherst, MA 01003, USA Erik D. Reichle Department of Psychology, University of Masachusetts, Amherst, MA 01003, USA Geoffrey Underwood Department of Psychology, University of Nottingham, Nottingham NG7 2RD, England Paul M. J. van Diepen Laboratory of Experimental Psychology, University of Leuven, Tiensestraat 102, B-3000 Leuven, Belgium Johan Van Rensbergen Laboratory of Experimental Psychology, University of Leuven, Tiensestraat 102, B-3000 Leuven, Belgium Frangoise Vitu Laboratoire de Psychologic Experimentale, Universite Rene Descartes, 28 rue Serpente, 75006 Paris, France Martien Wampers Laboratory of Experimental Psychology, University of Leuven, Tiensestraat 102, B-3000 Leuven, Belgium David Zola Department of Educational Psychology, University of Illinois at Urbana-Champaign, IL 61801, USA

This page intentionally left blank

CHAPTER 1 1

Eye Guidance and Visual Information Processing: Reading, Visual Search, Picture Perception and Driving Geoffrey Underwood University of Nottingham

and Ralph Radach Technical University of Aachen

Abstract What determines the location of the next fixation as we move our eyes around the display in front of us? This question is addressed with reference to the specific activities necessary for reading, for the inspection of two-dimensional pictures and line drawings, and when seeing movement in a three-dimensional display such as a video or in moving around a real world scene in a vehicle. An underlying question here is whether the non-fixated parts of the text or scene can be processed to the extent that the information that is extracted can be used to guide future fixations to useful parts of the display. When we make decisions to skip over part of the display, or when we selectively fixate informative parts or regions, the evidence may be interpreted as suggesting that the movements of our eyes are under cognitive control to an extent that is determined by constraints of task and stimulus configuration. Eye guidance appears as low-level as needed and as cognitive as possible for a given set of circumstances. The evidence and the interpretations of the evidence are the subject of debate, however, and this chapter provides an introduction to this argument.

Eye Guidance in Reading and Scene Perception/G. Underwood (Editor) © 1998 Elsevier Science Ltd. All rights reserved

2

G. Underwood & R. Radach

Introduction Our eyes do not move randomly over a page of text, but move in synchrony with the course of the ideas represented in the sentences that we read. The progressive left-to-right movements are demanded by the way that our writing system has developed, with other orthographies requiring, and receiving, patterns of top-tobottom or right-to-left fixations. But other than moving our eyes in the same direction of the words as they might be presented to us in spoken language, what else determines where we place our eyes? We do not move simply from word to word, we do not move a fixed distance with every saccadic movement, and we do not move invariably from left to right, so what is constraining selection of the next fixation? The difficulty of processing will influence the frequency with which right-to-left regressive saccadic movements are made, and we tend not to fixate the spaces between words or the very ends of words. These are examples of high-level cognitive processes and low-level visual influences, and many of the discussions in this volume debate the conditions under which one or other influence emerges. What determines whether we make progressive vs. regressive fixations, and if we do make a progressive fixation, then which of the words to the right should we fixate? If we skip the next word, is the decision made on the basis of purely visual information — the word is very short, perhaps — or on the basis of early word identification? If skipping is guided rather than random, then here is a use for parafoveal processing. Only if we can identify some of the features of an unfixated item can the guidance mechanism move our eyes towards more informative words or important objects in a scene, and so part of our task is to establish the extent to which parafoveal vision can deliver useful information. The contrast between cognitive and visual influences is also apparent in the discussions of scene perception — do we look at meaningful parts of the scene or at areas of high visual conspicuity ? And if we look at the semantically important parts of a scene early in the sequence of fixations, the question arises of how we knew that the object was important before we pointed our eyes at it. In both scene perception and in reading the contentious candidate that may be informing us where to move our eyes next is the processing of information from parts of the display that are not being fixated. Parafoveal vision may be able to deliver sufficient information to the mechanism that decides about the location of the next fixation, and many of our discussions are concerned with the question of when we can make use of this information, and when we must rely upon low level processes and pre-programmed sequences of movements. Demonstrations of parafoveal and peripheral processing are a requirement of any account of selective fixation, and occupy a number of discussions here, but before we can come to this point, in answering our question of what controls selection of the next fixation location, we must first consider a number of choices that must be made in describing eye movements themselves.

Eye guidance

3

Measurement-related and methodological issues The question of how we decide where to look next is likely to be answered with finely controlled displays and finely executed analyses of spatial and temporal data streams. We can describe the visual processing given to a visual display with any number of measures taken from our eye movement recorders, and so the question requires that we consider with care the measures that are available. This is the starting point in our collection of discussions. There are several methodological decisions to be made in the definition and computation of basic oculomotor events: the selection of an algorithm to detect saccades, the setting of cut-off values for saccade size and fixation duration, the treatment of saccadic overshoot, blink time and saccade duration. Inhoff and Radach (Chapter 2) discuss a number of measurement-related and methodological issues that so far have received little attention in the literature, although they play a crucial role whenever oculomotor data are collected and analysed. It is evident, that these choices can alter measured oculomotor effects. For example, when a higher value for minimal fixation duration is chosen, linguistic processing is given a greater chance to influence the first fixation duration in a word. At present, no optimal solutions for these problems can be offered, but increasing awareness of their presence among researchers in the field may be a first step toward more useful discussion on the topic. Inhoff and Radach further discuss the merits and potential problems of oculomotor measures commonly used to index cognitive processes in reading, including fixation probability, single, first and second fixation duration, gaze duration, total viewing duration and "repair time". Special emphasis is given to the consideration of deviations from the assumption of a direct correspondence between current fixation position and unit of actual cognitive processing. For example, local fixation patterns indicate that word processing is often distributed over several fixations and spill-over effects suggest that there is a variable time-span between eye and mind. Such phenomena, together with accumulating evidence on inter-relations between "foveal" and "parafoveal" information are likely to play a dominant role not only in theoretical discussions but also in the emerging general debate on methodological problems of eye movement research. The computation of complex reading time measures is one of these important methodological issues, and is addressed in detail in the discussions by Liversedge, Paterson and Pickering (Chapter 3). They start by noting that the usual technique in psycholinguistic eye movement research is to design sentences that cause specific processing difficulties and to compare the induced eye movement patterns with those recorded for control sentences. There are three possible scenarios when readers encounter processing difficulties in these experiments: the eyes can stay within the difficult region until the problem is resolved, they can proceed with a progressive saccade (perhaps with some reading time increase in the next region),

4

G. Underwood & R. Radach

or, alternatively, a regressive saccade could be initiated to allow re-reading of earlier segments of text. The measures usually reported in psycholinguistic studies are first fixation duration, first pass reading time and total reading time for a given critical region. These spatially contiguous measures are likely to handle the first two of the scenarios outlined above. But they may fail to give an adequate account of the spatiotemporal dynamics when one or several regressions occur and a reanalysis is initiated. Therefore, Liversedge et al. propose to combine them with two temporally contiguous measures. Regression path reading time is defined as the sum of all fixations from the first fixation in a region up to but excluding the first fixation to the right of this region. It provides an index of the time spent detecting a problem and re-reading the text prior to fixating novel linguistic material. Subtracting from this the first pass reading time for the critical region yields the second measure, re-reading time. Liversedge et al. apply these measures to a specific experiment and show that they can indeed capture effects that may be overlooked when using only the established reading time measures. There have already been some studies discussing the merits of complex reading time measures, but this chapter is perhaps the first that does this, not in order not to advocate a certain theoretical position, but to make a general methodological point. We hope that in the near future there will be more discussion of this kind leading to the general acceptance of new useful reading time measures but preventing an inflation of measures. It is important to decide upon the most appropriate measures, but it is just as important that we come to a consensus, otherwise each study may use a different set of measures to the previous study and comparability would then be weakened.

Visual processing and eye guidance in reading The question of what determines the location of the next fixation is addressed in a number of discussions, and some of the most powerful determinants are discussed by Radach and McConkie in their report of analyses of saccade landing positions based on a corpus of German reading data (Chapter 4). They emphasise that eye movement control in reading is strictly word-based, that is, at any given point in time the eyes try to make a saccade to a specific word-target. Once a target has been selected, the determination of fixation positions can be explained by a set of simple low-level oculomotor principles. The most fundamental of these is what they call the landing position function, a linear relation between launch distance and landing position of an initial saccade into a word. They present evidence showing that this function extends to launch distances of up to -21 characters, thus virtually including all possible visually guided progressive saccades. Radach and McConkie further show that the landing position function continues smoothly when launch sites from within the same word (refixations) are added. Together with the fact that the

Eye guidance

5

"preferred viewing position" of refixations is just right of the word centre (rather than close to the word boundaries) this argues against any special status for refixation landing positions. This does not necessarily mean that a distinction between "inter-word strategies" and "within-word tactics" (O'Regan, 1990) is unjustified, but it does mean that exactly the same eye guidance principles apply in both cases. Another interesting point concerns the landing positions of inter-word regressions. When the landing position function for regressions is plotted, the part for regressive refixations behaves very much like progressive refixations. For every increment in launch distance there is a certain shift in landing position. But surprisingly, for regressive saccades coming back from the next or second next word, landing positions are always clustered at the word centre, with almost no variation due to launch distance or word length. This result implies that regressions are determined by a qualitatively different mode of control, perhaps with more cognitive mediation and certainly with more precision as compared to the control processes for inter- and intra-word progressions and refixations. Whether to move their eyes to the left or to the right is an important decision for readers. There are costs and benefits associated with each, and Vitu, McConkie and Zola investigate the conditions under which regressions occur in reading by analysing a corpus of reading data for fifth-grade children (Chapter 5). Regressions constitute an interruption in the default left-to-right scanning direction that can be caused by problems on several levels: processing difficulties on the semantic or syntactic level, difficulties in word recognition or problems on the level of low-level perceptual or oculomotor mechanisms. An example for the latter are inaccuracies in eye positioning, for example, when the eyes have "accidentally" skipped a word. Within-word regressive saccades may often result from a tendency to make an additional fixation when the initial fixation on a word is mislocated. Vitu, McConkie and Zola's analyses start with replicating earlier findings that regressive saccades are more likely to occur following longer forward saccades. They then consider the influence of characteristics related to the word passed before the regression, including word length, distance between word and current fixation position, fixation location in the word, word frequency and whether or not the word had been refixated. Most interesting are results relating to regressions to previously skipped words. When saccade length is controlled, skipped words are more often regressed to and regression frequency increases with skipped word length. Instead of being simply a function of the preceding progression amplitude, the likelihood of regressing to a skipped word is systematically related to fixation positions preceding and following skipping. The greater the chance to identify the word before or after skipping, the less likely is a regression. In addition, there is a tendency to regress to less frequent skipped words. The results are less clear for cases where no skipping occurs, which, as Vitu et al. point out, may be due to combined analyses of intra- and inter-word regressions. In summary, there is substantial support for a significant

6

G. Underwood & R. Radach

role of word processing factors in the determination of regressions and a good starting point for more specific investigations of why and how they occur. In the default case of moving left to right with a progressive movement the reader may choose to fixate the next word to the right, but on a notable number of occasions the decision will be made to not fixate the next word at all. Brysbaert and Vitu discuss the characteristics of saccade target selection and the resulting frequency of fixating vs. skipping words (Chapter 6). For a number of years, word skipping has been one of the controversial topics, because it is one of the focus points where "low level" and "processing" theories of eye guidance differ greatly in their predictions. From the viewpoint of a strong oculomotor model, there should be only marginal, if any, influences of cognitive factors on word skipping. On the contrary, a strong processing (e.g., attention-based) model would predict that word skipping is related to parafoveal word recognition and a subsequent decision to avoid fixating a word that is already identified. The most important determinants of word skipping are saccade launch distance and target word length. This is not surprising, since word skipping frequencies represent the portion of a saccade amplitude (or landing site) distribution that falls on the letters of the target word. However, these effects can not be taken to indicate that word skipping is determined entirely by low-level processes, as larger and more distant words are also more difficult to process and therefore less likely to be parafoveally identified. In order to disentangle the contributions of low-level and word processing factors on word skipping, a number of experiments with carefully controlled stimulus sentences have been carried out. Brysbaert and Vitu present an elegant meta-analysis of these experiments, including results from seven studies manipulating the easiness of parafoveal words and eight studies manipulating contextual constraints. They establish that there is a significant cognitive influence on word skipping. When assessing the relative contribution of processing load against word length, however, it turns out that only a modest portion of the variance (4% for easiness and 11% for context) can be explained by cognitive factors. If this is the situation for specifically constructed sentence material, one is tempted to speculate how often equally high variation of contextual constraint and equally strong variation of word frequency is encountered when reading 'normal' text. Brysbaert and Vitu explain word skipping by referring to an "educated guessing" process based on the perceptibility of a parafoveal target word as a function of distance. This process provides an estimation of the chance to recognise the word from the current parafoveal position. The decision to skip (or to stay: fixating word-by-word must not be the default) is made on the basis of this quick and crude estimation. In this view word skipping is executed by a low-level process but based on cognitive (i.e., high-level) knowledge. If word skipping is a product of a decision based upon parafoveal processing, then we should be able to demonstrate that words a short distance from fixation can

Eye guidance

7

be identified, at least implicitly. A number of ingenious tasks have been created in the past as attempts to demonstrate effects of parafoveal processing, and in the latest experiment in this tradition Kennedy reports a task that used specific word comparisons (Chapter 7). Two words were presented to the right of a fixated prompt word, and the experiment asked whether the two parafoveal words could mutually facilitate recognition with fixation. The reader first looked at either the word "looks" or the word "means" and was asked to decide either about the visual identity or about the synonymy of the two eccentric target words. The critical experimental manipulation involved properties of the first parafoveal target word, which was shown to influence viewing duration measures while the eyes were still on the prompt word. Interestingly, gaze duration on the foveal word was shorter when the parafoveal target word was long (nine vs. five letters) and of low word frequency. For low word frequency targets, gaze duration was shorter when type trigram frequency is high, that is, when the word shared the same initial letters with many other words. This pattern of effects was largely due to a variation in refixations on the prompt word. However, an analysis restricted to cases where the prompt received only one fixation showed similar effects on fixation duration of parafoveal word length and initial trigram informativeness. Kennedy interprets these results by suggesting local process monitoring over a region larger than the current word. Saccade triggering is assumed to be sensitive to the extraction rate for parafoveal sublexical information. In the case of "difficult" parafoveal items an early saccade can be initiated when information acquisition is insufficiently rapid. The information acquired parafoveally is subsequently traded off when the word is directly fixated. Apart from its theoretical importance, this study provokes an interesting methodological question: to what extent can the results of simple word comparison tasks be generalised to normal reading? The rationale for choosing the task was to limit oculomotor complexities while still requiring stimulus words to be lexically processed. It allowed for an orthogonal variation of word frequency, initial trigram token and type familiarity that would have been difficult to achieve in a sentence reading task. This task may indeed provide a "microscope" to better look at subtle parafoveal effects that might otherwise be obscured by other, more potent, sources of variance. Once established as a working hypothesis, results should be re-investigated in a more ecologically valid reading situation. Kennedy made an attempt in this direction by presenting results of a corpus analysis that indeed provides some support for the generality of the reported effects. Another, more reading-like comparison task was used by Murray (Chapter 8). He presented two short sentences, one below the other, that could either be identical or differ by just one word. Subjects were asked to decide whether the second sentence was identical to the first, or not. Interestingly, in this task subjects do not use a simple word by word comparison strategy, but apparently build up a representation

8

G. Underwood & R. Radach

of the first sentence which they then compare to the second. There are two very important points about Murray's study. First, he extends the range of factors that can be shown to influence first fixation durations to the level of semantic plausibility. When a noun phrase is followed by a verb, the duration of the initial fixation on the verb is modulated by the plausibility of their semantic relation. This basically means that even the highest levels of processing in text reading can have an immediate effect on the duration of the ongoing fixation. Secondly, he shows that such effects can even operate when the critical information is available only at distances usually referred to as "parafoveal". When looking at the last fixation within the noun phrase, there is a clear effect of semantic plausibility: when the following verb is less plausible, fixation duration increases substantially. This effect is related to fixation position distance: it levels off when the preceding fixation was located more then four letters left of the critical word. On the basis of his data, Murray argues that there is no base for a functional distinction between information acquired from foveal vs. parafoveal words. Taken together, the studies by Kennedy and by Murray provide intriguing insights into the mechanisms of parafoveal-to-foveal crosstalk in eye movement control and are likely to stimulate more research on this interesting subject. If parafoveal words can be identified during reading, then the question arises of whether this processing is restricted to aiding recognition processes, or whether identification, or part identification, can be used by the processes that determine the landing position upon the previously parafoveal word. The discussions by Liversedge and Underwood (Chapter 9) and by Everatt, Bradshaw and Hibbard (Chapter 10) address the controversial issue of the landing position effect. They investigate the possibility that landing positions of initial saccades into words may be codetermined by cognitive properties of these words such as lexical informativeness, morphological composition or orthographic saliency. This hypotheses was investigated in a series of experiments carried out initially by Hyona, Niemi and Underwood (1989), by Underwood, Clews and Everatt (1990), and by Everatt and Underwood (1992). Stimuli for these experiments were obtained in pilot studies where subjects were asked to guess words on the basis of their first five or last five letters. The criterion value for words with informative beginnings vs. endings was that at least 89% of the respective guesses were correct. Stimulus words were presented within short passages of text which subjects were required to read. Of particular interest are two experiments, reported by Everatt et al., in which the sentence context prior to the critical word was manipulated. In one study, the general context either primed the informative beginning/ending words or was neutral. In another study, the critical words were preceded by a word with which they were either semantically related or unrelated. In both experiments there was a reliable effect of informativeness on landing position, but semantic manipulations did not mediate the effect. The authors see this as evidence against a role of parafoveal semantic pre-processing in eye

Eye guidance

9

guidance. Since results of other studies also make it unlikely that the landing position effect can be attributed to the morphemic composition of target words, it appears that orthographic properties of the letter sequences themselves give rise to the phenomenon. When comparing the landing position effect for adult dyslexic and non-dyslexic readers, Everatt et al. found that in the dyslexic group the average effect is less pronounced and there is more variability. Explanations for these results are discussed with respect to deficits in parafoveal visual processing or in eye movement control. The determination of the landing position is certainly not under strong control from parafoveal cognitive processing. There are stronger determinants such as launch distance and word length, and the landing position effect is not robust (see, for example, Rayner and Morris, 1992). The effect is seen in some experiments but not in others, and it is important for us to discover the conditions under which it does appear. One possibility is that it appears only when the foveal load is light, and when attentional resources may be allocated to parafoveal processing. To examine this possibility Liversedge and Underwood report two experiments that were intended to test whether a manipulation of foveal processing load would interact with the landing position effect. The rationale was that if foveal processing is more difficult, there should be less opportunity for processing of letter strings in the parafovea, leading to a weakening of the landing position effect. In the first experiment, processing load was manipulated by inserting a category word before the target word that referred to an antecedent noun phrase which was either a typical or atypical instance of the category. In the second experiment, the target word was preceded by a possessive pronoun referring to an antecedent noun phrase with a stereotypical gender which was either congruous or incongruous with the gender of the pronoun. Unexpectedly, in both experiments there was no effect of initial trigram on the landing position of the first saccade into the target word. However, when in a subsequent analysis of the second experiment only items with the shortest vs. longest gaze durations in the region before the target word were considered, an interesting pattern of results emerged. When foveal load was low, saccades landed further into words containing infrequent trigrams. The saccade into a word with a distinctive beginning was more extensive if the previous region of the sentence had received a greater amount of processing. A possible explanation for the weak evidence in these studies as compared to the earlier experiments on the landing position effect is the selection of the stimulus material. In most experiments where a cognitive landing position effect occurred, the target words had either a very high guessing probability or very salient beginning letter clusters (but note that in some of these studies the effect was also not replicated). A good example is the study by Hyona (1995), who used loan words whose initial trigrams almost only occur in these words, and are certainly the most extreme variation of orthographic saliency possible in a specific language. On the

10

G. Underwood & R. Radach

contrary, in the two experiments by Liversedge and Underwood, the differences in type initial trigram frequency were rather moderate. This, in turn, appears to indicate that to obtain cognitive landing position effects requires quite severe manipulation of experimental materials and, hence, these effects may reflect an exception rather than the rule in normal reading. On the other hand, the fact that the number of studies reporting cognitive landing position effects is increasing (e.g., Beauvillain, Dore and Baudouin, 1996; Inhoff, Briihl and Schwartz, 1996; Vonk, Radach and van Rijn, 1997) suggests that the effect is real, and that it can be observed only in specific conditions. But this does not necessarily mean that existing models of eye guidance need to be fundamentally changed. For example, when the eyes are thought to be driven by attention shifts, why should it not be possible that attention is occasionally captured by unusual letter clusters, leading to an extra fixation close to a word beginning? More generally, it may be that the cognitive landing position effect operates on the selection component rather than the amplitude computation component of the control system. Looking at the experiments on this topic presented in the current volume, there is one very interesting detail: when describing the results of earlier studies, Everatt et al. note that there was an effect of word beginning informativeness on the initial landing position, but that this was not caused by a difference in saccade length. If landing positions are different by more then one character and saccade length are identical the question arises: why were launch positions of the initial saccades different by more then one character? Since the configuration of preceding words in the experimental sentences was identical, something must have happened while the eyes were still on the prior word. The obvious candidate for this is that in some cases the salient vs. less salient parafoveal word beginning caused an additional fixation (refixation) in word n-1. This effect may be quite similar to what is called "refixations on the prompt" in Kennedy's chapter, providing a further example of parafoveal-to foveal crosstalk. Evidence for such an interpretation can also be derived from Hyona (1995) who reported that unusual letter clusters "attracted" fixations particularly to the space prior to the word. An alternative account for cognitive landing position effects would be to claim that they operate in a more graded fashion on the computation of saccade amplitude. (See Radach and McConkie's discussion in Chapter 4 for a more detailed description of discrete vs. graded effects). This is not in harmony with most current eye guidance models, but it appears to be not completely implausible, given the fact that some amplitude modulation is possible during late stages of saccade preparation, even when the decision to execute a saccade has already been taken (Becker, 1989). The overview of eye guidance during reading by Rayner, Reichle and Pollatsek (Chapter 11) starts with a discussion of evidence on the spatial and temporal components of eye guidance in reading. One interesting point they discuss is the controversy around the notion of an "optimal viewing position". They agree with

Eye guidance

11

the idea that saccades are in most cases intended to go to the centre of words, but raise doubts about whether the actual landing position has serious consequences for word recognition and eye movement control. In general, Rayner et al. see "where decisions" as co-determined by low level and cognitive mechanisms and "when decisions" as primarily determined by linguistic processing, though with some low-level modulation. After briefly reviewing recent modelling attempts by other authors, Rayner et al. present their new E-Z Reader model. What makes this model distinct from its predecessors is that it has been mathematically implemented and can be used as a simulation environment. From a theoretical perspective, the most important new feature is a decoupling of attention and saccade programming. There are now two distinct stages in the process of word recognition: a familiarity check, initiating the programming of saccades and the completion of lexical access, causing attention to shift to the next word. The speed of these processes is thought to be a function of word frequency, with modulation from both retinal eccentricity and contextual predictability of the target word. Saccade programming in the new model is divided into an early, labile stage, where saccade cancellation is possible (leading to word skipping) and a later, nonlabile stage that does not allow for cancellation (accounting for very short fixation durations between two successive saccades). This construction enables E-Z Reader to overcome one major limitation of the "classic" Morrison Model (Morrison, 1984; Rayner and Pollatsek, 1989) the handling of refixations. In the earlier model, there was no specific mechanism for dealing with multiple fixations on a word. The notion of a saccade programming deadline as proposed by Henderson and Ferreira (1990; see also the discussion by Henderson and Hollingworth in Chapter 12) turned out to be not in harmony with some aspects of the available data including fixation duration patterns (first vs. second of two) and fixation duration distribution data (e.g. Kennison and Clifton, 1995). In the new model, the labile saccade program will initiate refixations on the word as long as the lexical access stage in the word recognition module is not completed, and then provide a signal to go to the next word. As Rayner et al. point out, this mechanism is somewhat similar to what O'Regan (1990, 1992) has suggested to be the base for refixations. The chapter ends with a number of simulations run on a corpus of reading data which show that E-Z Reader is quite successful in simulating fixation duration, fixation probability and frequency of refixations while capturing specific effects related to word frequency, processing spillover and parafoveal preview benefit. The E-Z Reader model as described by Rayner et al. sets a new standard against which alternative ideas will have to compete. Its appeal is not only related to the fact that it can handle a large range of oculomotor phenomena with a limited set of provisions (although in the recent version the number of provisions has increased). On a more general level the popularity of attention-based eye guidance models may be related to the fact that they are situated in the "middle ground" between more extreme looking proposals at the oculomotor or cognitive end of the theoretical

12

G. Underwood & R. Radach

spectrum. In the following section we will briefly look at some theoretical issues that have received prominent attention in the reading-related chapters of this volume. Since virtually all authors present their own ideas in relation to the "standard model", this discussion can be seen not only as a summary of more or less common lines of reasoning, but to a certain extent also as a collection of alternative views and potential challenges to the Rayner et al. model. Is word recognition the engine that drives every saccade in reading? The findings by Radach and McConkie and by Vitu, McConkie and Zola that inter-word regression landing sites cluster at the word centre indicate that regressive saccades can be much more accurate than they usually are during progressive, left-to-right reading. What makes regressions distinct from other types of reading saccades is that in most cases the saccade is initiated several hundred milliseconds after the target word first appeared in parafoveal vision. At this point in time there is much more information available about the word (i.e., with respect to its spatial co-ordinates within the line of text) as compared to normal progressions, and, hence, the usual saccadic range error can be avoided. When a normal progressive saccade is prepared, the necessary parafoveal low-level (low spatial frequency) information can be obtained already during the first 50 ms of the current fixation (Pollatsek and Rayner, 1982). At the other end of the fixation, minimal saccade reprogramming time can be estimated to be in the order of 80 ms (Becker, 1989). Within these constraints, the time and resources effectively available during a normal fixation to determine the next saccade amplitude seem to be insufficient to achieve optimal accuracy within the target word. It is an open question whether similar considerations can also be applied to saccade target selection. Brysbaert and Vitu consider the timing of events during a fixation for the case of word skipping. In a simple additive model, they add 60 ms transfer time from the eyes to the brain, 100 ms saccade programming time and 90 ms extra processing time increase due to a (minimal) eccentricity of three letters, arriving at a sum of 250 ms. This estimate leaves virtually no time for operations like the completion of lexical access on the current word, the execution of a familiarity check on the next word, and, on the basis of the result, the cancellation of the labile saccade program to this word, which are all necessary to skip a word in the Rayner et al. model. This conclusion would change if some of the operations were executed in parallel, of course. On the other hand, as Brysbaert and Vitu's meta-analysis shows, lexical and contextual information does mediate at least some saccade target selection decisions. Also, in comparison to the classic attention-based model (e.g., Rayner and Pollatsek, 1989), the timing concerns raised above apply to a somewhat lesser extent to the E-Z reader model, because in the new familiarity check stage less cognitive processing is demanded to trigger saccade programming.

Eye guidance

\3

It is interesting to observe that while Rayner et al. take a more moderate position on the role of cognition in determining saccades, researchers who emphasise low-level factors try to make their scanning routines smarter. For example, the alternative mechanism for word skipping proposed by Brysbaert and Vitu involves an estimation of the chances to recognise the next word within the current fixation. The decision is not based on immediate lexical processing, but guided by long-term knowledge about parafoveal word recognition that has been acquired individually by a certain reader. Another use of cognition to aid a low-level mechanism has been proposed by Radach (1996), who suggested that readers may obtain a limited repertoire of parafoveal word length configurations that statistically correspond to psycholinguistic units and may be used to mediate eye placement decisions. Does attention move in discrete, sequential steps from word to word? One of the cornerstones of current thinking about eye movement control in reading is the notion of the perceptual span, denoting the region from which useful information is acquired during a certain fixation. More specifically, Rayner and Pollatsek (1987), define a letter identification span encompassing a region 3-4 letters to the left to 6-10 letters to the right of the current fixation position (sometimes more) and a word identification span including the current word, and possibly the next or the next two words. The left boundary of the perceptual span is thought to be tied to the left boundary of the current word. All variants of attention-based eye movement control models claim that within the perceptual span attention moves in discrete steps from word to word in a left-to-right direction. This has been called the sequential attention hypothesis (e.g., Henderson and Ferreira, 1990). Although in the new E-Z reader model attention and saccade programming are decoupled, the serial attention shift claim is maintained, as the familiarity check on the next word is only initiated after lexical access on the current word has been completed. Although Rayner et al. provide convincing evidence to support this claim, there is also evidence accumulating that is not in harmony with their model. In the following we will discuss some problems that have been raised in the reading-related chapters of the this volume. The first problem is related to the possibility that the region from which letterlevel information is extracted can be extended over the left word boundary. Vitu, McConkie and Zola report that after a saccade, when the eyes land at positions just following a skipped word, the probability of making a regression back to this word is very low. Obviously, when input from the skipped word is needed and this word is not far away from the current fixation position, it is often preferred to get this input via a parafoveal "back-view" rather than by programming and executing a regression. Converging evidence comes from Inhoff and Radach's discussion which also suggests that readers may sometimes obtain information (including lexical inform-

14

G. Underwood & R. Radach

ation) from locations left of the current word beginning. In principle, the current attention-based models could accommodate these findings by postulating something like an "attentional regression" (see Duffy, 1992, for a discussion of "mental regressions") but this would lead to complications. For example, shifting attention back to a previous word would mean interrupting or delaying processing of the currently fixated word and should lead to substantial costs in processing time. However, increased fixation durations after word skipping have so far never been reported. This leads to a more general problem, which concerns the possibility that linguistic information from neighbouring words is processed in parallel with information from the currently fixated word. As Kennedy points out in Chapter 7, if attention is allocated sequentially, properties of a parafoveal word will not influence foveal processing time, because an attention shift only occurs when foveal processing is complete. Contrary to this view, he demonstrates that sublexical properties of a parafoveal word can influence fixation durations and refixation rates on the foveal word. Similar effects are shown by Murray on the semantic level. At present there is only a small number of studies that directly contradict the sequential attention hypothesis, but the number of studies directly supporting the hypothesis is also limited. If more evidence on parafovea-to-fovea crosstalk accumulates in future research this would seriously challenge the sequential attention hypothesis, and, more generally, any qualitative difference in information extraction from within and beyond a word boundary. An alternative model that could account for parafoveato-fovea crosstalk could include an attention gradient, where letter information is acquired in parallel from all letters within the letter identification span with the rate of extraction being a function of letter eccentricity. This possibility has been discussed, among others, by Inhoff, Pollatsek, Posner and Rayner (1989) and by Kennison and Clifton (1995). Within such an alternative framework, saccade processing and the extraction of text information would be less tightly coupled as in the current attention-based processing models. Eye guidance in picture viewing and visual search The development of attention-based processing models has been a major step forward towards an understanding of eye movement control in reading and has also influenced theoretical discussions in neighbouring domains such as scene perception. We are not yet in a position to describe such a complete model of fixation patterns on scenes, partly because we have not yet found an adequate way of describing the pictorial information, but the remaining discussions in this volume go some way to describing what it is that such a model will have to explain. What controls the locations of our fixations when viewing scenes? Henderson and Hollingworth ask whether we select components of a scene on the basis of

Eye guidance

15

purely visual characteristics, or whether eye guidance can be influenced by the semantic characteristics of an object (Chapter 12). This is one of the recurring questions asked in the discussions of reading elsewhere in this volume, of course. As Henderson and Hollingworth point out, most of our knowledge of eye-movement control has been derived from studies of reading lines of text, but a complete theory must take into account the processes that are observed in scene perception. We would go further than this, and argue that the complete theory will also describe the placement of fixations when the observer is watching dynamic scenes such as in cinema or video film, and when the observer is moving within the environment that is being inspected. Only then will the theory be ecologically valid, and there are discussions of these extensions to the scope of our question in Chapters 16 by d'Ydewalle, Desmet and Van Rensbergen, Chapter 17 by Chapman and Underwood, Chapter 18 by Crundall, Underwood and Chapman, Chapter 19 by Dishart and Land, and Chapter 20 by Liu. The relatively sparse literature suggests that informative areas receive more fixations than others. This result is consistent with the idea that we allocate visual attention, through a greater number of fixations and through longer fixations, when we encounter material that requires cognitive processing. The principle holds for reading and also for scene inspection. But the interesting question is how we know that some areas of a scene are informative and that others are not worth a fixation? The early literature suggests that uninformative areas are often not fixated at all, suggesting that the guidance decision is influenced by the processing of extrafoveal information1. The debate about the significance of word skipping that is addressed specifically in the discussions by Brysbaert and Vitu and by Vitu, McConkie and Zola, re-emerges here as a question about object-skipping. A secondary similarity appears in Henderson and Hollingworth's distinction between visual and semantic informativeness. Do observers avoid fixating certain areas of a picture because they have detected few discontinuities of texture, for example, or because the object in an area is uninformative? If we are guided by object informativeness then we can conclude that fixation patterns are determined by cognitive operations as well as more restricted perceptual processes. This question has been approached by using scenes in which objects are congruous, and therefore have lower semantic informativeness, or incongruous. For example, a microscope on the counter in a hotel bar would be more informative than a tumbler with a straw in the same location. When first shown a picture, do observers fixate an informative object earlier than they do a more predictable object? There are inconsistencies in the literature 1 The distinction between parafoveal and extrafoveal is generally made in the discussions of scene perception and text inspection here because the non-fixated objects in scene perception studies are usually located beyond the parafovea. The "extrafovea" does, of course, include the parafovea.

16

G. Underwood & R. Radach

reviewed by Henderson and Hollingworth, but the balance of replies to the question is against the idea that the semantic properties of objects are able to attract early fixations. What then does influence fixation locations during scene inspection? The factors that are considered to be potent here are image size, viewing task, display duration, and image content and type. Henderson and Hollingworth incorporate these influences in their view of eye-movement control during scene inspection. The "saliency map framework" suggests that attention is always allocated to the region with the greatest saliency value and these regions are in turn fixated. The framework predicts which factors will be influential during the early stages of inspection, and describes how saliency will change as information about the scene is collected, such that the early stages of inspection will be driven by predominantly visual features, and later stages by predominantly semantic features. We are not only making progress in the investigation of scene inspection, but we are also seeing the development of theoretical frameworks that will help focus future investigations. Findlay and Gilchrist also address the question of the relationship between the direction of our eyes and the direction of attention, asking whether we can redirect attention without moving our eyes (Chapter 13). There is well established evidence that this dissociation is possible, but these demonstrations may be possible only in highly constrained and highly artificial laboratory situations. Distinctions between serial and parallel processes in visual search have relied upon the assumption that observers can search an array of items with rapid attentional shifts that are covert in that they are made in the absence of eye movements. Findlay and Gilchrist challenge this assumption, and explore the characteristics of the overt and covert attentional processes. There is also a methodological consideration here, of course, because if attention can disconnect from the point of regard from which we inspect scenes, then an analysis of fixations will give us a misleading account of search patterns. By inspecting the first saccade in a visual search task, Findlay and Gilchrist focused on the processes that result in selection of the location of the first fixation in the display. Is there evidence of a covert movement of attention to the target prior to the eye movement? There was no difference in the saccadic latency whether the target stimulus was defined on the basis of a simple feature (object shape) or on the basis of a conjunction of features (shape and colour together). Attention is thought to be necessary for the identification of targets composed of conjunctions of features, but not for simple targets, and so the absence of difference in saccadic latency suggests that the attentional processes necessary for the conjunction search are not operating prior to the first saccadic movement. However, with one of their tasks Findlay and Gilchrist were able to show a use for covert attentional shifts followed by delayed saccades, but this was with training over the course of several hours. In general, they conclude that we do not move attention and eyes at different times, in contrast with the assumptions of the Henderson and Hollingworth account of scene perception in which attention is released once the object being fixated is processed. At this point

Eye guidance

17

attention is attracted by the area with the greatest saliency, and the eyes programmed to move to that point in space. The relationship between attention and the point of fixation remains a matter of some debate. Part of the argument used by Findlay and Gilchrist to distinguish between processes that require attention and those that do not is based upon the so-called "pop-out" of objects in displays when processing does not require attention. If the target in a visual search task is a solitary circle in a display otherwise consisting of squares and triangles, then the search time will be low and the target will stand out from its background. Also, the number of irrelevant items does not influence the time taken to find the target. It appears that this kind of target is detected without the need of attentional resources. In contrast, a target defined by a conjunction of features does not pop-out, and has a search time that is a function of the number of irrelevant items, and it is concluded that identification requires attention. Henderson and Hollingworth examine some of the evidence concerning the pop-out of objects in line drawings and other scenes, and De Graef's discussion focuses on this phenomenon in an examination of the capture of attention by objects that are schematically inconsistent with their contextual scene. This takes up Henderson and Hollingworth's discussion of the use of extrafoveal information in guiding attention and the observer's eyes to objects that form part of the scene. At what point in the processing of a scene can semantic processes start to have an influence on the eye guidance mechanism? Does an object need to be fixated before its meaning is appreciated, or is there a pop-out effect for objects that are incongruous with their surroundings? De Graef points out that incongruous objects variously present processing enhancement and processing difficulty. In one task these objects illustrate a pop-out, and in another they are the source of processing delays. In reading tasks, for example, a semantically anomalous word will cause a break in the otherwise smooth flow of ideas through the reader's mind. When should we expect facilitation from perceptual pop-out and when should we expect inhibition from schema-driven perception? De Graef accepts the view that the two effects are not logically inconsistent, and uses a mismatch principle (Johnston and Hawley, 1994) by which items that occur together repeatedly become unitized. Recognition of any component item that matches expectations will then be facilitated, and processing of any item that does not match expectations will also be facilitated and will result in rejection, or pop-out, from the unitized representation. To answer the question of whether implausible objects gain an attentional advantage that may result in early fixation, De Graef uses an ingenious display in which target objects wiggle (Chapter 14). An object that wiggles (moves up and down rapidly) may elicit attentional capture, and when the wiggling object is a target in a search task it is then possible to set up questions about the relationships between attention and the processing of implausible objects. Implausible objects were fixated earlier and more frequently than their plausible counterparts, and this

18

G. Underwood & R. Radach

evidence must be added to the review by Henderson and Hollingworth in that it goes against the current flow which suggests that there are no fixation advantages to be gained by semantics. This argument about the potency of extrafoveal items to attract fixations has many of the characteristics of the debate elsewhere in this volume, over the matter of whether properties of a parafoveal (target) word can have a significant influence on viewing time and fixation position measures on a foveated word. Low target plausibility reduced fixations on the priming stimulus in De Graef s experiments, consistent with the idea that implausible targets present a processing difficulty while they remain in the extrafovea, and it is this difficulty that attracts attention and fixation. An interesting conclusion from the wiggle experiments comes from a manipulation of the distance between the priming object (location of the first fixation on the scene) and the target (wiggled) object. In this study implausible targets were fixated more than plausible targets, as before, and near targets (3° distance) were fixated more than far targets (8°). It is the need for foveal analysis that determines whether an object will be fixated, and this need is greater for implausible objects and for objects that are too far into the periphery for analysis. De Graef concludes that we can consider these effects as demonstrating that plausible objects, which are by definition most of the objects that we encounter, enjoy a larger useful field of view. It may be that the context of the scene reduces the possible candidates and thereby increases the importance of features that can be detected extrafoveally, an account that can be adopted quite readily as an explanation for context effects in word skipping during reading. The longer our eyes dwell upon one object or word, not only the greater is the information extracted from the point of regard, but also the greater is the amount of information that can be collected about extrafoveal objects. The discussion by van Diepen, Wampers and d'Ydewalle also focuses on the necessity of foveal inspection (Chapter 15). The technique used in the work reviewed in their discussion uses a display that changes according to the current location of the observer's eyes — a contingent display. The parts of the display around the point of fixation are degraded in some of the studies, to simulate the effect of a scotoma, and the area of degradation can be varied to create a blind-spot that can vary in size. The "mirror image" of this procedure gives the observer a moving window with extrafoveal information degraded. A technical appendix to their discussion explains how these displays are engineered. Contingent degradation at the point of fixation allows a direct approach to the question of how much information can be extracted from extrafoveal vision, for the simple reason that the participant observer in these experiments, in effect, has no foveal information available. However, it should be kept in mind that this technique provides an estimate of maximal rather then normal extrafoveal information acquisition as under normal conditions foveal and extrafoveal visual processing will always have to compete for shared resources.

