No title

JOURNAL OF SEMANTICS Volume 24 Number 4

Special issue on PROCESSING MEANING (Part 1) Edited by: JULIE SEDIVY ROBYN CARSTON IRA NOVECK BART GEURTS

CONTENTS Editors’ Preface

305

GIOSUE´ BAGGIO AND MICHIEL VAN LAMBALGEN The Processing Consequences of the Imperfective Paradox

307

GUY POLITZER The Psychological Reality of Classical Quantifier Entailment Properties

331

PAULA RUBIO FERNA´NDEZ Suppression in Metaphor Interpretation: Differences between Meaning Selection and Meaning Construction

345

FLORIAN SCHWARZ Processing Presupposed Content

373

Editor’s Note

417

Please visit the journal’s web site at www.jos.oxfordjournals.org

Journal of Semantics 24: 417 doi:10.1093/jos/ffm015

Editor’s Note

Ron Artstein Sigrid Beck Brady Clark Ariel Cohen Seana Coulson Paul Egre´ Zachary Estes Richard Gerrig Martin Hackl Michela Ippolito Michael Israel Lara Jones Kathrin Koslicki Emiel Krahmer Beth Levin Jeff Lidz Giorgio Magri

Emar Maier Ken Manktelow Lisa Matthewson Friederike Moltmann Rick Nouwen Guy Politzer Anne Reboul Hotze Rullmann Magda Schwager Roger Schwarzschild Wendy Simons Penka Stateva Wolfgang Sternefeld Simon van de Kerke Rachel Yang Malte Zimmermann

Ó The Author 2007. Published by Oxford University Press. All rights reserved. For Permissions, please email: [email protected].

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

We wish to express our gratitude to the following colleagues who are not members of the Editorial Board of the Journal but have kindly helped us in the past year with the refereeing of papers submitted:

Journal of Semantics 24: 305–306 doi:10.1093/jos/ffm014

Editors’ Preface

Ó The Author 2007. Published by Oxford University Press. All rights reserved. For Permissions, please email: [email protected].

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

It is evident that, in theory, researchers in the cognitive sciences and semantics have nothing to lose and a lot to gain from working together. In practice, however, contacts have for a long time been lukewarm, at best. A few decades ago, this could have been blamed on the fact that semantics was viewed, by many of its practitioners and outsiders alike, as being concerned with a formal characterisation of language rather than what goes on in the minds and/or brains of speakers and hearers. But this has changed: since the late 1970s, theoretical semantics has become increasingly ‘‘cognitivist’’ in its orientation, and notions like information states, speakers’ beliefs, and (mental) discourse representations have come to play a key role in many semantic theories. Although there may still be an anti-mentalist undercurrent, it has weakened dramatically, and is surely not the prevailing trend anymore. The true reason why there has been relatively little cooperation between (neuro-) psychologists and semanticists may well be that, when tried, it often turns out to be remarkably hard. Although experimentalists and theoreticians share the same general outlook on language and meaning, and broadly speaking have the same interests, carrying through joint projects can be quite challenging. In part this is because expectations are often too high. For all the mentalistic terminology they use, it is rare for semantic theories to directly yield predictions that can be tested in the lab without further ado, while on the other hand a substantial portion of the data reported in the experimentalists‘ journals are of marginal relevance to issues in theoretical semantics. In part it is also because, when push comes to shove, theoreticians and experimentalists have different interests, after all. There is enough overlap for cooperation to be worthwhile, but it requires considerable effort to appreciate each other’s concerns and understand each other’s vocabularies. Fortunately, however, there are people who are willing to make the effort, and their numbers are going up. In recent years, quite a few semanticists have begun to take an active interest in experimental work, while experimental psychologists are increasingly aware that semantic literature, though often abstruse, is closer to their interests than they may have thought. This special issue brings together a handful of papers that investigate how the mind and brain process and represent linguistic meaning. The

306 Editors’ Preface authors’ ‘‘core businesses’’ range from logic and semantics to psychology and neuroscience, but all papers are interdisciplinary to the point that it would be hard to guess the authors’ backgrounds from the content of their papers alone. This is how it should be, of course, and we hope to see more work like this in the near future. JULIE SEDIVY ROBYN CARSTON IRA NOVECK BART GEURTS Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

Journal of Semantics 24: 307–330 doi:10.1093/jos/ffm005 Advance Access publication July 16, 2007

The Processing Consequences of the Imperfective Paradox GIOSUE` BAGGIO F.C. Donders Centre for Cognitive Neuroimaging, Radboud University Nijmegen

Abstract In this paper we present a semantic analysis of the imperfective paradox based on the Event Calculus (van Lambalgen & Hamm 2004), a planning formalism characterizing a class of models which can be computed by connectionist networks. We report the results of a questionnaire that support the semantic theory and suggest that different aspectual classes of VPs in the progressive give rise to different entailment patterns. Further, a processing model is outlined, combining the semantic analysis with the psycholinguistic principle of immediacy in the framework of recurrent networks. The model is used to derive predictions concerning the electrophysiological correlates of the computations described by the Event Calculus.

1 INTRODUCTION Recently, a number of studies have brought experimental data to bear on semantic theories (see, for instance, the work on quantifiers by Geurts & van der Slik 2005 and McMillan et al. 2005, 2006). However, formal semantics and psycholinguistics have reached their most important results independently, even when the issues at stake could have been addressed more completely by joining efforts. One example is the study of discourse-based inference in psycholinguistics, where formal notions of truth, entailment and veridicality have often been neglected (see Cook et al. 2001 and Frank et al. 2003 among others). The assumption behind this paper is that there exists a relatively unexplored territory in which the two disciplines can interact productively. Here we consider a small portion of this territory: the imperfective paradox and its processing consequences. In the remainder of this section we provide a minimal methodological background against which semantic theories can be combined with processing models and we introduce the semantic paradox which we shall work with. In section 2 we present an Ó The Author 2007. Published by Oxford University Press. All rights reserved. For Permissions, please email: [email protected].

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

MICHIEL VAN LAMBALGEN Department of Philosophy, University of Amsterdam

308 Processing Consequences of the Imperfective Paradox analysis of the paradox based on the Event Calculus (van Lambalgen & Hamm 2004), a planning formalism characterizing a class of models which can be computed by connectionist networks. In section 3 we report the results of a questionnaire that support the semantic theory and suggest that different aspectual classes of VPs in the past progressive give rise to different entailment patterns. In section 4 a processing model is outlined, combining the formal semantic analysis with the psycholinguistic principle of immediacy in the framework of recurrent networks. The model is used to derive predictions concerning the electrophysiological correlates of the computations described by the Event Calculus.

From a semantic analysis of a given linguistic structure it is usually not possible to derive predictions concerning the complexity and time course of the processes involved in producing or comprehending utterances in which the structure occurs. Although there may be cases in which processing hypotheses can be formulated on the basis of the semantic theory alone (Geurts & van der Slik 2005; McMillan et al. 2005), this approach is unlikely to work in general. The reason is that processing largely depends upon the particular algorithms and neural mechanisms that, in a physical system such as the human brain, compute the linguistic structures posited by the theory—or functionally equivalent structures. Because semantics typically does not describe algorithms and neural mechanisms (nor perhaps it should), there appears to be no direct way to relate semantic theories to what is observed in language-processing experiments. The solution is to adopt a theoretical framework that allows formal semantic analyses to be explicitly related to processing algorithms inspired by the available psycholinguistic evidence and ultimately to mechanistic models of the kind investigated in neuroscience. Marr’s (1982) scheme for the analysis of cognitive systems seems an appropriate choice in this regard. Marr argued that information-processing systems should be understood at three nearly independent levels of analysis. The first level contains a description of the computations to be performed by the system and more precisely a characterization of the goals that have to be attained in order to solve the information-processing problem. For example, a sentence S can be seen as introducing a specific goal, namely, the construction of a cognitive model making S true. In the semantic analysis proposed below, the tense and aspect of VPs like the past progressive are treated as instructions to update the current

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

1.1 Methodological preamble

Giosue` Baggio and Michiel van Lambalgen 309

discourse model so as to achieve that goal. We regard it as the task of semantics to describe information-processing goals and update instructions. The actual processing steps are described at the intermediate level, where constraint satisfaction algorithms, implemented in artificial neural networks (Marr & Poggio 1976), and processing principles such as immediacy and incrementality are combined. Marr completes the picture with a third level, at which the neurobiological architecture is described. We discuss issues relevant for the computational level in section 2 and for the algorithmic level in section 4.

1.2 The imperfective paradox

(1) The girl was writing a letter when her friend spilled coffee on the tablecloth. From (1) the reader would typically conclude that, barring unforeseen circumstances, the girl attained the desired goal and would thus assent to the statement ‘The girl has written a letter’. Such inference is based on the assumption that spilling coffee on the tablecloth is usually neutral with respect to the writing activity, that is it is not a typical immediate cause leading to its termination. It is possible to imagine scenarios in which writing is temporarily interrupted or even terminated by the accident. Nonetheless, as the data reported in section 3 will demonstrate, failing to explicitly mention a disabling condition in the discourse is sufficient for leading the reader to assume that there was no such obstacle to attaining the intended goal. Inferences to a goal state are non-monotonic, that is they can be suppressed if the discourse describes an event which terminates the relevant activity: (2) The girl was writing a letter when her friend spilled coffee on the paper. Assuming that writing was intended to occur on the same paper sheets on which coffee was spilled, the accident is sufficient for terminating the activity and it is therefore a disabling condition for obtaining a complete letter. Accordingly, on the basis of (2) the reader would be more likely to assent to ‘The girl has written no letter’.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

Vendler (1957) famously classified VPs as states (‘know’, ‘love’ etc.), activities (‘write’, ‘run’ etc.), accomplishments (‘write a letter’, ‘run a mile’ etc.) and achievements (‘finish’, ‘reach’ etc.). We are concerned with the inferences licensed by sentences containing activities and accomplishments in the past progressive. The following example involves the accomplishment ‘write a letter’:

310 Processing Consequences of the Imperfective Paradox One important observation is that suppression can obtain only with accomplishments and not with activities (Rothstein 2004). Since a sentence containing an activity in the past progressive, such as ‘writing letters’, does not involve a canonical goal, it is interpreted as entailing that ‘The girl has written one or more letters’ regardless of the consequences of the second event on the writing activity: (3) The girl was writing letters when her friend spilled coffee on the tablecloth. (4) The girl was writing letters when her friend spilled coffee on the paper.

. . . the progressive picks out a stage of a process/event which, if it does not continue in the real world, has a reasonable chance of continuing in some other possible world. (de Swart 1998) They differ however in the (largely informal) descriptions of the possible worlds used. For example, Dowty (1979) would claim that the following are equivalent: a. ‘The girl is writing a letter’ is true in the actual world; b. ‘The girl will have written a letter’ is true in all so-called ‘inertia worlds’, worlds which are identical with the present world until ‘now’, but then continue in a way most compatible with the history of the world until ‘now’. These approaches are intensional in the formal sense of using possible worlds. In fact, most authors (but not all) would agree that the progressive creates an intensional context: even though the accident in (2) may have terminated the writing activity at a stage in which it was unclear whether the girl was writing a letter or, say, a memo, still only one of

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

Upon reflection, there is something paradoxical about examples (1) and (2) which is not found in (3) and (4). Although it belongs to the meaning of the accomplishment ‘writing a letter’ that the writing activity is directed towards the goal state of a complete letter, the actual occurrence of that goal state can be denied without contradiction. But how can a seemingly essential component of the meaning be denied without affecting the meaning itself? This is known as the ‘imperfective paradox’. The semantic literature is replete with attempted solutions of the paradox, ranging from explaining the problem away (Michaelis 2001) to various invocations of possible worlds (Dowty 1979; Landman 1992; de Swart 1998). It is impossible to review all proposed solutions here. Instead, we will focus on some representative claims. Possible worlds analyses are based on the idea that

Giosue` Baggio and Michiel van Lambalgen 311

(5) The girl was writing a letter. (6) The girl was writing a memo. is true of the situation described by (2). Explicitly denying that the progressive creates an intensional context, Michaelis (2001) argues that1

We find this questionable, for without access to a person’s intention it may be very hard to tell initially whether she is drawing a circle or a square, writing a letter or a memo. But that person’s intention in performing an activity is characterized precisely by the associated consequent state, even though the latter cannot yet be inferred from the available data. Here the Event Calculus will come to our rescue because the notion of goal or intention is built into the semantics from the start. In particular, the meaning of a VP is represented by a scenario which describes a plan for reaching the goal. However, unlike approaches such as Parsons’ (1990) where one quantifies existentially over events, the scenario is a universal theory and does not posit the occurrence of the intended consequences. Although the plan is appropriate for that purpose, attaining the goal is guaranteed only in a minimal model (in which no unforeseen obstacles occur) of the scenario plus the axioms of the Event Calculus. The meaning of an accomplishment (as embodied in the scenario) involves a culminating event type, which therefore must exist; but no existential claims are made concerning the corresponding event token, which, as in example (2), may also fail to occur. Type and token are handled by different mechanisms in the Event Calculus. These notions form the basis of our semantic analysis of the imperfective paradox, to which we now turn. 1 Replace ‘drawing a circle’ with ‘writing a letter’ and ‘drawing a square’ with ‘writing a memo’ in Michaelis’ examples to see the connection with our examples.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

. . . the Progressive sentence ‘She is drawing a circle’ denotes a state which is a subpart not of the accomplishment type She–draw a circle but of the activity type which is entailed by the causal representation of the accomplishment type. Since this activity can be identified with the preparatory activity that circle drawing entails, circle drawing can in principle be distinguished from square drawing etc. within the narrow window afforded by the Progressive construal [and] does not require access to culmination points either in this world or a possible world.

312 Processing Consequences of the Imperfective Paradox 2 SEMANTIC ANALYSIS

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

The reference to goal states in the preceding section suggests that a semantic analysis of the progressive can be based on a planning formalism which is able to talk about goals and actions and which includes a theory of causality together with a principle of inertia (Hamm & van Lambalgen 2003). Such a formalism is presented in van Lambalgen & Hamm (2004). It consists of an Event Calculus which has found applications in robotics, reformulated using the computational machinery of Constraint Logic Programming. The reader may wonder why planning can provide a source of inspiration to linguistics. The reason can be found in the nature of planning computations, which typically proceed as follows. First a goal is specified, which may be an action to be performed at a particular location (e.g. pick up outgoing mail in an office building). Next a plan is computed, that is a sequence of actions to get to the required location, derived by backward chaining from the goal to obtain a sequence of subgoals, the last one of which can be executed in the agent’s initial position and state. Planning requires a situation model (including a map of the building, a causal theory of the agent’s actions, a specification of values of variables such as ‘door open/closed’, the agent’s initial position and state and a record of its past and current actions), a repertoire of possible actions (‘follow wall’ and ‘go through door’) and observations (‘door open/closed’). While the agent executes the plan, it also registers its observations and actions in the situation model; knowledge of its actions may be important for the agent to estimate its current position. Particularly relevant for our purposes is that a plan might have to be re-computed in mid-course when the initial situation model is updated due to new observations (for instance, a closed door which was expected to be open on the basis of the initial model or a wrong estimate of the agent’s current position). Later on we shall see how this ‘re-computation’ relates to the imperfective paradox. This short description should be sufficient for enabling the reader to see the connection with language processing. The comprehender starts with an initial discourse model, in which an incoming sentence or clause must be integrated. Suppose the main verb of the sentence is non-stative, for instance, an activity. If the sentence is in one of the simple tenses, it is unpacked in the relevant action and its participants, and the discourse model is updated accordingly. This is the analogue of updating the situation model with representations of individuals and actions in planning. In more complex cases, such as sentences involving accomplishments like (1) and (2), the VP is taken to express the

Giosue` Baggio and Michiel van Lambalgen 313

existence of a goal-directed plan. If, on the contrary, the main verb of the sentence is stative, the sentence can be viewed as analogous to an observation report, and the discourse model is updated with a property.

2.1 The Event Calculus

2

The predicate Releases is used to reconcile the two notions of causality: while instantaneous change leads to one form of inertia, where properties do not change their value between the occurrences of two events, continuous change requires that variable quantities may change value without concomitant occurrences of events. The solution of this conflict is to exempt, by means of the predicate Releases, those properties which we want to vary continuously from the inertia of the first form of causation.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

The Event Calculus is a planning formalism which allows one to talk about actions, goals and causal relations in the world. Its main function is to return a plan given a goal, the initial state and causal relations. Formally, the Event Calculus is a many-sorted predicate logic. It has two different sorts for events, viewed either perfectively or imperfectively. The former are called event types and are symbolized by e, e#, . . ., e0, . . .. The latter are called fluents (Newton’s name for timedependent variables) and are symbolized by f, f#, . . ., f0, . . .. One may think of event types as action types, such as, for example ‘break’ or ‘ignite’; fluents can be thought of as time-dependent properties, such as ‘being broken’ or ‘writing’; the time parameter in fluents is implicit, but they can have further parameters (for instance, for the subject of ‘writing’). The real distinction between event types and fluents comes from the different roles they play in the axioms of the Event Calculus. The universe also contains sorts for individuals (‘the girl’), real numbers interpreted as instants of time, and various other real quantities (such as position, velocity and degree of some quality). The primitive predicates comprise the minimum necessary to talk about two forms of causality, instantaneous (as in a collision) and continuous (as when a force is acting): (7) a. Initially(f): ‘fluent f holds at the beginning of the discourse’ b. Happens(e, t): ‘event type e has a token at time t’ c. Initiates(e, f, t): ‘the effect of event type e at time t is the initiation of f ’ d. Terminates(e, f, t): ‘the effect of event type e at time t is the termination of f ’ e. Trajectory(f1, t, f2, d): ‘if f1 holds from t until t + d, then at t + d f2 holds’ f. Releases(e, f, t): ‘event type e releases the continuously-varying fluent f’2 g. HoldsAt(f, t): ‘fluent f holds at at time t’

314 Processing Consequences of the Imperfective Paradox

1. The discourse model contains only those occurrences of events forced to be there by the discourse and the axioms; 2. The interpretation of the primitive predicates is as small as is consistent with the discourse and the axioms. These two requirements define minimal models. They imply that no unforeseen events are allowed to happen and that all causal influences are as expected. The choice to work with minimal models instead of all models leads to non-monotonicity in discourse interpretation: adding a new sentence or clause to the discourse may invalidate a conclusion derived from the initial model. In van Lambalgen & Hamm (2004) it is argued that it is precisely the possibility to retract previously inferred 3 See the related notions of ‘frame’ in Minsky (1977) and ‘script’ in van Dijk (1980). For a more explicit connection with planning, see Schank & Abelson (1977).