Eye guidance

19

Studies of the recognition of simple line drawings in these artificial scotoma experiments suggest that foveation is not a requirement for identification, although gaze durations did increase. This increase suggests that we can use extrafoveal vision to determine the identity of objects, but that this takes extra time. Interestingly, individual fixation durations do not vary, suggesting that it is difficult to fixate in one (degraded) location while attending to an object in another (clear) location. By varying the onset of the degradation relative to onset of the fixation, such that the observer receives a clear image of fixated material for a brief interval before it is obscured, it is possible to establish the time-course of extraction of information from foveated images. This procedure establishes that observers can extract the information they need within the first few centiseconds of fixation. The remaining fixation time may be used for task processing, saccadic programming, or perhaps extrafoveal processing. Evidence of the early identification of foveated objects during long fixations, combined with evidence that suggests that extrafoveal information can serve object identification if the inspection is sufficiently prolonged, would result in the prediction that extrafoveal identification will be most readily observed in cases of prolonged fixation. This is the pattern of results reported in the discussion by Liversedge and Underwood, for the case of parafoveal effects during reading. Consistent with the conclusions of Findlay and Gilchrist, the studies reviewed by van Diepen, Wampers and d'Ydewalle suggest that observers can divide fixation and attention when the necessity arises, but that this does not come easily. Presumably if we made this dissociation regularly when reading and when viewing scenes it would be well-practised and thereby readily observable as a skill in laboratory experiments. Saccades to the attended object are unavoidable, in support of the framework for fixation guidance outlined by Henderson and Hollingworth, in which our eyes follow shifts in the location of attention. Eye guidance while watching dynamic scenes Our working assumption is that recognition of part of a scene benefits from foveation but is not impossible without it. Extrafoveal information can be used, if foveal information is unavailable, and we have demonstrations of object recognition without fixation. The extent of extrafoveal processing becomes increasingly significant, of course, when inspecting scenes containing moving objects. The efficient detection of movement in the extrafovea was a property of the human visual system exploited by De Graef by having objects of interest move up and down while other parts of the scene were being inspected. These wiggling objects were very successful in attracting attention. When whole objects are moving within a scene together, and when the scene itself is apparently moving, by virtue of the

20

G. Underwood & R. Radach

relative movement of the observer, then we can expect extrafoveal processing to become more important. Different types of apparent movement result in different perceptions, and d'Ydewalle, Desmet and Van Rensbergen demonstrate the importance of taking account of the type of motion for those of us who are concerned with the recognition of moving objects in dynamic scenes. The approach taken involves the presentation of moving scenes in cinematic ("movie") format, and in which different types of editing cuts are made. What do the observer's eye movements tell us about the perception of change in these dynamic real-world scenes? Three types of editing cut were investigated. At the first level there were small changes in the image portrayed, implemented by changing the camera position for example, and these changes have little effect upon fixation patterns. In contrast, at the second level the position of the camera is changed to give the view from a different direction plus a changed background, and these changes prompt eye movements. The measure of movement here is an increase in the variance in horizontal fixation locations, peaking a few hundred milliseconds after the change. This is an unconventional measure of fixation behaviour that Chapman, Crundall and Underwood also found useful in their attempts to understand the perception of dynamic scenes in their discussions. A number of the discussions in this collection illustrate not only the huge range of measures that are available once we start collecting an observer's eye movements, but also the different implications associated with changes in these measures. D'Ydewalle et al. concluded that this influence on the variance of fixation locations was an indication of post-perceptual, cognitive effects, and the third type of editing cut, involving a change to action sequence of the narrative, also produced increased variance. The observers in the study reported by Chapman and Underwood also watched moving images (digitised video clips, in this case) of real-world scenes, but in this study there were no artificial changes to the camera position to disturb the understanding of the narrative. The changes of interest involved potentially dangerous events. The video clips were recorded from a moving vehicle as it moved along various types of roads, and the events of interest involved the actions of other road users, such as a cyclist who rides out in front of the vehicle from a side road, and a parked vehicle ahead that is obscuring a pedestrian who is about to cross the road. For the actual driver of the vehicle, these are hazardous events that would require an evasive act. When drivers watch these video clips in the laboratory, their fixation behaviour indicates the extent to which the event captures their attention, and thereby indicates differences between drivers of varying ability and between events of varying saliency. When a dangerous event unfolds, the object that is the source of the danger collects extended fixations suggesting that an object to be monitored requires greater attention — the observers are monitoring these objects to determine whether the potential for danger inherent in a pedestrian moving from behind a

Eye guidance

21

parked car does in fact become an actual hazard that would require an action. There were also differences due to driving experience consistent with this interpretation, with novice drivers having longer fixation durations than experienced drivers. For the experienced driver, as for the experienced reader encountering a story for the first time, understanding of the unfolding events is facilitated through having previously encountered other configurations of the components of the scene or the sentence. There may be a further relationship between reading and driving here, in that when we move through a sentence there are limited options for the next word. Syntax and semantics restrict the range of words that can occur at any point, and these constraints are acquired by readers and used to predict what can happen next. The same may be said of driving, and if we stop the video at any point, experienced drivers will be able to give the range of events that could happen. Perhaps the reduced fixation durations of both experienced drivers and experienced readers reflects the ease of recognition of the objects that are presented in the sequence of road or sentence events, and perhaps in both cases contextual facilitation can influence recognition. A difficulty with the simple account of Chapman and Underwood's results is that fixation behaviour varied according to the type of road being presented on the video clips. Urban roads collected shorter fixations than rural roads, and perhaps this is a reflection of the overall difficulty of processing, but a more difficult sentence attracts longer gazes, not shorter ones. This is where the analogy between watching driving videos and reading collapses, of course, because reading is usually a self-paced task whereas the video is always shown at the same pace. When a fixed-speed video contains many potentially dangerous events (e.g., a crowded street scene) rather than very few objects (e.g., a rural lane) then the observer must scan from object to object searching in case one of them becomes the hazard to be avoided. When the observers are not watching driving videos, but driving themselves, we return them to a self-paced task in which an increase in perceived difficulty can be accommodated by a reduction in speed. The final three discussions present studies of fixation behaviour in drivers of varying experience and in varying environmental situations. Chapman and Underwood found that, at least for certain types of events with an abrupt onset (e.g., a pedestrian steps out from behind a parked car), there is evidence of attentional capture. Measures of fixation location and of the variance of fixation location indicate that at these moments most observers were fixating the same point on the screen. This is the equivalent of one of De Graef's line-drawings wiggling and then popping-out of the surrounding display — it cannot be avoided, and it is difficult to look anywhere else on the screen when one of these dangerous events is unfolding. Crundall, Underwood and Chapman also consider variations in the attentional field induced by roadway events and by prior experience, in studies in which drivers

22

G. Underwood & R. Radach

control a vehicle through a set route involving a range of roads, and in which their ability to detect extrafoveal targets is assessed in a laboratory task. Different roadways present different difficulties, and novice drivers respond to them in different ways. The variance of horizontal search proved to be a successful indicator here, as it did in the study reported by d'Ydewalle, Desmet and Van Rensbergen. In the driving study, novices showed little variation in their fixation behaviour regardless of the difficulty of the road being driven, whereas experienced drivers were sensitive to task demand. Specifically, experienced drivers had a huge increase in the variance of their horizontal fixation locations on a multi-lane road on which traffic merging was frequent. These drivers were looking around them to check on the locations of adjacent vehicles. The novices tended to look straight ahead, with longer fixation durations, possibly in response to the increased perceptual load of a crowded urban motorway on which many vehicles were changing their traffic lanes. Drivers were later presented with a laboratory task to investigate their responses to varying perceptual loads, using video clips similar to those described in Chapman and Underwood's discussion in combination with a peripheral target detection task. Low perceptual demand was associated with saccades of shorter latency in the direction of targets, and reduced fixation durations on the targets, but target detection was uniformly influenced by perceptual load for near targets (<6°) and for far targets (>6°), suggesting that perceptual load has a general effect of interference rather than any attentional narrowing that we could describe as tunnel vision. Comparisons between the fixation behaviour of novice and experienced drivers are also made in the discussion by Dishart and Land, but the novices here are totally inexperienced in that they are in the early stages of receiving tuition prior to attempting their driving tests. It is therefore difficult to make direct comparisons between this discussion and the previous studies, and the novices here must be regarded at a very different stage in the acquisition of skill relative to the recently qualified drivers described in the previous discussions. The question asked by Dishart and Land is how fixation patterns reflect the earliest stages of learning to drive, when controlling the vehicle is still a problem. On the basis of a study reported by Land and Lee (1994) we have the suggestion that fixation of the tangent point of a curve is important for steering control. The tangent point is the area of a curve where the road appears to change its direction, and constant fixation will provide the driver with information about the extent to which the steering wheel should be turned. There are other features of the roadway to inspect however, as in Chapman and Underwood's video clips portraying dangerous events, and this is why the Land and Lee suggestion deserves further consideration. Their study, involving three experienced drivers, used a road with no on-coming traffic and very few side roads. There were few obstacles and not only was there little need to look ahead to check that the road was clear, but the road was embanked and therefore provided little opportunity for the driver to look ahead. Regular

Eye guidance

23

fixation of the tangent point in the Land and Lee study may have been a consequence of the absence of alternatives. Dishart and Land extend the investigation of tangent points in their discussion to look for changes in the early stages of driving, with evidence, at least from one driver, of increased use of the tangent point after the first four hours of driving tuition. This suggests a change from using the immediately adjacent road kerbside for lane maintenance, to use of the more distant tangent point for anticipatory steering control. This may be stage-specific change in behaviour, however, and restricted to the case of driver's in the initial stages of learning to control the vehicle in order to keep it on the road. When ten experienced drivers are compared with ten novices who have recently passed a driving test, it emerges that with some curves it is the experienced drivers who tend to not look at the tangent point (Underwood et al., 1997). Indeed, many of the drivers in this study failed to fixate near to the tangent point during at least one of the curves while they were driving, leading to the question that if tangent point fixation is essential for steering, then how were they able to keep the vehicle in lane when they were looking elsewhere? Presumably extrafoveal vision is used extensively by more experienced drivers. The driver in the Dishart and Land study also showed an interesting increase in fixation durations over the first few hours of driving experience, and they interpret this as an indication that the driver has started to process more useful information. When first introduced to the vehicle's controls this driver may have been searching the view ahead in an attempt to learn what it is that will help with steering control, but after a few hours the relationship between road curvature and required steerage is in the process of being acquired. When drivers are placed in a steering simulator, however, any form of active road-related experience (such as cycling and computer driving games) predicts the ability to perform in the task. Furthermore, longer fixations are usually associated with increased perceptual difficulty, however, and if this increase proves to be a reliable result it may be that this driver has started to realise some of the difficulties of driving safely. A methodological decision to be made when studying driving behaviour is whether to use a simulator or an on-road vehicle. Dishart and Land used both, Crundall et al. used an instrumented vehicle with an on-board PC and eye-tracker, and Liu gave his experienced drivers a journey in a laboratory simulator. The advantage of the simulator over the natural roadway, of course, is that all participants can be presented with the same driving conditions and with the same traffic problems. The advantage of the roadway is that it is authentic — the quality of the display is usually better than that in a simulator in all possible respects including size, colour, texture and resolution. In all cases the drivers are in control of their own motion, real or apparent, and receive complex visual displays that vary according to their own position in the scene. Liu applies an analysis to his fixation data that lends itself very readily to a modelling process. Liu's Markovian analysis asks whether

24

G. Underwood & R. Radach

each fixation location is dependent upon the location of the previous fixation. Given that the driver has just looked at the tangent point of the approaching curve, what part of display is inspected next? The leading car, the nearest visible part of the road ahead, or the right-hand side of the road. And having then looked at the nearest part of road, on the second fixation in this sequence, does the driver look at the leading car, the right-hand side of the road, or the left-hand side of the road. Each fixation made during the simulated driving task can be entered into this sequence analysis to find a fixation transition matrix for different driving tasks. Liu instructed his drivers either to maintain the vehicle in the centre of the road, or to maintain a selfdetermined distance from a lead car. Well formed fixation transition matrices were discovered for each of these driving tasks, which themselves generated different matrices. Liu proceeds to use his analysis of fixation sequences to develop a model of the driver's cognitive processes. Given that a specific sequence of eye fixations has been identified, is it possible to determine the cognitive representations that produced that sequence? The question asks whether we can determine a driver's intentions by first recognising a sequence of fixations, for if we can then the builders of the intelligent vehicles of the future will be able to modify the information provided to the driver to help achieve these goals. For example, we know that glances into the side mirrors precede overtaking and lane-changing manoeuvres, and if the intelligent vehicle's perimeter field detectors have identified an obstructing vehicle that would render the manoeuvre unsafe, then a warning could be sounded, or the action could even be prohibited. While this may sound like science fiction it should be borne in mind that Liu's investigations are made on behalf of an international car manufacturer, and further discussions of intelligent vehicle control can be found in Parkes and Franzen (1993) and Underwood et al. (1994). Concluding remarks The question that we started with was "what determines where we look next?" and most of the discussions have focused on the reduced form of "what determines what we look at?" For the case of driving Liu has provided a very direct answer: what we look at depends upon what we looked at last. The story is complicated by the use of displays that contain moving objects of course, with De Graef s wiggling drawings and Chapman and Underwood's moving hazards having a pronounced saliency in attracting attention. If we performed Liu's sequence analyses on the fixation patterns from these experiments, we would find that whatever was being looked at now, a moving object would be fixated next. The transition matrices found in his driving simulator suggest that there are fixed patterns of fixations, or scanpaths, that predict where a driver will look given a previous fixation in any other part of the visual field. The equivalent conclusion from studies of reading would be that having

Eye guidance

25

looked at one word in a sentence the next fixation will be upon the next word. A transition matrix for reading would have the word immediately to the right of the current word as the most likely object of the next fixation (for a transition matrix analysis of reading data, see Stark, 1994). Just as with scene inspection, there are other sources of variation to be taken into account. Discussions in this volume have focused upon just three of these sources — the conditions under which a regressive fixation will be made, the conditions under which word skipping will occur, and the variation of the landing position within the next word. Some similarities are reported across the domains discussed here. For example, the debate over whether semantic information can be collected with extrafoveal vision is represented both in the discussions of reading (e.g., by Murray, by Kennedy, by Liversedge and Underwood, and by Everatt, Bradshaw and Hibbard), and in discussions of scene perception (e.g., by Henderson and Hollingworth, by Findlay and Gilchrist, by De Graef, and by van Diepen, Wampers and d'Ydewalle). In both cases the data are equivocal. The answer is probably that semantic information is potentially available, but that it is only collected under very specific circumstances. If these conditions are not met, then the experiment will conclude that no influence is present. A similarly unresolved argument concerns variations in landing positions that can be attributed to anything more than the gross visual features of the word or object that is about to be fixated. At one end of the spectrum there is Radach and McConkie's emphasis on low level determinants of fixation positions leaving little room for cognitive influences. At the other end there is Liversedge and Underwood's claim for cognitive landing position effects, at least when the prior fixation is long and there is a greater opportunity for an extrafoveal word or object to be part-recognised and for the saccadic programming that will lead to fixation of a selected part of that word or object. More evidence will be needed to resolve this specific controversy. But on a more general level the appropriate answer may be that eye guidance is as low-level as needed and as cognitive as possible for a certain task and stimulus configuration. To determine the details of what kind of information is acquired and processed when and how during the time course of visual and cognitive analysis is a very complex research enterprise. The interpretation of the evidence, including the new data presented here, illustrate the complexity of the enterprise, but they also demonstrate the progress that is being made. References Beauvillain, C., Dore, K. and Baudouin, V. (1996). The 'center of gravity' of words: Evidence for an effect of the word-initial letters. Vision Research, 36, 589-604. Becker, W. (1989). Metrics. In: R.H. Wurtz and M.E. Goldberg (Eds.), The Neurobiology of Saccadic Eye Movements (pp. 13-61). Amsterdam: Elsevier.

26

G. Underwood & R. Radach

Duffy, S.A. (1992). Eye movements and complex comprehension processes. In: K. Rayner (Ed.), Eye Movements and Visual Cognition: Scene Perception and Reading (pp. 333-354). New York: Springer. Everatt, J. and Underwood, G. (1992). Parafoveal guidance and priming effects during reading: A special case of the mind being ahead of the eyes. Consciousness and Cognition, 1, 186-197. Henderson, J.M. and Ferreira, F. (1990). Effects of foveal processing difficulty on the perceptual span in reading: Implications for attention and eye movement control. Journal of Experimental Psychology: Learning, Memory, and Cognition, 16, 417-429. Hyona, J. (1995). Do irregular letter combinations attract readers' attention? Evidence from fixation locations in words. Journal of Experimental Psychology: Human Perception and Performance, 21, 68-81. Hyona, J., Niemi, P. and Underwood, G. (1989). Reading long words embedded in sentences: Informativeness of word parts affects eye movements. Journal of Experimental Psychology: Human Perception and Performance, 15, 142-152. Inhoff, A.W., Pollatsek, A., Posner, M.I. and Rayner, K. (1989). Covert attention and eye movements in reading. Quarterly Journal of Experimental Psychology, 41, 63-89. Inhoff, A.W., Briihl, D. and Schwartz, J. (1996). Compound word naming in reading, online naming and delayed naming tasks. Memory and Cognition, 24,466-476. Johnston, W.A. and Hawley, K.J. (1994). Perceptual inhibition of expected inputs: The key that opens closed minds. Psychonomic Bulletin and Review, 1, 56-72. Kennison, S.M. and Clifton, C. (1995). Determinants of parafoveal preview benefit in high and low working memory capacity readers: Implications for eye movement control. Journal of Experimental Psychology: Learning, Memory, and Cognition, 21, 68-81. Land, M.F. and Lee, D.N. (1994). Where we look when we steer. Nature, 369, 742-744. Morrison, R.E. (1984). Manipulation of stimulus onset delay in reading: Evidence for parallel programming of saccades. Journal of Experimental Psychology: Human Perception and Performance, 10, 395-396. O'Regan, J.K. (1990). Eye movements and reading. In: E. Kowler (Ed.), Eye Movements and Their Role in Visual and Cognitive Processes (pp. 395-453). Amsterdam: Elsevier. O'Regan, J.K. (1992). Optimal viewing position in words and the strategy-tactics theory of eye-movements in reading. In: K. Rayner (Ed.), Eye Movements and Visual Cognition: Scene Perception and Reading (pp. 333-354). New York: Springer. Parkes, A.M. and Franzen, S. (1993). Driving Future Vehicles. London: Taylor & Francis. Pollatsek, A. and Rayner, K. (1982). Eye movement control in reading: The role of word boundaries. Journal of Experimental Psychology: Human Perception and Performance, 8,817-833. Radach, R. (1996). Blickbewegungen beim Lesen: Psychologische Aspekte der Determination von Fixationspositionen (Eye Movements in Reading: Psychological Aspects of the Determination of Fixation Positions). Miinster/New York: Waxmann. Rayner, K. and Morris, R.K. (1992). Eye movement control in reading: Evidence against semantic preprocessing. Journal of Experimental Psychology: Human Perception and Performance, 18, 163-172. Rayner, K. and Pollatsek, A. (1987). Eye movements in reading: A tutorial review. In: M. Coltheart (Ed.), Attention and Performance XII: The Psychology of Reading (pp.

Eye guidance

27

327-362). London: Erlbaum. Rayner, K. and Pollatsek, A. (1989). The Psychology of Reading. Englewood Cliffs, NJ: Prentice Hall. Stark, L.W. (1994). Sequences of fixations and saccades in reading. In: J. Ygge and G. Lennerstrand (Ed.), Eye Movements in Reading. Oxford: Elsevier. Underwood, G., Chapman, P., Crundall, D., Cooper, S. and Wallen, R. (1997). The visual control of steering and driving: Where do we look when negotiating curves? Paper presented at the Vision in Vehicles VII Conference, Marseilles. Underwood, G., Clews, S. and Everatt, J. (1990). How do readers know where to look next? Local information distributions influence eye fixations. Quarterly Journal of Experimental Psychology, 42, 39-65. Underwood, G., Sommerville, F., Underwood, J.D.M. and Hengeveld, W. (1994). Information Technology on the Move: Technical and Behavioural Evaluations of Mobile Telecommunications. Chichester: John Wiley. Vonk, W., Radach, R. and van Rijn, H. (1997). Eye guidance and the saliency of word beginnings. Paper presented at the Ninth European Conference on Eye Movements, Ulm.

This page intentionally left blank

29

CHAPTER 2

Definition and Computation of Oculomotor Measures in the Study of Cognitive Processes Albrecht Werner Inhoff State University of New York at Binghamton and Ralph Radach Technical University of Aachen

Abstract Oculomotor measures provide distinct methodological advantages in the study of cognitive and perceptual processes. Derivation and use of these measures is not, however, straightforward. This chapter reviews methodological choice points that need to be considered when raw data are used to define basic oculomotor events, such as fixations and saccades. The chapter also considers the usage of these oculomotor events for the indexing of perceptual and cognitive processes. Particular attention is given to viewing duration measures as these are used as the primary indicator of cognitive processes in current eye movement research. The discussion is limited in that the raising of measurement-related and methodological issues is not followed by the presentation of specific solutions. Nevertheless, we consider the discussion of these issues a crucial step in the eventual development of oculomotor measurement standards.

Eye Guidance in Reading and Scene Perception/G. Underwood (Editor) © 1998 Elsevier Science Ltd. All rights reserved

30

A. W. Inhoff & R. Radach

Introduction Oculomotor measures provide distinct methodological advantages in the study of cognitive and perceptual processes. They appear sensitive to a wide range of 'cognitive processes' and they can be obtained under relatively natural task conditions; hence, the growing acceptance and popularity of these measures. However, measurement of oculomotor activity in the cognitive and perceptual sciences is far from being straightforward. There has been some discussion pertaining to the measurement, evaluation, and reporting of oculomotor data (e.g., McConkie, 1981; Heller, 1983; McConkie, Wolverton and Zola, 1984; Mueller et al., 1993; Harris, Abramov and Hainline, 1984; Nodine et al., 1992) but no measurement standards have evolved and many methodological issues have remained either unaddressed or unresolved. As a result, there is still considerable ambiguity in the definition of basic oculomotor events and the use of different types of oculomotor measures in the study of conceptually similar issues. An informal survey conducted by us (Inhoff, Radach, and Heller), to which thirty-two researchers using oculomotor measures responded, indicated that all of them believe that there is a need for increased discussion of measurement-related and methodological issues. More specifically, two thirds of them consider the definition of functional oculomotor events controversial, and more than one third wish to appraise the specification and interpretation of extant measures. This uncertainty has led to considerable differences in the specification and interpretation of oculomotor measures, hampering the appraisal, comparison, and exchange of scholarly findings. The main goal of this chapter is to provide a review of likely methodological choice points that need to be considered when oculomotor data are collected. The chapter is not concerned with questions related to the hard- and software setup, nor with the assessing of the accuracy and reliability of the eye tracker output (see McConkie, 1981, for an excellent review). Instead, our considerations focus on issues that are encountered 'down-stream' when raw data are used to define basic oculomotor events, such as fixations and saccades, and when oculomotor data are used to index perceptual and cognitive processes. We will show that different procedures can be (and have been) used to specify saccades and fixations and that the researcher using oculomotor measures has a relatively large degree of freedom in choosing — or defining — oculomotor measures of cognitive processes (see also Chapter 3 and Konieczny, 1996). Our discussion is limited in that it focuses on the use of oculomotor measures in the study of perceptual and cognitive processes during reading. We presume, however, that similar methodological issues are encountered in other domains of study. It is further limited in that discussion of measurement dilemmas is not followed by the specification of measurement solutions. Nevertheless, we consider the discussion a crucial step in the eventual development of oculomotor measurement standards.

Oculomotor measures

31

Defining fixations and saccades Measurement of basic oculomotor events, comprising fixations and saccades during text reading, is often considered straightforward, as there appears to be little ambiguity as to whether the eyes are fixated or in motion. However, a close look at the movement-time profile shown in Fig. 1 (recorded in Binghamton using a dual-Purkinje eye tracker) indicates that the saccade-fixation transition is not necessarily clear-cut. The figure shows the ending 20 ms of a fixation near character location 9 (on a line of print containing 80 characters of text), followed by a right-directed saccade of about seven character spaces. The high-velocity right directed movement is followed by a shifting to the left, with high-and low-velocity movement components extending across 8.4 and 1.2 character spaces and lasting 16 ms and 14 ms, respectively. As can be seen, the measured (right) eye appears to move too far, then shifts (or converges) to its final position. The low-velocity shift is followed by a 60 ms period during which the eye is relatively stable. After this, the trajectory shows another high- and low-velocity movement sequence, with movement sizes of 1 and 0.2 character spaces and

Fig. 1. Space-time diagram of eye movements recorded over a period of 200 ms with a sampling rate of 1000 Hz using a dual-Purkinje eye tracking system. Left-to-right eye movements correspond to a bottom-to-top change on the ordinate (see text for details).

32

A. W. Inhoff & R. Radach

movement durations of 8 ms and 4 ms, respectively. This is followed by another period during which fixation location remained 'almost' unchanged. Brief fixations which are followed by short saccades of approximately one character space are relatively common in some readers. Records from the Binghamton laboratory indicate that approximately 5-15% of their fixations follow this pattern. Figure 1 indicates that the onset of a stable fixation is preceded by two different types of oculomotor events: a high-velocity change of fixation location and a reversed-direction, lower-velocity shifting toward a resting location. Is the reversedirection movement part of the saccade? According to Deubel and Bridgeman (1995a,b), it reflects a corresponding lens movement in the dual-Purkinje's eye movement recording. During saccade offset, deceleration of the lens lags deceleration of the eyeball so that the lens movement overshoots the eyeball movement until it is 'pulled' back by its attachments to the eyeball tissue. Smaller reverse-direction movements can also be seen, however, when a pupil reflection method is used (see Fig. 3). Presumably they reflect the onset of vergence movements and, perhaps, small movements of the pupil relative to the eyeball. Deceleration of the eyeball fluids and the pupil may lag behind the deceleration of oculomotor muscle forces, possibly moving the pupil briefly in the direction of the saccade after muscle movements were terminated. Deubel and Bridgeman (1995a, p.532) suggest that the "maximum of the Purkinje tracker overshoot represents a practicable estimator for the veridical end of the saccade". More generally, it may be assumed that the time at which the eye movement reaches is largest extent coincides with the veridical end of the saccade. If useful information was obtained immediately after the veridical saccade was completed, then fixation onset should be defined as the point in time at which the recorded saccade reached its largest value. According to this view, the 14 ms and 4 ms reverse-direction shifts should be considered part of the fixation. If, however, effective perception of text was suppressed during postsaccadic movements, then fixation onset should be defined as the point in time at which the complete eyeball system was relatively stationary. According to this view, the 14 ms and 4 ms shifts should be excluded from the fixation duration. Deubel and Bridgeman (1995b) showed that brief flashes can be perceived during the lens movement, but it remains unclear whether useful information can also be gleaned from a steady image that is visible during the postsaccadic movement period. On the one hand, postsaccadic movement will smear the retinal image. On the other hand, postsaccadic movements are relatively small and principled, and the reader may be able to extract useful information during that movement period. Looking at Fig. 1, other methodological issues arise. The relatively small change in fixation location, extending across one character space, could indicate that the reader terminated the brief 74/60 ms fixation duration to seek new information. Alternatively, the small change in location may not have any functional significance

Oculomotor measures

33

Fig. 2. Space-time diagram of eye movements recorded over a period of 160 ms using an AmTech ET4 pupil reflection eye tracking system with a sampling rate of 500 Hz. The recording shows a null-saccade, consisting of a rapid lateral jerk of both eyes that extend across approximate 2 character spaces, followed by an immediate return to the previously occupied fixation location.

for the linguistic analysis of fixated text, and the short fixation period should be considered part of a single fixation duration. The movement-time profile shown in Fig. 1 thus can be used to arrive at radically different descriptions of oculomotor activity: it may depict one functional saccade of average length which was followed by a relatively long fixation duration. Alternatively, it may depict a sequence of two saccades, their mean extent being considerably shorter than the size of an average saccade, a mean fixation location that is located at a point where the eyes never went, and two fixations, their mean duration also being shorter than the duration of a typical fixation. Figure 2 shows yet another movement-time profile (also obtained in the Binghamton laboratory), according to which the eyes begin to move one to two character spaces away from a fixation but then return to the original fixation location. These displacements need not be lateral but can occur in any direction. They may constitute a functional oculomotor event; hence fixation durations prior to the 'jerk'

34

A. W. Inhoff & R. Radach

may need to be discriminated to from the fixation duration following the jerk. Alternatively, these jerks may be low level oculomotor 'irregularities' that are completely unrelated to the perceptual and linguistic analysis of fixated text. The fixation period of a target area can also be punctured by one or several blinks. Blink time can be excluded from the computation of fixation duration, as text is not visible during this period. However, it appears likely that cognitive analyses of text continue during the blink period, unless the blink occurs at the very beginning of a fixation; hence, blink time may be added to the fixation duration. Another methodological option is to exclude fixations with blinks from analysis. The methodological treatment of saccade durations adds yet a further option to the computation of fixation durations. Most researchers do not include saccadic movement in their computation of fixation durations. Yet it is plausible to assume that the cognitive analysis of a fixated segment of text continues during the saccade that follows the fixation. Hence postfixation movement time may need to be included in the fixation duration measure when it is used to measure cognitive processes. Selecting cutoff values Small microsaccades of the eyes, extending across a fraction of a character space, occur frequently when the eyes are fixated. They are, however, not necessary to maintain stable fixation (Kowler and Steinman, 1980). Interestingly, during prolonged fixation, the intersaccade interval for microsaccades is in the order of normal saccade latencies (Cunitz and Steinman, 1969). McConkie (1983) considers two possible functions of these very small movements: they could indicate fine attentional adjustments or reflect an intrinsic tendency to move the eyes after some delay has passed. So far, no consistent functional explanation has emerged and there is no empirical evidence for relations to perceptual or cognitive processes (Kowler and Steinman, 1980). When detected, small saccades may be excluded from the set of data by selecting a minimum saccade size value. Saccades smaller than the criterion value can either be considered part of a larger saccade or they can be excluded from analysis. Similarly, the functional role of brief fixations is also poorly understood. In recent studies, conducted in the Binghamton laboratory, we have excluded fixation durations of less than 50 ms from the data set, assuming that shorter duration fixations are not determined by on-line cognitive processes (e.g. Inhoff, Briihl and Schwarz, 1996). Other researchers (e.g., Vitu, O'Regan and Mittau, 1990; Hyona, Niemi and Underwood, 1989) have used cutoff values in the order of 70-100 ms. Again, fixation durations that are shorter than the criterion value can either be considered part of a preceding or following fixation or they can be excluded from analysis. Possibly, the optimal cutoff value is determined by the procedure used to

Oculomotor measures

35

compute fixation durations and by effect sizes. Cutoff values may be higher when movement time is included in the computation of fixation duration. Furthermore, relatively high cutoff values may be advantageous when linguistic processes are studied, as McConkie, Reddix and Zola (1992) showed that fixation durations of 140 ms or less are unaffected by the lexical properties of fixated text. In addition, some — but not all — researchers exclude fixation durations above and below two or three standard deviations from the mean. Use of different saccade size and fixation duration cutoff criteria may have a profound effect on the description of readers' eye movements during reading. For instance, setting the minimum fixation duration to either 50 ms or 100 ms will yield radically different descriptions of the saccade-fixation sequence depicted in Fig. 1. Cutoffs may also determine effect sizes and their reliability. In psycholinguistic studies, effect sizes that are not reliable when a 50 ms cutoff is used may be reliable when a 140 ms cutoff is used. Currently, there are no comparisons of the effectiveness of different cutoff values. Binocular measurement The prevailing view, according to which "both eyes pretty much move in synchrony with each other across the page" (Rayner and Pollatsek, 1989, p. 113), implies that binocular measurement should not create any additional measurement dilemmas. However, recent findings indicate that there can be marked saccadic and postsaccadic asymmetries in the movements of the eyes. A typical binocular asymmetry is shown in Fig. 3. Colleweijn, Erkelens and Steinman (1988) provided a detailed description of the temporal and spatial dynamics of binocular coordination of horizontal saccades in a simple scanning task and found movements of the abducting (temporally moving) eye to be somewhat larger than corresponding movements of the adducing (nasally moving) eye. Similar observations were reported by Heller and Radach (1995) for reading and other complex visual tasks. The magnitude of absolute amplitude differences between the eyes can be quite remarkable, with means in the order of 5% of total movement amplitude when large (10-12 letter) saccades are considered and of approximately 15% when small (2-3 letter) saccades are considered. During fixation, the initial disparity is reduced — but generally not offset — by a slow, relatively uniform convergence shift. When reading a line of text, the residual disparity at the end of many fixations can lead to an accumulated binocular difference in fixation position of up to several characters. The relative contribution of each eye to postsaccadic convergence is modulated by its landing position. In a simple scanning experiment both eyes converged symmetrically when they landed on the target (consisting of a parafoveally presented 0.4 degree wide rectangle). In case of under- or overshoot, however, it was

A.W.lnhoff&R.Radach

36

0

50

100

150

200

250

300

350

400

Time (in ms) Fig. 3. Space-time diagram of eye movements recorded using an AmTech ET4 pupil reflection eye tracking system running at 500 Hz. Reading direction is from bottom to top on the ordinate. The recording shows a disparity in the landing position of the regressive saccade made by the two eyes and an asymmetry in the subsequent slow vergence movement.

the eye that landed farther away from the target centre that performed the larger vergence movement, indicating that vergence is not reflexive, as commonly thought, but to a certain degree 'goal directed' (Radach, Heller and Jaschinski, 1996). Asymmetry in postsaccadic vergence indicates that fixation durations, when defined as the period of time during which the eyes are relatively stable, may differ for the two eyes. When fixations are short, instances can occur in which the fixation duration of one eye is slightly above the criterion value whereas the fixation duration of the other eye is slightly below. Other measurement dilemmas may occur. The two eyes may be fixated at different characters, requiring a choice as to which one of the two fixation locations defines the initial fixation position within a word, a measure that has been used in several studies of reading (see below). Instances may even occur in which the two eyes fixate different linguistic units, e.g., the line of regard of one eye may fall on the ending character of one unit (word, morpheme, syllable, or other) and the line of regard of the other eye may fall on the first character of the next unit.

Oculomotor measures

37

Using oculomotor viewing duration measures After fixations and saccades have been identified they can be used to index cognitive processes. For the last fifteen years, this indexing has been guided by two processing assumptions, often referred to as eye-mind and immediacy assumption (Just and Carpenter, 1980). According to these assumptions, the location of a fixation coincides with the cognitive processing of text at that location, and an ensuing fixation duration is determined by the area's (e.g., word's) perceptual and cognitive analysis. According to the two assumptions, the viewing duration of an area of text, or some derivative of it, can be used to index its perceptual and linguistic analysis. Methodological issues arise, however, when an effective viewing duration measure is to be defined. Readers may not move the eyes from word to word in the left-to-right direction (in left-to-right ordered text). Some words, notably short words, are skipped (i.e., not fixated) during sentence reading (e.g., Vitu et al., 1995). Conversely, some words, notably long words, receive more than one fixation. Sentence and passage reading often involves regressions (eye movements to previously read text) and multiple word readings. Furthermore, there is now accumulating evidence indicating that the linguistic processing of a word can be initiated before it is fixated (see Rayner and Pollatsek, 1989, for a review) and that the processing of a word can continue after it has been fixated (see Chapter 11). In the following section, we will describe several oculomotor measures that have been reported in the literature. One or several of these measures are generally used as dependent variables in experimental studies and as criterion variables in nonexperimental regression studies (e.g., Just and Carpenter, 1980). Particular attention will be given to the computation of (word) viewing durations, as processing time has been used as the primary indicator of linguistic processes in the cognitive sciences (the listing of measures is not exhaustive). To simplify the discussion, we will not consider the question of whether mean effects of experimental manipulations on viewing duration measures are produced by similar effects in all relevant cases or by combinations of unequal effects in subsets of the data (see McConkie, Zola and Wolverton, 1985 for a discussion of theoretical and methodological implications of this "frequency of effects problem"). We will further assume that the spatial properties of to-be-studied text is matched across conditions. Various types of scaling have been used when this matching is not possible, e.g., the viewing duration of a target area may be divided by the number of characters in the target area. At the outset of our discussion will be a seeming 'non-event': instances in which the word of interest (also referred to as target word) is skipped.

38

A. W. Inhoff & R. Radach

(1) Skipping rate (fixation probability being the inverse, but less common, measure) A target is considered skipped when neither the space preceding it nor any of its constituent letters receives a fixation during the initial right-directed reading of text. When computed, skipping rates generally constitute a supplementary measure that, when analysed as a function of the skipped target's visuo-spatial and linguistic properties, may provide a converging source of evidence for a particular hypothesis. Virtually all studies have shown that target length and the launching distance of a saccade from the target are the main determinants of target skipping rate, with skipping being particularly common when the target is short (Vitu et al., 1995) and when the launch site is near the target's beginning letter (see, e.g., McConkie, Kerr and Dyre, 1994, who propose a 'skipping formula' to describe skipping rates mathematically). There is also considerable evidence showing that linguistic variables modulate target skipping. When matched for length, high-frequency words are more likely to be skipped than low-frequency words (Inhoff and Topolski, 1994), especially when words contain seven or fewer characters (Rayner, Sereno and Raney, 1996), and contextually constrained words are more often skipped than unconstrained words (Rayner and Well, 1996). The fact that lexical effects on word skipping are sometimes not found in analyses of large data corpuses (e.g., Radach in O'Regan et al., 1994) may be due to the less extreme variation of word frequency and contextual constraint in natural text material as compared with experimental sentences. In an elegant meta-analysis on a large number of published studies, Brysbaert and Vitu (Chapter 6) show that although linguistic variables influence skipping, they account only for a relatively modest portion of the total variance, hence, our suggestion that they constitute a supplementary source of evidence in the examination of cognitive processes. Skipped words are generally excluded from viewing duration analyses. However, attempts have been made to estimate their processing time. This was accomplished either by setting the fixation duration of skipped words to zero (Just and Carpenter, 1978; 1980), or by basing means only on subjects with viewing durations above a certain minimum (Carpenter and Just, 1983). Another technique is to substitute pretarget viewing durations for target viewing durations when the pretarget fixation is reasonably close (e.g., within three characters of the target, Folk and Morris, 1995), or by using the fixation nearest to the target as target viewing duration, irrespective of whether the fixation is to the right or left of the target (e.g., Garrod et al., 1990, who used a pre- or post- target window of four and one character spaces, respectively). On the one hand, the assignment of processing time to skipped words will increase the number of observations that can be used for the computation of target viewing durations. The advantage may be offset, however, by methodological shortcomings. Assignment of a zero processing duration to a target word (target

Oculomotor measures

39

region) is likely to underestimate the demands of the target's linguistic analyses. Using the duration of a pre- or post-target fixation is also problematic, as it is likely to confound the target's linguistic analyses with the linguistic analysis of either preor post-target words. Currently, no data are available to determine the advantages and disadvantages of viewing duration substitution measures when target skipping occurred. (2) Single fixation durations Computation of target viewing durations is uncontroversial when the target is the recipient of a single fixation during the initial left-to-right sentence reading. Instances in which the target receives more than one fixation or instances in which the word is fixated after text to the right of it was viewed (following a regressive saccade) are excluded. The exclusion of instances in which the word is a recipient of more than one fixation from the single fixation duration measure rests on the assumption that qualitatively different processes control eye movements when a target receives a single fixation and when it receives multiple fixations (e.g., O'Regan, 1992). Single fixations that follow a regression are excluded because it remains unclear what type of information had been obtained prior to the initial skipping of the word. Computation of single fixations may not yield a meaningful measure when target refixations are common. The exclusion of a large number of target vie wings is likely to increase the variability and decrease the sensitivity of the measure. Furthermore, assumptions underlying the computation of single fixation durations remain controversial (see below). (3) The durations of the first and second of two target fixations Instances in which the target receives more than one fixation, which is common when long words are encountered, can be used to determine whether initial fixations on a target are functionally distinct from subsequent fixations. Differential effects of linguistic target characteristics on the first fixation duration and on refixation durations would recommend the separation of instances in which the target receives a single fixation and of instances in which the target receives more than one fixation. Comparison of first and second fixation durations, as a function of the location of these fixations on a target word reveals a distinct tradeoff. Initial fixations that occupy beginning or ending word positions tend to be relatively short and fixations near the word centre are relatively long. Conversely, the duration of the second fixation is relatively short when the first had landed near the target centre and increases as first fixation position is shifted toward the target's beginning or ending letters (O'Regan, Pynte and Coeffe, 1986; O'Regan et al., 1994; Pynte, Kennedy and Murray, 1991). However, it remains unclear whether this tradeoff is due to the computation of functionally distinct linguistic representations during a target's initial fixation and its refixation.

40

A. W. Inhoff & R. Radach

Hyqna, Niemi and Underwood (1989) examined this issue by manipulating the degree to which a long word's beginning and ending letter sequences constrained the set of compatible candidate words (this was referred to as informativeness). Their results showed an effect of informativeness on the duration of the second but not of the first fixation during sentence reading. In contrast to this, informativeness determined the duration of initial fixations, especially when the eyes landed near the word's beginning, in Pynte, Kennedy, and Murray's (1991) study. Taken together the two studies suggest that neither first nor second fixations are 'code specific'. This conclusion is consistent with Rayner et al.'s recent (1996) findings. In their analyses effects of word frequency were determined separately for the first and second of two target fixations (when the target received two fixations), and for single fixation durations (when the target received a single fixation). The comparison revealed virtually identical effect sizes, amounting to 29, 31, and 25 ms, in the first and second of two intratarget fixation and in the single fixation durations, respectively. These results indicate that the separation of single fixation from multiple target fixations may not be empirically justified. Still, this conclusion must be considered tentative, as word properties other than word frequency and 'informativeness' may have differential effects on the first and second of two target fixations. (4) First fixation durations First fixations comprise the duration of the first fixation on a target word during first pass sentence reading, irrespective of whether the target word receives one or several fixations. It thus provides a measurement of word viewing durations whenever the target is a recipient of a fixation. The measure was first used by Inhoff (1984) and is now one of the most commonly reported linguistic measures. It is sensitive to a range of linguistic computations including orthographic properties (Lima and Inhoff, 1985), phonological properties (Inhoff and Topolski, 1994; Pollatsek et al., 1992), lexical properties (Inhoff and Rayner, 1986; Rayner and Duffey, 1986), metaphorical status (Inhoff, Lima and Carrol, 1984) and contextual constraints (Inhoff, 1984). The measure has been criticised, however, for confounding instance in which the target receives a single fixation and instances in which it receives multiple fixations (O'Regan, 1990; 1992). For the criticism to be meritorious, it has also to be shown that single fixation and the first of several target fixations are sensitive to different types of linguistic computations during sentence reading, which does not appear to be the case (see above). (5) Gaze durations This measure was first used by Just and Carpenter (1978; 1980) and is perhaps still the most commonly reported index of cognitive processes. Computationally, gaze

Oculomotor measures

41

durations are defined as the time spent viewing a target word until another word is fixated. Similar to first fixation durations, gaze durations include all instances in which the target was fixated. Similar to first fixation durations, gaze durations have been shown to be sensitive to a wide range of pre-lexical, lexical, and post-lexical word properties. One typical computation of gaze durations is to cumulate all the viewing durations on a word during its initial reading. Since the measure is to index cognitive processes, it could be argued that the cumulation should also include the time spent moving the eyes within a word, and perhaps even the time spent moving the eyes from the target word to the next word in the text. Currently, there are no data available that compares the diagnostic value of different types of gaze computations. It appears unlikely that the inclusion of movement time in the gaze measure will lead to notable changes in effect sizes as movement time appears to constitute a small fraction of the time spent on a target word (Blanchard, 1985). (6) Mean fixation durations A mean fixation duration is defined as the mean duration of all fixations that occupy a target during its first pass reading. Mean fixation durations were computed in some of the earlier studies of reading where they were found to be sensitive to linguistic processes, yielding, for instance, shorter durations when a contextually constrained target was viewed than when an unconstrained target was viewed (e.g., Ehrlich and Rayner 1981). Like other oculomotor measures, mean fixation durations confound instances of single and multiple target fixations. However, in contrast to first fixation durations and gaze durations, this confounding can lead to a serious distortion of experimental effects. For instance, words that are difficult to recognise are likely to receive more than one fixation (e.g., O'Regan et al., 1994; Rayner, Sereno and Raney, 1996; Radach, Heller and Inhoff, 1997; Vitu et al., 1995); hence, large effects of difficulty on gaze durations. However, since initial fixation durations and refixation durations also tend to be shorter than single fixation durations, the mean of multiple fixations of difficult targets can be shorter than the mean of single fixations of easy targets. (7) Total viewing durations Total viewing durations are computed by cumulating all fixation durations on a target, irrespective of whether it is fixated during first pass sentence reading or not. Total target viewing durations were first used by Inhoff (1983) to determine effects of text structure. Recently, they have also been used to determine the linguistic processing of individual words (e.g., Daneman, Reingold and Davidson, 1995; Perea and Pollatsek, 1998). Interpretation of the measure appears to be difficult. Perea and Pollatsek (1998) suggest that it may be sensitive to lexical analyses that may occur relatively late in

42

A. W. Inhoff & R. Radach

the word recognition process. According to Daneman et al. (1995), the measure is particularly sensitive to processes that occur after a word has been identified. Difficulties in the interpretation of the measure may be avoided by distinguishing between 'forward gaze durations1, i.e., gazes that accrue during first pass reading, and 'regressive gaze durations', i.e., gazes that accrue during second and subsequent word readings (see also Carpenter and Daneman, 1981). First pass reading can then be related to the ease of the word's linguistic analysis, and regressive gazes can be related to another linguistic process, e.g., the context-dependent adjustment of word meaning (Carpenter and Daneman, 1981) or the re-assignment of syntactic structure (Rayner and Frazier, 1987; 1989). (8) Repair time (or total repair time) Repair time is another example of an oculomotor measure that has been used to study 'complex' linguistic processes that follow the accessing of word meanings. It is computed by cumulating all consecutive fixations in between the first and last fixation on a word, irrespective of whether they are first or second pass readings and irrespective of whether these fixations are on the target or not (Daneman et al., 1995; Daneman and Reingold, 1993; Daneman and Stainton, 1991). For instance, the total repair time on the false homophone 'hare' in the phrase 'she had her long hare permed and ....' would comprise fixations 1 to 7, given the following fixation sequence: the initial fixation of 'hare' (1) followed by 'long' (2), 'had' (3), 'hare' (4), 'long' (5), 'permed' (6), 'hare' (7), 'blowing' (8), and no further fixations on 'hare'. According to Daneman et al. (1995) repair time is particularly sensitive to post-recognition processes that are invoked to discern sentence meaning when a context-inconsistent word is encountered ('hare' in the example). Consistency of experimental effects across measures Effect sizes may also differ across viewing duration measures. For instance, effects of a word's linguistic properties on total repair time may not be evident in first fixation or gaze durations. Since repair time includes the time spent re-reading the target and surrounding text, this discrepancy of effects may indicate that a particular linguistic property becomes relevant after the target's lexical representation had been activated (see also Chapter 3 for a discussion of discrepancies between viewing duration measures). In other instances, discrepancies between measures are more difficult to reconcile. Inhoff (1984) obtained different effects of contextual constraints on first fixation durations and gaze durations. Other studies reported effects of linguistic manipulations on first fixation durations but not on gaze durations (Inhoff and Rayner, 1986; Lima and Inhoff, 1985; Pollatsek et al., 1992) or the reversed pattern of effects with reliable effects in the gaze durations but not first fixation durations

Oculomotor measures

43

(Balota, Pollatsek and Rayner, 1985; Kennison and Clifton, 1995; Sereno and Rayner, 1992). In view of Rayner et al.'s (1996) findings, and other recent results, effect sizes should generally be larger when gazes are measured than when first fixations are measured, as effect size should increase with the number of fixations on a fixated target. From this perspective, effects of first fixation durations in the absence of robust effects on gaze durations appear particularly problematic. However, the inclusion of intra-word refixations in the gaze measure is likely to increase the variability of this measure. Consequently, effects that are robust in the first fixation durations may not be robust in the gaze durations.

Movement-related measures The reporting of viewing durations is often accompanied by the reporting of one or several of the following movement measures: the size of saccades that lead to a target fixation, the landing position(s) on the target, refixation rates, the size of the saccade that led to a target's refixation, the position of the refixation, and the size of the saccade leaving the target. A saccade to a target can be followed by a single fixation or by the first of several target fixations. Hence, similar to the viewing duration measures, different types of saccade size measure can be computed. Eye movements that lead to the re-reading of text can be measured and reported as regression rate, regression size, and regressive fixation location. Progressive and regressive saccades are usually reported in character units and landing position in terms of the fixated character position (for a specified word length). In spite of this plethora of potential measures, we confine our discussion to two measures: fixation locations and refixation rates (see Chapters 4, 6, 9 and 11 for a more detailed discussion of movement related processes and measures). Fixation position measures may be of diagnostic value for the study of cognitive processes. As noted before, there is now some consensus that eye movements resulting in target skipping are co-determined by the linguistic analysis of the skipped word, but the more general claim, that the planning of saccades to a subsequently fixated word is co-determined by linguistic analyses has remained controversial. The original claim made by Underwood and co-workers (Underwood, Bloomfield and Clews 1988; Underwood, Clews and Everatt, 1990) of an effect of parafoveally available semantic or lexical information on initial landing positions was contradicted failures to replicate (e.g., Rayner and Morris 1992, Radach et al., 1995). On the other hand, there is now evidence for small effects on the orthographic level (Inhoff, 1989), indicating that very unusual letter clusters may lead to modulations of saccade amplitudes (Hyona, 1995, for details, see Underwood, this volume). Inhoff, Briihl and Schwartz (1996; see also Hyona et al., 1989) also obtained larger saccades and a right-shifted fixation location when

44

A.W. Inhoff& R. Radach

morphologically complex targets were fixated than when length-matched monomorphemic control words were fixated. Some researchers have also recommended measurement of fixation locations because the linguistic analysis of a word may be a function of its initial fixation location (O'Regan and Jacobs, 1992). According to this view, any study of linguistic processes in which target viewing durations are not analysed as a function of target fixation location is likely to confound linguistic effects with fixation location effects. Even though the location of a fixation affects the duration of the ensuing fixation, there is currently no support for the view that this confounds linguistic processes. On the contrary, O'Regan et al. (1994) and Rayner, Sereno and Raney (1996), who determined the size of the word frequency effect as a function of fixation location, found that the two effects were completely additive; similarly, effects of morphological complexity on first fixation durations and gaze durations were completely independent of fixation location in our recent work (Inhoff, Briihl and Schwartz, 1996). Refixation rates constitute a movement related measure that has proven to be sensitive to cognitive variables. They are generally defined as the ratio of the frequency of multiple target fixations over the frequency of multiple and single target fixations during first pass reading. Refixations are of profound methodological importance, since the main constituent of linguistic gaze duration effects is the increase in refixation rate rather than a prolongation of individual fixation durations (Blanchard, 1989; O'Regan, 1990). Examination of refixation rates has yielded a number of important insights into oculomotor activity during reading. Specifically, refixation rates are relatively low when the initial fixation is near the target word centre (e.g., McConkie et al., 1989; O'Regan et al., 1994) and increases as the eyes move off the centre. Refixation rates are also sensitive to the linguistic properties of fixated words, occurring more often when a low-frequency word is fixated than when a high-frequency word is fixated. Refixations may be related to first fixation durations. First fixation durations on a difficult item may, for instance, be shorter than first fixations on a easy item but when this occurs, the difficult item is more likely to be refixated (DeGraef, personal communication). McConkie et al. (1989) have first described the relation between initial fixation position and refixation rate in terms of a U-shaped quadratic function, where word frequency influences the height of the curve but not its steepness. This dissociation of visuomotor and lexical influences on refixations has subsequently been confirmed by others (Vitu, O'Regan and Mittau, 1990; O'Regan et al., 1994; Rayner, Sereno and Raney, 1996). Radach and McConkie (Chapter 3) show that the positioning of refixations within words is guided by the same mechanisms as for initial fixations. Cognitive variables like word frequency have a substantial effect on the proportion of refixations within words but very little influence on refixation positions

Oculomotor measures

45

(Hyona, Niemi and Underwood, 1989; Radach, Inhoff and Heller 1997). Differing results in Pynte, Kennedy and Murray (1991) and Pynte (1996) may be due to the fact that in these studies initial fixation position was imposed experimentally. In view of our earlier discussion of binocular activity, it should be noted that vergence velocity may be co-determined by cognitive processes. Hendriks' (1996) recent results point to the possibility that vergence velocity during reading increases as the difficulty of a cognitive task increases.

The eye-mind and immediacy assumptions reconsidered Following McConkie and Rayner (1975) and Rayner (1975), a large number of studies have shown that useful linguistic information is obtained from the fixated word and from the next (parafoveal) word in the text (see Rayner and Pollatsek, 1989, for a review). Hence, the linguistic analysis of a target does not commence with its fixation but before that. Processing of a target is thus spatially distributed, making it often difficult to ascertain the extent to which a word is analysed while it is fixated and before it was fixated. Though there have been some attempts to provide measurement solutions to this problem (McConkie et al., 1979), effects of prefixation analyses on target viewing durations are generally neglected, the justification being that the linguistic analysis of a fixated word is independent of the linguistic analysis of the next word in the text and that the duration of a target word's (parafoveal) analyses, i.e., its analysis prior to fixation, constitutes a constant (Morrison, 1984; Inhoff et al., 1989). These assumptions may not be warranted, however. Recent results from the eye-movement laboratory at Binghamton (Inhoff, Briihl, and Starr, in preparation) indicate that viewing durations of a fixated target word can be affected by its semantic relationship to the next word in the text. In the critical experiment, readers fixated a word that was followed either by an unassociated word (e.g., fixation of 'mother's' followed by 'garden') or by an associated word ('mother's' followed by the fixation of 'father'). Under these conditions, fixation durations on the identical target words (mother's) were shorter in the associated condition, presumably because of parafovea-to-fovea cross-talk. Kennedy (Chapter 7) also obtained parafoveal-tofoveal cross-talk in a reading-like word recognition paradigm when fixated and parafoveal words shared word-initial orthographic letter sequences. Other results indicate that the duration of a target word's parafoveal analyses is variable, and that the extent to which a parafoveally visible target is analysed is determined by the ease of linguistic analysis during the pretarget fixations. For instance, readers obtained less useful linguistic information from a parafoveally visible target when a lowfrequency word was fixated than when a high-frequency word was fixated (Henderson and Ferreira, 1990; Kennison and Clifton, 1995). Eye movements

46

A. W. Inhoff & R. Radach

during target viewing can thus be affected by the target's prior parafoveal analysis and by parafovea-to-fovea crosstalk when the target is subsequently fixated. Conversely, readers may seek linguistic information from a previously fixated word. In one recent study (Briihl and Inhoff, 1995), eye-movement-contingent display changes were used to control the parafoveal visibility of a seven-character target word so that it was either fully visible, partially visible, or completely masked. In all instances, the target was completely visible when it was either directly fixated or when the eyes moved to the right of it. Under these conditions, readers skipped approximately 10% of the targets, irrespective of whether a target had been fully visible, partially visible, or fully masked prior to its skipping. Furthermore, inspection of oculomotor activity during the reading of the posttarget sentence fragment indicated that readers rarely — if ever — moved the eyes back to target words, even when they had been partially or fully masked prior to their skipping. Nevertheless, they were able to report sentence content when asked to do so. These observations suggested to us that readers obtained visual information from the word to the left of fixation and that word skipping was not conditional on the prior identification of the next word in the text. This view was corroborated by results of a recently completed experiment (Inhoff, Radach and Starr, in preparation; see also Binder, Pollatsek and Rayner, in preparation). In our experiment, target sentences were used that contained a critical 3-, 4-, or 5-character word pair. The two words of a pair were identical except for their first letter, e.g., 'hat' and 'mat'. During reading, one member of the pair, e.g., 'hat', was visible to the right of fixation until it was either fixated or skipped. As soon as the eyes moved to the right of the critical word, an eye-movement-contingent display change was implemented that replaced the previously visible member of the pair, 'hat', with its alternate, 'mat'; i.e., the identity of the critical word was changed after it had been fixated. After sentence reading was completed, readers were asked to identify the critical word(s) by choosing among three visually similar candidates, e.g., 'hat', 'mat', 'cat'. If readers obtained useful information solely to the right of fixation, then they should always select the word that had been visible to the right of fixation, hat in the example. However, the corresponding decision was made on only 60% of the trials. Furthermore, if readers obtained useful information only when the target was either fixated or to the right of fixation, then selection of the two remaining alternatives, mat or cat, in the example, should be equally likely. However, this was not the case. They selected the word that was visible after the critical word had been fixated — or skipped — on approximately 30% of the trials; the word that had never been shown during sentence reading was selected on approximately 10% of the trials. Converging evidence comes from an extensive corpus analysis of saccade landing positions. Radach and Kempe (1993) have shown that the initial saccade landing position on a word is substantially shifted to the left if a preceding short word was

Oculomotor measures

47

skipped. To account for this effect, Radach (1996) plotted distributions of fixation positions on the last three and the first three letters of two consecutive words. In the case of two medium-sized (5-7 letter) words the distribution was U-shaped with a minimum fixation frequency near the empty space between the two words. When the transition from a 3-letter word to a 5-7 letter word was plotted, the distribution was inverted U-shaped and did not show a local minimum. A possible explanation for this pattern of data is that specific parafoveal word combinations (e.g. short and long word, a parafoveal pattern likely to represent a noun phrase) can form a unified saccade target. Given a distribution of saccade landing positions over such a two-word target, readers may regularly obtain useful information to the left of the 'fixated' word within the target region. Other oculomotor effects have been reported that complicate the computation of target-specific oculomotor measures, be they first fixation durations, gaze durations, single fixation durations, regressive gaze durations, or some other measure. Fixation durations that precede the skipping of a word have been claimed to be longer than fixation durations that occur when the next word in the text is fixated (Pollatsek, Rayner and Balota, 1986; Hogaboam, 1983; see, however, McConkie, Kerr and Dyre, 1994, who did not obtain such an effect). If fixation durations that precede skipping were increased, then gaze durations and fixation durations for words that are followed by word skippings would overestimate processing demands. Conversely, spillover effects have been noted in the literature, where effects of the linguistic analysis of one word are found in the fixation duration of the following word (Balota et al., 1985). When this occurs, gaze durations and fixation durations will underestimate the demands of cognitive/linguistic analyses. A recent study (Perea and Pollatsek, 1998) reported a particularly striking example of spillover effects. Specifically, post-target fixation durations — but not target fixations — were longer when the previously fixated target word had an orthographic neighbour with a higher frequency of occurrence than when all of its orthographic neighbours had lower word frequency counts. Use of eye-movement-contingent display changes While eye movement measures provide distinct advantages over more traditional measures they are also subject to one particular methodological disadvantage: the reader rather than the experimenter has control over the viewing of text. This disadvantage can be offset, however, by the use of eye-movement-contingent display changes, pioneered by McConkie and Rayner (1975) and Rayner (1975). In this procedure, the location of a fixation is determined and used to show an experimenter defined segment of text. For instance, the method has been used to control the visibility of parafoveally available text by showing the reader the

48

A. W. Inhoff & R. Radach

currently fixated word and by masking all other words in the text. During reading, the masking of each word is removed as soon as it is fixated. These eye-movementcontingent display change techniques can also be used to control the visibility of fixated text (Rayner et al., 1981), the temporal availability of parafoveally visible or fixated text (Rayner et al., 1981; Rayner and Sereno, 1992), or fixation locations (for instance, by centering a fixated word around the current fixated location). In spite of its methodological promise, the technique has been used in relatively few laboratories. In a well-known article, O'Regan (1990) argued that the change of text displays during fixations will disrupt oculomotor activity because eyemovement-contingent display changes generate noticeable flicker and because phosphoric persistence of replaced characters decreases the contrast of subsequently shown characters. In response to these concerns, Inhoff et al. (1998) compared effects of eyemovement-contingent display changes as a function of display-change flicker frequencies. Display changes were implemented on a phosphor-free electroluminescent panel and on a conventional (phosphor-based) CRT. The results indicated that gaze durations, first fixation durations, and saccade size remained unaffected by flicker and phosphor persistence, affirming the methodological promise of the technique. Conclusion In spite of the increasing popularity and methodological promise of eye-movement measures, there are currently no measurement standards that define basic oculomotor events (fixations and saccades). Furthermore, processing assumptions that underlie the translation of oculomotor events into measures of cognitive processes are increasingly challenged. Nevertheless, there has been a distinct paucity of methodologically oriented studies, and it remains unclear whether some definitions of oculomotor events and some oculomotor measures are more effective than others. The lack of research on methodological issues is of particular concern in view of the marketing of ready-to-use oculomotor data-processing programs. The danger exists that the 'measurement solution' offered by one of these programs becomes a de facto standard merely by virtue of its ease of use, its affordability, its marketing, or its use by established laboratories. Acknowledgements Preparation of this chapter was supported by grant No 503870 from the NIMH and Grand No BMH1-CT94-1441 from the European Union under the BIOMED Programme. The authors are indebted to Dieter Heller for many helpful discussions and

Oculomotor measures

49

to Alan Kennedy, Peter De Graef and Simon Liversedge for valuable comments on earlier drafts of the manuscript.

References Balota, D.A., Pollatsek, A. and Rayner, K. (1985). The interaction of contextual constraints and parafoveal visual information in reading. Cognitive Psychology, 17, 364-390. Binder, K.F., Pollatsek, A. and Rayner, K. (in preparation). Extraction of information to the left of the fixated word in reading. Blanchard, H.E. (1985). A comparison of some processing measures based on eye movements. Acta Psychologica, 58, 1-15. Briihl, D. and Inhoff, A.W. (1995). Integrating information across fixations during reading: The use of orthographic bodies and of exterior letters. Journal of Experimental Psychology: Learning, Cognition and Memory, 21, 55-67. Carpenter, P.A. and Daneman, M. (1981). Lexical retrieval and error recovery in reading: A model based on eye fixations. Journal of Verbal Learning and Verbal Behavior, 20, 137-160. Carpenter, P.A. and Just, M.A. (1983). What your eyes do while your mind is reading. In: K. Rayner (Ed.), Eye Movements in Reading. Perceptual and Language Processes. New York: Academic Press, pp. 275-307. Collewijn, H., Erkelens, C.J. and Steinman, R.M. (1988). Binocular co-ordination of human horizontal saccadic eye movements. Journal of Physiology, 404, 157-182 Cunitz, R.J. and Steinman, R.M. (1969). Comparison of saccadic eye movements during fixation and reading. Vision Research, 9, 683-693. Daneman, M. and Reingold, E. (1993). What eye fixations tell us about phonological recording during reading. Canadian Journal of Experimental Psychology, 47, 153-178. Daneman, M., Reingold, E.M., and Davidson, M. (1995). Time course of phonological activation during reading: Evidence from eye fixations. Journal of Experimental Psychology: Learning, Memory and Cognition, 21, 884-898. Daneman, M. and Stainton, M. (1991). Phonological receding in silent reading. Journal of Experimental Psychology: Learning, Memory and Cognition, 17, 618-632. Deubel, H. and Bridgeman, B. (1995a). Fourth Purkinje image signals reveal eye-lens deviations and retinal image distortions during saccades. Vision Research, 35, 529-538. Deubel, H. and Bridgeman, B. (1995b). Perceptual consequences of ocular lens overshoot during saccadic eye movements. Vision Research, 35, 2897-2902. Ehrlich, K. and Rayner, K. (1981). Contextual effects on word perception and eye movements during reading. Journal of Verbal Learning and Verbal Behavior, 20, 641-655. Folk, J.R. and Morris, R. (1995). Multiple lexical codes in reading: Evidence from eye movements, naming time, and oral reading. Journal of Experimental Psychology: Learning, Memory and Cognition, 21, 1412-1429. Garrod, S.O., O'Brien, E.J., Morris, R. and Rayner, K. (1990). Elaborative inferencing as an active or passive process. Journal of Experimental Psychology: Learning, Memory and Cognition, 16, 250-257.

50

A. W. Inhoff & R. Radach

Harris, C.M., Abramov, I. and Hainline, L. (1984). Instrument considerations in measuring fast eye movements. Behavior Research Methods, Instruments and Computers, 16, 341-350. Heller, D. (1983). Problems of on-line processing of EOG-data in reading. In: R. Groner, C. Menz, D.F. Fisher and R.A. Monty, (Eds.), Eye Movements and Psychological Functions: International Views. Hillsdale, NJ: Erlbaum, pp. 43-52. Heller, D. and R. Radach (1995). Binocular coordination in complex visual tasks. Perception 24, Suppl., 72. Henderson, J.M. and Ferreira, F. (1990). Effects of foveal processing difficulty on the perceptual span in reading: Implications for attention and eye movement control. Journal of Experimental Psychology: Learning, Memory, and Cognition, 17,417-429. Hendricks, A. (1996). Vergence eye movements during fixations in reading. Acta Psychological, 131-151. Hogaboam, T.W. (1983). Reading patterns in eye movement data. In: K. Rayner (Ed.), Eye Movements in Reading: Perceptual and Language Processes. New York: Academic Press. Hyona, J., Niemi, P. and Underwood, G. (1989). Reading long words embedded in sentences: Informativeness of word halves affects eye movements. Journal of Experimental Psychology: Human Perception and Performance, 15, 142-152. Hyona, J. (1995). Do irregular letter combinations attract reader's attention? Evidence from fixation locations in words. Journal of Experimental Psychology: Human Perception and Performance, 21, 68-81. Inhoff, A.W. (1983). Attentional strategies during the reading of short passages. In: K. Rayner (Ed.), Eye Movements in Reading: Perceptual and Language Processes. New York: Academic Press, pp. 181-192. Inhoff, A.W. (1984). Two stages of word processing during eye fixations in the reading of prose. Journal of Verbal Learning and Verbal Behavior, 23, 612-624. Inhoff, A.W., Lima, S.D. and Carroll, P.J. (1984). Effects of context on metaphor comprehension. Memory and Cognition, 12, 558-567. Inhoff, A.W., Briihl, D. and Schwarz, J. (1996). Compound words effects differ in reading, on-line naming, and delayed naming tasks. Memory and Cognition, 24, 466—476. Inhoff, A.W., Briihl, D., and Starr, M. (in preparation). Parafovea-to-fovea priming during eye fixations in reading. Inhoff, A.W., Starr, M., Liu, W., and Wang, J. (in preparation). Eye-movement-contingent display changes are not compromised by flicker and phosphor persistence. Psychonomic Bulletin and Review. Inhoff, A.W., Pollatsek, A., Posner, M.I. and Rayner, K. (1989). Covert attention and eye movements during reading. Quarterly Journal of Experimental Psychology, 4 la, 63-89. Inhoff, A.W. and Rayner, K. (1986). Parafoveal word processing during eye fixations in reading: Effects of word frequency. Perception and Psychophysics, 40, 431—439. Inhoff, A.W. and Topolski, R. (1994). Use of phonological codes during eye fixations in reading and in on-line and delayed naming tasks. Journal of Memory and Language, 33, 689-713. Inhoff, A.W., Radach, R., and Starr, M. (in preparation). Use of information to the left of fixation during reading.

Oculomotor measures

51

Just, M.A. and Carpenter, P.A. (1978). Inference processing during reading: Reflections from eye fixations. In: J.W. Senders, D.F. Fisher, R.A. Monty (Eds.), Eye Movements and the Higher Psychological Functions. Hillsdale, NJ: Erlbaum. Just, M.A. and Carpenter, P.A. (1980). A theory of reading: From eye fixations to comprehension. Psychological Review, 87, 329-354. Kennison, S.M. and Clifton, C. (1995). Determinants of parafoveal preview benefit in high and low working memory capacity readers: Implications for eye movement control. Journal of Experimental Psychology: Learning, Memory and Cognition, 21(1), 68-81. Konieczny, L. (1996). Human sentence processing: a semantics-oriented parsing approach. IIG-Reports 3/96, Universitat Freiburg, Institut fur Informatik und Gesellschaft. Kowler, E. and Steinman, R.M. (1980). Small saccades serve no useful purpose. Vision Research, 20, 273-276. Lima, S.D. and Inhoff, A.W. (1985). Lexical access during eye fixations in reading: Effects of word initial letter sequence. Journal of Experimental Psychology: Human Perception and Performance, 13, 272-285. McConkie, G.W. (1981). Evaluating and reporting data quality in eye movement research. Behavior Research Methods and Instrumentation, 13, 97-106. McConkie, G.W. (1983). Eye Movements and Perception during Reading. In: K. Rayner (Ed.), Eye Movements in Reading: Perceptual and Language Processes. New York: Academic Press. McConkie, G.W., Hogaboam, T.W., Wolverton, G.S., Zola, D. and Lucas, P.A. (1979). Toward the use of eye movements in the study of language processing. Discourse Processes, 2, 157-177. McConkie, G.W., Kerr, P.W. and Dyre, B.P. (1994). What are 'normal' eye movements during reading: Toward a mathematical description. In: J. Ygge and G. Lennerstrand (Eds.), Eye Movements in Reading. Oxford: Elsevier, pp. 315-327. McConkie, G.W. and Rayner, K. (1975). The span of the effective stimulus during reading. Perception and Psychophysics, 17, 578-586. McConkie, G.W., Reddix, M.R., and Zola, D. (1992) Perception and cognition in reading: Where is the meeting point. In: K. Rayner (Ed.), Eye Movements and Vision Cognition: Scene Perception and Reading New York: Springer, pp. 293-303. McConkie, G.W., Wolverton, G.S. and Zola, D. (1984). Instrumentation considerations in research involving eye-movement contingent stimulus control. In: A.G. Gale and F. Johnson (Eds.), Theoretical and Applied Aspects of Eye Movement Research. Amsterdam: Elsevier/North Holland. McConkie, G.W., McConkie, G.W., Zola, D. and Wolverton, G.S. (1985). Estimating frequency and size of effects due to experimental maniplations in eye movement research. In: R. Groner, G.W. McConkie and C. Menz (Eds.), Eye Movements and Human Information Processing. Amsterdam: Elsevier, pp. 137-147. Morrison, R.E. (1984). Manipulation of stimulus onset delay in reading: Evidence for parallel programming of saccades. Journal of Experimental Psychology: Human Perception and Performance, 10, 667-682. Mueller, P.U., Cavegn, D., d'Ydewalle and Groner, R. (1993). A comparison of a new limbus tracker, corneal reflection technique, Purkinje eye tracking and electro-oculography. In: G. d'Ydewalle and Van Rensbergen, J. (Eds.), Perception and Cognition.

52

A. W. Inhoff & R. Radach

Advances in Eye Movement Research. Amsterdam: Elsevier/North Holland. Nodine, C.F., Kundel, H.L., Toto, L.C. and Krupinski, E.A. (1992). Recording and analyzing eye-position data using a microcomputer workstation. Behavior Research Methods, Instruments and Computers, 24, 475-485. O'Regan, J.K. (1990). Eye movements and reading. In: E. Kowler (Ed.), Eye Movements and Their Role in Visual and Cognitive Processes: Reviews of Oculomotor Research. Amsterdam: Elsevier. O'Regan, J.K. (1992). Optimal viewing position in words and the strategy-tactics theory of eye movements in reading. In: K. Rayner (Ed.), Eye Movements and Visual Cognition: Scene Perception and Reading. New York: Springer. O'Regan, K., Pynte, J. and Coeffe, C. (1986). Comment le regard explore un mot isole (How does the eye inspect isolated words?). Bulletin de Psychologic, 39, 7-10. O'Regan, J.K. and Jacobs, A.M. (1992). Optimal viewing position effects in word recognition: A challenge to current theory. Journal of Experimental Psychology: Human Perception and Performance, 18, 185-197. O'Regan, J.K., Vitu, R, Radach,R., and Kerr, P.W. (1994). Effects of local processing and oculomotor factors in eye movement guidance in reading. In: J. Ygge and G. Lennerstand (Eds.), Eye Movements and Reading. Oxford: Pergamon, pp. 329-348. Perea, M. and Pollatsek, A. (in preparation). The effects of neighborhood frequency in reading and lexical decision. Journal of Experimental Psychology: Human Perception and Performance. Pollatsek, A., Rayner, K. and Balota, D.A. (1986). Inferences about eye movement control from the perceptual span in reading. Perception and Psychophysics 40, 123-130. Pollatsek, A., Lesch, M., Morris, R. and Rayner, K. (1992). Phonological codes are used in integrating information across saccades in word identification and reading. Journal of Experimental Psychology: Human Perception and Performance, 18, 148-162. Pynte, J. (1996). Lexical control of within-word eye movements. Journal of Experimental Psychology: Human Perception and Performance, 22, 958-969. Pynte, J., Kennedy, A. and Murray, W.S. (1991). Within-word inspection strategies in continuous reading: Time course of perceptual, lexical and contextual processes. Journal of Experimental Psychology: Human Perception and Performance, 17, 458-470. Radach, R. (1996). Blickbewegungen beim Lesen. Psychologische Aspekte der Determination von Fixationspositionen. (Eye movements in reading. Psychological aspects of fixation position control). Miinster/New York: Waxmann. Radach, R., Heller, D. and W. Jaschinkski (1996). Binocular coordination, fixation disparity and ocular dominance. Perception, 25, Suppl., 67. Radach, R., Heller, D. and Inhoff, A. (1997). Blickbewegungen and kognitive Prozesse: Stand und Perspektiven (eye movements and cognitive processes: current issues and developments). In: H. Mandl (Ed.), Bericht ueber den 40. Kongress der Deutschen Gesellschaft fuer Psychology (Proceedings of the 40th Congress of the German Psychological Society). Goettingen: Hogrefe. Radach, R. and Kempe, B. (1993). An individual analysis of initial fixation positions in reading. In: G. d'Ydewalle and J. van Rensbergen (Eds.), Perception and Cognition. Advances in Eye Movement Research. Amsterdam: Elsevier/North-Holland, pp. 213225.

Oculomotor measures

53

Radach, R., Krummenacher, J. Heller, D. and J. Hofmeister (1995). Individual eye movement patterns in word recognition: perceptual and linguistic factors. In: J. Findlay et al. (Eds.), Eye Movement Research. Amsterdam: Elsevier/North-Holland. Rayner, K. (1975). The perceptual span and peripheral cues in reading. Cognitive Psychology, 7, 65-81. Rayner, K. and Duffey, S.D. (1986). Lexical complexity and fixation times in reading: Effects of word frequency, verb complexity and lexical ambiguity. Memory and Cognition, 14,191-201 Rayner, K. and Frazier, L. (1987). Parsing temporarily ambiguous complements. Quarterly Journal of Experimental Psychology, 39a, 657-673. Rayner, K., Inhoff, A.W., Morrison, R., Slowiaczek, M.L. and Bertera, J.B. (1981). Masking of foveal and parafoveal vision during eye movements in reading. Journal of Experimental Psychology: Human Perception and Performance, 7, 167-179. Rayner, K. and Morris, R. (1992). Eye movement control in reading: Evidence against semantic preprocessing. Journal of Experimental Psychology: Human Perception and Performance, 18, 163-172. Rayner, K. and Pollatsek, A. (1989). The Psychology of Reading. Englewood Cliffs, NJ: Prentice Hall. Rayner, K., Sereno, S.C. and Raney, G.E. (1996). Eye movement control in reading: A comparison of two types of models. Journal of Experimental Psychology: Human Perception and Performance, 22 (5), 1188-1200. Rayner, K. and Well, A. (1996). Effects of contextual constraints on eye movements in reading. Psychonomic Bulletin and Review, 3, 504-509. Sereno, S.C., and Rayner, K. (1992). Fast priming during eye fixations in reading. Journal of Experimental Psychology: Human Perception and Performance, 18, 173-184. Underwood, G., Bloomfield, R. and Clews, S. (1988). Information influences the pattern of eye fixations during sentence comprehension. Perception, 17, 267-278. Underwood, G., Clews, S. and Everatt, J. (1990). How do readers know where to look next? Local information distributions influence eye fixations. Quarterly Journal of Experimental Psychology, 42a, 39-65. Vitu, F., O'Regan, J.K. and Mittau, M. (1990). Optimal landing position in reading isolated words and continuous text. Perception and Psychophysics, 47, 583-600. Vitu, F., O'Regan, J.K., Inhoff, A.W. and Topolski, R. (1995). Mindless reading: Eye movement characteristics are similar in scanning letter strings and reading text. Perception and Psychophysics, 57, 352-365.

This page intentionally left blank

55

CHAPTER 3

Eye Movements and Measures of Reading Time Simon P. Liversedge University of Durham Kevin B. Paterson University of Nottingham and Martin J. Pickering University of Glasgow

Abstract In this chapter, we consider the use of reading time measures that sum fixation durations in order to gain an understanding of the difficulty experienced when reading texts. We draw a distinction between two approaches to summing the duration of fixations. One approach is to sum the duration of fixations that are spatially contiguous in the text, meaning that the fixations neighbour each other in a specified region of space. The other approach is to sum fixations that are temporally contiguous, meaning that they occur in a sequence over a specified period of time. It is argued that both types of reading time measure are needed if the experimenter is to understand the time course of the influence of a linguistic variable on readers' processing of text. We first discuss a number of hypothetical eye movement records in order to illustrate the differential sensitivities of qualitatively different eye movement measures. We then report an eye movement experiment investigating how people process reduced relative clause sentences with and without the focus operator only in order to examine the utility of different reading time measures. The results showed that measures summing temporally contiguous fixations can make an important contribution to the experimenters' understanding of the precise pattern of eye movements which occur when a problem is encountered in the text.

Eye Guidance in Reading and Scene Perception/G. Underwood (Editor) © 1998 Elsevier Science Ltd. All rights reserved

56

S.P. Liversedge, K.B. Paterson & M.J. Pickering

Introduction Over the last century researchers have monitored people's eye movements in order to investigate various aspects of vision. One area in which eye movement methodology has proved of particular benefit is that of reading research. Psychologists working in this area often conduct experiments in which they record the eye movements subjects make as they read text (usually from a computer screen). They then compute from the eye movement record how long people took to read different portions of the text and use this information to draw inferences about the underlying psychological processes. Importantly, there is an underlying assumption that, on the whole, there is a close correspondence between the pattern of eye movements made by a reader and the mental processes needed to understand the text that they are currently inspecting. Thus, the direction of gaze indicates (to a large extent at least) what part of the text is being currently processed, and the time taken to process the text is indicative of the ease with which processing occurred. These assumptions are warranted by the considerable evidence that the linguistic properties of text have a direct influence on the time it takes to read that text (see Rayner and Pollatsek, 1989, for a review). However, there is not a perfect yoking of eye and mind. Research suggests that readers are able to at least partially process text that they have not yet encountered (see Balota and Rayner, 1991, for a review; also, other chapters in this volume). Furthermore, if a portion of text causes the reader difficulty, then this difficulty can "spill over" and affect processing of subsequent text (e.g. Ehrlich and Rayner, 1983). One of the major advantages of eye movement methodology over alternative reading time measures (e.g. self-paced reading) is that the experimenter can separate those fixations that were made when a region was first read from fixations that were made later in the eye movement record. This is important as it allows experimenters to determine when a characteristic of the text first influenced processing and therefore permits them to make inferences about the time course of processing during text comprehension. When text is read, the eye makes a sequence of fixations that are separated by saccadic movements. Studies have shown that information is extracted from the text during the fixation, but not during the saccades (again see Rayner and Pollatsek, 1989, for a review). Consequently, most (though not all) experimenters analysing eye movements treat the time spent fixating the text, and not the time spent making saccades, as a measure of reading time. Very often researchers sum fixations that readers make on portions of text according to several different algorithms. Thus, it is quite usual for researchers to consider a number of different reading time measures rather than considering only individual fixations. In this chapter we will focus on the different types of measures that have been widely used in psycholinguistic experiments. Our principle claim will be that reading time measures may retain the property of either spatial or temporal contiguity (or both),

Reading time measures

57

Fig. 1. Hypothetical eye movement record one.

and when analysing a reader's eye movement record it is important to consider measures which sum temporally contiguous fixations and also measures which sum spatially contiguous fixations in order to avoid failing to detect effects (Liversedge and Pickering, 1995; Liversedge, Pickering and Traxler, 1996). We will explain what we mean by spatial and temporal contiguity below and use hypothetical data to illustrate our point. We will then present data from an experiment conducted by Paterson, Liversedge and Underwood (1998) to demonstrate that simply using measures which retain the property of spatial contiguity results in a failure to detect certain effects. Figure 1 shows a hypothetical eye movement record containing a sequence of nine fixations. Each of these fixations had a duration of between 196 and 320 milliseconds, and the sum of the duration of these fixations represents the overall amount of time during spent reading the sentence. The location of each fixation is represented by a small circle and the lines between circles indicate the trajectories of saccades made between each successive fixation. The sentence is the same as one used by Paterson et al., and has been divided into six regions, each spanning one or two words of the sentence. As would be expected, the reader starts by fixating the first noun phrase of the sentence. Paterson et al. were principally interested in the time subjects spent reading the verb invited in Region Four as this is the verb that syntactically disambiguates the sentence and is likely to be the first point at which disruption might be detected. Therefore we will often refer to Region Four as the critical region. In most eye movement experiments investigating reading, researchers construct experimental sentences which are designed to cause a reader difficulty as they are read, and counterpart control sentences which are usually very similar but are constructed such that they do not cause the reader problems. The experimenter then compares reading times for regions of the experimental sentences with reading times for equivalent regions of control sentences. In this way, the experimenter is able to observe the degree of disruption to normal reading the experimental manipu-

58

S.P. Uversedge, K.B. Paterson & M.J. Pickering

lation induced. When measuring processing difficulty, researchers generally report three reading time measures: first fixation duration in a critical region along with first pass reading time and total reading time for the critical region, and other regions of interest (see Rayner et al., 1989, for a general discussion). The first fixation duration in a particular region of text, as we might expect, is simply the time the reader spent initially fixating the region. This measure is generally taken to be the very earliest point at which we might expect to see an effect due to the experimental manipulation, as this is the first time the reader has directly fixated the region in which disruption to processing is anticipated. In Fig. 1, the first fixation duration in Region Four, the critical region, is 270 ms. It is a common finding that the duration of the first fixation is sensitive to processing difficulty experienced immediately on reading that word. For example, studies have shown that subjects have a longer first fixation for words that have a low frequency of occurrence in the language than words with a high frequency of occurrence, when word length is controlled (e.g. Inhoff and Rayner, 1986; Rayner and Duffy, 1986; Raney and Rayner, 1995). Similarly, readers have a longer first fixation for a word that disambiguates a sentence in favour of an initially dispreferred syntactic analysis, as compared to the same word in an unambiguous version of the sentence (e.g. Rayner, Carlson and Frazier, 1983; Murray and Liversedge, 1994). However, the first fixation duration measure does not always provide an indication of initial processing difficulty. Quite often the region of text that is predicted to cause difficulty contains either a long word, or several words, and is likely to be fixated more than once during the first sweep of the eyes through the sentence. In such a situation all of these fixations can contribute to initial processing, in which case a more sensitive measure may be first pass reading time. First pass reading time (or gaze duration, if the region for which the measure is computed contains a single word) is defined as the sum of all the fixations made in a region until the point of fixation leaves the region either to the left or to the right. From Fig. 1 it can be seen that the gaze duration for Region Four is 513 ms (obtained by summing 270 ms and 243 ms). If effects are found on measures of either first fixation duration or gaze duration in one condition relative to another, the experimenter will usually conclude that difficulty was experienced immediately on processing that region of text. Total reading time sums all fixations made within a region of text, including those fixations made when re-reading the region. The total reading time for Region Four in Fig. 1 is 748 ms, obtained by summing the first pass fixations in the region (i.e. 270 + 243 ms), and the duration of the fixation made on the region following a regressive saccade from Region Five (i.e. 235 ms). If an effect is observed for total reading time on a region, but not for earlier measures such as first fixation duration or first pass reading time, then this is generally taken as an indication of the manipulation having a relatively late effect on processing.

Reading time measures

59

In addition to these reading time measures, researchers often provide a measure of the number of regressive saccades a reader makes from the critical region. This is usually done by computing the probability of a reader making a first pass regression (i.e. a regressive saccade that terminates the summation of fixations for first pass or gaze duration reading time). Such a measure is also usually taken to indicate that the reader is experiencing difficulty when processing the critical word. However, the probability of making a first pass regression does not provide an index of the time a reader spends on earlier portions of a sentence after making a regression. For instance, it is possible that a reader may regress to an earlier region of text an equal number of times under different conditions of an experiment but spend more time re-reading text under one condition than another. To a large extent, the reading time measures (as defined above) have been widely adopted as "industry standards" within the psycholinguistic community. However, twenty years ago there was far less standardisation of measures than there exists today. In fact, early studies frequently failed to provide unambiguous definitions of the measures which were used. In recent years, there has also been some debate concerning the relative merits of different measures of disruption to processing during reading (e.g. Kennedy et al., 1989; see also Altmann, Garnham and Dennis, 1994; Altmann, 1994; but see Rayner and Sereno, 1994a,b). However, it is important to note that all these studies conducted comparative analyses of different measures in an attempt to resolve questions regarding the interpretation of data with respect to an underlying psycholinguistic theory. By contrast, in this article we make a comparison between reading time measures in order to address a theoretical question concerning the properties of the measures themselves. To our knowledge, there are few, if any, published articles to date that explicitly set out to do this. Reporting a combination of measures, such as first fixation duration, first pass and total reading times minimises the possibility that an experimenter may fail to detect an effect. However, it does not rule out this possibility (see Konieczny, Hernforth and Scheepers, 1997). The effectiveness of this approach critically depends upon what eye movement behaviour occurs when a reader detects a problem. Consider the options that are available to a reader (in terms of eye movements) when they encounter a word in a sentence which causes them difficulty. First, the reader could remain fixating the problematic word until their difficulty is resolved. Secondly, the reader could make a regressive saccade to permit them to re-read earlier parts of the text in an attempt to work out where they went wrong. Finally, the reader might make a rightwards saccade in order to read the next region of the sentence in the hope that this may help them resolve their difficulty. Most discussion about the pattern of eye movements made on encountering processing difficulty during reading has been concerned with what happens when reading sentences that are difficult to parse. However, many linguistic phenomena other than garden path sentences induce processing difficulty, and since we wish to

60

S.P. Liversedge, K.B. Paterson & M.J. Pickering

Fig. 2. Hypothetical eye movement record two.

keep the discussion of reading time measures in this chapter at a more general level we will refrain from associating particular mental processes with specific patterns of eye movement behaviour. Instead, we will consider the differential sensitivities of each of the measures in relation to different patterns of eye movement behaviour that may occur as a result of experiencing processing difficulty. The simplest situation is where the reader makes a saccade to the right after detection of disruption. In such a situation, assuming that fixations in the critical region are not inflated due to detection of disruption, no measure of initial processing will detect an effect for that region as in such a situation there is no observable cost to processing. In such a situation any effects would presumably be detected on the subsequent region. Figure 2 shows a situation in which the reader has remained fixating the region that caused them difficulty. In this case the reader makes three successive fixations on the critical region. In such a situation it seems likely that all three measures described above would detect the disruption the reader experienced. Summing the fixations on the region would inflate the reading times for that region relative to reading times for the same region of the control sentences where presumably the reader would spend less time. It is important to note that under these circumstances the increased reading time may be due to both detection of the problem and reanalysis processes that permit recovery from the processing difficulty. It is the third scenario, in which a reader makes a leftward saccade upon encountering the problematic word, that we find most interesting. In such a situation, it is possible that upon reading the problematic word, the reader may make a long fixation, or even a series of fixations prior to making a regression. If this occurs, then as with the situation depicted in Fig. 2, the reading time measures described above will detect the effect. However, in the situation where the reader detects a problem and makes an immediate regression, such measures will fail to detect disruption because there may be no difference in the first pass reading time for the critical region of experimental and control sentences. Indeed, there may even

Reading time measures

61

Fig. 3. Hypothetical eye movement record three.

be a shorter first pass reading time for the critical region in the condition that introduced processing difficulty than in the control condition. Importantly, in such a situation, the reading time measures described above would not detect disruption to processing and researchers may fail to detect an effect. After a reader has made a regression, there are a large number of possible patterns of eye movements they may make. Below we will consider several of these possibilities and point out important differences between them. This discussion should illustrate that if a researcher considers only first fixation, first pass and total reading time measures, then they may be unable to make claims about when an experimental manipulation influenced processing after detection of a misanalysis. That is to say, these measures alone provide only limited information regarding the time course of processing. Consider Fig. 3 which depicts one of the most simple patterns of eye movements a subject could make after a regression from a problematic region. Having made a leftwards saccade from the critical region, the reader makes a single fixation on text in Region Two, prior to making a rightwards saccade in order to fixate as yet uninspected text. In such a situation first fixation duration, gaze duration and total reading time for the verb will be no different to the control sentences. The only way in which an experimenter may detect disruption to processing is by computing the total time for Region Two which will be slightly inflated due to the fixation of 243 ms. At this stage, two important points should be made. First, computing total reading times for each of the regions of a sentence should ensure that effects such as these are detected. They would only be missed if gaze duration and total reading times were computed for Region Four alone. Secondly, it might be argued that if a researcher computed the number of regressions a subject made from Region Four during the first pass, then they would also detect such an effect. However, our point is slightly more subtle than this. We do not dispute that the effects depicted in Fig. 3 could be detected without recourse to measures of reading time defined in novel ways. What we do claim, however, is that spatially

62

S.P. Uversedge, K.B. Paterson & M.J. Pickering

Fig. 4. Hypothetical eye movement record four.

contiguous measures of reading time and regressions do not necessarily provide an indication of the first point at which a variable has an influence on reading time after a subject has made a regression. To expand upon this point, consider Fig. 4. The eye movement record in Fig. 4 shows that upon encountering the disambiguating word, the reader makes a long regressive saccade to the first noun phrase of the sentence and made a series of fixations to re-read this portion of the sentence. After the fourth fixation during re-inspection (214 ms), the reader then makes a long saccade to Region Five and continues to read the sentence normally. This situation is interesting as, presumably, all the four fixations in Region One reflect the processes involved in recovery after detection of misprocessing. As with the situation shown in Fig. 3, disruption of this type would be detected if the total reading times for each of the regions was computed. However, note that the total time measure would include fixations made during the first pass of the eyes through the sentence along with the fixations made after the regressive saccade from Region Four. That is to say, the total reading time measure is not a measure of recovery time alone, but instead is a mixture of initial processing time and recovery time. This is an important point as will become clear if we consider the hypothetical data shown in Fig. 5. Figure 5 shows a pattern of data where the reader makes a series of fixations, each in a different region after reading the problematic word. In this situation, the problem of failing to detect an effect is exacerbated. The problem here is twofold in that not only are the fixations made during reanalysis grouped with fixations made during the initial analysis of the sentence, but also, because fixations made during reanalysis were not made within the same region, they will never be grouped together to provide a measure of recovery from disruption. Thus, effects which might have occurred during reanalysis could be substantially weakened due to noise from first pass fixations included in the total time measure for a region and also due

Reading time measures

63

Fig 5. Hypothetical eye movement record five.

to reinspection fixations being considered in isolation rather than being grouped together. An important characteristic of the reading time measures discussed so far is that they sum spatially contiguous fixations. That is, the fixations which are summed for first pass reading time and total reading time all occur within the same region of a sentence. It is the property of spatial contiguity which can result in measures of this kind failing to provide an indication of the time course of a variable having an effect on processing during recovery, and possibly even failing to detect an effect, because in such a situation each subsequent fixation is not spatially contiguous with its predecessor. In order to circumvent this possibility, we require a measure which sums not spatially contiguous fixations, but temporally contiguous fixations. That is to say, we need a measure that groups fixations in terms of when they occur in relation to each other in time rather than in spatial location. Two measures that retain the property of temporal rather than spatial contiguity are regression path reading time (Konieczny, 1996) and re-reading time1. We define regression path reading time as the sum of all the fixations from the first fixation in a region up to but excluding the first fixation to the right of this region. This measure provides an index of the time a subject spent detecting a problem and then re-reading the text prior to fixating novel linguistic material. The regression path reading time for Region Four in Fig. 5 is 1337 ms (comprising the fixations of 270,243,214,310 and 300 ms). Re-reading time is defined as the regression path reading time for a region less the first pass reading time for a region. This measure simply provides an index of the time a subject spent re-reading the text after encountering a problem, but before they make an eye movement to fixate words to the right of the problematic 1 A number of measures exist which are the same or similar to regression path reading time. See for example, Brysbaert and Mitchell, 1996; Clifton, Kennison and Albrecht, 1997; Liversedge, 1994; Konieczny et al., 1997.