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

While (7a–d) are the predicates required by instantaneous change, (7e–f) are used to model continuous change. Three further notions will be used in the semantic analysis presented below. The first is that of scenario.3 Whereas the axiom system EC provides a macro-theory of causality, scenarios provide micro-theories stating the specific causal relations obtaining in a given situation. Scenarios can be used to describe the causal environment of actions and events at the level of granularity expressed in natural language, such as, for instance, writing a letter or drawing a circle. Scenarios can be taken to represent, in a tenseless fashion, the meaning of VPs. For instance, the scenario for the accomplishment ‘writing a letter’ specifies that the writing activity is causally related to the amount of letter produced, that the termination of the activity is temporally contiguous to the completion of the letter and so on. The second notion is that of integrity constraint. As we said above, we regard a sentence as introducing an information-processing goal (‘Make S true’) to be achieved by updating the current discourse model or constructing a new one. Integrity constraints regiment such updates. They can take the form of either obligations ‘?u succeeds’, requesting an update of the discourse model satisfying u, or prohibitions ‘?u fails’, blocking updates of the discourse model satisfying u. While scenarios describe the meaning of VPs in a tenseless fashion, integrity constraints specify the contribution of tense, aspect and temporal adjuncts to the semantics of VPs. The third notion is that of minimal model. The axioms of the Event Calculus constitute a correct theory of causality if and only if the following two conditions are satisfied:

Giosue` Baggio and Michiel van Lambalgen 315

conclusions which allows a rigorous treatment of the imperfective paradox. The most important meta-theorem concerning the EC formalism is that minimal models exist and can be computed efficiently (van Lambalgen & Hamm 2004; Stenning & van Lambalgen 2005, 2007). Furthermore, minimal models can be computed (or approximated, depending on the expressiveness of the logical language) by connectionist networks—a topic to which we return in section 4.3.

2.2 The semantics of the progressive

(8) a. b. c. d. e. f. g.

Initially(letter(a)) Initiates(start, write, t) Initiates( finish, letter(c),t) Terminates( finish, write, t) HoldsAt(write, t) ^ HoldsAt(letter(c), t) / Happens( finish, t) Releases(start, letter(x), t) HoldsAt(letter(x), t) / Trajectory(write, t, letter(x + g(d)), d)

Here write is the activity fluent, letter(x) is a parametrized fluent for the completion stage x of the letter, a is a constant for the stage at which writing is initiated, c is a constant for the stage at which the letter is considered finished and letter(c) is the fluent for the goal state. The dynamics of the scenario is formalized by (8g), which says that, if the stage of completion of the letter at time t is x, then the writing activity, lasting from time t until t + d, will result in a letter whose stage of completion at time t + d is x + g(d), where g is a monotone increasing, time-dependent, real-valued function relating the activity to the completion stage.5 Accomplishments are distinguished from activities by the statements regimenting the behaviour of the goal state, here (8c) and (8e): because no canonical goal is involved in activities such as ‘write letters’, their scenario will not contain these clauses. 4

See Andrade-Lotero (2006) for the algorithms deriving scenarios from Discourse Representation Structures in Discourse Representation Theory (Kamp 1981) and from constructions in Construction Grammar (Goldberg 1995). 5 The way g is defined depends upon the particular VP (or, equivalently, the particular plan) at issue. For instance, ‘erasing the blackboard’ or ‘dismantling the engine’ may be represented using a monotone decreasing function.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

The semantics of the VP ‘was writing a letter’ can be decomposed into the tenseless lexical meaning of ‘write a letter’ and the temporal and aspectual contribution of the past progressive. The lexical meaning is represented by the following scenario:4

316 Processing Consequences of the Imperfective Paradox The contribution of the past progressive is represented by the integrity constraint (9) ?HoldsAt(write, R) ^ R < S succeeds

Proposition 1. Let D be a discourse consisting of scenario (8) given above. Suppose D is extended to D# so that the query ?HoldsAt (write, R) ^ R < S succeeds in D#. Suppose also limt/N gðtÞ > c. Then there is a unique minimal model of D# and in this model there is a time t > R for which HoldsAt(letter(c),t). By virtue of the stipulation Letter(letter(c),t), there will be a letter at time t.6 If write holds at R, as required by (9), it must either hold initially or have been initiated. The latter requires an event start which initiated the writing activity. Since the starting event is not provided by discourse, we must assume that write holds initially. The clause (8g) states that the writing activity will increase the stage of completion of the letter. As time tends to infinity, the latter will be at least equal to c (the final completion stage). We have stipulated that a letter whose stage of completion is c is a finished letter. Hence, there must be a time at which the letter is considered finished. The writing activity will stop as soon as a complete letter is obtained, as is guaranteed by (8d–e). Notice that this holds for accomplishments but not for activities: if the VP is the activity ‘writing letters’, (8c) and (8e) are not part of the scenario and writing will continue also after one letter has been completed—in fact, it will continue indefinitely, if no terminating event is described. We have just seen how the Event Calculus deals with the interpretation of VPs in the past progressive. We will now discuss the contribution of subordinate ‘when’ clauses to the meaning of (1)–(4). Because spilling coffee on the tablecloth usually does not terminate the writing activity, the dynamics of the scenario will lead to a complete 6

See van Lambalgen & Hamm (2004) for a more rigorous statement and a computational proof.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

which forces an update of the discourse model such that the activity fluent write holds at the reference time R located in the past of the moment of speech S (Reichenbach 1947). As we noted above about examples (1)–(4), different inferences can be drawn from activities and accomplishments in the past progressive. In the Event Calculus, this follows from the fact that updating the initial model according to (9) leads to different models depending on whether the scenario represents an activity or an accomplishment. Proposition 1 provides information on the inferences licensed by a discourse containing an accomplishment in the past progressive.

Giosue` Baggio and Michiel van Lambalgen 317

letter in (1) and to an indefinite number of letters in (3). The argument mirrors the one given above as a gloss of Proposition 1. The situation is different for examples (2) and (4). Spilling coffee on the paper is typically sufficient for terminating the writing activity. This bit of world knowledge can be expressed by the following addition to scenario (8): (10) Terminates(spill, write, t) But what is more important here is the integrity constraint introduced by the ‘when’ clause (11) ?Happens(spill, R) ^ R < S succeeds

3 BEHAVIOURAL DATA In the preceding sections we have argued that accomplishments and activities in the past progressive behave differently as regards entailment.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

which requires the accident to occur at R, while the writing activity was taking place. Since during the writing process there is no complete letter (yet), spilling coffee on the paper will terminate the activity before the letter is finished. Therefore, there will be no complete letter in the final discourse structure. As for (4), the theory is consistent both with the situation in which the writing activity was terminated before even a single letter had been completed and with the case in which one or more letters were finished when the accident happened. The proposed semantics predicts that readers would assent to ‘The girl has written a letter’ in (1), to ‘The girl has written no letter’ in (2) and to ‘The girl has written one or more letters’ in (3). This is consistent with received semantic wisdom about entailments of activities and accomplishments in the progressive. Behavioral data supporting this claim are presented in the next section. To summarize, the attainment of the consequent state can be derived in a minimal model of a discourse containing an accomplishment in the progressive. However, the computation of discourse models is non-monotonic: if the minimal model is extended with a sentence or a clause describing an event which terminates the relevant activity (what we called a disabling condition), the derivation of the goal state is blocked. Non-monotonicity affords a neat solution of the imperfective paradox, as there is no contradiction in assuming that the representation of the goal state is both essential to the meaning of the progressive VP (as an event type in the scenario) and suppressible on the basis of additional discourse information (as an event token in the minimal model).

318 Processing Consequences of the Imperfective Paradox

3.1 Subjects Thirty-two native speakers of Dutch (mean age 22.5, 27 female) completed the cloze probability test (see below) and 36 (mean age 22.5, 28 female) the entailment questionnaire. Participants were selected from the database of the F.C. Donders Centre for Cognitive Neuroimaging in Nijmegen. The two sets of subjects were disjoint.