64

S.P. Liversedge, K.B. Paterson & M.J. Pickering

region. Note that these measures will usually group together the fixations a reader makes after they have encountered difficulty. Additionally, re-reading time does not include fixations made as the sentence is initially processed. By using measures which sum temporally contiguous fixations in parallel with measures such as first pass and total reading time which sum spatially contiguous fixations, we should ensure that we do not fail to detect subtle effects during recovery. Also, if recovery from disruption is an ongoing process, then by summing temporally contiguous fixations the experimenter might obtain a clearer picture of the nature of this process over time. We stress that we are not advocating that measures summing temporally contiguous fixations will always be better at detecting effects which occur during recovery than measures which sum spatially contiguous fixations. Rather, we are suggesting that in some circumstances measures summing temporally contiguous fixations may be more sensitive to the time course of effects than measures summing spatially contiguous fixations. We therefore suggest that the two types of measure should be used together. This will maximise the possibility that an experimenter will detect any effects occurring as the reader reinspects text. It will also permit experimenters to examine the nature of eye movement behaviour that occurs after a reader experiences difficulty during sentence comprehension. Consider Fig. 6: in this scenario, upon encountering the disambiguating region, the reader makes a regressive saccade and a brief fixation on the preceding word before making a rightwards saccade to continue reading the sentence its entirety. Only when they have inspected the whole of the sentence does the reader spend a substantial period of time re-reading portions of the sentence that preceded the disambiguating verb. This may be contrasted with the scenario depicted in Fig. 5, in which the reader spends a substantial amount of time re-reading portions of text preceding the disambiguating verb before inspecting the remainder of the sentence. The point is that the total reading times for Regions Two, Three and Four for the hypothetical eye movement records depicted in Figs. 5 and 6 will be the same, yet the patterns of eye movements made in the two scenarios are qualitatively different. It is our contention that such differences are likely reflect differences in the processes occurring during recovery. Such differences would not be detected if the experimenter relied solely upon reading time measures which sum spatially contiguous fixations. In advocating the use of additional measures, we are aware that a preponderance of statistically dependent measures can result in inflated Type I error rates, and the possibility that different measures could produce conflicting results. We thank Alan Kennedy and Wayne Murray for bringing these points to our attention during the Chamonix workshop. This is certainly true of the existing measures. First fixation, first pass reading time and total reading time are all statistically dependent measures, as a summation of the same fixations contribute to all three measures of reading time. Also, Altmann, Garnham and Dennis (1992) report two studies in which

Reading time measures

65

Fig. 6. Hypothetical eye movement record six.

measures of first pass reading time and the probability of a reader making a first pass regression favoured competing accounts of initial sentence processing. Of the two new measures discussed in this paper, only the regression path reading time includes a sum of fixations that also contribute to other measures of reading time for a region. The re-reading time measure is statistically independent of first fixation, first pass and total reading times for the region of interest. Furthermore, because regression path and re-reading times sum temporally contiguous aspects of the eye movement record, any discrepancy between these measures and measures of spatially contiguous fixations will be informative about the pattern of eye movements that occurred. Our central claim is that reading time measures which sum spatially contiguous fixations (first fixations, first pass reading times and total reading times), do not necessarily provide data to permit the experimenter to make strong claims about the time course of disruption to processing and possible recovery. We therefore advocate the use of two further reading time measures in psycholinguistic experiments: regression path reading time and re-reading time. We now demonstrate the utility of this approach by considering a recent experiment by Paterson, Liversedge and Underwood (1998). Experiment In this experiment, Paterson et al. monitored eye movements as subjects read a series of reduced and unreduced relative clause sentences. Reduced relative clause sentences like (1) are temporarily ambiguous between two syntactic analyses: a reduced relative clause reading in which the phrase allowed a party modifies the subject noun-phrase (i.e. the teenagers), and a simple active reading in which the

66

S.P. Liver sedge, K.B. Pater son & M.J. Pickering

noun-phrase a party is a direct object argument of the verb allowed. Such sentences are disambiguated in favour of the reduced relative clause reading on encountering the verb invited. Unreduced relative clause sentences (e.g. 2) are unambiguous due to the inclusion of a relative pronoun and auxiliary verb (i.e. who were). (1) The teenagers allowed a party invited a juggler straightaway. (2) The teenagers who were allowed a party invited a juggler straightaway. It is a well established finding that readers often experience processing difficulty when reading reduced relative clause sentences in isolation (e.g. Bever, 1970; Frazier and Rayner, 1982). Most accounts assume that readers initially adopt a simple active reading of the reduced relative clause sentence, and on encountering the disambiguating verb must reanalyse the sentence in terms of the dispreferred relative clause reading. Readers are said to have been 'garden-pathed' when they are forced to reanalyse the sentence in favour of an initially dispreferred syntactic analysis. Our experiment tested a recent claim by Ni, Grain and Shankweiler (1996) that the inclusion of the focus operator only will guide the initial processing of reduced relative clause sentences such as (3), and so enable the reader to avoid the garden path that is normally experienced when reading such sentences. (3) Only teenagers allowed a party invited a juggler straightaway. Ni et al.'s claims are derived from the Referential theory of sentence processing (Grain and Steedman, 1985; Altmann and Steedman, 1988), according to which the nature of the referential context can guide initial processing decisions regarding structural ambiguities. Ni et al propose that when readers process the subject noun-phrase only teenagers in sentences like (3), they construct a mental representation in which a focused set of teenagers is contrasted with some other set. However, as no contrast set is explicitly mentioned in the preceding discourse (because the sentence is presented in isolation) the reader must infer one. Ni et al. argue that for reasons of parsimony, the reader will infer that there are two sets of teenagers, and must anticipate further modifying information, such as a relative clause, in order to specify the nature of the difference between these two sets. Furthermore, they claim that anticipating modifying information will influence the initial processing of sentences like (3), such that readers will preferentially adopt the reduced relative clause reading of the ambiguous phrase allowed a party, and will not be garden pathed when encountering the disambiguating verb. While not disputing that the referential properties of only will influence the processing of the sentence, in contrast with Ni et al., we predicted that only would not guide initial processing decisions (see Paterson, Liversedge and Underwood, 1998, for a full discussion). We anticipated that the reader would be garden-pathed immediately on encountering the disambiguating verb of a sentence like (3).

Reading time measures

67

However, we hypothesised that the referential properties of only would facilitate reanalysis, so that readers would find it easier to recover from the garden path in (3) than in (1). We will report the duration of the first fixation and gaze duration as measures of initial processing that is spatially localised at the disambiguating verb. Method Subjects Thirty-two students from the University of Nottingham participated in this experiment. All subjects had normal and uncorrected vision and were paid £4. Materials and design Subjects read a series of thirty-six sentences that were either temporarily ambiguous reduced relative clause sentences or unreduced relative clause sentences which were disambiguated by the inclusion of a relative pronoun and auxiliary verb (i.e. who were). The sentences began with one of two determiner types: either the definite article (the) or a focus operator (only). Examples of these sentences are given in sentences (4) to (7) below. (4) The teenagers/ allowed a/ party/ invited/ a juggler/ straightaway. (5) Only teenagers/ allowed a/ party/ invited/ a juggler/ straightaway. (6) The teenagers who were/ allowed a/ party/ invited/ a juggler/ straightaway. (7) Only teenagers who were/ allowed a/ party/ invited/ a juggler/ straightaway. Prior to analysis the sentences were divided into six regions that spanned one or two words in the sentence. Region divisions are indicated by a slash in sentences (4) to (7). The critical region of interest was Region Four, which contained the disambiguating verb. Apparatus and procedure Eye movements were monitored using a SRI Dual-Purkinje Generation 5.5 eyetracker produced by Fourward Technologies. The tracker monitored subjects' gaze location every millisecond and the software sampled the tracker's output to establish the sequence of eye fixations and their start and finish times. Before the start of the experiment, subjects read an explanation of the eye-tracking procedure and a set of instructions. They were instructed to read at their normal rate and to read to comprehend the sentences as well as they could. Subjects were then seated at the eye-tracker and placed on a bite-bar and under forehead restraint to minimise head movements. Subjects then completed a calibration procedure.

68

S.P. Liversedge, K.B. Paterson & M.J. Pickering

Before each trial, a fixation cross appeared near the upper-left-corner of the screen. Immediately subjects fixated this cross, the computer displayed a target sentence, with the first character of this sentence replacing the fixation cross. This also served as an automatic calibration check, as the computer did not display the text until it detected a stable fixation on the cross. If subjects did not rapidly fixate the cross, the experimenter re-calibrated the eye-tracker. The experiment was conducted in two blocks, with a short intervening break while the experimenter set-up the equipment for the second block, and subjects were calibrated at the beginning of both blocks, with other re-calibrations performed every eight materials to maintain a high level of accuracy. This meant that the eyetracker was calibrated a minimum of 10 times during the experiment. Once subjects had finished reading each sentence, they pressed a key, and the computer displayed a comprehension question. Comprehension questions followed all of the experimental and filler trials. Half of these questions had "yes" answers, and half had "no" answers. Subjects responded to the comprehension questions by fixating either the word "yes" or "no". These words were presented on the left and right hand sides of the screen below the comprehension question. Subjects' responses were recorded by the experimenter without feedback. The computer displayed each experimental list in a fixed Latin Square order, together with 32 fillers that were materials for an unrelated experiment, and an additional 11 items that appeared at the beginning of the two halves of the experiment, and following each of the breaks for re-calibration. Results and discussion Prior to analysis, an automatic procedure pooled short contiguous fixations. Fixations of less than 80 ms were incorporated into larger fixations found within one character, and fixations of less than 40 ms that were not within three characters of another fixation were deleted. We removed those trials where either subjects failed to read the passage properly, or where there had been track loss. More specifically, those trials were removed in which a zero first pass reading time was recorded for two consecutive regions of text. This accounted for 3.0% of the data. The data were analysed for all six regions of text. As the disambiguating verb provided the point at which we expected to detect garden path effects, the reading time results for Region Four are reported below. We also report the total reading times for Regions Two and Three. The data were analysed using 2 (Determiner) x 2 (Sentence Structure) ANOVAs, treating subjects and items as random variables. The mean first fixation duration, gaze duration, total reading times, regression path reading times and re-reading times for the disambiguating verb, along with the total reading times for Region Two and Three are given in Table 1.

Reading time measures

69

Table 1 First fixation, gaze duration, total reading times, regression path reading times, and re-reading times for regions of reduced and unreduced relative clause sentences beginning with either the or X1W/A)

The

Only

Measure

Reduced

Unreduced

Reduced

Unreduced

First fixation duration (ms) for disambiguating verb

219

192

191

173

Gaze duration (ms) for disambiguating 272 verb

238

250

213

Total reading time (ms) for Region 2

549

337

444

328

Total reading time (ms) for Region 3

386

296

321

307

Total reading time (ms) for disambiguating verb

446

310

355

261

Regression path reading time (ms) for 426 disambiguating verb

270

318

278

Re-reading time (ms) for disambiguating verb

31

68

65

155

First fixation duration There was a significant main effect of Sentence Structure (Fj(l,31) = 6.5, p < 0.05, MSe = 2626; F2(l,35) = 4.4, p < 0.05, MSt = 3868), with a longer first fixation on the disambiguating verb when it was part of a reduced as compared to an unreduced relative clause sentence. There was also a significant main effect of Determiner (F,(l,31) = 14.0, /?< 0.001, MSe = 1292; F2(l,35) = l 1.6, p<0.01,MS e = 1550), with a longer first fixation at this region when sentences beginning with the as compared to only. Yet, crucially, there was no interaction of Sentences Structure and Determiner (F < 1). The results demonstrated that subjects encountered a garden path on the first fixation on the disambiguating verb of reduced relative clause sentences, and that this effect occurred regardless of whether the sentence began with the or only. The garden path effect on the duration of the first fixation on the critical verb is in line with previous findings (Rayner, Carlson and Frazier, 1983; Murray and Liversedge, 1994). For present purposes, the main effect of Determiner is not important. However, the lack of an interaction demonstrates that the inclusion of only did not guide initial parsing of reduced relative clause sentences.

70

S.P. Liversedge, K.B. Paterson & M.J. Pickering

Gaze duration There was a significant main effect of Sentence Structure (F,( 1,31) = 11.0, p < 0.01, MSC = 3605; F2( 1,35) = 6.3, p < 0.05, MSe - 6217), with a longer gaze duration on the disambiguating verb of reduced than unreduced sentences. There was also a significant main effect of Determiner (F,(l,31) = 5.6, p < 0.01, MSe = 3216; F2(l,35) = 7.6, p< 0.01, M5e = 2776), such that there was a longer gaze duration for sentences beginning with the compared with only. There was no interaction of Sentence Structure and Determiner (F< 1). The results were very similar to those of the first fixation duration, and showed that subjects were garden-pathed when reading reduced relative clause sentences despite the inclusion of the focus operator only. The first fixation and gaze duration results are clear. Subjects experienced processing difficulty on reading the disambiguating verb of reduced relative clause sentences, regardless of the inclusion of the focus operator. However, it was possible that the inclusion of only may have facilitated reanalysis for the reduced relative clause sentences at a later point in the reading process. If this was the case then we should observe an interaction between Sentence Structure and Determiner on the total time measure. We examined the total readings times for Regions 2 and 3 along with the disambiguating region. Total reading time At the disambiguating verb, there was a significant main effect of Sentence Structure (F,(l,31) = 38.2, p < 0.001, M5e = 11022; F2(l,35) = 31.8, p < 0.001, M5e = 14083), with a longer total reading time for the disambiguating verb of reduced as compared to unreduced sentences. There was also a significant main effect of Determiner (F,(l,31)= 13.2,p<0.01,MSe= 11828; F2( 1,35) = 27.8, p< 0.00 l,AfSe = 6471), such that there was a longer total reading time for sentences beginning with the as compared to only. However, there was no reliable interaction of Sentence Structure and Determiner (F,(l,31) = 1.8, p > 0.05, MSe = 7824; F2(l,35) = 1.8, p > 0.05, MSe = 8203). The total reading times for the disambiguating verb have same pattern as the first fixation duration and gaze duration results. There was a longer total reading time for reduced than unreduced sentences, and a longer reading time for sentences beginning with the as compared to only. Yet, importantly, there was no interaction of Determiner and Sentence Structure that might be expected if only had facilitated recovery from the garden path effect experienced when reading reduced relative clause sentences. At Region Two, there was a significant main effect of Sentence Structure (F,(l,31) = 69.4, p< 0.001, M5e= 12477; F2(l,35) = 81.7, p< 0.001, MSC= 12075), with a longer total reading time for reduced than unreduced sentences. There was also a significant main effect of Determiner (F,( 1,31) = 20.6, p < 0.001, MSe = 4993; F2(l,35) = 12.5, p < 0.001, MSC = 8541), with a longer total reading time for

Reading time measures

71

sentences beginning with the than only. There was a significant interaction of Sentence Structure and Determiner (F^l,31) = 9.0,p < 0.01, M5e = 8175; F2(l ,35) = 11.3, p < 0.01, M5e = 7771). An analysis of simple effects showed that there was a longer total reading time for reduced as compared to unreduced sentences beginning with the (F,(l,31) = 88.3, p < 0.001, M5e = 8175; F2(l,35) = 106.9, p < 0.001, MSe =7771), and for reduced as compared to unreduced sentences beginning with only (F,(l,31) = 26.6, p < 0.001, M5e = 8175; F2(l,35) = 31.3, p < 0.001, MSe = 7771). A similar pattern of effects occurred for total time on Region Three. There was a significant main effect of Sentence Structure (F,(l,31) = 13.4, p < 0.001, MSe = 6492; F2(l,35) = 12.8, p < 0.01, MSe = 7531), with a longer total reading time for reduced than unreduced sentences. There was an effect of Determiner that was significant by subjects and marginal by items (Fj(l,31) = 5.0, p < 0.05, MSe = 4607; F2(l ,35) = 3.8, p < 0.06, MSe = 6164), such that there was a longer total reading time for sentences beginning with the than only. Finally, there was a significant interaction of Sentence Structure and Determiner (F,(l,31) = 4.3, p< 0.05, MSe = 10462; F2( 1,35) = 8.1, p < 0.01, MSe = 7007). Simple effects showed that there was a longer total reading time for reduced as compared to unreduced sentences beginning with *MF,(1,31) = 12.2, /?<0.01,M5e= 10462; F2(l,35) = 21.5,/?<0.001,MSe = 7007), but no difference between reduced and unreduced sentences beginning with only (F < 1). The results for Region Two and Three produced the predicted interaction of Sentence Structure and Determiner. There was a longer total reading time for Regions Two and Three of reduced as compared to unreduced relative clause sentences beginning with the. There was also a longer total reading time for Region Three of reduced as compared to unreduced sentences beginning with only. At this point we have evidence to suggest that while initial processing was not guided, reanalysis was facilitated by the inclusion of only in reduced relative clause sentences. More time was spent re-reading text contained in Regions Two and Three of reduced relative clause sentences beginning with the, and text contained in Region Three of the reduced relative clause sentences beginning with only, after encountering a problem at the disambiguating verb at Region Four. It seems likely that readers re-read these earlier portions of text in order to recover from the misanalysis experienced at the disambiguating verb, but these total reading times alone are difficult to interpret in two respects. It is not possible to determine if, upon encountering the disambiguating verb, subjects immediately made a regressive saccade and spent considerable time re-fixating earlier portions of the text prior to reading as yet uninspected text. Or alternatively, if subjects read beyond the disambiguating region before making a regressive saccade to re-read earlier portions of the text. This is exactly the point we highlighted in our discussion of the hypothetical eye movement records in Figs. 5 and 6. To remind the reader, regression path reading times indicate the time subjects spend reading the disambiguating region, and re-reading text prior to the dis-

72

S.P. Liver sedge, K.B. Pater son & M.J. Pickering

ambiguating region, before inspecting text to the right. Also, re-reading times indicate the time subjects spent inspecting text immediately after making a regression from the disambiguating region, but before fixating text to the right of the disambiguating region. Using these measures we should be able to establish whether readers spent time re-reading the beginning of the sentence immediately upon encountering the disambiguating verb, or whether they returned to re-read earlier portions of the sentence after they had read the sentence in its entirety. If subjects did re-read immediately upon encountering the verb, then we should obtain a significant interaction between Sentence Structure and Determiner for both regression path and re-reading time measures. Alternatively, if subjects read the sentence in its entirety before re-reading the text, then there should be no interaction. Regression path and re-reading time The regression path measure showed a significant main effect of Sentence Structure (F,(l,31) = 27.1, p< 0.001, M5 e =l 1418; F2( 1,35) = 22.2, p< 0.00 l,MSe = 14262), with a longer regression path reading time for reduced as compared to unreduced sentences. There was also significant main effect of Determiner (F,(l,31) = 19.6, p < 0.001, MSe = 4077; F2( 1,35) = 11.9, p < 0.01, MSe = 7363), such that there was a longer regression path reading time for sentences beginning with the than only. Finally, there was a significant interaction of Sentence Structure and Determiner (F,(l,31) = 10.8, p < 0.01, M5e = 10098; F2(l,35) = 12.3, p < 0.01, MSe = 9781). There was a significantly longer regression path reading time for reduced as compared to unreduced sentences beginning with the (F,(l,31) = 38.9, p < 0.001, MSe = 10098; F2(l,35) = 42.3, p < 0.001, MSt = 9781), but no difference in reading time for reduced and unreduced sentences beginning with only (F,(l,31) = 2.5, p > 0.05, MSe = 10098; F2(l,35) = 2.4, p > 0.05, MSe = 9781). The re-reading time showed a significant main effect of Sentence Structure (F,(l,31) = 20.4, p < 0.001, MSe = 6252; F2(l,35) = 13.8, p < 0.001, MSe = 9661), such that more time was spent re-reading earlier portions of reduced relative clause sentences. There was a main effect of Determiner (F,(l,31) = 9.7, p < 0.01, MSe = 2241 \ F2(l,35) = 5.8, p < 0.05, MSe = 3945), with more time spent re-reading to earlier portions of sentences beginning with the than only. There was also a significant interaction of Sentence Structure and Determiner (F,(l,31) = 8.3, p < 0.01, MSe = 13965; F2(l,35) = 13.9, p < 0.001, MSe = 9415). Simple effects showed there was a longer re-reading time for reduced as compared to unreduced sentences beginning with the (F,(l,31) = 17.4, p < 0.001, MSC = 13965; F2(l,35) = 28.0, p < 0.001, MSC = 9415). However, there was no difference in the re-reading time for reduced and unreduced sentences beginning with only (F < 1). The interaction between Sentence Structure and Determiner was significant for both these reading time measures, and we obtained the same pattern of results as for the total reading time for Region Three. This pattern of reading times suggests that

Reading time measures

73

upon encountering the disambiguating region of reduced relative clause sentences containing the, subjects made a regressive saccade in order to spend more time re-reading preceding portions of the sentence than they did when they read reduced relative clause sentences containing only. The results show that recovery was initiated immediately upon detecting a problem at the disambiguating region, and furthermore, that during the recovery process readers spent time re-inspecting text they had already read prior to fixating as yet unread text. We suggest that for these garden path sentences subjects attempted structural reanalysis immediately upon encountering the syntactically disambiguating verb. The analysis of the temporally contiguous fixations made an important contribution to our understanding of the pattern of eye movements which occurred immediately after subjects read the disambiguating verb. Only by examining reading times which summed temporally contiguous fixations were we able to determine when during the course of processing subjects spent time re-reading the sentence. In this case re-reading occurred immediately upon a reader encountering the problem, rather than occurring after the sentence had been read in full. Conclusion The results of our experiment confirmed that the focus operator only did not guide initial parsing decisions when reduced relative clause sentences were read. The inclusion of only did, however, facilitate reanalysis after detection of the initial misanalysis. Our conclusion is that the inclusion of a focus operator does not cause a reader to anticipate modifying information and therefore the focus operator does not guide parsing decisions for reduced relative clause sentences. Turning to the reading times measures, we have provided an argument in favour of the use of reading time measures which sum temporally contiguous fixations. We advocate the use of these measures in conjunction with existing reading time measures which sum spatially contiguous fixations. By using measures which sum spatially contiguous fixations and measures which sum temporally contiguous fixations, the experimenter is better able to determine the time course of the influence of linguistic variables on processing, and distinguish between qualitatively different types of eye movement behaviour which may occur when processing difficulty is experienced. This approach enabled a fuller understanding of the data obtained in the present experiment. We suggest that the use of these measures will prove valuable in the interpretation of results obtained in other eye movement experiments.

74

S.P. Liversedge, K.B. Paterson & M.J. Pickering

Acknowledgements Special thanks are due to Keith Edwards of the Psychology Department at the University of Glasgow who developed software to support the reading time measures described in this paper. We also thank Wayne Murray, Franchise Vitu and an anonymous reviewer for providing comments on an earlier version of this paper. This research was completed when the first author was a Research Fellow at the University of Nottingham.

References Altmann, G.T.M. and Steedman, M. (1988). Interaction with context during human sentence processing. Cognition, 30, 191-238. Altmann, G.T.M. (1994). Regression-contingent analyses of eye during sentence processing: Reply to Rayner and Sereno. Memory and Cognition, 22, 286-290. Altmann, G.T.M., Garnham, A. and Dennis, Y. (1992). Avoiding the garden path: Eye movements in contexts. Journal of Memory and Language, 31, 685-712. Balota, D.A. and Rayner, K. (1991). Word recognition processes in foveal and parafoveal vision: The range of influence of lexical variables. In: D. Besner and G.W. Humphreys (Eds.), Basic Processes in Reading, Hillsdale, NJ: LEA. Bever, T.G. (1970). The cognitive basis for linguistic structures. In: J.R. Hayes (Ed.), Cognition and the Development of Language. New York: Wiley. Brysbaert, M. and Mitchell, D.C. (1996). Modifier attachment in sentence parsing — Evidence from Dutch. Quarterly Journal of Experimental Psychology, 49, 664-695. Clifton, C., Kennison, S.M. and Albrecht, J.E. (1997). Reading the words her, his, him: Implications for parsing principles based on frequency and on structure. Journal of Memory and Language, 36, 276-292. Grain, S. and Steedman, M. (1985). On not being led up the garden path: The use of context by the psychological syntax processor. In: D.R. Dowty, L. Kartunnen and A.M. Zwicky (Eds.), Natural Language Parsing: Psychological, Computational, and Theoretical Perspectives. Cambridge: Cambridge University Press. Ehrlich, K. and Rayner, K. (1983). Pronoun assignment and semantic integration during reading: Eye-movements and immediacy of processing. Journal of Verbal Learning and Verbal Behavior, 22, 75-87. Ferreira, F. and Clifton, C. (1986). The independence of syntactic processing. Journal of Memory and Language, 25, 75-87. Frazier, L. and Rayner, K. (1982). Making and correcting errors during sentence comprehension: Eye movements in the study of structurally ambiguous sentences. Cognitive Psychology, 14, 178-210. Inhoff, A.W. and Rayner, K. (1986). Parafoveal word processing during eye fixations in reading: Effects of word frequency. Perception and Psychophysics, 40, 431-439. Kennedy, A., Murray, W.S., Jennings, F. and Reid, C. (1989). Parsing complements: Comments on the generality of the principle of minimal attachment. Language and Cognitive Processes, 4, S151-S176.

Reading time measures

75

Konieczny, L. (1996). Human sentence processing: a semantic oriented parsing approach. Freiburg: IIG Berichte 3/96. Konieczny, L., Hemforth, B. and Scheepers C. (1997). Head position and clause boundary effects in reanalysis. Paper presented at the 10th Annual CUNY Sentence Processing Conference, March, 1997, Santa Monica, CA, USA. Liversedge, S.P. (1994). Referential context, relative clauses and syntactic parsing. Unpublished PhD thesis. University of Dundee. Liversedge, S.P. and Pickering, M. (1995) A comparison of eye movement measures in language processing. Paper presented to the Eighth European Conference on Eye Movements, Derby, September 1995. Liversedge, S.P., Pickering, M.J. and Traxler, M.J. (1996). A comparative analysis of qualitatively different eye movement measures. Poster presented at the Ninth Annual CUNY Conference, New York, March, 1996. Murray, W.S. and Liversedge, S.P. (1994). Referential context effects on syntactic processing. In: C. Clifton Jnr., L. Frazier and K. Rayner (Eds.), Perspectives on syntactic sentence processing. Hillsdale, NJ: Erlbaum. Ni, W., Grain, S. and Shankweiler, D. (1996). Sidestepping garden paths: The contribution of syntax, semantics and plausibility in resolving ambiguities. Language and Cognitive Processes, 11,283-334. Paterson, K.B., Liversedge, S.P. and Underwood, G. (1998). The influence of focus operators on syntactic processing of short relative clause sentences. Quarterly Journal of Experimental Psychology, in preparation. Raney, G.E. and Rayner, K. (1995). Word frequency effects and eye movements during two readings of a text. Canadian Journal of Experimental Psychology, 49, 151-172. Rayner, K. and Duffy, S.A. (1986). Lexical complexity and fixation times in reading: Effects of word frequency, verb complexity, and lexical ambiguity. Memory and Cognition, 14, 191-201. Rayner, K. and Pollatsek, A. (1989). The Psychology of Reading. Englewood Cliffs, NJ: Prentice Hall. Rayner, K. and Sereno, S.C. (1994a). Regressive eye movements and sentence parsing: On the use of regression-contingent analyses. Memory and Cognition, 22, 281-285. Rayner, K. and Sereno, S.C. (1994b). Regression-contingent analyses: A reply to Altmann. Memory and Cognition, 22, 291-292. Rayner, K., Carlson, M. and Frazier, L. (1983). The interaction of syntax and semantics during sentence processing: Eye movements in the analysis of semantically biased sentences. Journal of Verbal Learning and Verbal Behavior 22, 358-374. Rayner, K., Sereno, S.C., Morris, R.K., Schmauder, A.R. and Clifton, C. (1989). Eye movements and on-line language comprehension processes. Language and Cognitive Processes, 4, 21-50.

This page intentionally left blank

77

CHAPTER 4

Determinants of Fixation Positions in Words During Reading Ralph Radach Technical University of Aachen George W. McConkie University of Illinois at Urbana-Champaign

Abstract This chapter begins by reviewing previous findings concerning where the eyes land (i.e., landing positions) in words following progressive saccades during reading. It then reports the results from an examination of landing positions of German readers in two previously unexamined situations: intraword saccades (refixations) and interword regressive saccades (regressions). No evidence was found to support the common distinction between intra-and inter-word progressive saccades; landing positions in refixations are continuous with those in interword progressive saccades. In contrast, interword regressive saccades do not show the normal linear relation between launch site and mean landing position in the word that is observed in other conditions. In all cases, eye movement control during reading appears to be word-based.

Eye Guidance in Reading and Scene Perception/G. Underwood (Editor) © 1998 Elsevier Science Ltd. All rights reserved

78

R. Radach & G. W. McConkie

Introduction Although the eye-movement pattern made while reading a passage is infinitely complex, at one level it can be represented simply as a series of individual decisions about when and where to move the eyes. The where decision, at least in reading languages having spaces between words, appears to be word-based: that is, each saccadic eye movement is intended to take the eyes to a specific word location (McConkie and Zola, 1984).1 If this is so, then understanding where the eyes actually land in the text requires two theories: first, a selection theory which indicates how one word rather than another is selected as the target of a saccade, and, second, a performance theory which indicates where the eyes land given the selection of a target word for a particular saccade. In this chapter we will deal primarily with the latter of these two issues: the landing positions of the eyes within words. McConkie, Kerr, Reddix and Zola, (1988) conducted a quantitative analysis of the landing positions of the first saccades into a word (initial fixation positions), using a corpus of 40,000 fixations based on a sample of 66 readers. Included in the analysis were three factors of potential influence, word length, saccade launch distance and prior fixation duration. These authors concluded that early visual processes must parse the text into word-units, probably on the level of low spatial frequency objects that are separated by empty spaces. This provides candidates that could serve as the target of the next saccade. On some basis, which they did not investigate, a word-unit is selected and a saccade is launched to take the eyes to it. Where the eyes actually land (i.e., the distribution of saccade landing positions, or initial fixation positions) can be accounted for by a small set of basic visuomotor principles, which they attempted to explicate. McConkie et al. (1988), like many others (see reviews in O'Regan 1990, 1992), assume that there is a functional target location near the word center to which the eyes are being sent. O'Regan (1990) suggests that this is a result of learning that the center of the word is an "optimal viewing position" for word identification. He suggests that the position in each word from which it can be most readily identified varies according to the information structure within the word, but that, on average, the word center is optimal. However, the actual distribution of landing positions in words of different lengths, often referred to as the "preferred viewing position" (Rayner, 1979), shows an inverted U shape, with a maximum somewhat to the left of

1 There are exceptions to this generalization. For example, Hofmeister (1997) has data indicating that landing positions following return sweeps (saccades to the next line of text) are distributed relative to the beginning of that line, without influence of word locations. Also, it has been proposed (Radach, 1996) that some saccades may be sent to word groups, particularly when a relatively long word is preceded by a short function word. Further research is necessary to explore this possibility.

Fixation positions in reading

79

the word center. Thus, on most fixations the eyes are not at the optimal location. McConkie et al. proposed that this is the result of two sources of error in the visuomotor system. A "saccadic range error" leads to a linear relation between saccade launch distance and the mean landing position, which we will refer to as the "landing position function". When saccades are launched from a distance of about seven letters relative to the word center (a figure close to the average saccade length in reading), the resulting normal distribution of landing positions has its maximum close to the center of the word. When saccades are launched from locations closer to the target word, landing sites are shifted to the right (overshoot); when they are launched from more distant locations, landing positions are shifted to the left (undershoot). The second source of error is a "random placement error" that causes the spread in landing positions and that leads to an increase in this spread or variance with distance of the center of the target word from the launch site.2 In summary, the "preferred viewing position" in a word is seen, not as a basic oculomotor phenomenon, but as the result of several combined factors of influence (McConkie et al., 1988; see also McConkie, Kerr, Grimes and Zola, 1990; McConkie, Kerr and Dyre, 1994). This chapter begins with a summary of factors that determine the locations of initial fixation positions on words following forward intra-word saccades during reading, as obtained from an extensive analysis of a corpus of German reading data (Radach, 1996). It then presents analyses of the same corpus, examining initial fixations on words following saccades that lead to refixations of words, and regressive intra-word saccades, to see if they show the same properties. It ends with a discussion of theoretical issues regarding eye movement control during reading, arguing that the eyes are sent to a selected word (discrete decision) rather than a certain distance (continuous decision), and that where the eyes land in the word is determined by oculomotor error factors rather than by certain alternatives that have been recently proposed. Methodology Eye movement data were collected from four participants (German-speaking graduate students of physics) as they read a German translation of the first two parts of the book Gulliver's Travels (about 160 book pages).3 The text was presented in

2 Saccade target undershoot, saccadic range effect and increase in variation for more distant targets are common phenomena in the literature on basic oculomotor and other motor control processes (Becker, 1989; Kapoula and Robinson, 1986; Poulton, 1981). 3 The data used in this study were collected by the first author while on a Fulbright scholarship at the University of Illinois in Champaign/Urbana. The authors gratefully acknowledge the support of Gary Wolverton, Paul Kerr and John Grimes.

80

R. Radach & G. W. McConkie

screen pages of five to seven double-spaced lines of up to 72 ASCII characters each on a 15-inch VGA monitor in negative polarity. At a viewing distance of 80 cm each letter corresponded to approximately 0.25° of visual angle. Participants were instructed to read the text at their normal pace in order to comprehend the main ideas and to be able to answer questions at the end of each of 32 text segments. Eye movements were recorded using a Generation 5 Dual-Purkinje eye tracker with a sampling rate of 1000 Hz. The algorithm used for saccade identification is described in McConkie, Wolverton and Zola (1984), and details on the calibration routine are reported in McConkie (1981). After excluding blinks, cases of track loss and fixations outside the page, matrices of 47989, 47826, 59857 and 64226 valid saccade-fixations pairs were available for the four participants, respectively. Orthographic and lexical variables (e.g., letter frequency and word frequency measures) were either generated on the basis of the text itself (48334 words) or imported from the German CELEX corpus (Celex, 1995). The method employed in this study is quasi-experimental, involving the analysis of a complete set of normal reading data. The key technique is the method of orthogonal sampling (Kliegl, Olson and Davidson, 1983) where a number of variables is controlled (held constant) within a subset of data while a target variable is systematically investigated by comparing cases with different values of that variable. This method has certain weaknesses. For example, effects may sometimes be influenced by variables not considered in the sampling scheme or mediated by hidden interactions. The method also has a number of strengths. In cases where aspects of behavior being studied cannot be experimentally manipulated, such as where the eyes go during free viewing, this method makes it possible to separate out the effects of different variables on behavior. Also, when the corpus includes very large data sets from different individuals, similarities and differences among participants can be investigated and problems commonly arising in group statistics can be avoided. This is true in particular for the present corpus, because it includes enough data to carry out a complete replication of the original study by McConkie et al. (n = 66) for each individual reader. Where the eyes go when initially fixating a word On the basis of the data corpus described above, we have replicated McConkie et al.'s analysis and extended it to include a broader range of variables (Radach, 1996; see also Radach and Kempe, 1993). In the following section we will concentrate on those aspects of our studies on initial fixation positions in words that are most relevant for ongoing discussions of theoretical issues and which also lay the ground for the analyses of refixations and regressions presented in the later part of the chapter.

Fixation positions in reading

81

As indicated above, we are convinced that initial saccades are being sent to specific words during reading. Furthermore, the control system seems to act as if a particular location in the word, near its center, were acting as the target for the saccade. At the same time, there is an "optimal viewing position" within a word from which the word is most easily identified (O'Regan, 1990). One way to identify this is to compute the probability of refixating the same word as a function of the initial landing position. The basic result is that the likelihood of making a second fixation on a word is minimal when the eyes first land close to the word center (O'Regan et al., 1984). The effect was first observed in normal reading by McConkie et al. (1989), who found that the U-shaped refixation curve can be well specified by a quadratic equation of the form: Y = A+B(X-Q2 where X is a fixation location within a word, Y is the proportion of fixations at that location that are followed by a refixation of the word, A is the vertical offset of the curve, B indicates the slope and C is the minimum point. This minimum point is the horizontal offset of the curve, the location where refixation frequency is minimal. The analysis of the German corpus data found considerable interindividual variability with respect to the actual distributions of landing positions (the "preferred viewing position"). For three participants, the landing positions vary around locations about halfway between word beginning and word center, whereas for one participant they are much closer to the word center. However, there is no comparable variation in the refixation curves. For all four individual participants and at all word lengths studied (5 to 9), the minimum of each refixation curve is within one letter position of the word center. Therefore, we consider the word center to be an estimate of an interindividually stable optimal viewing position. The actual frequency distribution of saccade landing positions for a given word length (the "preferred viewing position") is Gaussian in shape and can be fit well by a normal curve. The data in this global distribution can be partitioned into launch site contingent landing position distributions. As an example, one such distribution would indicate landing positions within a seven-letter word following saccades that come from a distance of ten letters left of the word center. These distributions are also normally distributed and, most importantly, there is a linear relation between launch distance and mean landing position, the "landing position function". In the 206 particular combinations of word length and launch site considered in our analysis, the average shift of mean landing position with each increment in launch distance (the slope of the landing position function) is 0.35, 0.33, 0.55 and 0.38 letters for our four participants. In regression analyses, performed on the individual subject's mean landing positions as a function of launch distance for word length 4-15, more then 90% of the variance is accounted for by a simple linear model.

82

R. Radach & G.W. McConkie Center based landing position

-21 -20 -19 -18 -17 -16 -15 -14 -13 -12 -11 -10 - 9 4 - 7 - 6 - 5 - 4 -3 -2 Launch site relative to word beginning Fig. 1. Estimated mean landing positions of initial saccades as a function of launch site relative to the word center for 9-11 letter words. Each data point represents between 1054 and 133 observations. At about launch site -15 the peaks of the underlying landing site distributions start to move off the target word and the estimated means are more variable. For this data set the slope of the linear regression function is 0.53 (r2 = 0.96). Launch distances of 5 or less include cases where the saccade was launched from within the same word.

This key finding, which has been observed in several analyses of different sets of reading data in both English and German (see also Kerr, 1992; McConkie, Kerr and Dyre, 1994), is illustrated in Fig. 1. This figure was prepared in order to explore how far from the word center the linear landing position function is found. We have plotted the landing position function for words of length 9-11, based on pooled data from our four subjects. Data pooling was required in order to have a large enough sample size for the analysis. This word length range was selected for two reasons. First, words had to be relatively long to counter the tendency of landing site distributions to move off the left word boundary for very distant launch sites. Second, words had to be short enough to provide a sufficient number of observations. German text is well-suited to meet these constraints. As is apparent from Fig. 1, the mean landing positions become less stable at launch distance of-15 due to reduced sample sizes but even at -21 (n = 133) the linear trend is still present. We conclude from this analysis that for inter-word progressive saccades the landing position function is linear over the entire range of launch distances that is found in normal reading. Unlike McConkie et al. (1988), but in agreement with McConkie, Kerr and Dyre (1994), we also found a small effect of word length on mean landing positions for a given launch distance. The individual average increase (relative to the word beginning) in mean landing position was 0.16,0.18,0.14 and 0.11 letter position for each one-letter increase in word length. This is a much smaller effect than the launch

Fixation positions in reading

83

position effect reported above. In linear regression analyses, performed on the mean word length contingent landing positions for each launch site and participant, only a modest portion of the variance can be accounted for (r2 > 0.34, see Radach, 1996 for further details). The effect can be interpreted as a "center of gravity" phenomenon (Findlay, 1982), modulating the launch position effect. Such an interpretation would follow O'Regan's (1990) suggestion that the landing position saccades can be influenced by the presence of further elements in the critical visual configuration, in the present case additional letters belonging to the same target word (see also below our discussion of the relative importance of range-effect vs. "center of gravity" effect). A curious observation made during the analysis of the word length effect was that not only the length of the target word but also the length of the preceding word had an influence on initial landing positions. When further investigating this effect, we found that this was not caused by prior word length per se, but largely due to the fixation pattern on the preceding word (Radach and Kempe, 1993). With saccade launch position held constant, there is a substantial rightward shift in mean landing position in a word when the prior word received more than one fixation (effect size in the order of 0.5 letters) and a large leftward shift when the preceding word was skipped (effect size up to 1.5 letters). Radach (1996) suggests that the first effect may indicate that in the case of refixations it is more likely that a parafoveal word or letter cluster is being recognized and that the subsequent saccade is lengthened in a "skipping-like" fashion. The second effect, a leftward shift of initial fixation position after skipping a short word is more puzzling and has led Radach (1996) to consider the possibility that saccades may sometimes be aimed at units of two words in which case a small function word is not "skipped" but remains unfixated because it is part of the larger two-word target unit (see Chapter 2 for a discussion of the related issue of information acquisition from locations left of the actual fixation position). A further variable of potential influence on saccade landing positions is the duration of the previous fixation, the 'latency' of the initial saccade. This variable is of particular interest, because it provides a link to temporal aspects of eye movement control that may well significantly modulate spatial saccade parameters. Several hypotheses on the issue are feasible. McConkie et al. (1988) and O'Regan (1990) have proposed that when a preceding fixation is long, the following initial fixation locations converge towards the optimal viewing position. An alternative hypothesis could be that, due to increased parafoveal preprocessing, there should generally be a rightward shift of fixation positions after longer fixations (Pollatsek, Rayner and Balota, 1986).4 Although Pollatsek et al. do not explicitly state this hypothesis, it might be derived from their more general proposal that"... there are places in text where more complete processing of the material fixated (and that just to the right of fixation) takes more time but then leads to longer saccades".

84

R. Radach & G.W. McConkie

We studied the role of the preceding fixation duration for a total of 40 conditions obtained by orthogonal sampling within the German data corpus: five launch sites from -10 to -1 (number of letter positions to the left of the center of the target word, in two-letter increments), two word length ranges (5-6 vs. 7-9 letter words) and four participants. For each of these conditions, mean initial landing positions were compared for the condition-specific quartiles of prior fixation duration. These comparisons used non-parametric Kruskal-Wallis tests (equivalent to ANOVA's) and r-tests of the extreme quartiles in each condition. Interestingly, our analysis supported neither of these two hypotheses. Of the 12 conditions with significant effects of prior fixation duration on saccade landing positions, 11 showed an unexpected pattern. In these conditions landing positions were shifted to the right following shorter preceding fixations. This result is likely to be related to the fact that refixations tend to be of shorter duration as compared to single fixations (e.g. Kliegl, Olson and Davidson, 1983; Underwood, Clews and Everatt, 1990; O'Regan et al., 1994; Rayner, Sereno and Raney, 1996), and that saccades following refixations tend to be longer (see the fixation pattern effect described above). With these considerations in mind, one could suspect that the observed rightward shift after shorter fixation durations is equivalent to a rightward shift of landing position following longer gaze durations which typically result from cases that include more refixations. Indeed this hypothesis was confirmed as part of recent explorations into interrelations between spatial and temporal aspects of eye movement control (Radach and Heller, in preparation). An additional source of variation in initial fixation positions within words is the position of the target word within the current line of text. Mean landing positions in words at the beginning of the line are substantially shifted to the right and those in words at the end of the line are shifted to the left. These are relatively large effects, producing up to one letter position difference in mean landing position, depending on word length and launch distance. This underscores the importance of line-level information (line length, line distance, layout) for spatial navigation through a page of text (Heller, 1982). A comprehensive model of eye guidance in reading will need to consider these aspects, especially the functional significance of fixations preceding and following return sweep saccades (Hofmeister, 1997). The variables discussed so far (with the exception of prior viewing duration) were all "low-level" variables operating on the oculomotor and/or perceptual level. It is now generally agreed that these explain most of the variance in initial saccade landing position distributions (see, e.g., Chapter 11 for a converging perspective). However, in recent years there has been considerable controversy around the question of whether parafoveally available cognitive (semantic, lexical or sublexical) information can also influence initial landing positions. One popular variation of this idea is the "parafoveal guidance hypothesis" stating that saccades go

Fixation positions in reading

85

further into words that have a less informative (more redundant) word beginning (Underwood, Bloomfield and Clews, 1988; Underwood, Clews and Everatt, 1990; but see Chapter 9 for a more cautious view). We have tested the original parafoveal guidance hypothesis by looking at the same 40 conditions (5 launch distances x 2 word length ranges x 4 participants) as described above. For each condition we compared mean landing position as a function the quartiles of initial trigram 'informativeness' (i.e., number of word forms that share the same initial trigram). Among the 40 conditions there were only four significant differences and only two of these were in the direction predicted by the hypothesis. Two similar analyses was carried out with token trigram frequency and word frequency as the dependent variables, again with negative results (for a similar corpus analysis, see Rayner, Sereno and Raney, 1996). It could be argued that in our corpus analyses we may not have given cognitive variables a fair chance to show their influence on saccade landing positions. However, when we recomputed the range of trigram frequencies used by Radach (1996) and Radach and Kempe (1993), and compared it to values given by other researches (e.g. Liversegde and Underwood, Chapter 9) we found that the differences between our high and low trigram items are in the same range. Our results are in line with other failures to replicate word beginning informativeness and trigram frequency effects (Rayner and Morris, 1992; Radach et al., 1995). However, other studies exist that have found evidence for a small effect of orthographic manipulations on the mean landing position (e.g., Hyona, 1995). In summary, we have found that in the eye-movement data of four German readers the landing positions on words following progressive inter-word saccades is primarily determined by the locations from which the saccades originated, or launch sites, with smaller effects due to word length and position of the word on the line of text. In these data, neither the duration of the prior eye fixation, nor the informativeness of the initial trigram of the word affected these landing positions. The purpose of the study described below was to determine whether launch-sitecontingent intraword and regressive saccades show landing position distributions having the same properties as those obtained from interword progressive saccades. This is of particular interest given the strong distinction often made between interword and intraword (refixation) saccades, and between progressive and regressive saccades. O'Regan's (1990, 1992) Strategy-Tactics theory assumes a general scanning routine containing two components: (1) the eyes are sent to successive words in the forward (rightward in English and German) direction, and (2) there is a within word tactic of repositioning the eyes (refixating) when their initial fixation position is at a non-optimal location. Averaged over words, the word center is considered to be the optimal location for their viewing. Given that interword and intraword saccades are assumed to be controlled on quite different bases, we might expect their landing position distributions to show different characteristics.