3.2 Materials The set of Dutch materials used in the tests included 160 items. Each item comprised two context sentences (C) providing a neutral setting for the events described by critical sentences,7 four critical sentences (S1)–(S4) and two probe pairs (P1)–(P2): (C) De deur van de woonkamer was gesloten. Binnen speelde de radio klassieke muziek. The door of the living-room wasPST closedPRT. Inside playedPST the radio classical music. ‘The door of the living room was closedPST PRT. Inside the radio playedPST classical music’. (S1) Het meisje was brieven aan het schrijven toen haar vriendin koffie op het tafelkleed morste. The girl wasPST letters on the to-writeINF when her friend coffee on the tablecloth spilledPST. ‘The girl was writingPST PRG letters when her friend spilledPST coffee on the tablecloth’. 7 That is, no disabling condition for the event described by the progressive VP was introduced by context items.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

An accomplishment such as ‘writing a letter’ implies that ‘The girl has written a letter’ provided that no obstacles are described in discourse. Failing to mention a disabling condition is sufficient for inferring that there is no such obstacle to attaining the goal. If, however, a disabling condition is introduced, the accomplishment will imply that ‘The girl has written no letter’. An activity such as ‘writing letters’ implies that ‘The girl has written one or more letters’ regardless of further discourse information. Since disabling conditions affect the possibility of attaining a predefined goal, and such predefined goals are part of the meaning of accomplishments but not of activities, accomplishments will be sensitive to the presence of a disabling condition in the discourse context, whereas activities will not. We administered an entailment questionnaire aimed at testing this claim.

Giosue` Baggio and Michiel van Lambalgen 319

Critical sentences were constructed manipulating the aspectual class of the VP in the past progressive (activity or accomplishment) and the causal type of the condition introduced by the subordinate ‘when’ clause (neutral or disabling with respect to the event described in the progressive clause). Activities and accomplishments differed only in the object NP: an indefinite (‘een brief ’) was used for accomplishments, a bare plural (‘brieven’) was used for activities. Disabling and neutral conditions differed only in the prepositional or object NP, for instance, ‘papier’ for the former and ‘tafelkleed’ for the latter. The distinction between neutral and disabling conditions was made according to the authors’ judgment. The probes (P1) were used with activities, (P2) with accomplishments.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

(S2) Het meisje was brieven aan het schrijven toen haar vriendin koffie op het papier morste. The girl wasPST letters on the to-writeINF when her friend coffee on the paper spilledPST. ‘The girl was writingPST PRG letters when her friend spilledPST coffee on the paper’. (S3) Het meisje was een brief aan het schrijven toen haar vriendin koffie op het tafelkleed morste. The girl wasPST a letter on the to-writeINF when her friend coffee on the tablecloth spilledPST. ‘The girl was writingPST PRG a letter when her friend spilledPST coffee on the tablecloth’. (S4) Het meisje was een brief aan het schrijven toen haar vriendin koffie op het papier morste. The girl wasPST a letter on the to-writeINF when her friend coffee on the paper spilledPST. ‘The girl was writingPST PRG a letter when her friend spilledPST coffee on the paper.’ (P1) Het meisje heeft een of meer brieven geschreven. The girl hasPRS one or more letters writtenPRT. ‘The girl has writtenPRS PRF one or more letters’. Het meisje heeft geen brief geschreven. The girl hasPRS no letter writtenPRT. ‘The girl has writtenPRS PRF no letter’. (P2) Het meisje heeft een brief geschreven. The girl hasPRS a letter writtenPRT. ‘The girl has writtenPRS PRF a letter’. Het meisje heeft geen brief geschreven. The girl hasPRS no letter writtenPRT. ‘The girl has writtenPRS PRF no letter’.

320 Processing Consequences of the Imperfective Paradox

3.3 Procedure Copies of the questionnaire were distributed to all subjects in the database meeting the following criteria: they had to be native speakers of Dutch with an age between 18 and 50 and with no history of neurological, psychiatric or cognitive disorders. The first 36 subjects who returned a completed questionnaire were included in the analysis. The first page of each booklet contained the test instructions. Participants were informed that the questionnaire consisted of 160 short texts and that each comprised three sentences (the two context sentences and the critical sentence) and was followed by a pair of probes. Subjects were instructed to read each sentence carefully and select the probe which they deemed correct on the basis of their expectations (‘verwachtingen’) about the continuation of the text (‘over het verlog van de tekst’). Participants were asked to respond as quickly and accurately as possible and to write a brief comment on their answer in the blank space following the probes. The reference to ‘expectations’ in the test instructions requires some explanation. Notice that we are not interested in what ‘subjects semantically know’ given a progressive clause—presumably that the writing activity was taking place some time in the past and that only a part of the letter was then completed. For this would amount to asking subjects what is true at the reference time, which is captured by the integrity constraint (9) above and requires no inference based on the relevant models. Rather, we are interested in what subjects infer about the outcome of an action described using the progressive. That is, we want to know how subjects reason about goals and, more precisely,

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

The following linguistic properties of critical sentences were normed. The mean length and the raw frequency of the differing words in the NP of subordinate ‘when’ clauses were matched using the CELEX corpus (Baayen et al. 1996). To determine the cloze probabilities of the verbs in the subordinate clauses, context items followed by a critical sentence with the final word blanked were presented to subjects. Participants were asked to fill in the blank with the first word that came to their mind. Four versions (40 items per condition) randomized and balanced across conditions were constructed. Mean cloze probabilities are not different across conditions (T-tests, all comparisons P > 0.05). This was established for each version as well as for the entire set of experimental items. Therefore, the same materials and test versions were used in the entailment questionnaire.

Giosue` Baggio and Michiel van Lambalgen 321

3.4 Data analysis Subject-based and item-based statistical analyses were carried out. For the former, we used a repeated-measures analysis of variance (ANOVA) model with Subject as the random effect, Aspectual Class (Activity or Accomplishment) and Condition Type (Neutral or Disabling) as fixed effects and the mean number of negative responses (i.e. of negative probes chosen as responses) as the dependent variable. To generalize over both subjects and test items,8 we employed a parallel repeatedmeasures ANOVA model in which Test Item (as defined above, i.e. as a set of context, critical and probe sentences) was the random effect, Aspectual Class and Condition Type were the fixed effects and the mean number of negative respondents (subjects giving a negative response) was the dependent variable. Univariate F-tests were computed in both cases.

3.5 Results The subject-based analysis revealed significant effects of Aspectual Class and Condition Type and a significant interaction between the two factors (see Table 1). Neutral activities (S1) had the lowest mean of negative responses (M ¼ 2.72, SD ¼ 3.22), followed by disabled activities (S2) (M ¼ 8.06, SD ¼ 7.05), neutral accomplishments (S3) (M ¼ 10.03, SD ¼ 9.23) and disabled accomplishments (S4) (M ¼ 25.14, SD ¼ 8.02). Except for neutral activities (S1), the distribution of 8

See Clark (1973) for the main motivation and Wike & Church (1976) for a critical discussion.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

what is the projected course of events after the reference time. In order to accomplish this, we constructed the probe pairs (P1) and (P2) using the Dutch present perfect (which focuses on the present consequences of a past event), we avoided probes of the form ‘The girl has written a part of the letter’ and we asked subjects to decide which of the two probes (positive or negative) matched their expectations about the continuation of the narrative. Participants’ written explanations provide no evidence that (P1) and (P2) were insufficient for giving accurate responses such that, for instance, a probe of the form ‘It is unclear whether the girl has written a letter’ was necessary. Rather, subjects’ comments suggest that, if the discourse implied that only a part of the letter was completed, as in (S4), then the negative probe in (P2) had to be selected.

322 Processing Consequences of the Imperfective Paradox

Aspectual Class Condition Type Aspectual Class 3 Condition Type

Subject-based analysis

Item-based analysis

F(1,35) ¼ 64.763 P < 0.001 F(1,35) ¼ 112.560 P < 0.001

F(1,159) ¼ 619.240 P < 0.001 F(1,159) ¼ 237.270 P < 0.001

F(1,35) ¼ 61.832 P < 0.001

F(1,159) ¼ 100.210 P < 0.001

Table 1 Summary of ANOVA statistics for the questionnaire data

the data in the different conditions appears rather similar, as indicated by standard deviations (SDs) and box height and whisker length in Figure 1a. Figure 1b shows that the pattern of responses appears as predicted by the theory: accomplishments are more sensitive than activities to the presence of a disabling condition in the discourse context. A similar pattern of effects was revealed by the item-based analysis (Table 1). Neutral activities (S1) had the lowest mean of negative respondents (M ¼ 2.35, SD ¼ 4.13), followed by disabled activities (S2) (M ¼ 7.13, SD ¼ 7.24), neutral accomplishments (S3) (M ¼ 9.54, SD ¼ 7.09) and disabled accomplishments (S4) (M ¼ 22.62, SD ¼ 9.04).