86

R. Radach & G. W. McConkie

Similarly, Morrison's (1984) model of eye movement control during reading, as modified e.g. by Henderson and Ferreira (1990) and Pollatsek and Rayner (1990), makes a clear distinction between refixations and progressive interword saccades. The original version of this model included no mechanism for accounting for refixations; it has been necessary to suggest other mechanisms to account for these. Again, if refixations and interword progressive saccades are being generated by different mechanisms, it would be expected that they would show different characteristics. Finally, neither O'Regan's model nor Morrison's model include a clearly defined mechanism for generating interword regressive saccades; typically, they are explained as resulting from processing difficulties at higher cognitive levels that require a reconsideration of earlier text, thus resulting from quite a different set of circumstances than the 'normal' progressive interword saccades that are the basis for existing theories. If differences were found in the properties of the landing position distributions of refixations or regressions, as compared to progressive interword saccades, this would provide evidence for the psychological reality of the processing distinctions described above. On the other hand, if no differences are found, this would not eliminate the possibility of these distinctions, since it may be that the distinctions lie at some higher level, where selection of a saccade target is made, with the mechanism that produces the landing position distributions (the control system that gets the eyes to a selected target) being common to all of these processes. The present study examined our corpus of German reading eye movement data to determine whether the landing position distributions following interword progressive saccades, refixations, and interword regressive saccades do or do not show similar properties. It also examines the issue of whether eye movement control is discrete or graded, and considers two alternative hypotheses of how landing positions in selected words are determined. Refixations vs. progressive interword saccades In the literature on eye movement control in reading two alternative views on refixations have been proposed. According to the first view, refixations are initiated when cognitive processing is difficult in one or another respect. Such a processing difficulty can take several forms: it may be that lexical access is hampered or that due to cognitive overload at modules of syntactic or semantic analysis the oculomotor system is ordered to "slow down" and park the eyes on the current word for some additional time (for a review, see Rayner and Pollatsek 1989). To account for

Fixation positions in reading

87

reflations, Henderson and Ferreira (1990) supplemented Morrison's attentionbased eye guidance theory with the notion of an oculomotor deadline, in which a refixation is initiated if attention has not shifted to a new word. Contrary to this prediction, the first of two fixations on words has been shown to be shorter than single fixations (Kiegl, Olson and Davidson, 1983; Underwood, Clews and Everatt, 1990; O'Regan et al., 1994; Rayner, Sereno and Raney, 1996; Radach, Heller and Inhoff, 1997). Pollatsek and Rayner (1990) discuss a another possibility in relation to parallel interactive models of word recognition (e.g. Paap et al., 1982). They propose that the total "level of excitation" in the lexicon (including excitation from parafoveal preprocessing during prior fixations) may provide a cognitive base for refixation decisions (see also Chapter 11 for a new processing mechanism to account for refixations). According to the second view, refixations are based on visuomotor factors: deviations of the initial fixation position from the generally optimal viewing position result in an increase of refixation frequency (see above). In his "strategy and tactics theory" O'Regan (1990) proposes that the eyes, driven by a global, preprogrammed scanning routine, attempt to go to the "generally optimal" viewing position located close to the center of the word. If the eyes do not land at the optimal position, "lexical processing will begin, but may not be able to terminate, since information about certain letters in the word is lacking" (O'Regan 1990, p. 427). Therefore, a refixation will be initiated that will bring the eyes to the opposite end of the current word. "When the eye is on one side of the word, it goes to the other. When it is near the middle, it goes to either one or other end. In other words, the eye is not attempting to get to the optimal viewing position. Rather, it is attempting to spread its fixations evenly over the word." (O'Regan 1990, p. 427). Although O'Regan (1990, 1992) does not make exact quantitative predictions with respect to landing positions of within-word saccades, it is evident that he proposes a qualitative difference between saccades across word boundaries (based on the "strategy") and within-word saccades (based on "within-word rescue tactics"). Such a difference should be evident when saccade landing positions are plotted as a function of launch site relative to the word (the landing position function), allowing a comparison between cases with launch sites lying to the left of the target word (initial fixations) and cases in which the eyes are launched from within the word (refixations). Specifically, there should be some type of discontinuity in mean landing position at the transition point between inter-word to intra-word launch sites. The left panel in Fig. 2 shows the frequency with which progressive intraword saccades (refixations) are launched from different letter positions of 9- and 11-letter

88

R. Radach & G.W. McConkie

Fig. 2. Left panel: Frequency distributions of progressive refixation launch sites and refixation landing positions for 9-11 letter words. Please note that a refixation launch site is equivalent to the landing position of the prior initial fixation. The scale on the abscissa indicates letter positions relative to the center of the word. Right panel: Mean landing positions of saccades as a function of launch site. The abscissa indicates letter positions relative to the space immediately to the left of the word. Negative numbers indicate launch sites of initial progressive saccades going into the target word, whereas positive numbers indicate launch sites from within the target word (for these refixations observations are identical to those in the left panel). Pooled data for four participants. Notice that the curves appear continuous from progressive interword saccades to intraword (refixation) saccades.

words combined, relative to the center of the same word, and the frequency with which the eyes land at these locations following the refixation saccades.5 The great majority of progressive refixations are initiated from positions at or to the left of the center of the word. The right panel shows the landing position function on 8- to 5 There are two features of the right panel of Fig. 2 that deserve comment. First, the vertical displacement of the curves from one another is primarily due to the fact that they are plotted relative to the space before the word (letter position 0 here), which was done to preserve continuity with Fig. 3. If plotted relative to the centers of the words, the curves are largely superimposed on each other. Second, the curves appear to diverge somewhat in the refixation range. This is primarily due to the fact that in this figure means of the observed landing positions are given as opposed to fitting a normal curve to the data as in Fig. 1. This has the effect of underestimating the central tendency of the distribution when it migrates toward the far end of the word, because saccades landing off the word are not included in the calculation. At any given letter position this error in estimation is greater for shorter words, hence causing a separation between curves for different word lengths, as observed here.

Fixation positions in reading

89

12-letter target words following saccades launched from different locations (launch sites) relative to the space before the target words. This figure illustrates the linear landing position function that has been observed in previous studies. However, it also shows that this relation continues smoothly from interword to intraword progressive saccades. There is no discontinuity at the transition point between the two types of saccades, as one might anticipate from current theories distinguishing these categories. Thus, this analysis provides no evidence for a difference in the basis on which landing positions are determined for initial and refixation progressive saccades.

Regressive inter-word saccades The frequencies of saccades made in a right to left direction (regressions) among our participants ranged from 16.7 to 36.1% (including only saccades that start and land on the same line of text), which is within the range typically observed among German readers. About 2/3 of these are inter-word saccades ranging from 10.6 to 29.6 percent of all saccades made by the four participants studied. Our goal in the following analysis was to determine whether the linear mean landing position function observed for data following progressive saccades is also observed in data for interword regressive saccades. The frequency of regressions varies considerably as a function of factors like text difficulty and reading instruction (Heller, 1982). Regressive saccades have recently received much attention in the psycholinguistic literature, having been shown to be related to language processing difficulties. Several regression-related eye movement measures can be considered: the frequency with which the fixation of a target word results in the initiation of a regression, the frequency with which a word is the recipient of a regression and the duration of fixations after regressive saccades. Regression-contingent analyses of syntactic parsing mechanisms have been discussed (Altman, Garnham and Dennis, 1992; Rayner and Sereno, 1994) and complex eye movement measures involving regressions have been developed (e.g., Daneman, Reingold and Davidson, 1995; see also Chapter 3). However, there is currently only limited understanding of whether and how spatial information about word locations within a sentence is being acquired and used for the subsequent planning and execution of regressions (Kennedy, 1992). The question addressed here concerns the landing position function following these regressions and, particularly, whether it shows the same linear functions of launch site as to progressive saccades. An analysis of launch sites of regressive saccades in our corpus of reading data shows that most regressions originate from positions relatively close to the target word. Of all regressive saccades made within one line of text to words of length 5 to 10, 26.0% come from within the same word

90

R. Radach & G.W. McConkie

(regressive refixations), 49.4% come from the immediately following word and only 24.6% from more distant locations. Figure 3 shows the landing position function for regressive saccades. The abscissa on this figure, which shows the launch site, indicates letter position relative to the space following the fixated word. Thus, negative numbers indicate refixation cases, where -1 means that the eyes were launched from the final letter in the word (-1 with respect to the space following the word), while positive numbers indicate launch sites to the right of the word, or interword refixation cases. As can be seen in the figure, for every word length the refixation regressions (negative numbers on the abscissa) show the same linear relationship between launch site and the mean of the landing position distribution that progressive saccades show (see Figs. 1 and 2). The mean of the slopes for the different word lengths, for this part of the data, is 0.39. However, for interword regressions (positive numbers on the abscissa), the result is quite different. Here, for every word length, the curves go flat (mean slope of 0.06). For these interword regressions there is essentially no relationship between launch site and the mean of the landing position distribution; no matter from where the saccades are launched, the mean landing position in the word on which they land is the same. Thus, the mean landing position function shows a sharp bend in the region of the space before the word, which separates intraword from interword regressions.

Fig. 3. Mean landing positions of regressive saccades as a function of launch site. The abscissa is numbered relative to the space following the target word, with negative numbers indicating launch sites from within the word, and positive numbers indicating launch positions to the right of the word boundary. The ordinate, indicating mean landing position, is numbered with respect to the center of the word. Each graph represents between 3819 and 744 observations. Pooled data from four participants.

Fixation positions in reading

91

These results indicate that the control of the eyes in making interword regressions is functionally different in some way than it is in the other cases studied. McConkie et al. (1988), following Kapoula and Robinson (1986), attributed the linear relation between launch site and mean landing position found with progressive saccades to a range effect that is frequently observed in other muscular systems (Poulton, 1981): a tendency to overshoot near targets and undershoot far targets. Thus, it is assumed that a word string within a line of text, perceptually delimited by preceding and following spaces, functions as a stimulus unit or 'blob', with the eyes being drawn to the center of that unit when it is selected as the target for a saccade. The center serves as a functional target location for directing the saccade. From this perspective, it appears that in making interword saccades, there is no range effect. Rather, the eyes go consistently, with some random error that produces the Gaussian distribution observed, to their target. A collateral observation from examining Fig. 3 is that the mean landing position following interword regressions is very near the center of the word, regardless of the word length or the launch site. Thus, the two variables that appear to have the greatest effect on mean landing position in other cases, have little or no effect following these particular saccades. This has two implications. First, it provides further evidence that the center of the word is the functional target for saccades going to the word. Second, it indicates that the mechanism producing the range effect is not always operating, even in the reading task itself. It is curious that the range effect should occur in intraword regressions but not in interword regressions. It is not just the direction of the saccade that is critical, since intraword regressive saccades show its pattern. Neither is it the distance that the saccade target lies into the periphery; the range effect is observed in progressive saccades of similar lengths to the interword regressions that do not show its pattern. There is clearly something different about saccade control with interword regressive saccades. Research is now needed to better understand the mechanism that gives rise to the range effect, and the conditions under which it operates. Finally, there is one more observation that indicates the distinctiveness of interword regressions. Radach, Heller and Hofmeister (1998) have data indicating that certain readers show a distinct peak of very short fixations (80-120 ms) in fixation duration frequency distributions that these investigators explain in terms of minimal latencies of saccades that have become unnecessary but could not be canceled soon enough (Becker, 1989; Morrison 1984). Interestingly, these very short fixations were never found before inter-word regressions, suggesting that progressive and refixation saccades are part of a mode of eye control that may automatically generate, but sometimes cancel, saccades. Apparently the decision to make a regression to a previous word is the result of a more deliberate mode of eye control.

92

R. Radach & G. W. McConkie

Continuous vs. discrete control of saccades In the existing research literature on reading, a number of variables have been found to have a local effect on the lengths of saccades. As one example, there is a marked effect of word frequency on the likelihood of refixating the same word (e.g., McConkie et al., 1989; O'Regan et al., 1994, Rayner, Sereno and Raney, 1996). The effect is basically a vertical displacement of the refixation function, as depicted in Fig. 4 (left panel), which results from an increase in the amplitudes of saccades launched from the initial fixation within higher-frequency words as compared to saccades launched from lower-frequency words. In most cases the sizes of these saccade length effects are less than one letter position (e.g., 0.82 letters for the observations in Fig. 4, right panel). This can give the appearance of an eye movement control system that adjusts the lengths of saccades in a graded fashion, slightly increasing or decreasing the lengths of saccades according to the local processing requirements during reading. The alternative to this position, and the one we espouse, is that the lengths of saccades are determined in a more discrete fashion, resulting primarily from which words are selected as targets of those saccades. In harmony with Morrison (1984) and Rayner, Sereno and Raney (1996) and others, we propose that eye movement control is primarily a discrete process, with most variables on the perceptual and cognitive level influencing which word is selected as the target of the next saccade.6 Thus, the observed effect of a local variable on saccade length is interpreted as being probabilistic in nature, with that variable having an influence on some underlying attractiveness level of the words that serve as potential candidates as target for the next saccade. Since the eyes can be sent to only one of these candidates on a given saccade, a selection is made through some winner-take-all competition. For many saccades, one candidate is so strong that small changes in candidate attractiveness have no effect on the outcome; for other saccades where alternative words have attractiveness values that are sufficiently similar, such an influence can cause a different word to be selected than would otherwise occur. In such a case, the saccade length will change by several letter positions; in extreme cases, the direction may even change. Within the framework just described, finding that a variable decreases the lengths of saccades by an average of 0.5 letter positions essentially indicates that the lengths of the great majority of saccades were not affected at all by the variable, even though an influence of the variable may have been present. For only a small proportion of the cases was the influence of the variable being studied great enough, and the As noted earlier, some variables, such as those that affect the center of gravity in a word, may influence the landing position within the word, thus being a graded influence. We assume that variables of this type account for only a small part of the variance in saccade lengths in reading.

Fixation positions in reading

93

Fig. 4. Left panel: Frequency of refixating the same word as a function of the initial landing position within the word. Data for word length 8-12 from four participants were pooled. Negative numbers on the abscissa indicate initial landing positions left of the word center. The selection of 'infrequent' and 'frequent' words is based on a median-split of word form frequency (Celex, 1995). Right panel: Landing positions of progressive saccades following an initial fixation position of -3 relative to the word center (with observations identical to those for -3 in the left panel). Negative numbers on the abscissa indicate landing positions within the same word (refixations); positive numbers indicate landing positions of regressive inter-word saccades.

existing selection likelihood structure of alternative words similar enough, for the effect of the variable to be realized in behavior. In these cases, a large change in saccade length occurred, sending the eyes to one word rather than another. Theoretical and methodological problems related to this "frequency of effects" problem in reading have been outlined by McConkie, Zola and Wolverton, (1985). An example of this type of influence can be seen in the case of the effect of the frequency of a fixated word on the length of the following saccade. The two local variables that appear to have the largest effect on the likelihood that a given word will be selected as the target of the next saccade are its location with respect to the currently fixated letter and its length (Kerr, 1992; McConkie, Kerr and Dyre, 1994; see also Chapter 6). To see other influences it is necessary to control for these two variables, which can be done by selecting cases from a large corpus of reading data as we have described above. Figure 4 shows a frequency distribution of landing positions in all cases where the saccade's launch site was three letter positions left of the center of the currently-fixated word (word length 8-12). As can be seen, the frequency distribution is markedly bi-modal, suggesting two populations of saccades, one keeping the eyes on the currently-fixated word (refixations) and the other taking the

94

R. Radach & G. W. McConkie

eyes to a following word. The two curves are formed by taking a median split of the data based on the cultural frequency of the currently fixated word, including all words of our corpus that are listed in the German CELEX Corpus. Figure 4 shows that the frequency of the word does not simply shift the distribution left or right in a simple manner, as would occur if the effect occurred in a graded fashion. Rather, word frequency influences how many of the saccades remain on the current word vs. go to the next, suggesting a discrete choice between alternative possible saccade targets. The curves for high- and low-frequency words both appear to be the sum of the same two underlying distributions, having similar modes; word frequency affects the number of cases in the two component distributions, resulting from the frequency with which the two alternative candidate words are selected as the target of the saccade. When launch site and the lengths of immediately surrounding words are controlled in the manner illustrated above, the landing position distributions actually change shape as word length configurations change. Bimodal distributions of the type shown in Fig. 4 are frequently observed, suggesting that choices are being made among alternative potential saccade targets. This is consistent with an eye movement control mechanism that is discrete and probabilistic in operation, rather than one that adjusts saccade lengths in a graded fashion on the basis of local variables. Some theoretical issues We have argued that eye movement control during reading is a discrete process, involving the selection of a target word, and the somewhat error-prone process of moving the eyes to that word. In this section we discuss several issues related to the landing position distributions on words. Launch site contingent landing position distributions are very well behaved, being Gaussian in shape, with a mean that is a linear function of the launch site (with certain exceptions), and a standard deviation that increases with launch distance (McConkie et al., 1988; see also McConkie et al., 1990; McConkie, Kerr and Dyre 1994). We have attributed the linear landing position function to a saccadic range error. Vitu (199la, 1991b) has argued for an alternative possibility. As we do, she assumes that a word is selected to be the target of the following saccade. However, she then proposes that the eyes go to the center of gravity of an attended region extending about seven letters from the beginning of the selected word. The blank space at the end of the attended word is thought to have sufficient influence on this center of gravity to account for changes in mean landing position; the farther the space lies to the right, the further to the right lies the center of gravity within this attended region. This theory can successfully predict that landing positions within longer words that begin at the same distance to the right of the launch site tend to lie further from the words' beginnings (Vitu, 199la). However, it does provide a basis

Fixation positions in reading

95

for predicting the landing position function: changes in mean landing position within words of the same length lying at different distances from the launch site. In contrast, in our analyses if landing position is measured relative to the centers of words, then word length has very little effect on the landing position distribution, whereas distance of the launch site from the target word has a large effect. Thus, we do not see center of gravity within an attended area as an alternative to the concept of the range effect for accounting for the strong, linear landing position function that exists in reading data. Rayner and Morris (1992) as well as Rayner, Sereno and Raney (1996) have suggested that the landing position within words is determined by which letters in the word have been previously identified, with the eyes tending to go beyond the identified letters, similar to the proposal made by McConkie (1979). This raises the question of how landing position distributions are affected by cognitive factors. In discussing this issue it is critical to distinguish between factors that influence the selection of the saccade target, and those that influence where the eyes land with respect to that word. Clearly, cognition does affect where the eyes go during reading; the question here is whether it influences where the eyes land with respect to a selected word, or only affects the selection process itself. The preprocessing hypothesis could be the basis for an alternative explanation of the range effect: namely, the farther the launch site is from a target word, the fewer the letters that are likely to have been identified from it, thus resulting in the eyes landing closer to the front of the word. However, two aspects of our data are not in harmony with this explanation. First, the slope of the landing position function tends to be in the range of 0.35. Thus, for every letter position farther from a word a saccade to it is launched, the landing position in the word moves leftward by 0.35 letter positions. However, as distance of the eye from the target word is increased, the visibility of the word should fall off faster than this. Second, the landing position distribution continues to be linear for all launch site distances examined so far, whereas we would expect that a visibility-based effect would be negatively accelerated, especially at more distant launch sites where a word's visibility would be minimal (Rayner and Morris, 1981; Nazir, Heller and Sussmann, 1992). There is another possible interpretation of the preprocessing hypothesis in the current context; namely, that number of letters identified from a target word before sending the eyes to it might affect the variability that we have called the random placement error. This is the source of error that produces the Gaussian shape of the landing position distribution and is indexed by the standard deviation of that distribution. We assume that this error arises from a combination of sources: natural lack of precision in saccadic programming, especially with limited response times, error in the accuracy of the eyetracking device, and variation in the center of gravity within the word which results from interword variation in the distribution of letters of different levels in density, reflected here as the number of pixels intensified in

96

R. Radach & G.W. McConkie

creating each letter. However, it is also possible that there may be systematic variability in landing positions related to number of letters peripherally identified from the target word, a factor that itself may vary from saccade to saccade, thus contributing to the spread of the distribution. This raises the question of whether any of the variability in these distributions is due to momentary variation in aspects of the reader's cognitive state. This must be a matter for further investigation, but we doubt that such influence will be found. Conclusions In accounting for eye behavior during reading, McConkie et al. (1988) argued for a word-based control system, with a sharp distinction between selecting a word to serve as the saccade target, and the process of getting the eyes to the target word. They developed the beginnings of a quantitative model to account for data concerning where the eyes land in words, or landing position distributions. This model has been further confirmed and extended by Radach and his co-workers (Radach and Kempe, 1993; Radach 1996). The primary goal of the current investigation was to determine whether this model, which was originally developed to account for progressive inter-word saccades, can be extended to two other conditions: refixations and interword regressive saccades. The original model appears to account for refixations, which raises questions about the sharp distinction that is often made in theories of reading between progressive and refixation saccades. However, interword regressive saccades clearly show different landing position characteristics than do the other cases. Neither word length nor distance from launch site had much effect on the landing positions of these saccades, and the saccadic range effect, typically found with progressive saccades, was absent. Interword regressions appear to be sent to the centers of words, with landing position distributions in which "optimal" and "preferred" viewing positions are very similar. Thus, the original model must be modified to account for regressive inter-word saccades, and further research is needed to understand the basis for the difference. Interestingly, intraword regressive saccades (refixations) did not show these unusual characteristics. Several issues were explored regarding the basis for eye movement control during reading. We have argued that eye movement control is discrete, based on the selection of word targets, rather than being a graded type of control, and have explored some of the implications of this position. In particular, we pointed out that even when a variable produces a significant effect on saccade length, it is likely that the lengths of relatively few of the saccades may have actually been changed. We also examined and argued against alternative positions to our model: specifically, that the landing position function may be due to center of gravity effects, and that the landing positions in selected words may be due to parafoveal letter processing rather than visuo-oculomotor factors.

Fixation positions in reading

97

In sum, we have argued that this one aspect of eye behavior during reading, namely, where the eyes go with respect to selected saccade target words, is the result of low-level visuo-oculomotor control factors, almost completely unaffected by higher cognitive processes. However, this should not be taken to indicate that we think there is no role for cognition in eye guidance. After all, the purpose of reading is to understand text and eye movements are made to serve this purpose. Elsewhere in this book (Chapters 6 and 11) it is argued that cognition plays a significant role in selecting words as saccade targets. We agree with much of what is said on the issue of "word skipping" in these chapters (see also McConkie, Kerr and Dyre, 1994). In this chapter we have presented evidence for a cognitive component, in the form of a word frequency effect, in the decision of whether to refixate or make an inter-word saccade. Finally, in the case of regressive inter-word saccades, the saccade parameters we have looked at suggest a control mode different from the low-level default routines. Although we have not investigated this in the present chapter, it is very likely that not only the large 're-inspection' regressions (Kennedy and Murray, 1987) but also many regressions coming from an adjacent word are based on cognitive grounds. The determination of landing positions within selected saccade targets, is highly regular and can be quantitatively described using models that account for most of the variance in the data. We suggest that investigators who obtain effects of cognitive factors on saccade landing positions consider the possibility that these factors are operating on the target selection stage rather than directly modulating saccade amplitude computation. Separating out these two sources of variance can be a difficult task. The predominant low-level determination of saccade landing positions within selected target words can be good news for researchers studying psycholinguistics processes: while there must be controls for this aspect of eye behavior, it does not appear to be a source of useful dependent variables, thus reducing somewhat the complexity of data analysis. Acknowledgements Preparation of this chapter was supported in part by grant No BMH1-CT94-1441 from the European Union under the BIOMED Programme. The authors are indebted to Dieter Heller, Albrecht Inhoff, Wayne Murray and two anonymous reviewers for helpful discussions and comments on earlier drafts of the manuscript. References Altman, G., Garnham, A. and Dennis, Y. (1992). Avoiding the garden path: eye movements in context. Journal of Memory and Language, 31, 685-712.

98

R. Radach & G. W. McConkie

Becker, W. (1989). Metrics. In: R.H. Wurtz and M.E. Goldberg (Eds.), The Neurobiology of Saccadic Eye Movements. Amsterdam: Elsevier, pp. 13-61. CELEX German database. Release D25. Computer software. Nijmegen: Centre for Lexical Information, 1995. Daneman, M, Reingold, E.M. and Davidson, M. (1995). Time course of phologogical activation during reading: Evidence from eye fixations. Journal of Experimental Psychology: Learning, Memory and Cognition, 21, 884-898. Findlay, J.M. (1982). Global processing for saccadic eye movements. Vision Research, 22, 1033-1045. Heller, D. (1982). Eye movements in reading. In: R. Groner and P. Fraisse (Eds.), Cognition and Eye Movements. Berlin: Deutscher Verlag der Wissenschaften, pp. 139-154. Henderson, J.M. and Ferreira, F. (1990). Effects of foveal processing difficulty on the perceptual span in reading: Implications for attention and eye movement control. Journal of Experimental Psychology: Learning, Memory, and Cognition, 16 (3), 417-429. Hofmeister, J. (1997). Uber Korrektursakkaden beim Lesen und bei leseaehnlichen Aufgaben. (On corrective saccades in reading and reading-like tasks). Unpublished doctoral dissertation. Technical University of Aachen. Hyona, J., Niemi, P. and Underwood, G. (1989). Reading long words embedded in sentences: Informativeness of word halves affects eye movements. Journal of Experimental Psychology: Human Perception and Performances, 15, 142-152. Hyona, J. (1995). Do irregular letter combinations attract reader's attention. Evidence from fixation locations in words. Journal of Experimental Psychology: Human Perception and Performance, 21, 68-81. Inhoff, A.W. (1989). Parafoveal processing of words and saccade computation during eye fixations in reading. Journal of Experimental Psychology: Human Perception and Performance, 15, 544-555. Kennedy, A. (1992). The spatial coding hypothesis. In: Rayner, K. (Ed.), Eye Movements and Visual Cognition. Scene Perception and Reading. New York: Springer. Kerr, Paul. W. (1992). Eye movement control during reading: the selection of where to send the eyes. Unpublished doctoral dissertation, University of Illinois at Urbana-Champaign. Kliegl, R., Olson, R.K. and Davidson, B.J. (1983). On problems of unconfounding perceptual and language processes. In: K. Rayner (Ed.), Eye Movements in Reading and Perceptual and Language Processes. New York: Academic Press, pp. 333-343. Kapoula, Z. and Robinson, D.A. (1986). Saccadic undershoot is not inevitable: Saccades can be accurate. Vision Research, 26 (5), 735-743. McConkie, G.W. (1979). On the role and control of eye movements in reading. In: P. A. Kolers, M.E. Wrolstad and H. Bouma (Eds.), Processing of Visible Language, Vol. I. New York: Plenum Press, pp. 37-48. McConkie, G.W. (1981). Evaluating and reporting data quality in eye movement research. Behavior Research Methods and Instrumentation, 13,97-106. McConkie, G.W. and Zola, D. (1984). Eye movement control during reading: The effects of word units. In: W. Prinz and A.T. Sanders (Eds.), Cognition and Motor Processes. Berlin: Springer, 63-74. McConkie, G.W., Zola, D. and Wolverton, G.S. (1985). Estimating frequency and size of effects due to experimental manipulations in eye movement research. In: R. Groner,

Fixation positions in reading

99

G.W. McConkie and C. Menz (Eds.), Eye Movements and Human Information Processing. Amsterdam: Elsevier, pp. 137-147. McConkie, G.W., Grimes, J.M., Kerr, P.W. and Zola, D. (1990). Children's eye movements during reading. In: J.F. Stein (Ed.), Vision and Visual Dyslexia. McConkie, G.W., Kerr, P.W. and Dyre, B.P. (1994). What are 'normal' eye movements during reading: Toward a mathematical description. In: J. Ygge and G. Lennerstrand (Eds.), Eye Movements in Reading. Oxford: Elsevier, pp 315-327. McConkie, G.W., Kerr, P.W., Reddix, M.D. and Zola, D. (1988). Eye movement control during reading: I. The location of initial eye fixation on words. Vision Research, 28, 1107-1118. McConkie, G.W., Kerr, P.W., Reddix, M.D., Zola, D. and Jacobs, A.M. (1989). Eye movement control during reading: II. Frequency of refixating a word. Perception and Psychophysics, 46, 245-253. McConkie, G.W., Wolverton, G.S. and Zola, D. (1984). Instrumentation considerations in research involving eye-movement contingent stimulus control. In: A.G. Gale and F. Johnson (Eds.), Theoretical and Applied Aspects of Eye Movement Research. Amsterdam: North-Holland. Morrison, R.E. (1984). Manipulation of stimulus onset delay in reading: Evidence for parallel programming of saccades. Journal of Experimental Psychology: Human Perception and Performance, 10, 667-682. Nazir, T.A., Heller, D. and Sussmann, C. (1992). Letter visibility and word recognition: The optimal viewing position in printed words. Perception and Psychophysics, 52 (3), 315-328. O'Regan, J.K. (1990). Eye movements and reading. In: E. Kowler (Ed.), Reviews of Oculomotor Research, Vol. 4. Eye Movements and Their Role in Visual and Cognitive Processes. Amsterdam: Elsevier. O'Regan, J.K. (1992). Optimal viewing position in words and the strategy-tactics theory of eye movements in reading. In: K. Rayner (Ed.), Eye Movements and Visual Cognition: Scene Perception and Reading. New York: Springer Verlag. O'Regan, J.K., Levy-Schoen, A., Pynte, J. and Brugaillere, B. (1984). Convenient fixation location within isolated words of different length and structure. Journal of Experimental Psychology: Human Perception and Performance, 10, 250-257. O'Regan, J.K., Vitu, F., Radach, R., Kerr, P. (1994): Effects of local processing and oculomotor factors in eye movement guidance in reading. In: J. Ygge and G. Lennerstrand (Eds.), Eye Movements in Reading. New York: Pergamon Press. Paap, K.R., Newsome, S.L., McDonald, J.E. and Schvaneveldt, R.W. (1982). An activationverification model for letter and word recognition: The word-superiority-effect. Psychological Review, 89, 573-594. Pollatsek, A., Rayner, K. and Balota, D.A. (1986). Inferences about eye movement control from the perceptual span in reading. Perception and Psychophysics, 40, 123-130. Pollatsek, A. and Rayner, K. (1990). Eye movements and lexical access in reading. In: D.A. Balota, G.B. Flores d'Arcais and K. Rayner, (Eds), Comprehension Processes in Reading. Hillsdale, NJ: Erlbaum, pp. 143-164. Poulton, E.G. (1981). Human manual control. In: V.B. Brooks (Ed.), Handbook of Physiology, Sect. 1, Vol. II, Part 2, pp. 1337-1389, Bethesda: American Physiology Society.

100

R. Radach & G. W. McConkie

Pynte, J., Kennedy, A. and Murray, W.S. (1991). Within-word inspection strategies in continuous reading: time course of perceptual, lexical and contextual processes. Journal of Experimental Psychology: Human Perception and Performance, 17 (2), 458-470. Radach, R. (1996). Blickbewegungen beim Lesen: Psychologische Aspekte der Determination von Fixationspositionen (Eye movements in reading: psychological aspects of the determination of fixation positions). Miinster/New York: Waxmann. Radach, R. and Kempe, V. (1993). An individual analysis of fixation positions in reading. In: G. d'Ydevalle and J. van Rensbergen (Eds.), Perception and Cognition. Advances in Eye Movement Research. Amsterdam: Elsevier/North-Holland. Radach, R., Heller, H., Krummenacher, J. and Hofmeister, J. (1995). Individual eye movement patterns in word recognition: perceptual and linguistic factors. In: J.M. Findlay, R. Walker and R.W. Kentridge (Eds.), Eye Movement Research: Processes, Mechanisms and Applications. Amsterdam: Elsevier/North-Holland. Radach, R., Heller, D. and Inhoff, A. (1997). Blickbewegungen and kognitive Prozesse: Stand und Perspektiven (eye movements and cognitive processes: current issues and developments). In: H. Mandl (Ed.), Bericht iiber den 40. Kongress der Deutschen Gesellschaft fur Psychologic (Proceedings of the 40th Congress of the German Psychological Society). Hogrefe, Goettingen. Radach, R. and Heller, D. Interrelations between spatial and temporal aspects of eye movement control in reading. Manuscript in preparation. Rayner, K. (1979). Eye guidance in reading: Fixation locations within words. Perception, 8, 21-30. Rayner, K. and Morris, R.E. (1981). Eye movements and identifying words in parafoveal vision. Bulletin of the Psychonomic Society, 17, 135-138. Rayner, K. and Morris, R. (1992). Eye movement control in reading: Evidence against semantic preprocessing. Journal of Experimental Psychology: Human Perception and Performance, 18, 163-172. Rayner, K. and Pollatsek, A. (1989). The Psychology of Reading. Boston: Prentice-Hall. Rayner, K. and Sereno, S. (1994). Regressive eye movements and sentence parsing: on the use of regression-contingent analysis. Memory and Cognition, 22, 281-285. Rayner, K., Sereno, S.C. and Raney, G.E. (1996). Eye movement control in reading: a comparison of two types of models. Journal of Experimental Psychology: Human Perception and Performance, 22, No.5, 1188-1200 Underwood, G., Bloomfield, R. and Clews, S. (1988). Information influences the pattern of fixation during sentence comprehension. Perception, 17, 267-278. Underwood, G., Clews, S. and Everatt, J. (1990). How do readers know where to look next? Local information distributions influence eye fixations. Quarterly Journal of Experimental Psychology, 42, 39-65. Vitu, F. (1991 a). The existence of a center of gravity effect during reading. Vision Research, 31, 1289-1313. Vitu, F. (1991b). Against the existence of a range effect during reading. Vision Research, 31, 2009-2015.

101

CHAPTER 5

About Regressive Saccades in Reading and Their Relation to Word Identification Fran9oise Vitu Universite Rene Descartes George W. McConkie University of Illinois at Urbana-Champaign and David Zola University of Illinois at Urbana-Champaign

Abstract The purpose of the present study is to describe some of the conditions under which regressive saccades occur in reading for fifth-grade children. The results obtained replicate the prior finding that regressions are more likely to occur following larger progressive saccades (Andriessen and deVoogd, 1973; Lesevre, 1964), and additionally show that regressions are more frequent following a progressive saccade that skips a word. Further analyses indicate that when a word is skipped, the regression likelihood increases with the skipped word length and the distance from the word of the fixation that precedes or follows skipping. The regression probability also tends to be higher for low- than high-frequency skipped words. These results suggest that regressive saccades in reading are not just the result of a pre-determined oculomotor scanning strategy, but are related to the perceptibility of the words. However, the question of the extent to which regressions in reading reflect word processing problems remains open. Indeed, in cases where no word was skipped, several factors that supposedly affect the ease of processing words did not affect the likelihood of regressing, or had an effect opposite to what would be predicted.

Eye Guidance in Reading and Scene Perception/G. Underwood (Editor) © 1998 Elsevier Science Ltd. All rights reserved

102

F. Vitu, G. W. McConkie & D. Zola

Introduction When reading a text, people's eyes move with saccades of variable sizes which are alternated with fixations of variable durations. The most common pattern is the eyes moving forward from one word to the next, with intermediate fixations lasting for about 250 ms. On some occasions, variations in this general pattern might be observed, such as the eyes skipping the next word, or making an additional fixation on the word before moving on to the next word(s). Sometimes, the eyes might also return to previously read portions of the text, which correspond to what is commonly called 'regressive' saccades in opposition to the more frequent progressive saccades (see Lesevre, 1964; Buswell, 1920). For more than a century, and particularly during the last thirty years, research has been conducted to determine the extent to which this variability in ocular behavior, and particularly that associated with progressive movements, relates to on-going processing of the encountered words and sentences. This research suggests that a single mechanism cannot be responsible for the variability of the ocular behavior in reading. Both higher level lexical and linguistic characteristics of the text and lower level visuomotor constraints and predetermined oculomotor scanning strategies contribute to this variability (for reviews, see Chapters 6 and 11; O'Regan, 1990; Vitu and O'Regan, 1995). At the same time, previous research has put much less effort into the study of regressive saccades. Most theories of eye movements in reading, including recent ones, make only vague assumptions concerning the determinants of regressive saccades. According to several authors, regressive saccades are associated with high processing levels related to sentence and text comprehension (Bouma and deVoogd, 1974; Buswell, 1920; Hochberg, 1976; Just and Carpenter, 1980; Shebilske, 1975; Shebilske and Fisher, 1983). If, in the process of reading, a given comprehension stage requires 'a review of previously read text to reencode it or process it to deeper levels', a regressive saccade is initiated (Just and Carpenter, 1980 (p. 337); see also, Carpenter and Just, 1977). This view is supported by several studies which indicate that the likelihood of making a regressive saccade is a function of several different factors related to the difficulty of the text. First, it depends on the global text difficulty, the order of the words, the presence or not of clarifying punctuation, the presence of a change in word meaning from sentence to sentence, the position of key words for text comprehension, and the tense in which the sentence is written (Bayle, 1942; Klein and Kurlowski, 1974; Shebilske and Fisher, 1983; Wanat, 1976). Second, regressive saccades often occur in semantically or syntactically ambiguous sentences where a given word may have different meanings or syntactic functions. In ambiguous sentences, the eyes often regress from the disambiguating region back to the ambiguous region or to the beginning of the sentence (see Rayner and Pollatsek, 1989, for a review). Thus, regressive saccades are related to some extent

Regressive saccades in reading

103

to semantic and syntactic word integration processes. According to Frazier and Rayner's (1982) garden path model of sentence processing, such regressive saccades occur when the ambiguous word has been incorrectly interpreted during the eyes' first encounter with it. All regressive saccades that occur in reading might not be a result of high processing levels related to sentence and text comprehension, but some might result from lower-level processes related to word identification and/or saccadic programming. First, there is not a perfect temporal matching between the on-going processing of the encountered words and the ocular activity; processing of a word sometimes lags behind the ocular activity (Schroyens et al., 1998; see also Chapters 2 and 11). The fixation and gaze durations on a word are affected not only by the word's lexical/linguistic properties, but also by those of the previously fixated word. This finding, referred to as 'spillover effect', suggests that the processing induced by a word might not be complete when the eyes move on to another word. Increasing the time the eyes spend on a subsequently fixated word might not be the only example of delayed processing effects. Another way to complete the processing of a previously fixated word might be to bring the eyes back to it. Thus, at least some regressive saccades in reading might result from processing difficulties at the lexical level, such as word recognition failures or failures to complete the processing induced by a word before the eyes moved on to another word (Bouma, 1978; Bouma and deVoogd, 1974; Pollatsek and Rayner, 1990; Shebilske, 1975). Finally, progressive saccades in reading are subject to inaccuracies in eye positioning, which result in unintended skipping of a word, or landing at variable locations in the words (see O'Regan, 1990; and Chapter 6). Regressive saccades might occur in response to such oculomotor aiming errors, bringing the eyes back at the intended location (Taylor, 1971). As one example of lower-level influences, within-word regressive saccades are more likely to occur when the eyes initially land near the end of a word than when they land in the middle of the word (O'Regan and Levy-Schoen, 1987; McConkie et al., 1989; Rayner, Sereno, and Raney, 1996; Vitu, O'Regan and Mittau, 1990; see also Pynte, 1996). This result might support the notion that regressive saccades are initiated in response to word identification processes. Indeed, because of visual acuity limitations, word processing is more efficient when the eyes are near the center of the word, and an additional fixation is often required when the eyes do not initially land at that optimal location (Nazir, O'Regan and Jacobs, 1991; O'Regan, 1990). However, in subjects scanning meaningless letter strings where word and letter identification are not involved, the proportion of regressive saccades is very similar to that observed in normal reading (Vitu et al., 1995; see also Rayner and Fischer, 1996 for slightly different results). Furthermore, as in reading, there is a relationship between the eyes' initial location in a letter string and the occurrence of a regressive saccade within the string. This suggests that some within-word regressive

104

F. Vita, G. W. McConkie & D. Zola

saccades in reading result from an oculomotor strategy (O' Regan, 1990; Vitu and O'Regan, 1995) rather than from the necessities of on-going processing associated with the fixated words. The strategy would consist of making an additional fixation on a word when the first fixation is mislocated. It has also been noted that regressive saccades are more likely to occur following longer forward saccades (Andriessen and deVoogd, 1973; Lesevre, 1964). The basis for this relationship is not clear. A longer saccade is more likely to skip over a word, which may necessitate returning the eyes to that word later. Longer saccades are also less accurate, thus favoring the likelihood of ill-placed fixation locations. In addition, during the fixation prior to a long saccade, less visual information is available from the saccade target word, since it lies further into the visual periphery. Thus, regressive saccades after long forward saccades might result from oculomotor aiming errors, or word recognition failures. Finally, it is possible that these regressive saccades simply result from a strategy of moving backwards after making a large progressive saccade. The above review yields four possible bases for making regressive saccades in reading: (1) high-level comprehension failure, (2) failure to fully identify a fixated vs. skipped word prior to moving the eyes, (3) mislocation of the eyes, resulting in a corrective saccade, and (4) a pre-determined oculomotor strategy that produces regressive saccades under certain conditions, such as when the prior saccade was large, or the eyes land near the end of a word. It should be noted that these are not exclusive of one another; more than one of these could be operating at the same or different times during reading. The present study is an attempt to describe some of the conditions under which regressive saccades occur in normal reading in order to begin to determine to what extent each of the above explanations underlie the making of regressions. This was done through the analysis of a large corpus of eye movement data from fifth-grade children reading half a novel appropriate to their age. This corpus was selected because fifth-grade children make more regressions than do adults, while still showing patterns similar to those of adults (Vitu, McConkie, Kerr, and O'Regan, in preparation). For example, both fifth-grade children and adults are less likely to refixate a word following an initial fixation near its center than at its ends (the Refixation Optimal Viewing Position effect), and are less likely to refixate a highthan low-frequency word. Fixation durations also present similar patterns for both fifth-grade children and adults as a function of both the initial fixation location and the word frequency. Since we did not have access to language variables or independent behavioral data that could be used to indicate comprehension failure, we were unable to test the high-level comprehension failure hypothesis. Instead, the analyses were directed at testing between explanations of regressions that are based on word processing factors and those that assume strictly oculomotor strategies. To accomplish this,

Regressive saccades in reading

105

several factors were identified that would be expected, on the basis of past research, to affect word processing. This included visual factors which affect the visibility of the word's constituent letters (word length, distance of the word from the current fixation location, fixation location within the word), word lexical characteristics (word frequency), as well as eye movement related factors (whether or not a word had been refixated). If these word-processing-related variables are found to affect the likelihood of making regressive saccades, this stands as evidence against a strictly oculomotor strategy as a basis for producing regressions. If word lexical characteristics, such as the word frequency, affect the likelihood of regressing in appropriate ways, this clearly suggests that word processing difficulties may underlie regressions. Method Thirty fifth-grade children (approximately 12 years old) participated in the experiment which was run at the Center for the Study of Reading (University of Illinois at Urbana-Champaign). Subjects' eye movements were recorded as they read the first six chapters of 'Old Yeller', a children's novel, written by Frederick B. Gipson. Their eyes were monitored using an SRI Dual-Purkinje Image Eyetracker (Cornsweet and Crane, 1973), sampling eye position every millisecond. The text was presented on a computer screen, seven lines at a time in triple-spaced format. The reader advanced the page by pressing a button. Subjects were asked to read the story for meaning and were tested for comprehension after each chapter. A set of 395,415 eye fixations was obtained. For the following analyses, only the cases in which there was no blink or other signal irregularity present were considered. Furthermore, since within- and betweenline saccades might result from different mechanisms, we restricted the present analysis to within-line regressive saccades. All between-line saccades as well as the short regressive saccades that often occur after a return sweep were eliminated. Analyses that consisted of measuring the regression likelihood considered only cases where the prior saccade was progressive. In addition, both the origin word from which the progressive saccade was initiated and the destination word where the progressive saccade terminated had never previously been fixated. Finally, for the particular cases where the progressive saccade led the eyes to skip one or several words, only cases where the skipped word had never been fixated before being skipped were considered. After selection, the total number of saccades available for analysis was 76,561, with a total of 31,390 cases in which a word was skipped (3843, 8107, 11632, 5519, 1383, respectively for 1- to 5-letter skipped words), and 45,171 cases in which the saccade took the eyes to the immediately following word (1271, 4047, 11493, 11291, 6664, 5233, respectively for 1- to 5-letter origin words in non-skip cases).

106

F. Vitu, G.W. McConkie & D. Zola

For each dependent variable, means or proportions were calculated for each subject, and these were then averaged across subjects. In this way, the weight of individual subjects' contributions to the final values were not influenced by the number of fixations that qualified for a particular condition. Analyses of variance were run on means or proportions obtained from the different individual subjects. Results General characteristics For the population being tested, 28% of all within-line saccades were regressive. The average size of within-line regressive and progressive saccades was -4.6 and 6.8 letters and the average fixation duration was 252 ms. Forty percent of the progressive saccades skipped a word, and among the fixated words, 21 % received at least two consecutive fixations during their first encounter (first pass). Figure la presents the frequency distribution of the lengths of within-line regressive saccades. Saccades of 1 to 3 characters in length were most frequent, accounting for 54% of all regressions, but some are much longer. The proportion of

Fig. 1. Distributions of the size of within-line regressive saccades when expressed in letters (1 a) or words (1b).

Regressive saccades in reading

107

very long regressive saccades (more than 14 characters) is small, being less than 3%. Figure Ib shows the lengths of these saccades in terms of words, with 0 being refixations of the same word. Refixations and regressions to the immediately preceding word account for 85% of all regressions. Further analyses show that 80% of regressive saccades are preceded by a progressive eye movement, and in these cases, regressions are more likely to occur after the previous word was skipped (49%) than after the previous word was fixated (39%) or after a refixation of the same word (12%). Effects of prior saccade length and word skipping Investigators (Andriessen and deVoogd, 1973; Lesevre, 1964) have reported that regressive saccades are more likely to occur following longer progressive saccades. Figure 2a presents the likelihood of regressing following saccades of different lengths in the current data set, and indicates that this same relationship holds in the present data, F(l5,435) = 49.06, p < 0.0005. However, the rise in regression likelihood only begins with saccades of 6 letter positions; shorter saccades show very little relationship, though there is a small local maximum at 2 letter positions and minimum at 5. Since longer saccades are more likely to skip words, it is possible that it is actually the skipping of words that leads to regressions. Figure 2b shows the likelihood of regressing following saccades of different lengths in cases where a word is skipped or is not skipped. Ws for long saccades that do not skip words become quite low. The figure shows that, for most saccade lengths, the likelihood of regressing is indeed increased when a word is skipped, F(l,29) = 15.80, p < 0.0005. However, for both word skip and non-word-skip cases, there is still a strong relationship between saccade length and regression likelihood, F( 12,348) = 45.54, p < 0.0005 and F(ll,319) = 15.92, p < 0.0005, respectively, with a significant interaction, F(9,261) = 2.09, p < 0.05. Thus, while skipping words does increase the likelihood of regressing, this is not the sole explanation for the prior saccade length effect. Effects of skipped word length It is possible that the perceptibility of a word skipped by a progressive saccade might affect the likelihood of regressing after skipping it, either because this influences the likelihood of being able to identify it on the fixation prior to, or the fixation following, the saccade. A first way to test this hypothesis is to look at the effects of the skipped word length on the regression likelihood. Indeed, as a result of visual acuity limitations, the perceptibility of a word depends on its length: the longer a word is, the fewer letters from the word are visible when it is located in

108

F. Vitu, G.W. McConkie & D. Zola

Fig. 2. Probability of making a regressive saccade after a progressive saccade of variable length (2a), and separately for cases where the progressive saccade skips a word or not (2b).

parafoveal vision, and the lower the identification probability for that word (Brysbaert, Vitu and Schroyens, 1996). Since longer words are less frequently skipped by children, the current analysis was restricted to words of length 1 to 5. In addition, to increase the ATs underlying the proportions, prior saccade lengths were grouped into bins of two letter positions. Figure 3 shows that skipped word length does indeed affect the likelihood of a regression occurring, when saccade length is controlled (between 6 and 14 letters), F(4,116) = 12.42, p < 0.0005. The regression likelihood is higher following skipping of 2- and 3-letter words (0.45 and 0.47, respectively) than following skipping of 1-letter words (0.35), and it is higher following skipping of 4- and 5-letter words (0.53 and 0.54, respectively) than following skipping of shorter words. The effect of skipped word length remains significant when only tested on 2- to 5-letter skipped words, F(3,87) = 3.12, p < 0.05. Finally, progressive saccades that skip words of all lengths studied still show a significant effect of saccade length (at the 0.0005 level), longer saccades being more frequently followed by regressions, F(4,l 16) = 14.33, F(5,145) = 14.34, F(6,174) = 30.51, F(6,174) = 33.70, F(6,174) = 8.61, for words of length 1-5. There was no interaction between saccade length and skipped word length, F( 12,348) = 0.78, for prior saccade lengths between 6 and 14 letters.

Regressive saccades in reading

10,2]

]4.6]

]8,10]

]I2,14]

109

]16,20]

PRIOR SACCADE LENGTH (LETTERS)

Fig. 3. Probability of making a regressive saccade after a word is skipped with a progressive saccade, as a function of both the progressive saccade length and the skipped word length.

Effect affixation location with respect to a skipped word Another factor that affects the perceptibility of a skipped word is the distance of the eyes from that word: the further the eyes from the word, the lower the probability of correctly identifying that word (Brysbaert et al., 1996). We therefore investigated whether the location of a skipped word, with respect to the position of the eyes prior to the progressive saccade that skipped it, affects the likelihood of regressing following that saccade. Figures 4a-d show the relationship between prior saccade size and regression likelihood following progressive saccades initiated from different distances (here called 'launch sites') to the left of a skipped word (of two to six letters, the data for 5- and 6-letter words being grouped together). In order to increase the APs underlying the proportions, launch sites were grouped into bins of two letter positions. As Figs. 4a-d clearly indicate, regression likelihood increases with launch site when saccade length is controlled, but this applies only to words longer than 2 letters; data for 2-letter words show no difference between adjacent launch sites, or an effect opposite to what was predicted. Furthermore, there is a progression such that the longer the word is, the larger the effect of launch site is. While for 3-letter

110

F. Vitu, G. W. McConkie & D. Zola

words, the difference mainly occurs between cases where the eyes were launched from 0 to 2 letters from the space before the word, and cases where the eyes were launched from further away, for longer words, differences between the different adjacent launch sites start to emerge. To test whether the differences being observed were significant, pair comparisons were made for each word length between each two adjacent launch site intervals (the prior saccade length in a given comparison pair was held constant). The results obtained show that none of the comparisons made for two-letter skipped words are significant, Fs(l,29) < 3.45. In contrast, for all three other tested word lengths, there is a significant difference between launch site intervals [0,2] and [2,4], F(l,29) = 5.89, p < 0.05, F(l,29) = 5.56, p < 0.05, F(l,29) = 9.11, p < 0.005, respectively for 3-, 4-, and 5/6-letter skipped words, but no significant difference between other launch site intervals, Fs < 2.77. The interaction between launch site and prior saccade length was never significant, Fs < 2.24. In addition, for all three tested word lengths, saccade length does continue to produce a significant effect when launch site is controlled, except in a few cases where the analysis relies on a smaller N, F(2,58) = 3.08, p < 0.05, F(2,58) = 4.73, p < 0.05, F(2,58) = 8.52, p < 0.0005, F(2,58) = 1.88 (2-letter skipped words), F(2,58) = 5.59, p < 0.01, F(2,58) = 13.72, p < 0.0005, F(2,58) = 13.22, p < 0.0005, F(l,29) = 5.63, p < 0.05 (3-letter skipped words), and F(2,58) = 9.40, p < 0.0005, F(2,58) = 2.90, p < 0.05, F(2,58) =1.15, F(l,29) = 0.32 (4-letter skipped words), F(l,29) = 1.75, F(l,29) = 0.003, F(l,29) = 0.008 (5- and 6-letter skipped words). An interpretation of these data can be made by considering where the progressive saccade lands with respect to the skipped word. For instance, in Fig. 4b which presents data for 3-letter skipped words, the first data point of each curve represents data for cases where the eyes land on the space following the skipped word or the first letter of the immediately following word; the second data point represents data where the eyes land on the first, second or third letter of the immediately following word; the third data point represents data where the eyes land on the third, fourth or fifth letter of the immediately following word. Thus, it appears that the likelihood of regressing is depressed if the eyes land immediately following a skipped word, with this likelihood rising as this distance increases. These observations suggest that when a word is skipped, the likelihood of regressing is not just a simple function of the length of the progressive saccade that skipped the word. Rather, this likelihood is a function of the eyes' location preceding

Opposite page: Fig. 4. Probability of making a regressive saccade after a word is skipped with a progressive saccade, as a function of both the progressive saccade length and the launch site from which the progressive saccade is initiated, separately for 2- (4a), 3- (5a), and 4-(4c) letter skipped words. Launch sites are measured relative to the space before the prior skipped word and are grouped in four intervals: [0,2], [2,4], [4,6], [6,8].

Regressive saccades in reading

111

112

F. Vitu, G. W. McConkie & D. Zola

and following skipping. This pattern is compatible with a word perceptibility (or lexical) explanation of these regressions: regressions to a skipped word are needed if the launch site is further from it and the word is at least 3 letters long (resulting from less information being acquired from it), but not if the eyes land so close to it that it can then be identified on the fixation following the saccade. Thus, it appears that the saccade length effect on regression likelihood at least in skip cases, may actually be the result of word-based processes taking place: the longer the saccade, the less likely a skipped word might be identified in parafoveal vision (before and/or after skipping), and the more likely a regressive saccade is initiated. Effects of skipped word frequency As shown in several studies, the perceptibility of a skipped word is not only a function of factors that affect the visibility of its constituent letters, but it is also a function of its frequency of occurrence in the language. Indeed, the less frequent a word is, the smaller the benefit that results from the possibility of seeing the word in parafoveal vision (Inhoff and Rayner, 1986; Vitu, 1991). To further test the lexical hypothesis, we therefore tested the likelihood of regressing following skipping of a word with a progressive saccade, as a function of the skipped word frequency of occurrence in the language. For this analysis, two classes of word frequency were defined using the American Heritage corpus (Carroll, Davies, and Richman, 1971), such that approximately the same number of cases were available in each class, and with words of moderate frequency excluded. High frequency words were defined as having a frequency greater than 196 occurrences per million; lower frequency as having a frequency less than 118 per million. This analysis was done across all preceding saccade size intervals but only for 3- and 4-letter skipped words. The obtained median word frequencies were 8046 and 2087 for 3- and 4-letter words of high frequency, and 49 and 71 for 3- and 4-letter words of low-frequency. The number of data available for low-frequency words was much lower than that for high-frequency words, which resulted in a smaller range of prior saccade lengths for low-frequency words. Results show for 3-, and to a certain extent 4-letter words, that there is a tendency for regressive saccades to be more likely to occur after a low-frequency skipped word (see Figs. 5a-b). The effect of word frequency (tested only for prior saccade lengths between 6 and 10 letters) was globally significant, F(l,29) = 10.96, p < 0.005, as well as the interaction between word frequency and word length, F(\ ,29) = 11.40, p < 0.005. Further analyses reveal that the effect of word frequency was significant only for 3-letter words, F(l,29) = 20.65, p < 0.0005, but not for 4-letter words, F(l,29) = 1.47. No other interaction was significant. Since in the present analyses, the launch site prior to skipping was not controlled, the observed effects of word frequency on regression likelihood might have resulted

Regressive saccades in reading

113

Fig. 5. Probability of making a regressive saccade after a 3- (a) or 4-letter word (b) is skipped with a progressive saccade, as a function of both the progressive saccade length and the skipped word frequency.

from the eyes being launched from different locations for high- and low-frequency words. To ensure that this was not the case, the distributions of launch sites prior to skipping were plotted for low- against high-frequency skipped words, separately for each saccade length interval and word length (3 and 4 letters). Results show that while the distributions for both word frequencies overlap in the cases of 4-letter skipped words, there is a systematic tendency for closer launch sites in the lowfrequency condition for 3-letter words. Since closer launch sites result in less regressions (see Fig. 4b), the launch site might not be responsible for the fact that more regressions occur after low-frequency skipped words. Thus, the likelihood of making a regression following skipping of a word is not only a function of visual factors that affect the visibility of the word's constituent letters, but it depends also on the skipped word frequency (at least for 3-letter words). This result suggests that at least some regressive saccades are initiated when processing of a word is not complete before skipping of the word occurs. Since low-frequency words are less easily identified in parafovea than high-frequency words (Inhoff and Rayner, 1986; Vitu, 1991), they are therefore more likely to produce regressions in word skip cases.

114

F. Vitu, G.W. McConkie & D. Zola

Effect of origin word length when no word is skipped Most progressive saccades do not skip a word. We now ask whether the factors that affect the perceptibility of the word from which these saccades originate, which we will call the 'origin word,' influence the likelihood of regressing following the saccade. A first step consists of testing whether the likelihood of regressing varies as a function of the origin word length. Figure 6 shows the likelihood of making a regression following a saccade that originated from words of different lengths from 1 to 6 letters. It indicates a general tendency for more regressions to occur following saccades originating from shorter words, F(5,145) = 5.11, p < 0.0005, when prior saccade length is controlled (between 2 and 8 letters). Data for words of each length show significant effects of prior saccade length except 1 - and 2-letter words, F(2,58) = 2.98, p < 0.10, F(3,87) = 0.41, F(4,116) = 3.09, p < 0.05, F(4,116) = 3.84, p < 0.01, F(3,87) = 19.89, p < 0.0005, F(4,l 16) = 5.56, p < 0.0005, respectively for 1- to 6-letter origin words.

]0.2]

J4.6]

18.10]

)12,I4)

] 16.20)

PRIOR SACCADE LENGTH (LETTERS)

Fig. 6. Probability of making a regressive saccade after a progressive saccade that does not skip a word, as a function of both the progressive saccade length and the length of the origin word from which the progressive saccade is initiated.

Regressive saccades in reading

115

Effect of fixation location on the origin word Previous research suggests that a word is more likely to be identified with a single fixation when the eyes are located near the center of the word (a position from which most letters of the word can be seen) than when they are located at its beginning or end (Brysbaert et al., 1996; Nazir, et al., 1991; O'Regan, 1990). It might therefore be that regressive saccades are more likely to occur when the origin word was fixated only once with the eyes at a non-optimal location for word processing. To test this hypothesis, an analysis was conducted which consisted of measuring whether the eyes' position on the origin word, when no word was skipped, influenced the likelihood of regressing following a forward saccade, with saccade length held constant. This analysis was performed only on cases where the origin word was fixated with a single fixation. To ensure that the number of data available for each condition was large enough, fixation positions in the origin word were grouped into three classes: beginning (Letters 0 and 1 for 4-letter words, and Letters 0,1, and 2 for 5- and 6-letter words), middle (Letters 2 and 3, for 4-letter words, Letter 3 for 5-letter words, and Letters 3 and 4 for 6-letter words), and end (Letter 4 for 4-letter words, Letters 4 and 5 for 5-letter words, and Letters 5 and 6 for 6-letter words). Results indicate a tendency for regressions to be more likely following fixations located near the end of the origin word than after fixations located near the beginning or middle of the word, but this applies mostly for 4-letter origin words; the effect of fixation position for longer words is not very consistent. When prior saccade length was controlled (between 4 and 8 letters), the effect of eye position in the origin word was globally significant, F(2,58) = 4.56, p < 0.05, but it was significant only for 4-letter words, F(2,58) = 7.90, p < 0.0005, F(2,58) = 2.59, p < 0.10, F(2,58) = 1.97. Despite the present control of the eyes' location within the origin word, the origin word length presented again a pattern opposite to that predicted by the lexical hypothesis, the effect of word length being still significant, F(2,58) = 4.29, p < 0.05. None of the interactions were significant. Effect of eye behavior on the origin word An additional analysis examined whether making more than one fixation on the origin word when no word was skipped affected the likelihood of regressing following the progressive saccade originating from that word. Several studies indicate that a word is less likely to be identified if presented only during the initial fixation, this being the result of less letter information being extracted from the word (O'Regan, 1990). Thus, a regressive saccade might be more likely to occur following a word fixated with a single fixation than following a refixated word. Forward saccades originating from 4- to 6-letter words were divided into those following a single fixation on the origin word, and those following refixations of the word. With saccade length controlled, no effect of the number of origin word

116

F. Vitu, G. W. McConkie & D. Zola

fixations on the regression likelihood was found, F(l,29) = 2.38. The interaction between number of origin word fixations and prior saccade length was also never significant, F(2,58) = 1.85. These tests were made for prior saccade lengths between 4 and 10 letters. Thus, these analyses fail to support the prediction made by the lexical hypothesis that the likelihood of identifying the origin word has an effect on regressing following a progressive saccade from that word. Effect of origin word frequency Since the ease of processing associated with a word is a function of its frequency, an analysis of the effect of origin word frequency was carried out. This analysis which was done separately for different prior saccade length intervals, used the same frequency categories as described earlier, 4- to 6-letter words, and only cases where the origin word received a single fixation. The median of word frequencies were 1504,781, and 457 for 4-, 5-, and 6-letter words of high frequency, and 70,32, and 14 for 4-, 5-, and 6-letter words of low frequency. The results that are presented in Figs. 7a-c show, at least for 5- and 6-letter origin words, a tendency for regressions to be most likely following low-frequency origin words. Analyses of variance (made for prior saccade lengths between 4 and 8 letters) reveal a significant effect of word frequency, F(l,29) = 9.15, p< 0.01, with a marginally significant interaction with word length, F(l,29) = 2.39,p < 0.10. The word frequency effect was actually significant only for 6-letter origin words, F(l,29) = 0.47, F(l,29) = 2.29, F(l,29) = 5.75, p < 0.05. The three-way interaction between word frequency, word length, and prior saccade length was also significant, F(2,58) = 3.20, p < 0.05. To ensure that word frequency was not confounded with the eyes' position in the origin word, we compared the distributions of fixation positions in high- and low-frequency origin words, separately for the different prior saccade length intervals. The results obtained show that for most of the conditions, the distribution of fixation locations do not differ between high- and low-frequency words. The present results therefore provide only weak evidence in favor of the hypothesis that initiating a progressive saccade from less perceptible words results in more regressions following that saccade.

Opposite page: Fig. 7. Probability of making a regressive saccade after a progressive saccade that does not skip a word, as a function of both the progressive saccade length and the origin word frequency, for 4-(a), 5- (b), and 6-letter origin words (c).

Regressive saccades in reading

117

118

F. Vitu, G. W. McConkie & D. Zola

Discussion The goal of the above analyses has been to determine whether factors that are known to affect the processing of words (particularly the origin word and the skipped word) influence the likelihood of making a regression during reading. The factors investigated included visual factors that affect the visibility of the word's constituent letters such as, the word length and the location of the fixation on it (in the case of origin words) or to the left and right of it (in the case of skipped words), lexical characteristics of the word (frequency), and eye behavior on the origin word when there was no word skipped. The analyses started from the observation that longer progressive saccades are more frequently followed by a regression, a fact that was replicated in the current study. Further analyses showed that this effect is related to some extent on whether the progressive saccade skips a word. Indeed, regressions are more likely to occur in cases where a word is skipped than in cases where no word is skipped, and the effect of the progressive saccade length is accentuated for skipped cases. Given the fact that the presence of a skipped word affects the likelihood of regressing, it is reasonable to expect that the characteristics of the skipped word might also have an effect, with less perceptible words inducing more regressions. The results obtained are compatible within certain limits, with this hypothesis. First, the likelihood of regressing is higher for longer words, and when word length is controlled, it tends to be higher for low- compared to high-frequency words. Second, the regression likelihood is a function of the eyes' location preceding and following word skipping. For words longer than 2 letters, the further the skipped word lies from the saccade's launch site, the greater the likelihood of a regression occurring after the saccade. However, if the eyes land very near the end of the word, this reduces regressions, as if the skipped word could be identified in these cases without the need to return the eyes to it. These observations suggests that the easiness of processing associated with a word located in parafoveal vision and that is not fixated during the first eye pass may be controlling regression frequency. Indeed, as shown in prior studies, the visibility of a word decreases with distance into the periphery (Brysbaert et al., 1996), and the efficiency of processing a parafoveal word depends on its length and frequency in the language (Inhoff and Rayner, 1986; Vitu, 1991). However, most forward saccades do not skip words. In these cases, one might also expect that characteristics of the origin word, from which the forward saccade was launched, would affect the likelihood of regressing. The present results do not support this hypothesis. First, although the frequency of the origin word had an effect for at least one of the three tested word lengths (6-letter words), longer origin words led to fewer regressions rather than more, and there was no effect of whether the word received a refixation (a factor which is known to affect the number of

Regressive saccades in reading

119

letters that can be extracted on the word). Second, it is only for one of the three tested word lengths that the location of the fixation on single-fixation origin words tends to affect the regression likelihood, with fixations located near the end of the origin word producing slightly more regressions. With respect to the different explanations for regressive saccades that were stated initially, it appears therefore that they are not just the result of an oculomotor strategy that elicits regressions after long saccades. The regressive saccades that occur after a word was skipped would depend for a large part on the on going processing associated with those words. At the same time, several variables that would be expected to affect the ease of processing origin words in non-word-skip cases did not affect regression likelihood or had an effect opposite to that predicted. This might argue against a pure lexical hypothesis for the occurrence of regressions in reading. The current study is obviously only a beginning toward investigating this issue. However, it does support the need for further studies, indicating that certain word-related variables are indeed influencing the frequency of regressing. An important limitation of the analyses reported above is that no distinction was made regarding the word to which regressions were sent: no distinction was made between a regression that refixates a word, goes to the previous word, or goes beyond the previous word. We accept the position that the control of saccades is word-based; that is, that each saccade is intended for a particular word (see Chapter 4). Given this, the current analyses are only a crude beginning at investigating the hypothesis that regressions are made in response to specific word processing difficulties. This problem particularly applies to the analysis of non-skip cases which might have been biased by a very large proportion of within-word regressions. In non-word-skip cases, the eyes are launched from closer to the destination word (on average 3 letters away from the beginning of the destination word) than in skip cases (7 letters away), which results in the eyes landing further away in the destination word for non-skip (on average between the 2nd and 3rd letter) compared to skip cases (on the first letter) (see McConkie et al., 1988; Vitu et al., 1995 for similar results). At the same time, within-word regressive refixations are more likely to occur when the eyes land near the end of the word and in cases where the word is difficult to process (McConkie et al., 1989; O'Regan and Levy-Schoen, 1987; Rayner et al., 1996, Vitu, 1991; Vitu et al., 1990). It is therefore very likely that the regression likelihood measure was contaminated by a large number of regressions within the destination word, regressions which were produced in response to the difficulty of processing associated with only the destination word, and not the origin word. This would account for our failure to report consistent effects of word-related factors in non-skip cases. The need to distinguish between intraword and interword regressions is emphasized by Radach and McConkie (Chapter 4), who observe that the mean of the landing position distributions of intraword regressions tends to vary linearly with

120

F. Vitu, G.W. McConkie & D. Zola

Fig. 8. Distribution of landing sites in 4-letter words, following a interword regressive saccade initiated from different launch sites.

distance of the launch site with respect to the center of the word, whereas interword regressions do not. Figure 8 shows this same phenomenon in the current data set: the mean landing position for interword saccades going to the immediately preceding word is quite constant following saccades of very different lengths. Thus, intraword and interword regressions appear to be controlled on different bases. Despite the limitations that apply to the analyses reported in this chapter, it remains unequivocal that regressions are more likely to occur following word skipping, and that in those instances, regressions reflect to some extent word-related processes. As shown above, regressions are more likely to occur for long and low-frequency skipped words; they are also more likely to occur when the critical word was further from the center of vision preceding and following skipping. These results suggest that a word might be skipped while it has not been yet identified in parafovea. Thus, the size of progressive saccades would not always be appropriately defined to respond to the necessities of on going processing associated with the parafoveal words, a view which is compatible with several accounts for word skipping in reading (see Chapter 6). The present finding that a word skipping is frequently followed by a regression also raises important questions for the use of eye movements as a tool to investigate

Regressive saccades in reading

121

the cognitive processes involved in reading. In the study of sentence processing, several measures have been defined to track the effects of syntactical and/or semantic ambiguities (see Chapters 2 and 3). Among those are the regression likelihood and the first pass reading time. These measures might be subject to biases if no caution is taken as to whether a word skipping occurred in some sentences, or in some conditions. First, the regression likelihood might be enhanced in one condition as a result of word skipping, a fact which might not be attributable to high processing level influences, but rather to lower-level influences related to word identification or saccadic programming. This possibility is very likely at least in studies which compare the ocular behavior on sentences that differ by the presence or absence of one or two very short words before the critical word (see Chapter 3 for an example). Indeed, word skipping is much larger for shorter words (see Chapter 6). The first pass reading time measure might also be influenced by whether or not a word is skipped. Indeed, the time spent in a given region of text strongly depends on how many fixations occur in that region, a fact again which might have nothing to do with higher level influences (see Trueswell, Tanenhaus, and Kello, 1993, for a similar conclusion). In addition, as shown in Fig. 9, the duration of the fixation that precedes a regression is shorter in word-skip cases (x - 235 ms) than in nonword-skip cases (x = 245) (see Pynte, 1996, for similar results). Fixation durations are also shorter before a regression than before a progressive saccade (and this particularly for word skip cases). Thus, if a regressive saccade is initiated from the critical word, the first pass reading time for the region that contains the critical word will be shortened, and this particularly if the word preceding the critical word was

Fig. 9. Distribution of the duration of fixations that precede either a regressive or progressive saccade in word-skip and non-word skip cases.

122

F. Vitu, G. W. McConkie & D. Zola

skipped. This suggests the necessity, when using measures that sum the time spent on several different words, to ensure that the eye movement pattern which precedes or follows the critical word is similar in the different conditions (see Chapter 3 for a similar conclusion). Finally, there remains the question of what is being extracted from the destination word during the fixation that follows word skipping and precedes a regression toward the skipped word. Is processing of the destination word initiated during that fixation, or is it delayed until processing of the skipped word has been resolved? Acknowledgements

We would like to thank John Grimes who participated in the data collection for the present experiment, and Mark Brysbaert for his helpful comments on an earlier version of the manuscript. We would also like to thank R. Radach, K. Rayner and P. van Diepen for their helpful comments on the submitted version of the chapter. This work was supported by a NATO grant which was given to Fran9oise Vitu while she was a visiting researcher at the Beckman Institute, University of Illinois at Urbana-Champaign (1996-1997). References Andriessen, J.J. and DeVoogd, A.H. (1973). Analysis of eye movement patterns in silent reading. IPO Annual Progress Report, 8, 29-34. Bayle, E. (1942). The nature and causes of regressive movements in reading. Journal of Experimental Education, 11, 16-36. Bouma, H. (1978). Visual search and reading: Eye movements and functional visual field: A tutorial review. In: J. Requin (Ed.), Attention and Performance VII. Hillsdale, NJ: Erlbaum, pp. 115-145. Bouma, H. and De Voogd, A.H. (1974) On the control of eye saccades in reading. Vision Research, 14, 273-284. Brysbaert, M., Vitu, F. and Schroyens, W. (1996). The right field visual advantage and the optimal viewing position effect: On the relation between foveal and parafoveal word recognition. Neuropsychologia, 10 (3), 385-395. Buswell, G.T. (1920). An experimental study of the eye-voice span in reading. Supplementary Educational Monographs, 17, Chicago, University of Chicago. Carpenter, P.A. and Just, M.A. (1977). Integrative processes in comprehension. In: D. Laberge and S.J. Samuels (Eds.), Basic Processes in Reading: Perception and Comprehension. Hillsdale, NJ: Erlbaum. Carroll, J.B., Davies, P. and Richman, B. (1971). The American Heritage Word Frequency Book. New York: Houghton Mifflin. Cornsweet, T.N. and Crane, H.D. (1973). Accurate two-dimensional eye tracker using first and fourth Purkinje images. Journal of the Optical Society of America, 63, 6-13.

Regressive saccades in reading

123

Frazier, L. and Rayner, K. (1982). Making and correcting errors during sentence comprehension: eye movements in the analysis of structurally ambiguous sentences. Cognitive Psychology, 14, 178-210. Hochberg, J. (1976). Toward a speech-plan eye-movement model of reading. In: R.A. Monty and J.W. Senders (Eds.), Eye Movements and Psychological Processes. Hillsdale, New Jersey: Erlbaum. Inhoff, A.W. and Rayner, K. (1986). Parafoveal word processing during eye fixations in reading: Effects of word frequency. Perception and Psychophysics, 40, 431-439. Just, M.A. and Carpenter, P.A. (1980). A theory of reading: from eye fixations to comprehension. Psychological Review, 87, 329-354. Klein, G.A. and Kurlowski, F. (1974). Effect of task demands on relationship between eye movements and sentence complexity. Perceptual and Motor Skills, 39, 463^466. Lesevre, N. (1964). Les mouvements oculaires d'exploration. Etude electro-oculographique comparee d'enfants normaux et d'enfants dyslexiques. These, Paris. McConkie, G.W., Kerr, P.W., Reddix, M.D., Zola, D. and Jacobs, A.M. (1989). Eye movement control during reading: II. Frequency of refixating a word. Perception and Psychophysics, 46, 245-253. Nazir, T., O'Regan, J.K. and Jacobs, A.M. (1991). On words and their letters. Bulletin of the Psychonomic Society, 29, 171-174. O'Regan, J.K. (1990). Eye movements and reading. In: E. Kowler (Ed.), Eye Movements and Their Role in Visual and Cognitive Processes. Amsterdam/Oxford: Elsevier, pp. 395^53. O'Regan, J.K. and Levy-Schoen, A. (1987). Eye movement strategy and tactics in word recognition and reading. In: M. Coltheart (Ed.), Attention and Performance XII: The Psychology of Reading. Hillsdale, NJ: Erlbaum, pp. 363-383. Pollatsek, A. and Rayner, K. (1990). Eye movements and lexical access in reading. In: D.A. Balota, G.B. Flores d'Arcais and K. Rayner (Eds.), Comprehension Processes in Reading. Hillsdale, NJ: Erlbaum, pp. 143-163. Pynte, J. (1996). Lexical control of within-word eye movements. Journal of Experimental Psychology: Human Perception and Performance, 22 (4), 958-969. Rayner, K and Fischer, M.H. (1996). Mindless reading revisited: Eye movements during reading and scanning are different. Perception and Psychophysics, 58 (5), 734-747. Rayner, K. and Pollatsek, S. (1989). The Psychology of Reading. London: Prentice-Hall. Rayner, K., Sereno, S.C. and Raney, G.E. (1996). Eye movement control in reading: A comparison of two types of models. Journal of Experimental Psychology: Human Perception and Performance, 22 (5), 1188-1200. Schroyens, W., Vitu, F., Brysbaert, M. and d'Ydewalle, G. (in press). Visual attention and eye-movement control during reading: The case of parafoveal processing. Submitted to Quarterly Journal of Experimental Psychology. Shebilske, W. (1975) Reading eye movements from an information-processing point of view. In: D. Massaro (Ed.), Understanding Language. New York: Academic Press. Shebilske, W.L. and Fisher, D.F. (1983). Eye movements and context effects during reading of extended discourse. In: K. Rayner (Ed.) Eye Movements in Reading: Perceptual and Language Processes. Taylor, E. (1971). The dynamic activity of reading: A model of the process. Research Infor-

124

F. Vitu, G. W. McConkie & D. Zola

mation Bulletin No 9. New York: Educational Developmental Laboratories. Trueswell, J.C., Tanenhaus, M.K. and Kello, C. (1993). Verb-specific constraints in sentence processing: separating effects of lexical preference from garden-paths. Journal of Experimental Psychology: Learning, Memory and Cognition, 19(3), 528-553. Vitu, F. (1991). The influence of parafoveal preprocessing and linguistic context on the optimal landing position effect. Perception and Psychophysics, 50, 58-75. Vitu, F. and O' Regan, J.K. (1995). A challenge to current theories of eye movements in reading. In: J. Findlay, R.W. Kentridge and R. Walker (Eds.), Eye Movement Research: Mechanisms, Processes and Applications. Amsterdam, Lausanne, NY, Oxford, Shannon, Tokyo: Elsevier, pp. 381-392. Vitu, F., O'Regan J.K. and Mittau, M. (1990). Optimal landing position in reading isolated words and continuous text. Perception and Psychophysics, 47 (6), 583-600. Vitu, F., McConkie, G.W., Kerr, P.W. and O'Regan, J.K. Within-word eye behavior during reading: Effects of initial fixation location on the gaze duration and its components. In preparation. Vitu, F., O'Regan, J.K., Inhoff, A.W. and Topolski, R. (1995). Mindless reading: Eye movement characteristics are similar in scanning strings and reading texts. Perception and Psychophysics, 57, 352-364. Wanat, S.F. (1976). Language behind the eye: Some findings, speculations, and research strategies. In: R.A. Monty and J.W. Senders (Eds.), Eye Movements and Psychological Processes. Hillsdale, NJ: Erlbaum.

125

CHAPTER 6

Word Skipping: Implications for Theories of Eye Movement Control in Reading Marc Brysbaert University of Ghent

and Fran9oise Vitu Universite Rene Descartes

Abstract Eye movements in reading are characterized by short periods of steadiness (fixations) followed by fast movements (saccades). Saccades are needed to bring new information into the centre of the visual field where acuity is best; fixations are required to recognize words. Assuming that the central (foveal) word is identified during a fixation, it is tempting to forward the hypothesis that eye movements in reading essentially consist of word-to-word movements. Unfortunately, such a simple sequence of motion is rarely observed in empirical data. Some words are fixated more than once, some are initially not fixated but immediately afterwards regressed to, and some are not fixated at all. Ever since the first measurements of eye movements in reading, researchers have been puzzled by this complicated pattern of activity and have suggested various explanations for it. In this chapter, we will focus on one aspect, namely the fact that more than one third of the words are initially skipped during reading. First, we will discuss the explanations offered by different authors, then we will examine the empirical evidence more closely, and, finally, we will present an alternative account of word skipping.

Eye Guidance in Reading and Scene Perception/G. Underwood (Editor) © 1998 Elsevier Science Ltd. All rights reserved

126

M. Brysbaert & F. Vitu

Word skipping in different theories The loose relation between eye movements and text layout made many of the first researchers believe that eye movements were controlled by an autonomous oculomotor control centre (e.g., Buswell, 1920; Erdman and Dodge, 1898; Huey, 1908). According to this view, saccade sizes were more or less constant and only changed as a function of the global difficulty of the materials being read. Variations in saccade size resulted from noise in the oculomotor system and adjustments to the difficulty of the text. As for word skipping, this implied that the probability of skipping a word depended on the overall easiness of the text but not on the easiness of the word itself. The autonomous oculomotor control model remained the dominant model until the middle of the 1970s, although some refinements were added. For instance, Bouma and de Voogd (1974; see also Shebilske, 1975) attributed changes in saccade size to the limited capacities of an input buffer which could contain but a few activated word units. According to this view, it is only when the buffer gets full or empty that saccade sizes vary around their means. Other variants of the model speculated that variations in saccade size were determined by visual characteristics of the words, in particular the length of the upcoming, parafoveal, word. For example, having found that the length of inter-word saccades depended on the length of the parafoveal word, O'Regan (1980) hypothesized that the eyes were programmed to jump to the centre of the next word. Again, word skipping was not assumed to depend on the easiness of the word (but see O'Regan, 1979). Gradually, the possibility of linguistic influences on the probability of word skipping began to be taken into account. In his cognitive and peripheral search guidance theory, Hochberg (1975,1976) still maintained that eye movements were primarily determined by pre-established scanning routines independent of the linguistic information extracted at each fixation, but these routines depended on both the readers' knowledge of the language constraints and on the task they set to themselves when reading. If the task required subjects to pay attention to the letters and the spelling of the words without paying attention to the meaning (proofreading), readers were assumed to adopt a letter-by-letter or word-by-word scanning routine. If, however, the reading purpose was to extract meaning from text, then larger saccades were made in order to reduce the number of samples per line. In that case, the size of the saccades was controlled by peripheral search guidance mechanisms: At each fixation, the reader anticipated what he would find next on the basis of the meaning and the grammar status of the words previously read and on the basis of global visual information extracted in parafoveal vision (such as the next word's length). Readers directed their eyes to the portions of text which appeared as the most informative to test the predictions that had been made. So, according to Hochberg, the size of a saccade depended on (i) text difficulty and redundancy

Word skipping

127

(which affect a word's predictability and the size of the perceptual span), and (ii) the reader's ability to extrapolate upcoming information and to process parafoveal stimuli. Similar reasoning can be found in Shebilske (1975). Saccade sizes were mainly determined by the capacities of the input buffer (see above), but on some instances the on-going linguistic processing could intervene. If both the overall meaning of the words read on previous fixations and the visual information extracted from the parafovea (such as the next word's length) made the next word highly predictable, then the automatic oculomotor program could be interrupted and the word skipped. This could occur independently of the state of the internal buffer. So, linguistically induced word skipping was possible when the word was highly predictable from prior context and the parafoveal visual information compatible with this prediction. In the late 1970s, the idea that word skipping did not happen at random but was determined by the probability of the parafoveal word being identified during the previous fixation, was strongly promoted in a series of highly influential papers. McConkie (1979) hypothesized that saccade length was determined by the size of the perceptual span (see also Rayner, 1978). Saccades were tuned at each fixation in order to bring the eyes to the region of text which was not clearly visible from the current fixation location. Therefore, saccade sizes were determined by visual acuity limitations, but also by the ease of processing associated with the words to the right of the fixation location (i.e., their predictability and their frequency). Further support for the processing account of word skipping was provided by the different variants of the attentional theory of eye guidance in reading (Morrison, 1984; Pollatsek and Rayner, 1990; Henderson and Ferreira, 1990; see also Chapter 11 for the most recent developments). According to this theory, during reading there is an interplay between attentional processes and eye movement control. As originally proposed by Morrison (1984; see also McConkie, 1979), the sequence of events within a fixation is as follows. Initially, attention is focused on the word in foveal vision. As soon as this word is identified, the attentional beam shifts toward the next word and this word starts to be processed. The shift of attention triggers the programming of a saccade, which is executed after a more or less fixed delay. If processing of the parafoveal word is completed some time before the end of the delay, word skipping is possible. Because the parafoveal word has been identified, the attentional beam again shifts to the next word, and the saccade to the first parafoveal word will be cancelled and replaced by a saccade to the second word. As the probability of parafoveal word identification before saccade execution depends on the easiness of the parafoveal word, the attentional theory predicts that all word skipping should be linguistically based. Although the attentional theory has dominated eye movement research during the last decade, the idea that word skipping may be controlled by non-linguistic

128

M. Brysbaert & F. Vitu

strategies has not completely vanished. For instance, Just and Carpenter (1980, 1987) argued that saccade sizes were independent of on-going processing requirements, and were preprogrammed so that the eyes fixated every word and landed between the beginning and the middle of a word. The hypothesis was that the information extracted from the parafovea was not very detailed, and served mainly to locate the target word for the next saccade. This information was assumed not to be clear enough for recognition of the parafoveal word to take place and to influence the length of the next saccade. Only in exceptional cases when the eyes were located at the very end of a word and when the next word was very short and frequent was it possible that parafoveal word recognition occurred and led to word skipping. So, according to Just and Carpenter, word skipping was largely independent of the easiness of the parafoveal word, and the variability in saccade sizes mainly resulted from visual and oculomotor factors. Just and Carpenter's view on landing sites is particularly surprising, given how high-level the rest of their theory on eye movement control in reading was. A very similar argument can be found in O'Regan's (1990) strategy-tactics theory. According to this theory, the inter-word behaviour of the eyes was mainly determined by a preset oculomotor strategy that depended on the global difficulty of the text and on task demands. O'Regan basically made a distinction between a careful reading strategy (i.e., going from word to word) and risky reading strategies (e.g., skipping every second word, only looking at long words). A further distinction was made between the saccade target (i.e., the middle or the optimal viewing position of a parafoveal word) and the actual landing position. According to O'Regan, visuomotor constraints considerably reduce the accuracy with which a saccade can be executed during reading. So, in the strategy-tactics theory, there were two main determinants of word skipping. Words were skipped because of the strategy (e.g., because they were too short) or because of an oculomotor error (the eyes jumped too far). In addition, O'Regan (1990) accepted linguistic influences if the previous fixation exceeded a certain duration. In this case, parafoveal word processing could interfere with the oculomotor strategy. Finally, it should be noted that McConkie in more recent publications (e.g., McConkie, Kerr and Dyre, 1994) returned to the idea of independent oculomotor strategies during reading. Having found that the probability of word skipping can be predicted rather well by equations that only involve word length and launch site (see below), he started to question the viability of linguistic control theories. To put it in his own words: "...we have briefly outlined the approach we are taking in our attempt to produce a mathematical model of the eye movements of normal, skilled readers. Our greatest surprise thus far has been to observe how much of the variance in the data can be accounted for with a relatively few parameters, and these often reflecting such low-level variables as word length, eye position in word and launch site." (McConkie et al., 1994, p. 325).

Word skipping

129

Empirical data In this section we will look at the empirical evidence upon which the above theories are based. We will deal successively with the landing position distribution, the effects of word length and launch site, and the contribution of language factors such as word frequency and the extent to which a word is constrained by the sentence context. Finally, we will try to determine the relative importance of visual and linguistic variables, in order to decide which should come first in an acceptable model. Landing position distribution Although the impact of oculomotor errors on eye movements in reading has been defended most strongly by O'Regan and colleagues (e.g., Coe'ffe and O'Regan, 1987; O'Regan, 1990), it was McConkie and his group who presented the most convincing empirical evidence for such an effect (McConkie et al., 1988,1994; see also Radach and Kempe, 1993). A typical study involved a few participants reading a complete novel, so that analyses could be based on a large number of observations per person. One of the major analyses was the frequency distribution of initial fixations on a word as a function of the launch site (i.e., the letter position relative to the word centre from which the eye movement started) and the length of the target word. Figure 1 depicts the prototypical findings. There are two main effects. First, the distribution of landing positions is well captured by a Gaussian curve, and

Fig. 1. Frequency distributions of initial fixations on 7-letter words, following saccades launched from 5,10, and 15 letter positions to the left of the center of the word (reprinted from McConkie et al., 1994).

130

M. Brysbaert & F. Vint

second, the mean of the distribution is a function of the launch site. For each launch site one character position farther to the left, the mean of the landing position distribution moves leftward by about a third of a character position. McConkie attributes this to a range error, by which the system tends to overshoot near targets and undershoot far targets. So far, the landing position distribution has mainly been used by McConkie and colleagues to explain the distribution of initial fixations within words. However, as can be seen in Fig. 1, the pattern can easily be extended to the between-words case (i.e., word skipping), as has been asserted by O'Regan (see above). The strategytactics theory further draws attention to two other visuomotor constraints. The first refers to the fact that the landing error is likely to be larger after short fixation durations than after long fixation durations (Coeffe and O'Regan, 1987). The second points to the fact that eye movements to a target are influenced by the presence of other targets in the visual field (i.e., the global effect; Findlay, 1982; Vitu, 199la). Given that words of a text are not presented in isolation, the claim is that the properties of the words behind the target word will influence the saccade size towards the target word. This is particularly true for short target words. On the basis of the present evidence, it is beyond doubt that all comprehensive theories of word skipping should take the existence of involuntary word skipping due to oculomotor error into account (as well as the possibility that a word is involuntarily looked at due to a saccade undershoot). What is less sure, however, is whether all deviations from O'Regan's deterministic inter-word strategy can be explained by oculomotor errors. This question is especially relevant when one looks at the effects of word length and launch site on word skipping. Word length and launch site One of the most conspicuous aspects of word skipping is that it happens more often with short words than with long words. For instance, Vitu et al. (1995) reported skipping probabilities of about 80% for one-letter words, 60% for three-letter words, 30% for five-letter words, and 10% for words of seven letters or longer. Interestingly, these figures were virtually the same whether words in real sentences were presented, or meaningless z-strings that had the same layout (cf, the issue of mindless reading; Vitu et al., 1995; Rayner and Fischer, 1996). Kerr (1992) was the first to note that the average word length effect hides another, equally strong, effect of launch site. Figure 2 shows some typical data. They are drawn from a study in which 24 participants read 120 sentences (see Brysbaert and Mitchell, 1996, Experiment 3 for further details), resulting in a total of more than 22,000 observations for words from two letters to nine letters. Skipping rate is plotted as a function of word length (2-9 letters) and launch site (1-15 letters; operationalized as the distance in letter positions from the blank space in front of the

Word skipping

131

Fig. 2. Word skipping probability as a function of word length and launch site in letter positions relative to the blank space in front of the word.

word). Although a two-letter word on the average was skipped on 69% of the cases, this figure ranged from 90% at launch site one (the last letter of the previous word) to some 50% at launch site 15 (when the eyes already had skipped one or two words). Similar data have been reported by Kerr (1992), Vitu et al. (1995), and Rayner, Sereno, and Raney (1996). In addition, Vitu et al. (1995) showed that the effect of launch site on skipping rate also applies for meaningless z-strings. It should be noted that the word length and the launch site effect can equally well be explained by theories that are based on autonomous oculomotor scanning strategies and theories that are based on linguistic control. According to the oculomotor control theories, short words and words close to the launch site are skipped more often because they lie in an area of high visibility (cf. acuity drops away from the fixation location) and/or because they are less easy to land upon (e.g., due to the landing position error and the global effect). According to the linguistic control theories, short words and words close to the launch site are skipped more often because they are more likely to be identified before saccade onset. This is again due to the drop of visibility outside the fixation position and, in the case of word length, to the fact that short words in general are easier than long words (e.g., they usually have a higher frequency). In order to decide between both types of theories, researchers have tried to disentangle length and processing difficulty by looking at skipping rates for carefully controlled stimulus words and sentences.