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

Figure 1 Results of the questionnaire. (a) Boxplot for the subject-based analysis. Solid lines within the boxes represent the median, box height is equal to the inter-quartile range, whiskers indicate adjacent values and empty circles denote outliers. The maximum of potential negative judgments is 40. (b) Interaction plot for the subject-based analysis. Trace endpoints represent the mean number of negative responses computed across participants in each experimental condition. The maximum of potential negative judgments is 40.

Giosue` Baggio and Michiel van Lambalgen 323

4 PROCESSING MODEL

4.1 The principle of immediacy In a paper discussing language-processing models and their neural implementation, Hagoort (2006) proposed ‘six general architectural principles for comprehension beyond the word level’. One of these, also known as the ‘principle of immediacy’, is particularly relevant in this context. Immediacy has often been debated in the psycholinguistic literature (Marslen-Wilson 1989; Jackendoff 2002), where it is viewed as a hypothesis about the time course of the access and integration of lexical meanings in a sentence structure and is contrasted with syntaxfirst models (Frazier 1987; Friederici 2002). Immediacy is defined by Hagoort (2006) as . . . the general processing principle of unification. Semantic unification does not wait until all relevant syntactic information (such as word class information) is available, but starts immediately with what it derives on the basis of the bottom-up input and the left context. The corollary of immediacy is incrementality: output representations are built up from left to right in close temporal contiguity to the input signal. This general statement cannot be directly related to our Event Calculus analysis. The reason is that it refers only to lexical integration and not to the computation of denotations and discourse models. Hagoort’s notion of immediacy is sufficient for deriving processing predictions

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

In the preceding sections we have proposed an analysis of the imperfective paradox based on the Event Calculus and we have presented some behavioural data supporting the theory. Recall that we chose to design the semantic theory as a computational analysis in Marr’s sense. We are now ready to move on to the intermediate level of analysis and consider the processing steps subserving the computation of discourse models. The processing model is based on the combination of the Event Calculus and Constraint Logic Programming used in van Lambalgen & Hamm (2004). The algorithms presented there specify in an abstract manner the computational steps involved in satisfying integrity constraints or, more precisely, they spell out the derivation (the proof) of a given statement within the theory. In order to derive predictions about the complexity and the time course of the relevant computations, it is necessary to add an explicit processing component to the formal machinery of the Event Calculus: the principle of immediacy.

324 Processing Consequences of the Imperfective Paradox concerning semantic composition. However, to be relevant for describing the construction of discourse models, the principle of immediacy must be thus reformulated: Proposition 2. The algorithms updating a minimal model so as to satisfy an integrity constraint are executed as soon as the integrity constraint is given as input.

4.2 The re-computation hypothesis Let us now consider the processing steps leading to the construction of discourse models for (S3) and (S4). The first update instruction, in the form of integrity constraint (9), is introduced by the progressive VP. Since at that stage the information provided by discourse amounts to the context (C) and the main (progressive) clause of (S3) or (S4), no obstacle for the writing activity is described. Proposition 2 implies that (9) is satisfied as soon as the progressive VP is processed and Proposition 1 guarantees that a minimal model is computed such that there is a time at which the goal state holds. The upshot is that, when an accomplishment in the progressive is processed, a minimal model is immediately computed in which the goal state holds. Considering the subordinate clause, we must distinguish two processing steps. The first is carried out by adding (10) to the scenario (8), updating the discourse model according to (11) and deriving the existence of a time later than the initiation of the writing activity at which writing was terminated. The second step computes the suppression of the goal state inference by deriving the further statement that there is no time at which a complete letter was obtained. These are two distinct operations: for activities that are terminated by a disabling condition, such as (S2), only the first step is carried out when the subordinate is processed because there is no goal state to be cancelled; for disabled accomplishments, such as (S4), both steps are implemented. Recall that the satisfaction of the goal state is derived in a minimal model of the progressive VP. As a consequence, the hypothesis licensed by the theory is that, when an accomplishment in the progressive is followed by a subordinate clause describing a disabling condition,

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

The proposed definition of immediacy is a general hypothesis about the time course of the construction of discourse representations: it states that minimal models are computed as soon as update instructions, in the form of integrity constraints, are fed into the system by relevant lexical and morphosyntactic material. Propositions 1 and 2 form the core of our processing model.

Giosue` Baggio and Michiel van Lambalgen 325

4.3 Re-computation and perceptron learning The reader might wonder why computing a model that is incompatible with a previous structure (re-computation) would be different from, and in particular more complex than, computing a minimal model that monotonically extends the initial one. The answer can be found in the behaviour of the neural networks that compute minimal models. Due to the syntactic restrictions inherent in Logic Programming (Doets 1994), the models characterized by the Event Calculus can be viewed as stable states of the associated neural networks. It has been demonstrated that recurrent neural networks are sufficient and suitable for computing minimal models for propositional logic programs (Stenning & van Lambalgen 2005, 2007). It has also been shown that, for any propositional logic program, there exists a three-layer feedforward network of binary threshold units that computes the semantic operators on which the construction of minimal models is based (Hitzler et al. 2004). The language underlying the Event Calculus is not propositional but is a many-sorted predicate logic (see section 2.1), with matters being complicated further by the use of integrity constraints. Recent research suggests, however, that recurrent networks

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

the initial minimal model is re-computed to the effect that, in the new discourse structure, there is no time at which the goal state holds. Re-computation is thus the main processing consequence of the imperfective paradox. Predictions concerning the complexity and the time course of semantic computations for sentences like (S1)–(S4) can be derived from our processing model. The theory implies that the subordinate clause in (S4) involves the re-computation of the minimal model computed while processing the progressive clause, while in (S3) the initial model is simply extended (we return to the difference between ‘re-computation’ and ‘extension’ in 4.3). Furthermore, the activity cases (S1) and (S2) will also involve a monotonic extension of the initial model, such that the termination of the writing activity is computed for (S2), but no cancellation of the goal state (since a canonical goal is not involved in activities). In short, the subordinate clause in (S4) will be more complex to process compared to (S1)–(S3), as no re-computation is triggered in the latter conditions. As regards time courses, the Event Calculus requires the causal and temporal information carried by the verb ‘morste’ to activate the additional scenario clause (10), satisfy (11) and derive the failed occurrence of the consequent state. Therefore, re-computation is expected to surface only at the final word in (S4).

326 Processing Consequences of the Imperfective Paradox

9 Multilayer feedforward networks (Hornik et al. 1989) and recurrent networks (Scha¨fer & Zimmermann 2006) can approximate any Borel-measurable function to any desired degree of accuracy. Therefore, despite earlier claims advanced by Levelt (1990), connectionist networks can approximate arbitrary recursive functions as well, suggesting that they are suited for modelling linguistic processes. The fact that their computational power is restricted to that of finite-state machines (McCulloch & Pitts 1943), given that the human brain is itself a finite-state machine, adds to their plausibility in psycholinguistics and cognitive science more generally (van der Velde 1993; Petersson 2005).

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

can also approximate the semantic operators for first-order logic programs and their fixed points (Hitzler et al. 2004). Moreover, as shown in Stenning & van Lambalgen (2007), integrity constraints can be modelled via a form of back-propagation called ‘perceptron learning’ (Rosenblatt 1962; Rojas 1996).9 In this framework, any computation on a given minimal model, such as adding a logic program clause (a scenario clause), will somehow bring the network from its initial stable state to another stable state, corresponding to the new minimal model. Nevertheless, if the neural representation proposed in Stenning & van Lambalgen (2007) is approximately correct, there is a large difference in neural activity between, on the one hand, a monotonic extension of a minimal model, and a non-monotonic re-computation of a minimal model on the other. Consider first the case of a monotonic extension as envisaged by our processing model. Retrieving a clause such as (10) from semantic memory, assuming that spill denotes ‘spilling coffee on the tablecloth’, will result in the activation of a number of units (neurons) which were previously silent; but the activation state of the remaining units, including those representing the goal state (the complete letter), will remain the same. However, retrieving a different clause from semantic memory, for instance (10), where spill now denotes ‘spilling coffee on the paper’, will result not only in the activation of neurons which were silent but also in the readjustment of the activation patterns of units which were previously active. For instance, the neurons representing the consequent state (the complete letter) will no longer be active. This change in activation is achieved in the neural network by applications of perceptron learning. The difference between monotonic extension and non-monotonic re-computation can thus be found in the occurrence of the readjustment of connection strengths driven by perceptron learning. These considerations imply that non-monotonic re-computation in (S4) has more drastic consequences for neural processing as compared to the monotonic extension of a minimal model in (S3). We must now show that a processing model based on monotonic semantics does not predict a similar effect. In a strictly monotonic progression of structures, the goal state is necessarily not represented, for otherwise,

Giosue` Baggio and Michiel van Lambalgen 327

10 Formally, one never computes the completion of the scenario plus the EC axioms. For the notion of ‘completion’ see van Lambalgen & Hamm (2004).