132

M. Brysbaert & F. Vitu

Language factors The only way to find out whether parafoveal words are skipped because they were identified during the previous fixation, is to examine skipping rate for stimulus materials that are identical except for the target word. Otherwise, it is impossible to disentangle the effects of the scanning strategy (which can be a function of the difficulty of the text up to the target word and of the length of the words in front of the target word) from those of language processing. In addition, a difference may be drawn between studies in which the processing load of the target word was a function of properties of the word itself, and studies where the processing load was a function of the extent to which the word was constrained by the previous words in the sentence. As not all models predict an equal effect size for both cases (e.g., Hochberg and Shebilske predict a larger effect of contextual constraints than of lexical properties), we will discuss them separately. We were able to locate seven studies in which skipping rates were compared for easy and difficult parafoveal words (see Table 1). In five of these studies, the variable that was manipulated, was the frequency of the word. For instance, Rayner and Fischer (1996) compared sentences like "He invested his money to build a store and was soon bankrupt." with sentences like "He invested his money to build a wharf and was soon bankrupt". In the other two studies, the manipulation was whether the word was visible in parafoveal vision or not up to the moment the eyes crossed the blank space in front of the word. As can be seen, in all studies and for all word lengths, skipping rate was higher or equal in the easy condition than in the difficult condition. This establishes beyond doubt that lexical variables do influence the probability of word skipping. On the other hand, it should be noted that the overall difference in skipping rate is rather small (4%) and tends to be slightly larger for short words than for long words. Eight other studies looked at the effects of contextual constraints on word skipping (see Table 2). Contextual constraints are usually measured by examining how many participants fill in a particular word in a cloze task. For instance, given the sentence "The woman took the warm cake out of the ", participants are much more likely to fill in "oven" (93%) than "pantry" (3%) (Schwanenflugel, 1986). Again, in all studies we found, the difference was in the expected direction: predicted words were skipped more often than unpredicted words. The mean difference amounted to 9%. The largest effect (of 23%) was reported by Vonk (1984) who (in Dutch) compared sentences like "Mary was envious of Helen because she never looked so good", where the pronoun had no disambiguating value, with sentences like "Mary was envious of Albert because she never looked so good", where the pronoun did disambiguate (in an unexpected continuation of the sentence). Brysbaert and Vitu (1995) used the same materials, but compared sentences with an expected continuation "Mary was envious of Marc because he

Word skipping

133

Table 1 Skipping rate as a function of word characteristics Study

Manipulation

Word length (letters)

^easy

^diff

Diff.

Blanchard et al. ( 1 989) (exp 1 + 2)

parafov. preview

1-3

0.56

0.43

0.13

4-5

0.20

0.10

0.10

6-10

0.06

0.04

0.02

Pollatsek et al. (1992)

parafov. preview

3-8

0.12

0.10

0.02

Henderson and Ferreira (1993)

word frequency

4-7

0.18

0.18

0.01

Inhoff and Topolski ( 1 994) (exp 2 + 3)

word frequency

4-7

0.27

0.18

0.09

Rayner et al. (1996)

word frequency

5

0.20

0.14

0.06

6

0.19

0.16

0.03

7

0.13

0.12

0.01

8

0.09

0.08

0.01

9

0.06

0.06

0.00

10

0.08

0.07

0.01

Rayner and Raney (1996)

word frequency

6-9

0.17

0.11

0.06

Rayner and Fischer ( 1 996)

word frequency

5

0.18

0.08

0.10

6

0.12

0.05

0.07

7

0.05

0.01

0.04

8

0.02

0.02

0.00

9

0.03

0.02

0.01 0.04

always looked so good" and sentences with an unexpected continuation "Mary was envious of Marc because she never looked so good". They found a difference in skipping rate of 9%. Rayner and Well (1996) pointed out that the difference in contextual constraints between predicted and unpredicted words has to be rather large in order to obtain an effect on skipping rate. For example, in their first experiment, Ehrlich and Rayner (1981) had a difference between 93% and 15% continuations in the cloze task and

134

M. Brysbaert & F. Vitu

Table 2 Skipping rate as a function of contextual predictability Study

Manipulation

Ehrlich and Rayner (1981) context, constr. (exp 1 + 2)

Word length S^ (letters)

5difr

Diff.

5

0.41

0.35

0.06

Balotaetal. (1985)

context, constr.

4-8.0.11

0.02

0.09

Vonk (1984)

pronoun pred.

3

0.40

0.17

0.23

Schustack et al. (1987)

context, constr.

3-8

0.28

0.16

0.12

Hyona(1993)

context, constr.

7-10

0.04

0.00

0.04

4-7

0.13

0.07

0.06

Inhoff and Topolski (1994) word consistency (exp 1) Brysbaert and Vitu (1995)

pronoun pred.

3

0.49

0.40

0.09

Rayner and Well (1996)

context, constr.

4-8

0.22

0.10

0.12 0.09

reported skipping rates of 49% vs. 38%. In their second experiment, however, continuations were only 60% and 0%, and no difference in skipping rate was found (32% in both cases). Similarly, Hyona (1993) compared conditions of 65% and 32% continuation and found virtually no effect on word skipping (4% vs. 0%). To test the effect of context constraints directly, Rayner and Well (1996) compared three conditions: one in which the target word had been given by 86% of the raters in a cloze task, one in which the target word had been given by 41 % of the raters, and one in which the target word had been given by only 4% of the raters. Skipping rates were respectively 22%, 12%, and 10%; that is, virtually no difference was found between the 41% and the 4% condition. The relative importance of visual and linguistic variables Thus far, we have shown evidence for oculomotor, visual, and linguistic influences on word skipping. However, it should be noted that much of this evidence is not fully conclusive. For instance, the data of the landing position distributions (Fig. 1) have been obtained by aggregating over different word frequencies (and other lexical variables) and different text fragments before the eyes reached the launch site. Similarly, the effects of word length and launch site (Fig. 2) are usually

Word skipping

135

reported without reference to the frequency of the words in the different cells. An exception to the latter can be found in Rayner et al. (1996) (Fig. 2) who looked at the effects of launch site and word frequency on the skipping rate for five- and six-letter words. They reported independent effects of word length and launch site, together with a frequency effect at close launch sites (up to three letter positions in front of the target word). However, even in this study, the text fragments on which the data were based may have been quite different in the various conditions (e.g., it is not unlikely to assume that certain sequences of word lengths resulted in far launch sites and others in near launch sites, or that high-frequency and low-frequency words appeared in different sequences of word lengths). This unsatisfactory state of affairs has motivated theorists of eye movement control in reading to stay on their opposite positions. Much of the debate has centred around the presence or absence of statistically significant differences in skipping rate between easy and difficult conditions, and on imposing further constraints on what are good stimulus materials and empirical data (e.g., according to O'Regan, one should exclude cases of long fixation durations because in these exceptional instances the ongoing processing may interfere with the autonomous scanning strategy; unfortunately, O'Regan has never specified what are "long" fixation durations). However, there may be a way out of the debate by capitalizing more on the data of Tables 1 and 2. In these tables, we have skipping rates for words that were presented in identical sentences and which were constructed in such a way as to maximize the difference in processing difficulty (e.g., by having a large difference in frequency or in contextual constraint). So for these stimuli, word length should not have an effect on skipping rate if eye movements are entirely under the control of the ongoing processing. That is, for these studies we must not find that difficult three-letter words are skipped more often than easy five-letter words. The relative contribution of word length and processing difficulty can be determined by running a multiple regression analysis with both variables as predictors and skipping rate as dependent variable. This is the more easy because the predictors in Tables 1 and 2 are orthogonal. Although we could have taken the raw values of word length, we opted for an exponential transformation, which has the advantages that (i) larger weights are given to differences between short words than between large words, and (ii) that the function asymptotes to 1 (for word length zero) and to 0 (for very long words). Ideally, processing ease of the words should also be part of the exponential function (in order to retain the advantage of the asymptotic values), but for reasons of clarity we introduced this factor as a separate variable in the equation. So, skipping rate is predicted as a function of exp(word length) and easiness of the word (operationalized as -0.5 for the difficult condition and +0.5 for the easy condition). There are 36 data cells in Table 1, and 16 in Table 2.

136

M. Brysbaert & F. Vitu

Fig. 3. Skipping rate as a function of word length and word difficulty (circle = easy condition; square = difficult condition). Empirical data from Table 1. Fitted curve based on nonlinear regression with exp(word length) and contextual constraint as predictors.

Figure 3 shows the outcome for the data of Table 1 (processing load due to word characteristics). Parameters were estimated by means of nonlinear regression analysis. The resulting equation is: %skipping = 100 (e-0.36 word length ) + 4.3 processing load Both regression weights were significant (word length: f(34) = -29.4, p < 0.01; processing load: r(34) = 2.73, p < 0.01) and together accounted for 83% of variance, of which 79% was due to word length and 4% to processing load. Figure 4 shows the results for Table 2 (processing load due to context constraints). A problem here was that many studies did not give enough details to calculate the exact mean word length. In these cases, the average of the lengths reported was used as predictor (a better but more complex approach would have been to take the geometric mean). The resulting equation is: %skipping = 100 (e-0.34 word length ) + 10.0 processing load The regression weight of word length was highly significant, the one of processing load reached significance only for a one-tailed test (word length: f(15) = -11.6, p < 0.01; processing load: f(15) = 1.86, p < 0.10, two-tailed). The regression accounted for 55% of variance, of which 44% was due to word length and 11% to processing load. The main conclusion from Figs. 3 and 4 is that even for studies which specifically looked at the effects of processing difficulty on skipping rate, word length was a

Word skipping

3

4

5

6

137

7

8

9

word length Fig. 4. Skipping rate as a function of word length and contextual constraint (circle = easy condition; square = difficult condition). Empirical data from Table 2. Fitted curve based on nonlinear regression with exp(word length) and contextual constraint as predictors.

more important predictor of skipping rate than processing load. That is, to predict how often a word was skipped, it is better to know how long the word was than to know whether it was visible in the parafovea, of high frequency, or highly constrained by the preceding context. Of additional interest is the finding that the regression weight of word length is virtually the same for studies in which word properties were manipulated (0.36) as for studies in which context constraints were manipulated (0.34). So, the empirical evidence strongly points to the conclusion that short words are skipped more often because they are short and not because they are easy. The only thing processing difficulty does, is slightly modulate the basic underlying phenomenon. Towards a new model of word skipping In the previous section we have seen that word skipping is primarily determined by visual factors (such as the length of the upcoming word, and its distance from the launch site; Kerr, 1992). The impact of these factors is not secondary to the probability of recognizing a word with particular language characteristics at a particular distance in the parafovea, although linguistic factors may slightly change the a priori skipping probability. What we need, therefore, is a model of word skipping that is based on visual characteristics of the text. A problem at this point is that nearly all non-linguistic models of inter-word eye behaviour in reading are characterized by a deterministic scanning strategy. In older

138

M. Brysbaert & F. Vitu

models, this was a chain of more or less equidistant fixations; in more recent models, it basically is a word-to-word motion. The fact that eye movement patterns in reading rarely exhibit the simple characteristics predicted by these models, has to be explained by the assumption of a (considerable) oculomotor error and occasional intrusions of the linguistic processor. It may be asked, however, whether the assumption of a non-probabilistic oculomotor strategy has not been the major weakness of the oculomotor theories, as it only enables fairly simplistic and coarse predictions about inter-word eye behaviour. At least, it is tempting to see the regular patterns of Figs. 1, 3, and 4 as the result of an underlying stochastic process. According to this view, most words are skipped not because of error in the landing position, but because the system did not "intend" to look at them. Target selection based on the extended optimal viewing position effect To have a stochastic, non-linguistic model of eye movement control in reading, we must be able to predict the probability of word identification as a function of word length and stimulus distance from the fixation location. This is what Brysbaert, Vitu, and Schroyens (1996) investigated. They presented words of different lengths at various positions relative to the fixation location and for different periods of time. Figure 5 depicts their main finding. It shows the probability of recognizing a

Fig. 5. Probability of word recognition as a function of presentation duration and word position relative to fixation location. Empirical data and best fitting Gaussian distributions (reprinted from Brysbaert et al., 1996).

Word skipping

139

five-letter word on different positions in foveal and parafoveal vision and for time intervals ranging from 14 to 70 ms. For each presentation duration, the probability of recognizing a word at a certain distance could be described reasonably well with a Gaussian distribution that had the mode shifted slightly to the left of the word centre. A similar pattern was found for three-and seven-letter words. Brysbaert et al. (1996) called this the Extended Optimal Viewing Position (EOVP) effect. Given that the probability of recognizing a word presented at a certain distance and for a given period of time is described reasonably well by a normal curve, this information can be used by the eye guidance system to estimate the chances of identifying the upcoming parafoveal words within the time period of an average fixation (200-220 ms), and to select the most appropriate parafoveal target word. The estimates are based on (i) the length of the word blobs and the distance of the word blobs from the fixation location, and (ii) the standard deviation of the Gaussian EOVP curve. The latter is a function of text difficulty and task demands (e.g., it will be larger for easy texts and for cursory reading). Although the estimates are rather crude, they give a relatively good idea of what the chances are to recognize an "average word" of a particular length and at a particular distance in the parafovea during an "average fixation duration". It is our contention that this educated guess, which becomes available rather soon in the fixation because it is based on coarse visual information, is used in combination with the ongoing text processing to decide whether or not the next parafoveal word can be skipped. So, a parafoveal word with an estimated probability of 90% recognition at the end of the current fixation, can most of the time be skipped without implications for the ongoing text processing. Only in a few instances, when processing is hard, is it better not to skip the word. In the long run, this will lead to something like 10% fixation and 90% skipping (so that the skipping rate equals the identification probability). If the above reasoning is correct, then a straightforward prediction follows: We should be able to describe parafoveal target word selection with the use of a simple inverted Gaussian distribution [i.e., 1 - exp(-sqr(letter position - fixation location)/ standard deviation))], as shown in Fig. 6 (note that in this first approach, we disregard the shift of the EOVP curve to the left of foveal word centre). In Fig. 6, we depict the situation for a launch site on the middle of a three-letter word in front of a five-letter parafoveal word. There is a probability of pi that the foveal three-letter word will be refixated, a probability of p2 - pi that the parafoveal word will be fixated, and a probability of 1 - p2 that the parafoveal word will be skipped. Note that the function has only one parameter, the magnitude of the standard deviation. To test the plausibility of our model, we examined the first-pass forward saccades of Brysbaert and Mitchell (1996, Experiment 3; see above) and plotted the landing positions as a function of target word length and launch site. So, for each forward saccade we looked at the length of the words to the right of the launch position (up to 15 letter positions) and noted where the eyes landed relative to those words. The

140

M. Brysbaen & F. Vitu

Fig. 6. Graphical display of how an inverted EOVP curve can be used as a cumulative landing position distribution going from the launch site to several words in the right visual half-field.

Fig. 7. Fit of the mathematical model to empirical data from Brysbaert and Mitchell (1996, Experiment 3) as a function of word length and launch site. Left panel: 3-letter target words; middle panel: 5-letter target words; right panel: 7-letter target words. Left down corner of each panel: probability of landing sites in front of the word; right upper corner of each panel: probability of landing sites behind the word; middle section: probability of landing sites on the word. Lines: empirical values; closed circles: predictions by the model.

Word skipping

141

data are plotted in Fig. 7 (horizontal lines) for target words of three, five, and seven letters. So, when a saccade was launched 15 letter positions in front of a three-letter target word, there was 88% chance that the eyes would land in front of the word, 7% chance that they would land on the word, and 5% chance that they would land behind the word. For a five-letter target word, these probabilities were respectively 91%, 7%, and 2%. For seven-letter words, they were 88%, 11%, and 1% (see Fig. 7, first column of each panel). We then searched for the best fitting EOVP curve; which happened to have a standard deviation of 10 letter positions (indicated by the black circles in Fig. 7). As can be seen, the empirical data were described remarkably well by a mathematical equation that has but one unknown parameter. Fleshing out the model In the previous paragraph, we have described the general outline of a stochastic nonlinguistic eye guidance strategy that can account to a high degree for the observed word skipping rates as a function of word length and launch site. In the present paragraph, we will try to outline some of the assumptions underlying the model. First, it is important to realize that the probabilities depicted in Fig. 7 do not agree with the distribution of landing positions. They represent the probability that the landing site will be in front of the target word, on the target word, or behind the target word. It is perfectly possible (and quite likely) that the underlying landing site distribution is not normal and unimodal but multimodal with a peak near the middle of each parafoveal word. Otherwise, if the data of Figure 7 represented the landing site distribution, our findings would have been more in line with a model that postulates a constant saccade size and a normally distributed oculomotor error. This is not to say that we reject oculomotor errors as a factor contributing to word skipping (see above). As a matter of fact, simulations indicate that the fit of our model increases significantly if a normally distributed landing error with a standard deviation of two letter positions is added to the model. Second, it cannot be denied that for the model to work, it is not enough to calculate the identification probabilities of parafoveal words. It is also necessary to "decide" which word to pick out on the basis of these probabilities. At present, it is not clear which processes are involved in the decision and where the decision is made (leaving aside the possibility of a "homonculus problem"). On the one hand, we could think of a version that closely resembles a random control model. According to this version, at the beginning of a fixation, a random number is generated (between 0 and 1) and the word selected that corresponds to this probability (e.g., if the number falls between p\ and pi in Fig. 6, then the parafoveal five-letter word would be selected as target). On the other hand, one could think of a

142

M. Brysbaert & F. Vitu

version in which the selection depends on the amount of resources left open by the ongoing text processing. As suggested above, we tend to prefer the latter version. This implies that the decision is more likely to be taken by a system related to language processing than by an autonomous oculomotor control system. Given that we accept a decision based on language processing, it may be asked what the difference still is with prevailing linguistic control theories. According to these theories, a word is skipped because it was recognized during the previous fixation. According to our view, a word is skipped because the language system estimates chances high enough that it will be identified by the end of the current fixation or, at least, that bypassing the word will not hinder text understanding (e.g., because the text fragment is easy). So, the difference is essentially one between an evaluation of what has already been achieved and an educated guess of what will be achieved in the near future. We believe the latter forms the basis of eye guidance in reading, because an educated guess can be realized earlier in the time course of a fixation. In our view, attentional theories of eye movement control in reading have seriously underestimated the time it takes to identify a parafoveal word, although Rayner himself pointed to the fact that word identification times increase 90 ms per degree of eccentricity (Rayner and Morrison, 1981; see also Schiepers, 1980). This value should be added to the 60 ms stimulus transfer time from the eyes to the brain and the 100 ms needed for saccade programming (McConkie, 1983), so that it is virtually impossible for a parafoveal word beginning at an eccentricity of three letter positions to determine the size of the upcoming saccade, unless the fixation lasts considerably longer than 60 + 100 + 90 = 250 ms. What is possible, however, is that the system responsible for the programming of the next saccade at a certain point in time (i.e., before the programming of the saccade starts) estimates the likelihood that the parafoveal word will be recognized within the next 100 ms. In principle, the system could use whatever information is available at that moment (Vitu, 1991b), but most of the time this will not be very detailed information about the word, but rather some crude measures such as word length and distance in the parafovea (see also below). At this point, it is necessary to mention explicitly a third assumption of our model, namely that the parafoveal word is processed in parallel with the foveal word, but with a delay (see Schroyens et al., in press, for a more detailed description of our account and some empirical evidence for it). There is little point in estimating the probability that a parafoveal word will be recognized at the end of the current fixation, if processing only starts when the foveal word has been identified (as assumed in current attention theories). Furthermore, if the parafoveal word is processed in parallel with the foveal word, the educated guess can gradually be updated as a function of the duration of the fixation. As shown above, the basic parameters that determine word skipping are word length and launch site. However, the effect of these parameters is to some extent modulated by the difficulty of the

Word skipping

143

parafoveal word. One way of conceiving this, is to state that on a few occasions, the (easy) parafoveal word is recognized soon enough before the saccade is launched, so that the programmed saccade to this word can be cancelled and replaced by a saccade to the next word. This seems to be O'Regan's (1990) reasoning, when he claims that linguistic information can guide eye movements if fixation durations are exceptionally long. Another view is that the first educated target guess, based on visual factors, is constantly updated by incoming linguistic information about the parafoveal word until saccade programming begins. The advantage of the latter view is that the extra information need not be lexical identification, but can be any sub-lexical information (such as stimulus familiarity, fit of the word blob within the contextual expectations, frequency of the first bigram/trigram, and so on). The main purpose of this extra information is to improve the estimated probability that the parafoveal word will be recognized at the moment of saccade onset (which takes place some 100 ms later). Some new predictions There are three main predictions to be made on the basis of our model of word skipping. The first is that differences in inter-word eye behaviour as a function of text difficulty and expertise can be described by a change in the standard deviation of the Extended Optimal Viewing Position curve. This would mean that exactly the same mechanisms are used in different reading situations and that differences in eye behaviour are gradual rather than absolute (the latter is the case, for instance, in the distinction between a word-by-word strategy and a skip-every-second-word strategy). The second prediction is that because word skipping happens on the basis of partial information, in a non-neglectable number of instances words will be skipped erroneously. A word that on the basis of its length and distance looked like a safe bet to skip, may turn out to be more difficult and crucial for the understanding of the text than expected at first sight. In those cases, the skipping is likely to be followed by a regressive eye movement. Indeed, as documented in Chapter 5, the pattern of words initially skipped and immediately afterwards regressed to, occurs very regularly in reading and is probably too frequent to be ascribed to oculomotor error alone. In addition, many of these regressions are preceded by short fixations. For instance, in our eye movement data base of Brysbaert and Mitchell (see above), 35% of the fixation durations shorter than 200 ms were followed by a regression to the previous (skipped) word. This was more than the percentage of forward movements to the next word (23%) or to the second next word (27%). On the other hand, fixations of 300 ms and more had a chance of less than 15% to be followed by a regression to the previous word, and more than 30% to be succeeded by a forward movement to the next word or to the second next word. This finding clearly challenges the predictions

144

M. Brysbaert & F. Vitu

made by the attention theory (e.g., Rayner, 1995) that short fixations should be followed by a forward movement (because they are a sign that the parafoveal word was identified before the saccade started but after the execution deadline expired). The finding also presents problems for the "first-fixation-duration" measure often used in eye movement research, as the duration of a fixation on a target word followed by a regression to the previous skipped word, is unlikely to reveal much about the processing of the target word. The data on regressive eye movements after short fixation durations bring us to the third prediction of the model, namely that language influences on word skipping may be due to a shortening of an intended saccade rather than to a lengthening. According to most current models of eye movement control in reading, the initial target for a forward saccade is the next parafoveal word. So, any linguistic influence must be limited to a replacement of this saccade by a saccade to a more distant word. However, if the target of the initial saccade can be the second or even the third word in the parafovea, as our model claims, then linguistic influences can result in a cancelling of an intended skip and its replacement by a saccade to a less distant word. This would happen, for instance, when the eye guidance system originally estimated that a particular parafoveal word could safely be skipped, but in the time course of the fixation noticed that this was unlikely to be so (e.g., because the processing rate of the parafoveal word was slower than anticipated). Cancelling an intended skip may even be a more common language processing intervention than replacing a saccade by a saccade to a more distant word, as there is some evidence that shortening a saccade is easier than lengthening it (Becker, 1991, pp. 129-130). Conclusion It is tempting to think of eye movements during reading as an activity which is completely regulated by either a dumb oculomotor strategy or by the ongoing text processing. In both cases, the pattern of fixations within a line of text seems needlessly complicated and chaotic, and requires the assumption of a large landing position error. Another approach is to see eye guidance as hypothesis generation on the basis of incomplete information. In this view, errors (and, hence, corrective movements) are an inherent part of the model and should exhibit properties related to the hypothesis generator. We have tried to show that at least word skipping can be described quite well within such a framework.

Word skipping

145

Acknowledgements The collaboration leading to this text was made possible by grants from the Fonds voor Wetenschappelijk Onderzoek and the European Union (BIOMED BMHICT94-1441) to the first author, a grant from the Fyssen Fondation to the second author, and a Tournesol grant (T/94.046) fostering collaboration between France and Flanders. We would like to thank Keith Rayner, Ralph Radach, and an anonymous reviewer for helpful comments on an earlier draft.

References Balota, D.A., Pollatsek, A. and Rayner, K. (1985). The interaction of contextual constraints and parafoveal visual information in reading. Cognitive Psychology, 17, 364-390. Becker, W. (1991). Saccades. In: R.H.S. Carpenter (Ed.), Vision and Visual Dysfunction, Vol. 8: Eye Movements. London: MacMillan, pp. 95-137) Blanchard, H.E., Pollatsek, A. and Rayner, K. (1989). The acquisition of parafoveal word information in reading. Perception and Psychophysics, 46, 85-94. Bouma, H. and De Voogd, A.H. (1974). On the control of eye saccades in reading. Vision Research, 14, 273-284. Brysbaert, M. and Mitchell, D.C. (1996). Modifier attachment in sentence parsing: Evidence from Dutch. The Quarterly Journal of Experimental Psychology, 49A, 664-695. Brysbaert, M. and Vitu, F. (1995). Word skipping: Its implications for theories of eye movements in reading. Paper presented at the Eighth European Conference on Eye Movements. Derby, UK. Brysbaert, M., Vitu, F. and Schroyens, W. (1996). The right visual field advantage and the optimal viewing position effect: On the relation between foveal and parafoveal word recognition. Neuropsychology, 10, 385-395. Bus well (1920) An experimental study of eye-voice span in reading. Supplementary Educational Monographs, 17. Coe'ffe, C. and O'Regan, J.K. (1987). Reducing the influence of nontarget stimuli on saccade accuracy: Predictability and latency effects. Vision Research, 27, 227-240. Ehrlich, S.F. and Rayner, K. (1981). Contextual effects on word perception and eye movements during reading. Journal of Verbal Learning and Verbal Behavior, 20, 641-655. Erdmann, B. and Dodge, R. (1898). Psychologische Untersuchungen uber das Lesen aufexperimenteller Grundlage. Halle. Findlay, J.M. (1982). Global visual processing for saccadic eye movements. Vision Research, 22, 1033-1045. Henderson, J.M. and Ferreira, F. (1990). Effects of foveal processing difficulty on the perceptual span in reading: Implications for attention and eye movement control. Journal of Experimental Psychology: Learning, Memory, and Cognition, 16, 417-429. Henderson, J.M. and Ferreira, F. (1993). Eye movement control during reading: Fixation measures reflect foveal but not parafoveal processing difficulty. Canadian Journal of Experimental Psychology, 47, 201-221.

146

M. Brysbaert & F. Vitu

Hochberg, J. (1975). On the control of eye saccades in reading. Vision Research, 15,620. Hochberg, J. (1976). Toward a speech-plan eye-movement model of reading. In R.A. Monty and J.W. Senders (Eds.), Eye movements and psychological processes. Hillsdale, NJ: Erlbaum. Huey, E.B. (1908). The psychology and pedagogy of reading. New York: MacMillan. Hyona, J. (1993). Effects of thematic and lexical priming on readers' eye movements. Scandinavian Journal of Psychology, 34, 293-304. Inhoff, A.W. and Topolski, R. (1994). Use of phonological codes during eye fixations in reading and in on-line and delayed naming tasks. Journal of Memory and Language, 33, 689-713. Just, M.A. and Carpenter, P.A. (1980). A theory of reading: From eye fixations to comprehension. Psychological Review, 87, 329-354. Just, M.A. and Carpenter, P.A. (1987). The psychology of reading and language comprehension. Newtown, MA: Allyn and Bacon. Kerr, P. W. (1992). Eye movement control during reading: The selection of where to send the eyes. Doctoral Dissertation, University of Illinois at Urbana-Champaign, Urbana, IL. McConkie, G.W. (1979). On the role and control of eye movements in reading. In: P.A. Kolers, M.E. Wrolstad and H. Bouma (Eds.), Processing of Visible Language (Vol. 1, pp. 37^8). New York: Plenum. McConkie, G.W. (1983). Eye movements and perception during reading. In: K. Rayner (Ed.), Eye Movements in Reading: Perceptual and Language Processed. New York: Academic Press, pp. 65-96. McConkie, G.W., Kerr, P.W., Reddix, M.D. and Zola, D. (1988). Eye movement control during reading: I. The location of initial eye fixations. Vision Research, 28,1107-1118. McConkie, G.W., Kerr, P.W. and Dyre, B.P. (1994). What are 'normal' eye movements during reading: Toward a mathematical description. In: J. Ygge and G. Lennerstrand (Eds.), Eye Movements in Reading. Oxford: Elsevier, pp. 315-327. Morrison, R.E. (1984). Manipulation of stimulus onset delay in reading: Evidence for parallel programming of saccades. Journal of Experimental Psychology: Human Perception and Performance, 10, 667-682. O'Regan, J.K. (1979). Eye guidance in reading: Evidence for the linguistic control hypothesis. Perception and Psychophysics, 25, 501-509. O'Regan, J.K. (1980). The control of saccade size and fixation duration in reading: The limits of linguistic control. Perception and Psychophysics, 28, 112-117. O'Regan, J.K. (1990). Eye movements and reading. In: E. Kowler (Ed.), Eye Movements and Their Role in Visual and Cognitive Processes. Amsterdam: Elsevier, pp. 395^453. Pollatsek, A., Lesch, M., Morris, R.K. and Rayner, K. (1992). Phonological codes are used in the integration of information across saccades in word identification and reading. Journal of Experimental Psychology: Human Perception and Performance, 18,148-162. Pollatsek, A. and Rayner, K. (1990). Eye movements and lexical access in reading. In: D.A. Balota, G.B. Flores d'Arcais and K. Rayner (Eds.), Comprehension Processes in Reading. Hillsdale, NJ: Erlbaum, pp. 143-163. Radach, R. and Kempe, V. (1993). An individual analysis of initial fixation positions in reading. In: G. d'Ydewalle and J. Van Rensbergen (Eds.), Perception and Cognition: Advances in Eye Research. Amsterdam: North-Holland, pp. 213-226.

Word skipping

147

Rayner, K. (1978). Eye movements in reading and information processing. Psychological Bulletin, 85, 618-660. Rayner, K. (1995). Eye movements and cognitive processes in reading, visual search, and scene perception. In: J.M. Findlay, R.W. Kentridge and R. Walker (Eds.), Eye Movement Research: Mechanisms, Processes, and Applications. Amsterdam: North-Holland, pp. 3-22. Rayner, K. and Fischer, M.H. (1996). Mindless reading revisited: Eye movements during reading and scanning are different. Perception and Psychophysics, 58, 734-747. Rayner, K. and Morrison, R.E. (1981). Eye movements and identifying words in parafoveal vision. Bulletin of the Psychonomic Society, 17, 135-138. Rayner, K. and Raney, G.E. (1996). Eye movement control in reading and visual search: Effects of word frequency. Psychonomic Bulletin and Review, 3, 245-248. Rayner, K., Sereno, S.C. and Raney, G.E. (1996). Eye movement control in reading: A comparison of two types of models. Journal of Experimental Psychology: Human Perception and Performance, 22, 1188-1200. Rayner, K. and Well, A.D. (1996). Effects of contextual constraint on eye movements in reading: A further examination. Psychonomic Bulletin and Review, 3, 504-509. Schiepers, C. (1980). Response latency and accuracy in visual word recognition. Perception and Psychophysics, 27, 71-81. Schroyens, W., Vitu, F., Brysbaert, M. and d'Ydewalle, G. (in press). Visual attention and eye-movement control during reading: The case of parafoveal processing. The Quarterly Journal of Experimental Psychology. Schustack, M., Ehrlich, S. F. and Rayner, K. (1987). The complexity of contextual facilitation in reading: Local and global influences. Journal of Memory and Language, 26, 322-340. Schwanenflugel, P.J. (1986). Completion norms for final words of sentences using a multiple production measure. Behavior Research Methods, Instruments and Computers, 18, 363-371. Shebilske, W. (1975) Reading eye movements from an information-processing point of view. In: D. Massaro (Ed.), Understanding Language. New York: Academic Press, pp. 291-311. Vitu, F. (199 la). The existence of a center of gravity effect during reading. Vision Research, 31, 1289-1313. Vitu, F. (1991b). The influence of parafoveal preprocessing and linguistic context on the optimal landing position effect. Perception and Psychophysics, 50, 58-75. Vitu, F., O'Regan, J.K., Inhoff, A.W. and Topolski, R. (1995). Mindless reading: Eye movement characteristics are similar in scanning letter strings and reading texts. Perception and Psychophysics, 57, 352-364. Vonk, W. (1984). Eye movements during comprehension of pronouns. In: A.G. Gale and F. Johnson (Eds.), Theoretical and Applied Aspects of Eye Movement Research. Amsterdam: North-Holland.

This page intentionally left blank

149

CHAPTER 7

The Influence of Parafoveal Words on Foveal Inspection Time: Evidence for a Processing Trade-Off Alan Kennedy University of Dundee

Abstract The influence of properties of a 'target' word, presented in the parafovea, on the time to process one of two possible foveal 'prompt' words was examined using measures of gaze and fixation duration. Gaze duration on the prompt showed a sensitivity to the length and word frequency of the to-be-fixated target. There were no effects of the token 'familiarity' of the target's initial letters. More detailed analyses revealed that the obtained frequency effect arose from the inclusion of prompt refixations in the data set. A measure restricted to those cases where the prompt was fixated only once showed a sensitivity to the length of the target and its 'informativeness', defined by the number of words sharing its initial letters, but not to its word frequency or its familiarity. Parafoveal influences were paradoxically inverted, shorter foveal inspection time being associated with long targets and with targets sharing initial letters with many other words. Such parafoveal-on-foveal effects are incompatible with models of reading in which attention is allocated sequentially to successive words. The data are, however, consistent with the proposition that foveal and parafoveal processing occurs in parallel, implicating some form of local process-monitoring taking place over a region larger than the word. It is suggested that saccades can be triggered by a mechanism sensitive to the rate at which sub-lexical parafoveal information can be acquired. Information secured during parafoveal preview is traded-off when a word is eventually inspected foveally. Two responses are given to the objection that the results might be specific to the artificial laboratory task used. First, the same pattern of effects is obtained in an analysis of the filler materials. Second, the effects of parafoveal length and informativeness found under laboratory conditions are also present in a large corpus of eye movement measures obtained under conditions of normal reading. Eye Guidance in Reading and Scene Perception/G. Underwood (Editor) © 1998 Elsevier Science Ltd. All rights reserved

150

A. Kennedy

Introduction In two recent papers, Henderson and Ferreira (1990; 1993) examined the way foveal and parafoveal information combine and interact to determine the timing of eye movements on successive words in text. The topic is of central importance to theories of eye movement control and has been raised in one way or another since the very earliest days of eye movement research. It takes the form of a deceptively simple question: when fixating a particular word, what information is available from words not yet fixated? In their first (1990) set of experiments, Henderson and Ferreira demonstrated that processing load at the current point of fixation (i.e., 'foveal difficulty') changes the effective perceptual span. The experimental procedure looked at the size of the parafoveal preview effect using the boundary technique introduced by Rayner (1975). The difficulty of the foveal word was manipulated in different experiments, either by changing word frequency or local syntactic load (adding or removing a complementiser in a garden path sentence). Both word frequency and syntactic difficulty manipulations produced similar results: the time taken to process a word in the parafovea, when it was eventually fixated, varied as a function of the difficulty of the foveal word. The outcome can be illustrated using the data from the word frequency manipulation (although it was very similar for the manipulation of syntactic difficulty). When the foveal word was of low frequency, the preview advantage was 3 ms. In contrast, with a high frequency foveal word there was a 13 ms preview advantage. The theoretical background for the study of this preview effect is provided by Morrison's (1984) parallel programming model of eye movement control in reading, which is possibly the most clearly articulated example of a theory emphasising on-line processes exerting linguistic control over eye movements in real time. For Morrison, the when and where of eye movement control are influenced from moment to moment by linguistic properties of the word, or words, being processed. Initially, attention is focused on the foveal word n. When a certain criterion level of processing has been reached, attention shifts discretely towards word n + 1 and at the same time, while word n is still fixated, a saccade is programmed towards n + 1. Thus, for a period of time, the reader is fixating word n and processing word n + 1, the time employed by this processing being equal to the saccadic programming time. The reference to parallel programming in the model points to the possibility that a second saccade (to word n + 2) may be programmed before the saccade to n + 1 is initiated. Indeed, if parafoveal processing is particularly efficient, there may be time for two (or even more) attentional shifts, with the consequence that occasionally words may be skipped. The Morrison model is admirably explicit, but it faces a number of challenges. A particular problem arises from the notion of a 'criterion level' of information processing of the foveal word which triggers the attentional shift. If a constant

Influence ofparafoveal words on foveal inspection time

151

criterion is assumed, the model predicts a constant parafoveal preview effect. The amount of information collected from a word in the parafovea should not vary as a function of foveal processing load, for the simple reason that the attentional shift does not occur until the foveal criterion is exceeded. In this sense, the model is to be distinguished from one where parallel processing of foveal and parafoveal words takes place (McClelland and Mozer, 1986; Schroyens et al., 1996). The demonstration by Henderson and Ferreira (1990) that fixation on a difficult word reduces the amount of parafoveal information extracted poses a problem for the notion of a fixed perceptual (or attentional) span. Their suggested solution is to add to the model a maximum possible fixation time, or deadline, after which a saccade becomes inevitable regardless of the current level of processing. The model handles their data under this new assumption because of the way the saccadic deadline interacts with the parallel programming mechanism. If foveal processing is particularly difficult, the deadline will expire before attention has shifted to word n + 1. In this case, a saccade will be prepared, but with a target location within the current word (since that is where attention remains directed). It is worth noting, in the context of later discussion, that the model does not deal with the question of fixation location, but there is a strong presumption that these within-word refixations will invariably be in a left-to-right direction. The deadline assumption predicts a foveal-on-parafoveal interaction because the time spent shifting attention within the foveal word is effectively subtracted from the time attention can be allocated to word n + 1. Thus, if the saccadic deadline is exceeded, any pre-view advantage which might have accrued will be wiped out. It should be noted, however, that effects of foveal load on parafoveal preview are also consistent with a parallel processing model. If visual attention is to be thought of as a 'limited resource', allocating more to a foveal word will mean less available for processing in the parafovea1. Henderson and Ferreira (1993) present data which comment directly on the distinction between models in which attention is allocated in parallel and ones, like their modification of Morrison's, where saccade programming takes place in parallel, but attention is allocated in a strictly sequential fashion. They examined the way in which the difficulty of an as-yet unfixated word in the parafovea influenced processing time on a foveal word. If attention is allocated sequentially, properties of a parafoveal word will not influence foveal processing time, because an attentional shift only occurs when foveal processing is complete (or when the processing deadline expires). On the other hand, if processing occurs in parallel, parafovealon-foveal effects are predicted. Measuring eye movements and manipulating the 1 In fact, a contrary prediction can also be generated from this perspective: If the only relevant variable is the total time available, the more time spent inspecting a foveal word, the more time which can be allocated to parafoveal processing. See Schroyens et al. (1996) for a discussion.

152

A. Kennedy

frequency of successive words embedded in short sentences, they gauged the processing consequences for the currently-fixated word of varying the difficulty of words in the parafovea. The results showed no influence at all of parafoveal difficulty on foveal processing time (see also Carpenter and Just, 1983, for a similar finding). In fact, the time spent inspecting a foveal target was actually (nonsignificantly) shorter (244 vs. 252 ms) when the word in the parafovea was difficult (i.e., of low frequency compared to high). It is perhaps important to make clear that parafoveal words were processed during foveal inspection (that is, there was a pre-view advantage): the point is, parafoveal information was only extracted after the processing of the foveal word had been completed. Obviously, this outcome offers no support for the notion of parallel allocation of attention, but is perfectly compatible with the (modified) Morrison model. It is also compatible with the E-Z Reader Model discussed by Rayner, Reichle and Pollatsek (Chapter 11). This differs in significant ways from the original Morrison conception, in particular by postulating a semi-autonomous 'Motor Control' system controlled, not by attentional shifts, but by the completion of an early (pre-lexical) familiarity check on the inspected word. Nonetheless, despite these differences, the E-Z Reader Model postulates a process of sequential attention allocation, in this case, attention being shifted from word n to n + 1 when lexical access on n is achieved. Parafovealon-foveal effects are not predicted. The motivation for the work reported here rests on the claim that it is premature to attempt to resolve the nature of parafoveal-on-foveal interactions using studies of normal reading. A necessary first step is to examine the effects of a parafoveal target on foveal eye movement patterns and timing using experimental procedures with tighter control than is possible in reading itself. I will consider two general arguments in support of this claim and then turn to more specific issues. First, the dynamics of eye movement control in reading are highly complex and present a major source of variability in measured processing time. The analysis of eye movement measures derived from the unconstrained inspection of text (as distinct from fixations on isolated words) can be particularly problematic when trying to interpret experimentally-induced variation in fixation duration or saccade extent (Trueswell, Tanenhaus and Kello, 1994). In reading, each word is inspected by an initial fixation at a particular position resulting from an 'entry saccade' of a given size, launched from a particular location in another word. As McConkie et al. have convincingly shown (1988, 1989), there are a number of complex interactions between launch site, initial landing site and saccade size. In particular, within-word refixations act to frustrate the acquisition of 'clean' measures of preview effects. As a second general observation it should be noted that some of the more robust and reproducible eye movement effects found with studies of individual words viewed in isolation (e.g., the effects of word frequency on fixation duration or the effect of initial landing position on gaze or refixation rate), become diluted or statistically

Influence of parafoveal words onfoveal inspection

time

153

much less reliable when examined in the context of normal reading. Of course, it could be argued from this that experimental work on eye movements in reading should necessarily focus on that task. But it is equally reasonable to ask whether processing words in text actually engages different cognitive processes and if so, why? What precisely is it which dilutes laboratory-based effects? For example, why should relatively robust effects in a reading task vanish when the same words are scanned in a visual search task? These questions have become the source of a lively debate (Vitu et al., 1995; Rayner and Fischer, 1996), focusing on the contention that any laboratory procedure which does not demand lexical processing may punish the experimenter by delivering data of little relevance to the study of reading itself. A conservative stance on this question, however, must surely hold that the vast literature on lexical processing, almost wholly based on studies of relatively short words viewed in isolation at a fixed location, has provided fundamental insights into cognitive processes relevant to reading. These cannot easily be ignored. The present paper attempts to steer a middle course in the debate. It is obviously crucial to demonstrate how the dynamics of reading in practice modulate cognitive processes known to operate on isolated words, but a first step is to secure the necessary baseline data from a task less complex than reading itself, albeit demanding lexical processing. The more specific motivation for the present experiment relates to certain problematic aspects to the data provided by Henderson and Ferreira (1990,1993). In the first place, the refixation rate on the foveal word in the 1993 studies was extremely low (effectively zero) and it was not measured in the 1990 experiments. A refixation rate of 10-16% might have been expected for the relatively short words they employed (Radach, 1996), and it is rather unclear why the obtained rate was so low. In any case, the choice of materials appears to have been unfortunate in the context of a model where 'cancelled refixations' play such an important role (it will be recalled that whether an inter-word or intra-word saccade is triggered depends on a race between foveal processing load and a saccadic deadline). Even the 'difficult' items were, in absolute terms, apparently not very difficult to process. A further problem2 relates to the variation in difficulty itself. In both the 1990 and 1993 studies this involved a manipulation of word frequency. Word frequency is a property of a word as a whole and there are numerous studies suggesting that sub-lexical encoding processes, in particular relating to word-initial letters, influence the control of eye movements (Lima and Inhoff, 1985; Pynte, Kennedy and Murray, 1991; Rayner, 1978; Rayner, McConkie and Zola, 1980). Thus, for the manipulation of difficulty chosen by Henderson and Ferreira to operate, it would have been necessary for parafoveal words to be identified. The circumstances under This may be seen as more of a general than a specific problem, since the precise manipulation of word properties in continuous text is virtually impossible.