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

the model may have to be re-computed, which is not allowed by the monotonic logic underlying the theory. In the Event Calculus framework this means that the predicates and axioms for continuous change are not used, at most those for instantaneous change. Furthermore, because the progression of models is monotonic, one never actually computes minimal models.10 Under the assumption of monotonicity little semantic computation is going on. In particular (S3) and (S4) will both lead to monotonic extensions of the initial model and both extensions will be computed at the same time, that is when the final verb is processed: in (S3) the update will lead to a model in which (given the results of our entailment questionnaire) the consequent state is represented as being attained, in (S4) it will lead to a model in which the consequent state does not hold. Because there is no principled semantic or processing reason to assume that one final model would be more complex to compute than the other, it follows that a strictly monotonic semantics predicts no difference between the conditions. The re-computation hypothesis can be tested in a dedicated electrophysiological study. Event-related potentials (derived from electroencephalography (EEG) signals) or fields (derived from magnetoencephalography signals) and power changes in specific frequency bands provide direct insights into the complexity and the time course of neural processing, as opposed to reading times and eye movements, which are indirect and cumulative measures of processing load (Luck 2005). An EEG study using the materials of our questionnaire as stimuli would be able to determine whether the event-related potentials (ERP) correlate of model re-computation is the N400—the negative shift peaking around 400 ms after the onset of semantically anomalous words (Kutas & Hillyard 1980), words with lower cloze probabilities (Kutas & Hillyard 1984; Hagoort & Brown 1994) and words which provide information conflicting with the discourse context (van Berkum et al. 1999, 2003) or world knowledge (Hagoort et al. 2004). Since the computations underlying these phenomena pertain more to the domain of semantic composition than to the construction of discourse models, re-computation might evoke a different ERP effect. Preliminary data show that the final verb in (S4) elicits a sustained anterior negativity (SAN; thus different from the transient, centro-parietal N400) which is not observed in conditions (S1)–(S3). The amplitude of the anterior negative shift in (S4) is correlated with the number of negative responses in the probe selection

328 Processing Consequences of the Imperfective Paradox task, which is what should be expected if the SAN effect was a brain signature of re-computation: assuming that re-computation evokes a SAN in each (S4) trial in which a negative response is given, the larger the number of negative responses (the larger the number of trials in which re-computation took place), the larger the amplitude of the ERP component revealed by averaging. Current experimental work in our laboratory is aimed at corroborating these findings. Acknowledgements

GIOSUE` BAGGIO F.C. Donders Centre for Cognitive Neuroimaging Radboud University Nijmegen P.O. Box 9101 NL-6500 HB Nijmegen The Netherlands e-mail: [email protected] MICHIEL VAN LAMBALGEN Department of Philosophy University of Amsterdam Nieuwe Doelenstraat 15 NL-1012 CP Amsterdam The Netherlands e-mail: [email protected]

REFERENCES Andrade-Lotero, E. (2006), Meaning and Form in Event Calculus. Unpublished Master’s thesis, University of Amsterdam, The Netherlands. Baayen, R., R. Piepenbrock, & L. Gulikers (1996), CELEX2. Linguistic Data Consortium. Philadelphia. Clark, H. (1973), ‘The languageas-fixed-effect fallacy: a critique of language statistics in psychological

research’. Journal of Verbal Learning and Verbal Behavior 12:335–59. Cook, A., J. Limber, & E. O’Brien (2001), ‘Situation-based context and the availability of predictive inference’. Journal of Memory and Language 44:220–34. de Swart, H. (1998), ‘Aspect shift and coercion’. Natural Language and Linguistic Theory 16:347–85.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

This paper benefited from many discussions with Travis Choma, Bart Geurts, Peter Hagoort and Karl Magnus Petersson. We thank Ira Noveck and two anonymous reviewers for their comments. We are grateful to The Netherlands Organization for Scientific Research for support under grant NWOCOG/04-19.

Giosue` Baggio and Michiel van Lambalgen 329 Hitzler, P., S. Ho¨lldobler, & A. Seda (2004), ‘Logic programs and connectionist networks’. Journal of Applied Logic 3:245–72. Hornik, K., M. Stinchcombe, & H. White (1989), ‘Multilayer feedforward networks are universal approximators’. Neural Networks 2:359–66. Jackendoff, R. (2002), Foundations of Language: Brain, Meaning, Grammar, Evolution. Oxford University Press. Oxford. Kamp, H. (1981), ‘A theory of truth and semantic representation’. In J. Groenendijk, T. Janssen, and M. Stokhof (eds.), Formal Methods in the Study of Language. Amsterdam Mathematical Centre. Amsterdam. Kutas, M., & S. Hillyard (1980), ‘Reading senseless sentences: brain potentials reflect semantic incongruity’. Science 207:203–5. Kutas, M., & S. Hillyard (1984), ‘Brain potentials during reading reflect word expectancy and semantic association’. Nature 307:161–3. Landman, F. (1992), ‘The progressive’. Natural Language Semantics 1:1–32. Levelt, W. (1990), ‘Are multilayer feedforward networks effectively turing machines?’ Psychological Research 52:153–7. Luck, S. (2005), An Introduction to the Event-Related Potential Technique. MIT Press. Cambridge, MA. Marr, D. (1982), Vision: A Computational Investigation into the Human Representation and Processing of Visual Information. Freeman. San Francisco. Marr, D., & T. Poggio (1976), ‘From understanding computation to understanding neural circuitry’. MIT AI Memos 357:1–22. Marslen-Wilson, W. (1989), ‘Access and integration: projecting sound onto meaning’. In W. Marslen-Wilson (ed.), Lexical Representation and

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

Doets, K. (1994), From Logic to Logic Programming. MIT Press. Cambridge, MA. Dowty, D. (1979), Word Meaning and Montague Grammar. Reidel. Dordrecht. Frank, S., M. Koppen, L. Noordman, & W. Vonk (2003), ‘Modeling knowledgebased inferences in story comprehension’. Cognitive Science 27:875–910. Frazier, L. (1987), ‘Sentence processing: a tutorial review’. In M. Coltheart (ed.), Attention and Performance, Volume XII. Erlbaum Press. Hillsdale, 559–586. Friederici, A. (2002), ‘Towards a neural basis of auditory sentence processing’. Trends in Cognitive Sciences 6:78–84. Geurts, B., & F. van der Slik (2005), ‘Monotonicity and processing load’. Journal of Semantics 22:97–117. Goldberg, A. (1995), Constructions: A Construction Grammar Approach to Argument Structure. University of Chicago Press. Chicago. Hagoort, P. (2006), ‘The Memory, Unification, and Control (MUC) model of language’. In A. Meyer, L. Wheeldon, and A. Krott (eds.), Automaticity and Control in Language Processing. Psychology Press. Philaldelphia, 243–270. Hagoort, P., & C. Brown (1994), ‘Brain responses to lexical ambiguity resolution and parsing’. In C. Clifton, L. Frazier, and K. Rayner (eds.), Perspectives on Sentence Processing. Lawrence Erlbaum. Hillsdale, 45–81. Hagoort, P., L. Hald, M. Bastiaansen, & K. Petersson (2004), ‘Integration of word meaning and world knowledge in language comprehension’. Science 304:438–41. Hamm, F., & M. van Lambalgen (2003), ‘Event calculus, nominalisation and the progressive’. Linguistics and Philosophy 26:381–458.

330 Processing Consequences of the Imperfective Paradox Schank, R., & R. Abelson (1977), Scripts, Plans, Goals and Understanding. Lawrence Erlbaum. Hillsdale. Stenning, K., & M. van Lambalgen (2005), ‘Semantic interpretation as computation in nonmonotonic logic: the real meaning of the suppression task’. Cognitive Science 29: 919–60. Stenning, K., & M. van Lambalgen (2007), Human Reasoning and Cognitive Science. MIT Press. Cambridge, MA. van Berkum, J., P. Hagoort, & C. Brown (1999), ‘Semantic integration in sentences and discourse: evidence from the N400’. Journal of Cognitive Neuroscience 11:657–71. van Berkum, J., P. Zwitserlood, P. Hagoort, & C. Brown (2003), ‘When and how do listeners relate a sentence to the wider discourse? Evidence from the N400 effect’. Cognitive Brain Research 17:701–18. van der Velde, F. (1993), ‘Is the brain an effective turing machine or a finitestate machine?’ Psychological Research 55:71–9. van Dijk, T. (1980), Macrostructures. Lawrence Erlbaum. Hillsdale. van Lambalgen, M., & F. Hamm (2004), The Proper Treatment of Events. Blackwell. Oxford. Vendler, Z. (1957), ‘Verbs and times’. The Philosophical Review 66: 143–60. Wike, E., & J. Church (1976), ‘Comments on Clark’s ‘‘The language-asfixed-effect fallacy’’’. Journal of Verbal Learning and Verbal Behavior 14: 249–55. First version received: 29.08.2006 Second version received: 20.12.2006 Accepted: 08.01.2007

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

Process. The MIT Press. Cambridge, MA, 3–24. McCulloch, W., & E. Pitts (1943), ‘A logical calculus of the ideas immanent in nervous activity’. Bulletin of Mathematical Biophysics 5:115–33. McMillan, C., R. Clark, P. Moore, C. Devita, & M. Grossman (2005), ‘Neural basis of generalized quantifier comprehension’. Neuropsychologia 43:1729–37. McMillan, C., R. Clark, P. Moore, & M. Grossman (2006), ‘Quantifier comprehension in corticobasal degeneration’. Brain and Cognition 62:250–60. Michaelis, L. (2001), A Unification-Based Model of Aspectual Type-Shifting. University of Colorado at Boulder. Minsky, M. (1977), ‘Frame-system theory’. In P. Johnson-Laird and P. Watson (eds.), Thinking. Cambridge University Press. Cambridge, 104–116. Parsons, T. (1990), Events in the Semantics of English. MIT Press. Cambridge, MA. Petersson, K. (2005), ‘On the relevance of the neurobiological analogue of the finite-state architecture’. Neurocomputing 65-66:825–32. Reichenbach, H. (1947), Elements of Symbolic Logic. Macmillan. New York. Rojas, R. (1996), Neural Networks: A Systematic Introduction. Springer. Berlin, New-York. Rosenblatt, F. (1962), Principles of Neurodynamics: Perceptrons and the Theory of Brain Mechanisms. Spartan Books. New York. Rothstein, S. (2004), Structuring Events. Blackwell. Oxford. Scha¨fer, A., & H. Zimmermann (2006), ‘Recurrent neural networks are universal approximators’. In S. Kollias, A. Stafylopatis, W. Duch, and E. Oja (eds.), Artificial Neural Networks— ICANN 2006. Springer. Berlin, New York, 632–640.

Journal of Semantics 24: 331–343 doi:10.1093/jos/ffm012 Advance Access publication October 9, 2007

The Psychological Reality of Classical Quantifier Entailment Properties GUY POLITZER Institut Jean Nicod–CNRS

A test of directional entailment properties of classical quantifiers defined by the theory of generalized quantifiers (Barwise & Cooper 1981) is described. Participants had to solve a task which consisted of four kinds of inference. In the first one, the premise was of the form ‘Q–hyponym–verb–blank predicate’, where Q is a classical quantifier (e.g. ‘Some cats are [ ]’), and the question was to indicate what, if anything, can be concluded by filling the slots in ‘. . .–hyperonym–verb–blank predicate’ (e.g. ‘. . . animals are [ ]’). The second kind of inference was the same, except that the hyperonym was in the premise and the hyponym in the conclusion. The third and fourth kinds of inference differed from the first two by the position of the hyperonym (respectively hyponym) which occupied the place of the predicate (e.g. ‘some [ ] are animals’). It was observed that if the directional entailment holds people respond accordingly in most cases and that if the entailment does not hold they correctly fail to produce it. These results provide elementary, but essential empirical support to this semantic approach to quantification, and are a prerequisite for its application to the study of reasoning with quantifiers. The implications for the psychology of reasoning are discussed.

1 INTRODUCTION Generalized quantification theory (Barwise & Cooper 1981) assumes that a quantifier is a set of subsets of a given domain of individuals. This approach has offered a number of important concepts to semantic theory. One of them is directional entailingness (also called monotonicity). Informally, this concept captures the intuition that, for example, all dogs entails all spaniels, or that some spaniels entails some dogs. Formally, a quantifier QD is downward entailing if, for all domains D: QDY and X 4 Y 4 D entails QDX. This is instantiated by the former example. Similarly, a quantifier QD is upward entailing if, for all domains D: QDX and X 4 Y 4 D entails QDY (this is instantiated by the latter example). Directional entailingness can also be defined for binary relations between subsets (see Westersta˚hl 2001). QDXY (with X subject and Y predicate) Ó The Author 2007. Published by Oxford University Press. All rights reserved. For Permissions, please email: [email protected].

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

Abstract

332 Psychological Reality of Classical Quantifier Entailment Properties is left downward entailing if, for all D: QDXY and X# 4 X 4 D entails QDX#Y (for instance ‘All dogs bark’ entails ‘All spaniels bark’); it is right downward entailing if, for all D: QDXY and Y# 4 Y 4 D entails QDXY# (‘No amphibians are dogs’ entails ‘No amphibians are spaniels’). Analogous definitions obtain for left and right upward entailment. Viewed in these terms, the four classical quantified sentences have the following entailment characteristics (using the standard abbreviations A, I, E and O): 

 

Over the last two decades, the concept of directional entailingness has played an important role in semantic theory, for example in accounts of the distributional properties of negative polarity items (Ladusaw 1980) and the discussion of scalar implicature computation (these implications are suspended in downward-entailing environments: Horn 1989; Chierchia 2004). From the viewpoint of generalized quantifiers in semantic theory, an individual who knows the meaning of a quantifier must have the ability to carry out the directional entailments. If people are indeed proficient at this, then this skill might go some way towards explaining people’s reasoning with quantified sentences. Hence, Geurts (2003) proposed that these entailment properties are highly relevant to reasoning, and in particular to syllogistic reasoning. But psychologists unfamiliar with this conceptual framework may be puzzled, if not sceptical, with regard to the psychological reality of entailment relations that stem from work in philosophical logic and have been endorsed only on the basis of linguistic intuitions. Moreover, the notion that reasoning might be taken care of by formal properties of the semantic component of language rather than relying on some logical calculus (rule-based or model-based) is totally alien to most psychologists working in the field.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011



A (‘All X are Y’): left downward, for example, ‘All antiques are vases’ entails ‘All antiques for sale are vases’, and right upward, for example, ‘All antiques are Ming vases’ entails ‘All antiques are vases’. I (‘Some X are Y’): left upward, for example, ‘Some antiques for sale are Ming vases’ entails ‘Some antiques are vases’, and right upward (like all). E (‘No X are Y’): left downward (like all) and right downward, for example, ‘No antiques are vases’ entails ‘No antiques are Ming vases’. O (‘Some X are not Y’): left upward (like some) and right downward (like no).

Guy Politzer 333

2 EXPERIMENT

2.1 Materials and design The experiment was administered in written form. Four types of onepremise arguments were prepared. They will be described by way of an example, where the relational term are is used, and in which Q stands for one of the four classical quantifiers. Type 1.

Premise: Conclusion:

Q animals are [ ] . . . cats . . . [ ]

This type is defined by the following two characteristics: (i) the ‘subject’ (i.e. the restriction) is provided, but the predicate (here in square brackets) which is common to both sentences is left open (we will see in a minute how this was done) and (ii) the conclusion’s subject is a hyponym of the premise’s subject. The two dotted slots were prepared to accommodate the participant’s response, namely a quantifier and a verbal expression (is/is not/are/are not), respectively. An example of expected correct inference, with the quantifier all, would be Premise: Conclusion:

all animals are [ ] all cats are [ ]

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

A first step towards getting entailment properties on the psychologists’ agenda would be to ascertain that people do master these. So far, linguists have not felt it necessary to experimentally investigate the extent to which people do so, presumably because the existence of entailment properties of quantifiers is not in dispute. Geurts (2003, Geurts & van der Slik 2005) did carry out experiments on quantifier entailment but the aim of these investigations was more elaborate. In the former case, he tested an hypothesis about the relative complexity of various quantifiers, and in the latter he used fairly complex sentences containing two quantifiers and limited to right entailment. In brief, there is still need for an experimental demonstration that people master the basic entailment arguments described earlier, which is an elementary assumption to make if generalized quantification theory provides an adequate framework for the representation of, and reasoning with, quantifiers. The work that will be reported is but a first step limited to the quantifiers that have been studied most by psychologists, namely the classical Aristotelian quantifiers; there is, however, no reason why the method used could not be applied to other quantifiers, especially to those whose entailment properties are intuitively less clear.

334 Psychological Reality of Classical Quantifier Entailment Properties This is a test for left downward entailment. Type 2.

Premise: Conclusion:

Q [ ] are animals . . . [ ] . . . cats

In this type, (i) the common subject in square brackets is left unspecified and (ii) the conclusion’s predicate is a hyponym of the premise’s predicate. The dotted slots serve the same purpose as previously. This type tests for right downward entailment. Type 3.

Q cats are [ ] . . . animals . . . [ ]

This is a variant of type 1: It differs only in the hyponymy relation which now is reversed and serves to test for left upward entailment. Type 4.