154

A. Kennedy

which this might generally occur are little understood and rather controversial (see Rayner and Sereno, 1994, for a review). There is a real possibility, therefore, that the parafoveal 'difficulty' manipulation was simply ineffective, in which case it would be premature, on the basis of the 1993 null effect, to conclude that parafovealon-foveal effects do not occur. The experiment I shall be discussing was designed to examine parafovealon-foveal effects using a procedure with the level of control typical of single-word presentation methods, but allowing reading dynamics to influence the outcome (see Vitu, 1991, for a similar approach). The study made use of the following procedure. Following a fixation point, participants were confronted with a sequence of three words, each separated by a single space. The first word in the sequence, referred to as the 'prompt', was always either the word 'looks' or the word 'means'. The prior fixation point determined that the initial fixation on this was invariably on its third letter. The prompt determined the operation to be carried out on the other two displayed words, the first of which will be referred to here as the 'target'. When the prompt was the word 'looks', a yes-no decision was called for on the basis of the physical identity of the two adjacent target words; when the prompt was the word 'means', participants made a yes-no decision on the basis of the judged synonymy of the two targets. The decision demanded a sequence of fixations and saccades across the three words but, equally, maintained close control over initial fixation position. Method Participants These were 20 student volunteers who were paid £4.00 to take part. Some had previous experience of psycholinguistic eye movement experiments, but none had participated in experiments involving the present procedure or materials. All had normal (unconnected) vision and were naive as to the purposes of the experiment. Apparatus Eye movements were recorded from the right eye using a Dr Bouis Oculometer. This device presents an infra-red image of the eye to a dense array of photodetectors which computes the location of the pupil centre. Horizontal eye position was sampled every 5 ms and recorded for later analysis off-line. Viewing of the stimulus material was binocular. The Oculometer provided a near-linear output over the visual angle subtended by the displayed words (about 8° at most). The materials were presented using a high-resolution (8x16) monopitch font in negative polarity (i.e., white characters on a black background) on a Manitron

Influence of parafoveal words onfoveal inspection time

155

display driven at a refresh rate of 100 Hz. At the viewing distance of 525 mm, 3 characters subtended approximately 1° of visual angle. Responses were made using light-weight micro-switches which minimised head and shoulder movements. Materials and design The experimental materials comprised a set of 96 five-letter and 96 nine-letter words drawn from the Ku?era and Francis (1967) norms. For each word length, there were eight sets of twelve words, allocated randomly as the first target word to 'match' and 'mismatch' conditions in the 'looks' task. The second item in mismatch pairs was the same length and in the majority of cases had the same initial three letters. The eight sets of experimental items were selected to meet a criterion defining high and low values of word frequency; initial trigram frequency or 'informativeness' (that is, the number of words in the norms having that set of initial letters, regardless of length); and 'familiarity', or the token frequency of the initial three letters. Table 1 illustrates the composition of the experimental materials: for example, the word 'dense' in English, although relatively rare, shares its initial trigram with quite a large number of other words and is consequently relatively 'uninformative'. However, since none of these words is particularly frequent in the language, the trigram has low overall 'familiarity' (token frequency). By way of contrast, while the word 'arena' is rare, and few words share its initial letters, the trigram itself is very familiar. There were also 192 words allocated to the filler task (the 'means' decision), comprising four sets of 48 four-, five, eight- and nine-letter words. Target words demanding a 'yes' decision in this task were reasonably close synonyms. The total set of 384 word triples3 was presented in a different randomised sequence for each participant. It should be emphasised that the critical experimental words were all allocated to the 'looks' task and analysis (at least, initially) was restricted to the decisions made in that task. All the variables were manipulated within-Subjects, the design thus having factors of Decision Type (yes-no), Word Length, Word Frequency, Trigram Frequency and Trigram Token Frequency. Procedure On arrival, participants were given a sheet of instructions for the experiment asking them to observe a fixed routine on each trial. This began with fixation of the marker. After 500 ms the fixation marker was replaced by the display of the three experimental words, each separated by a single space. Participants were asked to look at 3 The complete set of materials is available from the author.

156

A. Kennedy

Table 1 Examples of the materials used. These were words of High and Low Word Frequency (WF), Initial Trigram Frequency (TF) and Initial Trigram Token Frequency (TK). The Table shows, for each case, Short (five-letter) and Long (nine-letter) words. The three numbers associated with each exemplar provide the cell means for word frequency (per million), 'informativeness' (i.e type initial trigram frequency indexed by the number of words in the lexicon sharing an initial trigram), and 'familiarity' (i.e., token initial trigram frequency indexed by the summed word frequency for the set of words sharing a particular initial trigram). HIWF

fflWF

fflTF

HITF

fflWF LOTF

HITK

LOTK doubt

HITK after

state 465 155 3209

103 51 399

766 13 1588

HIWF

LOWF

LOTF

HITF

LOWF fflTF

LOTK

HITK shout

LOTK dense

today 170 9 265

9 180

10 59

5209

290

LOWF

LOWF

LOTF

LOTF LOTK acute

HITK arena 5 14 1899

8 6 27

concerned executive building 146 75 105 310 40 19

emotional precedent diagnosis makeshift hierarchy 117 8 8 3 8 11 326 62 14 6

4041

276

406

870

4851

331

1734

40

each in turn and make a decision by pressing either a right- or left-hand button. As noted above, the nature of the decision to be made, and hence the form of processing called for on the target words, was determined by the prompt word itself. In the 'looks' task, participants had to decide whether the two target words were physically identical; in the 'means' task, whether they meant the same thing. Participants were initially given practice on the procedure and once this had been fully mastered a dental composition bite bar was prepared and the calibration technique demonstrated. This required the fixation of five points distributed evenly across the horizontal axis of the screen at the point where the experimental items would be displayed. The calibration points were restricted to 60 character positions, a region less than the full screen width, to increase resolution. After 24 practice items, participants worked through the experimental materials with a calibration after each set of twelve items. The experiment lasted about one hour, with rest periods given at regular intervals. Fixation duration and position was computed off-line. The effective resolution of the eye-tracking equipment was better than 0.5 characters.

Influence ofparafoveal words onfoveal inspection time

157

Results and discussion Gaze duration on the prompt The first analysis was on the sum of all fixations on the prompt (gaze), up to the point of the first excursion of the eyes outside its boundary, invariably to the right. It is sensible to consider all cases, regardless of errors, since analysis was restricted to behaviour on the prompt and the first of the two targets (i.e., prior to the possible commission of an error). Error-rate on the task was, in any case, extremely low (<1%). The relevant data are shown in Table 2. For the purposes of analysis, the space between the prompt and the target word was treated as the first character of the target. Analysis of variance showed two relatively strong effects of properties of the to-be-fixated word, both somewhat counter-intuitive. First, overall foveal gaze duration was reliably shorter when the parafoveal target was long, F,( 1,19) = 12.10, p = 0.003; F2(l,160) = 19.60, p < 0.001. As the table indicates, this main effect was entirely the result of low word frequency targets: the interaction between Target Length and Target Word Frequency was significant, F,(l,19) = 4.32, p = 0.05; F2(l,160) = 7.25, p = 0.008. There was no effect of Length for high frequency targets, but a highly significant effect for low frequency targets, F,(1,19) = 18.08, p = 0.007; F2(l,80) = 23.61, p< 0.001. There was a frequency effect for long targets, but in an unexpected direction: low frequency items were associated with shorter gaze durations on the prompt, 7^(1,19) = 4.48, p = 0.05; F2(l,80) = 9.41, p = 0.003. The trend in the opposite direction for short targets, was not significant, Fl(l,\9) = 2.81; F2(l,80) = 1.11. The data manifestly demonstrate that properties of the parafoveal target word can influence foveal gaze duration: indeed, they suggest that participants achieved, or were close to achieving, lexical access in the case of short words (albeit, the overall 10 ms 'frequency effect' was non-significant). It cannot be argued, however, that the obtained effects simply reflect parafoveal processing in the presence of a zero foveal load, since the more difficult parafoveal targets were associated with shorter, Table 2 Gaze duration (ms) on the prompt word as a function of the Length and Word Frequency of the parafoveal target word High target word frequency

Low target word frequency

Mean

Short target word

283

293

288

Long target word

280

266

273

158

A.Kennedy

not longer, prompt gaze duration. Moreover, the statistically significant effects all derive from the single case in which the target was a long, low frequency, word. In the other three cases, processing time on the prompt was more or less the same. A plausible, if slightly tendentious, account of this pattern of effects is that the low rate of information extraction from difficult targets triggered an early saccade, in effect prematurely aborting fixation on the prompt. There was also a significant interaction between Target Word Frequency and Target Initial Trigram Frequency, F,(l,19) = 4.57, p = 0.044; F2(l,160) = 3.84, p = 0.049. These data are shown in Table 3. The pattern of results parallels that shown in Table 2, with a single mean significantly shorter than the other three. In this case the 'difficult' targets can be identified as low frequency words with common (i.e., 'uninformative') initial trigrams. The effect of Trigram Frequency was highly significant for low frequency words, F,(l, 19) = 18.08, p = 0.007; F2(l, 160) = 23.61, p < 0.001. In this case, when the initial letters of the target were relatively uninformative an early saccade from the prompt was initiated. Equally, there was an effect of Word Frequency for targets with redundant initial trigrams, prompt gaze duration being shorter (not longer) for low frequency items, F,(l,19) = 4.50, p = 0.045; F2(l,160) = 5.00, p < 0.027. Again, the data are consistent with an early saccade triggered by difficult items in the parafovea. However, the crucial distinction here is that 'difficulty' relates to the rate of acquisition of sub-lexical information capable of disambiguating the target from other possible candidates. Neither the main effect of Target 'familiarity', nor any interactions involving that factor, approached significance. The surprising aspect of these data lies in their direction. For example, time spent on the prompt word was shorter when the target was long and presumably more difficult to process; similarly, prompt gaze duration was shorter (at least, for low word frequency targets) when the parafoveal word shared the same initial letters with many other words. Although these 'inverted' effects are intriguing, their interpretation is complicated by an obvious possible artifact. If, for some reason, the prompt word was refixated, this would act to change the launch position for the Table 3 Gaze duration (ms) on the prompt word as a function of the word frequency and frequency of the initial trigram of the parafoveal target word High target word frequency

Low target word frequency

High trigram frequency

282

272

Low trigram frequency

281

287

Influence ofparafoveal words onfoveal inspection time

159

Table 4 Probability of refixating the prompt word as a function of the length and word frequency of the parafoveal target word. The size of refixating saccades (in character positions) is shown in brackets. High target word frequency

Low target word frequency

Short target word

0.156 (0.243)

0.183 (0.287)

Long target word

0.167 (0.259)

0.132 (0.208)

saccade directed towards the target, making it more visible. The measure of gaze only permits firm conclusions to be drawn about parafoveal processing if all the saccades were launched from the same controlled position. The task was relatively simple to perform, and refixation within a five-letter prompt, presented many times, might appear unlikely, but there was, as Table 4 shows, an overall refixation rate (almost invariably involving right-going refixations) of around 16%. The probability data in Table 4 show a pattern of means similar to that found for the measure of gaze duration. That is, the condition associated with the shortest gaze (long, low frequency, targets) had the lowest prompt refixation rate. In fact, analysis of variance of the Target Length x Target Word Frequency interaction failed to achieve significance, F,(1,19) = 2.11, p = 0.16; F2 (1,160) = 4.22, p = 0.09, but sub-analyses are justified (obviously, to be treated with caution) in view of the significant effects found for the measure of gaze. In the case of low frequency targets, prompt refixation rate was significantly lower when the target was long, F,(l, 19) = 5.10,p = 0.034; F2(l, 160) = 8.65,p = 0.005. The Target Word Frequency effect was also significant for long words, 7^(1,19) = 4.34, p = 0.049; F2(l,160) = 4.80, p = 0.03. When the target was of low frequency the prompt was significantly less likely to be refixated (the apparent trend in the opposite direction for short targets was not significant). This pattern of refixations accounts for much of the variation in gaze duration, with the same rather paradoxical outcome: low frequency targets, or long targets, were less, not more, likely to trigger a refixation of the prompt. Obviously, the obtained 'frequency effect' on gaze duration (i.e., shorter prompt inspection times associated with long, infrequent, parafoveal targets) arose primarily as a result of changes in refixation rate. Table 4 also shows the average signed extent of within-prompt saccades, with left-to-right movements (comprising virtually all refixations) coded as positive. The average shift in position was about a quarter of a character, although it is important to bear in mind that this arose from a mixture of zero movement (the 85% of steady fixations on the prompt) and the occasional shift of one or two characters. The point of importance is that, in this

160

A. Kennedy

case, the eyes were much closer to the target word prior to launching the primary saccade. The Length x Target Word Frequency interaction was significant in these data, F,(l,19) = 4.16,p = 0.05; F2(l,160) = 4.32, p = 0.037, and sub-analyses parallel to those for the probability data gave broadly similar results. Average within-prompt movement was smaller when parafoveal targets were long and infrequent. There was a significant effect of Length for low frequency targets, F,(l,19) = 4.12, p - 0.05; F2(l ,160) = 6.81, p - 0.01 and a marginally significant effect of Frequency for long words, F,(l,19) = 3.12,p = 0.09; F2(l,160) = 3.40, p = 0.07. It is now possible to draw some initial conclusions concerning the apparent influence on prompt gaze duration of properties of words not yet fixated. Three aspects of the data should be emphasised: (1) both lexical and sub-lexical properties of a parafoveal target influence prompt gaze duration; (2) these effects stem largely from modulation of prompt refixation rate, targets which were easier to process being associated with higher (not lower) rates of prompt refixation; (3) the data are consistent with a monitoring mechanism which triggers an early saccade if the rate of information acquisition from the parafovea is insufficiently rapid. They are not easily reconciled with the class of model which predicts an attentional shift and/or an inter-word saccade only after the foveal word has been identified (or after the expiry of a deadline). Rather, the data suggest that properties of the parafoveal target act to modulate the rate of intra-word refixation. This poses severe problems for models which can only plausibly deal with within-word refixations by making them contingent on a foveal load high enough to exceed the saccade deadline, something which can safely be ruled out in the present task.4 The data are equally incompatible with the E-Z Reader, in which intra-word saccades result from a similar mechanism (failure to complete the familiarity check on a foveated word in time to cancel the 'labile program'). Since changes in prompt refixation rate contribute so substantially to changes in gaze duration, it may be argued that the data shown in Tables 2 and 3 do not represent a pure examination of parafoveal-on-foveal effects. Refixations radically changed the visibility of the target and, indeed, in some cases the final point of fixation on the prompt would have been within a character of the target word itself. To meet this objection, separate analyses were carried out on the approximately 85% of cases where the prompt was processed in a single fixation (i.e., a fixation location centred on the prompt throughout its inspection completely controlled Pollatsek and Rayner (1990) argue that refixations might be handled in the Morrison model without reference to a deadline. When processing text, integrative processes may contribute to local difficulty, causing a competition between a decision to stay and a decision to execute the next saccade. Interestingly from the point of view of the present paper, they also suggest that the level of sublexical excitation from neighbouring candidates may influence 'local difficulty'.

Influence of parafoveal words onfoveal inspection

time

161

launch position for the primary inter-word saccade). These 'Single Fixation Case' data provide an estimate of parafoveal influence, uncontaminated by variation in launch position. Gaze duration on the prompt (single-fixation cases) Analysis of variance of cases where the prompt was processed in a single fixation showed a highly significant main effect of Target Word Length, again with shorter durations associated with long targets (Short = 266 ms; Long = 256 ms), F,(l,19) = 9.86, p = 0.006; F2( 1,160) = 12.49, p < 0.001. The Target Length x Target Word Frequency interaction, which was significant in the analysis of gaze, failed to achieve significance, F,(l,19) = 2.47, p = 0.12; F2(l,160) = 2.74, p = 0.09. There was, however, a highly significant Decision Type ('match' vs. 'mis-match' items) x Length x Target Initial Trigram Frequency interaction, F,(l,19) = 15.67, p = 0.001; F2( 1,160) = 4.38, p - 0.04. It should be emphasised that no effects of Decision Type even approached significance in the analysis of gaze duration: in particular, analyses of the data in Tables 2 and 3, restricted to 'match' decisions only, produced virtually identical results. Nonetheless, this three-way interaction calls into question conclusions concerning the possible effects of target word frequency on prompt gaze duration. Prompt refixation appeared sensitive to global physical properties of the display as a whole (i.e., the contrast between a visually congruent 'yes' condition and a visually incongruent 'no' condition in the 'looks' task). The possibility that the obtained effects might, in fact, be peculiarly conditioned by the 'looks' task, which focuses on physical identity, will be taken up in a later section. In the present context, it is clear that one final analysis is demanded, concentrating on 'match' cases. That is, for this task, an uncontaminated estimate of parafoveal effects may be obtained from the condition in which a single saccade was made from the centre of the prompt towards a target configuration on the right involving two identical words. Gaze duration on the prompt (single-fixation, 'match' cases) Analysis of variance of prompt fixation duration showed no Length x Target Word Frequency interaction, but a highly significant interaction between Target Length and Target Initial Trigram Frequency, F,(l,19) = 15.68,p = 0.001; F2(l,80) = 6.89, p = 0.01. The form of the interaction, shown in Table 5, again arises from a single condition (associated with long, 'uninformative' or low-constraint parafoveal targets) where the mean fixation duration was significantly shorter than in the other three. The effect of Initial Trigram Frequency was significant for long targets, Fj(l,19) = 10.95, p = 0.004; F2(l,80) = 11.01, p - 0.002, but not for short. For targets with high frequency initial trigrams, long targets induced shorter prompt fixation times, F,(l,19) = 22.85, p = <0.001; F2(l,80) = 6.44, p = 0.01. This outcome may be compared with that of Lima and Inhoff (1985), showing initial fixation duration on a

162

A. Kennedy

Table 5 Fixation duration (ms) on the prompt word as a function of the length and frequency of the initial trigram of the parafoveal target word. The data are for cases where the prompt word was processed with a single fixation (85% of cases) and for 'match' decisions only. High initial trigram frequency

Low initial trigram frequency

Short target word

265

267

Long target word

244

272

low constraint word itself (i.e., in the terms adopted here, a foveal target) to be relatively short, although, of course, the present data relate to properties of an unfixated parafoveal word. For both types of short target word, and for long targets with highly informative initial letters, prompt fixation duration was roughly equal and relatively long. As noted above, this pattern of data suggests a mechanism in which the decision to make an early inter-word saccade is triggered when the rate of gain of word-initial information is particularly slow. Crucially, however, this arises from sub-lexical, rather than lexical properties of the target. When the target is long and its initial letters are shared with many other words, prompt fixation duration is short. There was no main effect of Target Token Familiarity and no significant interactions involving that manipulation. Interim summary The final analyses, on cases with identical targets in parafoveal and peripheral vision, lead to the conclusion that there is no parafoveal-on-foveal effect of Word Frequency in the absence of prompt refixation. In this respect, the outcome replicates the null effect reported by Henderson and Ferreira (1993). Nonetheless, it is not possible to conclude in favour of 'attentional autonomy' because of the strong effects of parafoveal target informativeness. Some form of process-monitoring appears to be taking place with three possible decisions: (1) to continue processing ('STAY'); (2) to refixate the prompt ('SHIFT'); or (3) to execute a saccade to the target ('GO'). The length of the target word and the overall stimulus configuration influence both the STAY and SHIFT decisions. A complex target configuration, involving a mis-match, which is presumably only visible when the target is short, triggers a within-prompt SHIFT. It is impossible to accommodate this outcome either with the expiry of a deadline based on foveal difficulty, as Henderson and Ferreira suggest, or with the completion of a familiarity check, as proposed in the E-Z Reader model. When the prompt received only a single fixation, the lack of

Influence ofparafoveal words on foveal inspection time

163

clear word frequency effects indicates that the target was not actually identified before the GO decision. Rather, the results point to the rate of acquisition of sub-lexical information (initial trigram frequency) from the parafovea as the major influence on processing time, with an early GO decision when the rate of gain of such information gain is low. It can, of course, be objected that the present task is a weak analogue of the normal reading situation the Morrison model was designed to treat and this point will be considered in some detail in the General Discussion. For the present, it may simply be noted that the task does provide controlled conditions under which the relevant hypotheses can be examined. It was necessary to process the prompt word for the task decision to be made and, in fact, prompt fixation durations clearly fall in the normal range. There are aspects to the data which are not at all congenial to models of eye movement control which appeal to attentional autonomy. The difficulties focus on foveal refixation rate and on sub-lexical parafoveal influences exerted from the target. Regarding refixations, it is evident that the addition of a deadline such as that suggested by Henderson and Ferreira (1990) cannot account for an obtained 18% overall prompt refixation rate on a five-letter word. Even if it could, within-word refixation in the present task appears to be associated with successful, rather than unsuccessful, processing of a parafoveal word. Fixation behaviour on the prompt is sensitive to the rate of acquisition of information from the target, suggesting a control system monitoring the rate at which the target 'candidate pool' reduces. Below a criterion level of information acquisition, a GO decision takes the eyes rapidly and directly to the target. Above this criterion level (at least for this laboratory task) the target is processed while the prompt is fixated (STAY). It remains an open question, however, as to what determines an intra-prompt SHIFT? Why should prompt refixation rate be lowest with the most difficult targets (e.g., long, infrequent words) when a processing advantage might be secured from being closer to the target? Similarly, what accounts for the fact that, in the case of short target words, there is a tendency for low frequency items to increase prompt refixation rate? The Morrison account (assuming this may be extended to a study involving single words) relates intra-word shifts to foveal load but, as noted above, any explanation appealing to the expiry of a deadline simply cannot be reconciled with the overall higher rate of refixation associated with easy (e.g., short) targets. An account more consistent with the obtained pattern of results simply suggests that if parafoveal information is available at all, a high priority is attached to securing it. That is, the cost of programming a saccade in the SHIFT decision can be set against the advantage of maintaining a negligible foveal load. Such a strategy would, of course, only be optimal if parafoveal processing gains could be traded against later, foveal, processing and to address this crucial question it is necessary to examine behaviour on the target word itself.

164

A. Kennedy

Fixations on the target The analysis strategy with regard to the target word is complicated by the need to distinguish single and multiple prompt fixations and match and mis-match cases. Measurements of inspection time on the first target word are also potentially subject to contamination from properties of the second word. For these reasons, and to avoid un-necessarily complicating the discussion, the analyses which will be reported relate to the full data set, with reference to other analyses only in the case of significant differences5. Analysis of eye movements on the target were conducted with the limited objective of gaining further understanding of parafoveal-on-foveal effects. Measures were made of both first fixation duration and gaze duration on the target word (the latter almost invariably resulting from no more than two fixations). In the analysis of first fixation duration, there was only one clear effect: durations were very reliably shorter for long targets (267 ms) than for short targets (286 ms), F,(l, 19)= 14.82, /? = 0.001 ;F2(1,160) = 22.29, p < 0.00. The same highly significant relationship was also found for the 'Single Fixation' case and in an analysis restricted to 'match' decisions. The direction of this difference is not, in itself, unexpected and simply provides indirect confirmation of the observation that long words are unlikely to be processed with a single fixation and the fact that the first fixation, when two or more are made, tends to relatively short (Vitu and O'Regan, 1995). It follows the result can only be sensibly interpreted along with the data on gaze duration and target refixation probability which are treated below. There was a tendency for first fixation duration to be longer on low frequency targets (HWF = 272 ms, LWF = 281 ms), but this narrowly failed to achieve significance by items, F,(l,19) = 4.86, p = 0.04; F2( 1,160) = 3.03, p = 0.08. Interestingly, the effect did not even approach significance in analyses of data derived from cases where the prompt was processed in a single fixation. Thus, the outcome is somewhat equivocal on the issue as to whether word frequency effects are found on first fixation duration with preview. Much seems to depend on whether the prompt was refixated. One explanation of this contrast with the results of Inhoff and Rayner (1986) might simply be that the experimental task is only likely to reveal modest effects of Word Frequency, given that the decision called for relates primarily to a physical match. An alternative account, consistent with the data on refixations, is that a processing trade-off is taking place, partial parafoveal identification of a target diluting any obtained foveal effect. Such a trade-off would be of great relevance in the on-going debate as to whether word frequency effects invariably occur on first fixation duration in normal reading and the issue is considered further below. There were no other significant main effects or inter5 A complete data set for all measures on both the prompt and target words is provided in Appendix I.

Influence of parafoveal words onfoveal inspection time

165

Table 6 Gaze duration (ms) on the target word as a function of its word frequency and the frequency of its initial trigram. High target word frequency

Low target word frequency

High trigram frequency

403

390

Low trigram frequency

402

426

actions in the analysis of target first fixation duration: in particular, there were no effects of either the 'informativeness' or the 'familiarity' of targets. Target gaze duration was very reliably longer (in all analyses) for long targets (442 ms) than for short targets (367 ms), F,(l,19) = 29.60, p < 0.001; F2(l,160) = 129.81, p < 0.001. There was also a robust effect of Word Frequency, (LWF = 414 ms, HWF = 396 ms), F,(l,19) = 6.63, p = 0.018; F2(l,160) = 5.85, p = 0.016 although this difference failed to achieve significance for cases when only a single fixation had been made on the prompt (possibly because of a lower probability of target refixation if the prompt itself had been refixated). The most consistent effect (apart from the ubiquitous Length effect) in all analyses was a highly significant interaction between Target Word Frequency and Target Initial Trigram Frequency, F,(l,19) = 8.34, p = 0.009; F2(l,160) = 6.96, p = 0.009, for the full data set (see Table 6). For low word frequency targets, items with rare initial trigrams were subjected to relatively prolonged inspection, F,(l,19) = 10.40, p - 0.004; F2(l,80) = 10.24, p - 0.002. A parallel sensitivity to trigram frequency was not evident in the case of high word frequency targets. The outcome represents an important qualification to the suggestion by Lima and Inhoff (1985) of an overall relationship between gaze duration and informativeness: this appears to hold true only if the inspected word itself is rare. For targets with low frequency initial trigrams the effect of Target Word Frequency was significant, F,(l,19) = 7.42, p = 0.012; F2(l,80) = 4.11, p = 0.043. The rate of within-word refixation on the target can be used to answer the question as to whether gaze differences stemmed from longer or from more fixations, or from a combination of both. The overall probability of refixating a target was relatively high (about 0.4). The vast majority were right-going, and analysis of these showed a strong effect of Target Length in all analyses: (Short = 0.27; Long = 0.53), F,(l,19) = 39.42, p < 0.001; F2(l,160) = 317.92, p < 0.001, for the full data set. The crucial relationship between Target Word Frequency and Target Initial Trigram Frequency is shown in Table 7. For high word frequency targets, refixation rate was greater in targets with 'uninformative' initial trigrams, F,(l,19) = 4.18, p = 0.05;

166

A. Kennedy

Table 7 Probability of right-going refixations on the target word as a function of it target word frequency and target initial trigram frequency High target word frequency

Low target word frequency

High trigram frequency

0.425

0.372

Low trigram frequency

0.392

0.398

F2(l,80) = 4.02, p - 0.046. A modest apparent trend in the reverse direction in the case of low word frequency targets was not significant. As these means suggest, targets with uninformative initial letters were more likely to be refixated if they were high frequency words, F,(l,19) = 6.92, p = 0.016; F2(l,80) = 6.75, p = 0.011. It should be noted, however, that the overall Target Word Frequency x Target Initial Trigram Frequency interaction narrowly failed to achieve significance in the byitems analysis, F,(l,19) = 7.37, p = 0.013; F2(l,160) = 3.38, p = 0.06, and this interaction was non-significant in analyses of the 'single fixation' and 'match' cases. But, in any case, the data provide no support for the notion that the longer gaze durations associated with any of the manipulated properties of target words (e.g., low frequency; rare initial trigrams; or long items) arose from more fixations alone. Rather, there was a higher incidence of right-going refixations when the initially-fixated region of the target was uninformative. This suggests a form of spatial control, with a higher probability of the eyes shifting towards the other, possibly more informative, end of the word. In contrast to the analysis of right-going refixations, which showed no overall frequency effect, data on the probability of left-going refixations showed a main effect of Target Word Frequency (HWF = 0.029, LWF = 0.046), F,( 1,19) = 5.59, p = 0.027; F2(l ,80) = 7.25, p - 0.008. There was also a significant main effect of Initial Trigram Frequency (HTF = 0.029; LTF = 0.046), F,( 1,19) = 4.16, p = 0.05; F2( 1,80) = 5.97, p = 0.015, (the means are coincidentally identical to those for Word Frequency). Some caution should be exercised in interpreting these findings, because the absolute number of left-going refixations was very low and a proportion may, in fact, have been mislocated fixations on the space between the prompt and target, reflecting attempts to refixate the prompt. The effects of Word Frequency simply confirm the observation that both right- and left-going refixation rate was somewhat higher for low frequency items. The effect of Initial Trigram Frequency is more interesting, since it offers further confirmation of a relationship between the direction of refixation and the spatial distribution of information in the word being inspected. As we have seen, there was an increased probability of right-going

Influence of parafoveal words onfoveal inspection time

167

refixations for targets with uninformative initial letters. The left-going refixation data show a parallel trend, with an increased probability for targets with informative initial letters. It is important to note that this directional selectivity cannot be accommodated in the Morrison Model or the subsequent modifications of it. Saccade extent Saccade size is a poor indicator of the intended landing position on the target in cases where prompt refixation acts to change launch position. Accordingly, analysis of the primary saccade-to-target was restricted to the 'single-fixation' case. The only significant influence on saccade extent was the length of the target word. Saccades directed towards 9-letter targets were 5.75 characters in extent and those towards 5-letter targets 5.08 characters, F,(l,19) = 33.94, p < 0.001; F2(l,160) = 169.64,/? < 0.001. Given the stimulus configuration, it follows that short words were initially fixated around the second letter and long words around the third. No other effects approached significance. The 'means' task I will now attempt to address the objection that the experimental procedure could have induced task-specific strategies. Since the prompt was invariably either the word 'looks' or the word 'means', it is at least possible that 'deep processing' would be discouraged. For example the two tasks can be discriminated on the basis of a single letter in the prompt. The Stroop Effect raises doubts as to whether a samedifferent word-judgement task could, in principle, be completed on the basis of 'minimal inspection' of this kind (see Murray, 1982, for an extended discussion of this issue), but in any case, the objection raises as many problems as it solves. This is because it implies that the task is formally equivalent to a situation where single (target) words are processed from an 'extremely sub-optimal' viewing position. That is, considering the prompt as simply a redundant prefix the eyes will typically be located six or seven letters to the left of the target's Optimal Viewing Position. But the whole pattern of results runs exactly counter to the expected outcome in such a situation, with several effects, and in particular, refixation probability taking an 'inverted' form. Notwithstanding these theoretical difficulties, it is possible to examine the objection by way of an analysis of behaviour on the 'means' task6. The materials employed in this task were filler words, not balanced with respect to any of the key experimental manipulations of word frequency, initial trigram 'informal!veness'

6 This analysis was suggested by Keith Rayner.

168

A. Kennedy

and 'familiarity'. However, a. post-hoc division of the materials is obviously possible and can be used to test the generality of the pattern of results obtained for the experimental materials (i.e., based on the 'looks' task). At the very least, this will rule out explanation in terms of the operation of a very task-specific strategy restricted to the case where a physical match is called for. Using median splits, the items used in the 'means' task were allocated to 'High' and 'Low' values with respect to measures of word frequency and the 'informativeness' and 'familiarity' of their initial trigram. To improve power in the analyses (since the procedure could not derive materials differing widely on the critical measures) values were then collapsed across measures of familiarity (i.e., token initial trigram frequency: it will be recalled there were no significant effects attributable to this factor in the 'looks' task). Analyses were carried out on prompt gaze duration. Since there were some important differences between 'all' and 'single fixation' cases, their analyses will be dealt with separately. In the 'all cases' analysis, the most important determinant of gaze duration was the length of the to-be-fixated target. The same 'inverted' effect evident in the experimental data (i.e, the 'looks' task) was present, with foveal gaze significantly longer in the case of short parafoveal targets (Short Target = 323 ms, Long Target = 302 ms)7, F,(l,19) = 11.81, p = 0.003; F2(l,184) = 24.09, p < 0.001. The effect of Target Word Frequency was not significant by items, but the direction of the obtained difference was 'inverted', with high frequency targets associated with longer prompt gaze duration (HWF = 315 ms; LWF = 308 ms, F,( 1,19) = 4.35, p = 0.04; F2(l,184) = 2.34, p = 0.12). It should be borne in mind that the filler materials included short, four-letter, words, increasing the probability of parafoveal identification. In the 'single fixation' analysis (i.e., when the prompt was processed in a single extended fixation), Target Length was also highly significant (Short Target = 298 ms,LongTarget = 281ms),F,(l,19) = 9.83,/? = 0.005;F2(l,184)=13.49,p<0.001, and the 'inverted' effect of Target Word Frequency was much stronger (HWF = 295 ms;LWF = 284; F,(l,19) =16.39,p = 0.001; F2(l,184) = 8.93,p = 0.004). Although the Length x Target Word Frequency interaction was not significant, F,(l,19) = 3.10, p = 0.09; F2(l,184) = 2.73, p < 0.1, it is interesting to examine the effects for the two lengths separately. There was no effect of Target Word Frequency for long targets (HWF = 283 ms; LWF = 279; F < 1), and a strong effect for short targets, (HWF = 307 ms; LWF = 289; F,(l,19) = 10.68, p = 0.004; F2(l,92) = 7.72, p = 0.007). To repeat, this outcome was reliable for those cases where the prompt was not itself refixated. It is entirely consistent with the proposed process-monitoring account, parafoveal processing success being associated with relatively long prompt inspection time and an early GO decision triggered where the rate of information

7 Average times were considerably slower in the 'means' task.

Influence of parafoveal words onfoveal inspection time

\ 69

acquisition was slow. Further support for this position can be found in parallel breakdown analyses of the effect of Target Initial Trigram Frequency for short and long targets. There was no effect for short targets (HTF = 281 ms; LTF = 282), but a marked tendency for long targets with 'uninformative' initial trigrams to be associated with shorter prompt inspection time, (HTF = 278 ms; LTF = 284; F,(l,19) = 6.00, p = 0.03; F2(l,92) = 3.21, p = 0.07). Normal reading The analyses of the 'means' task are encouraging, but the question still remains as to whether this pattern of results would obtain in normal reading. There is ample evidence of global parafoveal influences in reading: for example, the studies of McConkie and Rayner (1975), Rayner and Bertera (1978) and Underwood and McConkie (1985) all show a deterioration in reading performance when parafoveal information is perturbed. But it is important to determine whether the curious pattern of trade-offs found in the present laboratory task manifests itself in normal reading, if only because this would set significant constraints on models of eye movement control. As a first attempt to answer this question, an approach was made to Ralph Radach in Aachen asking whether it would be possible to carry out appropriate constrained searches of his large corpus of eye movement data. It should, in principle, be possible to discover the influence on relatively high frequency short words (the equivalent of the experimental 'prompt') of the length, word frequency and trigram constraint of the succeeding word8. The corpus of eye movement recordings was derived from four German-speaking students as they read the first two parts of Gulliver's Travels. It comprises parameters of around 50,000 fixations and saccades for each participant, together with letter and word frequency information generated from the text itself and also from the German CELEX corpus (Celex, 1995). A more detailed description can be found in Chapter 4 and Radach (1996). Cases were identified in which a word (ri) was fixated once only and its fixation duration computed as function of the length, word frequency and trigram informativeness of word n + 1. For data to be entered into the analyses, the following conditions were true: 1. Word n was 5-8 letters in length. 2. In cases where the length of n was 3 letters, the initial landing position was on the central letter; when n was 4 letters, initial landing position was one of the central letters; when greater than 4 letters, it was on letter position 3 or 4. The naivete of this proposition became apparent in the many weeks of data analysis, largely committed to making the question tractable. The analyses which follow were carried out in collaboration with Ralph Radach and I am deeply indebted to him for his work on the question and for his many insightful comments on the data.

170

A. Kennedy

3. 4. 6. 7. 8.

Word n frequency was greater than 50 per million. Word n did not begin a line and was not followed by punctuation. Launch position into Word n was not more than 15 characters to the left. The saccade out of Word n was progressive. No fixation lay to the right of Word n prior to its initial fixation (i.e., the data were not associated with re-reads). Since word frequency and word length are highly correlated, the frequency data were computed as quartiles for each length separately. To deal first with the influence of the length of Word n + 1 on Word n fixation duration. It will be recalled that the 'looks' task showed a strong 'inverted' effect, with shorter durations prior to the fixation of a long word in parafoveal vision (273 vs. 288 ms), and this was also evident in the control, 'means', task. Table 8 shows the results of the corpus analysis, averaged over the four participants.9 Whether Word n is relatively short or of medium length the data show more or less the same effects: a pronounced reduction in fixation duration as the length of Word n + 1 increases. The size of the effect overall is remarkably similar to the experimental data (274 vs. 280 ms for words of directly comparable length). Surprisingly, there appear to be no previous published data on this parafoveal effect. The well-known direct effect of word length on foveal fixation duration (in cases where only one fixation is made) is evident in the Word n data, with fixation durations 24 ms longer on longer words. However, there is little doubt that the length of n + 1 modulates this in the opposite direction. It is true that in the natural reading task there is a possible uncontrolled contingency which should be considered in interpreting this effect. Given the correlation between length and frequency, longer n + 1 words will be, on average, of lower frequency than the words which precede them. But for this to account for the effect it would be necessary to postulate parafoveal word identification as the norm rather than the exception. The process monitoring account proposed for the laboratory task offers a more promising explanation: when the next event is 'difficult', parafoveal processing is slower and the incentive to make an early exit from Word n greater.10 Further investigation will be needed to unravel the mechanisms at work, but the evidence clearly suggests a high degree of parallel processing. Furthermore, the question as to how well the experimental data generalise to naturalistic reading has been answered positively, at least for the case of word length. 9 Separate analyses showed the same trend in the data of all four participants. 10 The 'word group' hypothesis (Radach, 1996) suggests another possibility. If some saccades are directed to word groups, a fixation on Word n could be construed as located well to the left of the 'unified Optimal Viewing Position' (see also, O'Regan, 1990). In which case a lowlevel oculomotor routine to re-locate to a more appropriate position might be released (Footnote 1, Chapter 4).

Influence ofparafoveal words onfoveal inspection

time

171

Table 8 Fixation duration (single fixation cases) on Word n as a function of the length of Word n + 1. Data are given for two lengths of Word n and three lengths of Word n + 1. The number of cases in the corpus is shown in brackets. n + 1 (4 letters)

n + 1 (5-6 letters)

n + 1 (7-10 letters)

n (3-4 letters)

268 (1022)

262 (893)

244 (447)

n (5-8 letters)

287 (788)

280 (897)

274(519)

Table 9 Fixation duration on Word n as a function of the Word Frequency and Trigram Frequency of Word n + 1. Note, words with highly frequent initial trigrams are defined as 'uninformative' and words with rare initial trigrams as 'informative'. The data forn + 1 are average length-contingent quartiles for words 5-10 letters in length. Word n was 5-8 letters in length with a frequency greater than 50 per million. Numbers of cases are shown in brackets. Very high

High

Low

Very low

Word frequency

268 (755)

277(517)

279 (665)

286 (749)

Trigram frequency

268 (686)

275 (535)

282 (707)

284 (758)

Turning now to the question of the influence of lexical and sub-lexical properties of words in the parafovea, separate quartiles were computed for the distributions of fixation durations on Word n for each length of Word n + 1. These were then averaged to provide a measure as a function of the word frequency of n + 1 (independent of length). The data are shown in the top line of Table 9. There is a pronounced frequency effect, with fixation duration on Word n some 18 ms shorter when n + 1 is a 'very frequent' (805 per million) rather than 'very infrequent' (20 per million) word. The outcome offers an intriguing gloss on the question as to whether frequency effects can be found in the durations of the first fixation on a word. Clearly, when reading continuous text, such effects can be found even before a word is fixated at all! However, a note of caution must be sounded. Although the experimental data showed no effects of the word frequency of Word n + 1, such trends as were evident ran in the opposite direction and to that extent the outcome here cannot be taken as confirming their generality. It should be borne in mind that although separate analyses at each word length were possible, the data set was not large enough to attempt a dissociated analysis of the word frequency and initial trigram informativeness. In view of this, a more appropriate test of the generality of the experimental data relates to the obtained highly significant effects of sub-lexical

172

A. Kennedy

properties of the target word. The relevant corpus data are shown in the second line of Table 9. Fixation duration on Word n is clearly influenced by the redundancy of the initial letters of Word n + 1. Parafoveal words with common initial trigrams (i.e., 'uninformative' words) lead to an average 10 ms decrease in processing time on Word n, compared to case when n + 1 is highly redundant. The contrast (284 vs. 268 ms) is similar in size and direction to the comparable overall effect in the experimental data (270 vs. 255 ms for the 'looks' task). General discussion It may be concluded that information acquired from a parafoveal target influences eye movement behaviour on a foveal prompt word. This is incompatible with a processing model in which attention is allocated sequentially. In the laboratory task, where word frequency and trigram informativeness could be dissociated, the analysis of prompt gaze duration shows that parafoveal target word frequency may influence foveal reinspection rate. The effect is inverted, suggesting that in three of the four cases (crossing length and word frequency) processing of a parafoveal target progresses by a combination of extended fixations and shifts in position, to take the eyes closer to the as-yet unfixated target word. The strategy is not deployed in the case of the most difficult targets, where an early decision is made to abort prompt inspection. More directly, the data support the notion that sub-lexical properties of an unfixated parafoveal target influence prompt fixation duration when the prompt is processed with a single fixation. As Inhoff and Rayner (1986) suggest, readers make use of target-initial letters. It is not, however, the token familiarity of the stimulus, but its power to constrain identification of the target which exerts an influence. Furthermore, the influence is evident before the stimulus is fixated: it is not merely cashed later as a preview benefit. In a recent paper, Kennison and Clifton (1995) examined the reverse of the situation discussed here, namely foveal-on-parafoveal interactions. They looked at eye movements as adjective-noun pairs were read, manipulating the four combinations of high and low word frequency of the two words. It was not the primary purpose of their experiment (and the means are not reported), but they report no effect of the second word's frequency on either fixation duration or gaze duration on the first word. This does not present a serious conflict with the present results for two reasons. First, as already argued, the manipulation of word frequency alone may be inadequate to produce parafoveal-on-foveal effects. Second, launch position on the first word in normal reading is necessarily uncontrolled. As Kennison and Clifton point out, refixations almost inevitably result in a confounding between the frequency of the first word and the visibility of the second. Since refixations are less

Influence of parafoveal words onfoveal inspection time

\ 73

probable in high frequency words, their effect, somewhat paradoxically in normal reading, may be to make the next word harder to process. The importance of this observation for the present paper lies in their suggestion that information gained parafoveally is not lost, but may be traded off in later, foveal, inspection. As Kennison and Clifton remark, "the extent to which boundary word ("prompt" in the terminology of the present paper) and target word frequency effects can be observed on parafoveal preview benefit will largely be determined by whether the target word is viewed in parafoveal vision from the same distance across experimental conditions." (p. 78). The data reported here point to interacting processes combining information in parallel across the perceptual span: they are difficult to reconcile with the operation of a sequential, word-by-word, process of attention allocation. The results of Henderson and Ferreira (1990) demonstrate that if a foveal word is difficult there is a processing penalty when a succeeding word is encountered. The present results put the other side of the picture: a difficult parafoveal word may act in a parallel way, with a penalty reflected in processing at the current fixation point. It would be excessive to claim that the laboratory task employed was a direct analogue of normal reading, but, the corpus analysis shows more or less the same pattern of effects. Mutual processing interactions across the perceptual span are present in both cases. One crucial question remains: the degree to which prompt and target processing might be traded off. Does an extended time spent processing the prompt reflect itself directly in shorter processing on the target? If the target word has been partly or wholly identified during prompt inspection, is this reflected in a lower processing load when it is directly inspected? Models of eye movement control have been relatively silent on this question. It cannot be answered by the simple correlation between the fixation duration on the two words because of the large spurious positive correlation which would result from massive individual variation in processing time. It is possible, however, to adopt an analysis strategy with regard to the 'looks' task data, normalising individual prompt and target fixation durations using the average total duration on prompt and target combined as a baseline. Individual fixation durations were scaled in this way, restricting the analysis to the crucial 'single fixation match' case. A trade-off clearly predicts a negative correlation, with the null hypothesis of no trade-off predicting zero correlation. The obtained outcome is a strong negative correlation, r = -0.82. The significance of this was computed directly using the method of Random Data Permutation (Edgington, 1987) and produced a value of/? < 0.001. This outcome strongly supports the idea of the parallel processing of two successive words, with processing gains or deficits inherited when a parafoveal word becomes a foveal target.

1 With 10,000 iterations. I am grateful to John Todman for suggesting this procedure.

174

A. Kennedy

I will conclude by trying to relate these inter-word trade-offs to parallel, intraword, effects. It is well-known that if a word is fixated exactly twice, curves describing the relationship between the initial landing position and fixation duration for each of the two