Premise: Conclusion:

Q [ ] are cats . . . [ ] . . . animals

Similarly, type 4 differs from type 2 by the reversal of the hyponymy relation and serves to test for right upward entailment. The four types appeared with all four quantifiers each, hence 16 different arguments, which will be designated by their traditional letter (A, I, E or O) and a number referring to the type. The example given above for type 1 is an A1 argument. Out of the 16 arguments, eight are entailing and the other eight are non-entailing. The present study focused on people’s recognition of quantifiers’ entailment properties, so it is the first eight arguments that specifically address this question. Their eight non-entailing counterparts (A3 for A1, E3 for E1, etc.) were also considered because they provide complementary evidence in the form of a control: If people produce a conclusion for, for example, an A1 argument showing that they recognize a left downward entailment from ‘All animals are [ ]’ to ‘All cats are [ ]’, then they should not produce a conclusion for the corresponding A3 argument from ‘All cats are [ ]’ to ‘All animals are [ ]’, because this would entail that the same argument position is upward as well as downward entailing. Unspecified subjects and predicates were used to make the task as context neutral as possible. This was to avoid that the premises be associated, or restricted, to specific categories or attributes that could affect participants’ inferences for emotional or cognitive reasons, and so to enable one to generalize the results safely. To make the task natural,

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

Premise: Conclusion:

Guy Politzer 335

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

the predicates in argument types 1 and 3 were not presented as empty spaces in brackets as above, but were occupied by a pronoun. As the experiment was administered in French, it was easy to take advantage of the very natural way that this language refers to an attribute already present in the context by use of a pronoun (e.g. ‘All cats are [ ]’ is rendered by ‘Quelques chats le sont’, where the neutral pronoun le occupies a position that is empty in English). Participants were told that the pronoun referred to some unspecified attribute which they need not be concerned about. Similarly, and even more straightforwardly, for the eight argument types 2 and 4, it was said in the instructions that the premises referred to some set of objects (the nature of which, again, participants need not be concerned about). A premise such as ‘All of them are cats’ was naturally formulated as ‘All are cats’ (French ‘Tous sont des chats’) with no need for a pronoun (them). As performance could be influenced by the relational term that was used (are), the arguments were also formulated with have, so that the previous examples in this second version became ‘All cats have [ ]’, ‘All [ ] have cats’, and so on. Again there was no specific predicate to occupy the empty slot as it is natural in French to use the pronoun en (e.g. ‘Tous les chats en ont’, ‘All cats have’). Finally, French has a very common way of expressing the particular quantifier, namely il y a. (In English, its closest equivalent there is/are occurs less frequently and cannot be used to paraphrase some sentences as routinely as il y a for quelques sentences in French.) This provided an additional condition to generalize the results beyond the case of one single way to express particular quantification. Four semantic domains were used: animals, vegetables, flowers and fruit, with four different and very common hyponyms in each case (e.g. cat, dog, horse, goat), hence 16 pairs of hyponyms–hyperonyms; each pair was randomly allocated to one argument, with the constraint that the four semantic domains should appear in each argument type. A sample of four items is given in the Appendix. The arguments were presented in booklets prefaced with the instructions. The eight arguments of types 1 and 3 constituted one block, as did the eight arguments of types 2 and 4. Within a block, the arguments were rotated with the constraint that no two identical quantifiers could be adjacent, and that the types should alternate. One block was presented first to half of the participants and the other was presented first to the other half. In brief, the design was a 2 (verbal form of relation: are v. have) 3 2 (formulation of the particular quantifier) 3 2 (blocks) 3 16 (argument type) factorial design, with repeated measures on the last two factors.

336 Psychological Reality of Classical Quantifier Entailment Properties

2.2 Participants and procedure

2.3 Results There was no effect of order and no effect of the formulation of the quantifier. There was a significant effect of the are v. have factor for only one of the 16 argument forms, to be discussed later. In brief, none of the three binary factors had an effect that was both systematic and significant, so the results were pooled across these factors. For any of the arguments, an entailment judgment was attributed to participants if the quantifier they entered in the conclusion reproduced that in the premise (and similarly for the relational term). Non-entailment was operationally defined by any answer showing the absence of the reproduction of the premise quantifier, in line with the logical definition of entailingness. Take for example ‘All flowers are [ ] therefore . . . roses are [ ]’. Participants were credited with a left downward-entailing answer if they filled in the dotted space with all. They were considered as giving a non-entailing answer in the other cases, that is, if (i) they indicated that nothing followed or (ii) gave a different quantifier, whether logically incorrect (e.g. no) or correct (e.g. some). Of course in the latter case, giving a correct conclusion is an indication that the participant masters at least some of the semantic content of the quantifier in question; in addition, even though such an answer is preferred to the entailing answer, it is not incompatible with

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

The participants were 89 second-year psychology students in the University of Paris8. All were native speakers of French and untutored in logic. The experiment was administered during classes. Participants were instructed to assume that the premises were true; they were then asked to enter into blank spaces (i) a quantifier that would be guaranteed to make the conclusion true; (ii) the appropriate verbal expression (are/are not/is/is not or the analog with have, depending on the experimental condition). They were told that it was not always possible to find a conclusion that was surely true; whenever they thought this applied, they had to place crosses in the empty spaces of the conclusion. In all cases, they were asked to indicate how confident they were about their answers on a four-point scale ranging from 0 (not sure at all) to 3 (very sure). A high level of certainty should accompany the identification or the application of a formal argument, whereas a lower level of certainty should be observed when the response follows from the absence of such an identification. Previous research on deduction with quantified sentences has established the reliability and validity of this measurement (Politzer 1991).

Guy Politzer 337

Response Entailing* Non-entailing Mean certainty

YA1 88.8 11.2 2.68

YE1 86.4 13.6 2.63

E2Y 83.1 16.9 2.66

O2Y 43.7 56.3 2.32

[I3 92.2 7.9 2.40

[O3 81.8 18.2 2.28

A4[ 93.3 6.7 2.65

I4[ 71.9 28.1 2.50

Response Entailing Non-entailing* Mean certainty

A3 8.0 92.0 2.32

E3 11.2 88.8 2.21

E4 16.9 83.1 1.99

O4 35.6 64.4 1.87

I1 33.7 66.3 1.88

O1 28.1 71.9 1.94

A2 9.0 91.0 2.14

I2 47.2 52.8 2.09

Table 1 Percentages of responses and mean certainty ratings associated with the correct response for the 16 argument forms *Correct responses. (Entailing quantifiers are marked with an arrow indicating side and direction)

1 The expression ‘non-entailing’ is used to qualify one member of a pair of arguments such as A1/A3; this would be improper if used to qualify the associated quantifier, which is always entailing.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

its recognition. However, because a conservative approach was adopted, these answers were discounted. In fact, they turned out to be very rare (13 occurrences for the four arguments where they could occur, that is less than 4% of the answers). Table 1 displays for each of the 16 argument forms the percentages of entailing answers and (by complementarity) of non-entailing answers. The pattern of results is very clear. First, for the eight entailing cases, all except O2 were identified as such. The comparison of the percentages for and against entailment in each case shows the magnitude of the preferences: The weakest significant majority was 72% (for I4) and the strongest 93% (for A4). These differences were all significant at the 0.01 level (chi-square test). The O2 case did not elicit any significant majority and constitutes an exception. Pooling the eight arguments together, the mean frequency of entailing responses was 80% (86% excluding the O2 case). A similar pattern was obtained for the eight non-entailing arguments.1 All except I2 elicited a strong majority against entailment (weakest: 64% for O4, strongest: 92% for A3) and the associated differences were all significant (at the 0.05 level, most being well beyond this level). The mean frequency of non-entailment responses for these arguments was 76% (80% excluding I2). In brief, in 14 out of 16 arguments, participants were strongly sensitive to the presence or to the absence of the entailment property of the four classical quantifiers, while two arguments, both of type 2, and both particular, seem to have remained opaque to participants.

338 Psychological Reality of Classical Quantifier Entailment Properties 3 DISCUSSION

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

With two exceptions (O2 and I2, which will be considered later), participants’ pattern of inferences conformed to the expected pattern determined by directional entailments, in strong support of the hypothesis. Interestingly, the analysis of the certainty ratings shows a pattern in perfect agreement with a well-known phenomenon in the psychology of reasoning, namely that individuals are better at identifying valid than invalid arguments. For the entailment answers to entailing arguments, the certainty ratings ranged from 2.28 (for O3) to 2.68 (for A1) with a mean equal to 2.52, whereas for the nonentailment answers to non-entailing arguments they ranged from 1.87 (for O4) to 2.32 (for A3) with a mean equal to 2.06. In brief, all but one of the entailing inferences were performed with greater confidence than the non-entailing ones; in addition, the difference between the means is considerable, as it is almost half a unit on a three-unit scale. To the extent that a high certainty of the response correlates with the operation of a rule or to the identification of a property, this result adds support to the notion that in solving the arguments participants demonstrated that they recognized the entailment properties of the quantifiers. The first exception to the overall pattern is the I2 argument; this should not be viewed as problematic for two reasons. One, although it indicates that participants who recognized its non-entailing property failed to constitute a significant majority, its counterpart I4 indicates that the majority correctly recognized the upward entailment. Now, it is not too surprising that the relatively poor performance on the valid argument I4 (actually the poorest apart from O3) is accompanied with an even poorer performance on the associated invalid I2 argument for reasons that have just been mentioned. Two, the present result for the I2–I4 pair may be not entirely robust as it was confirmed in only one of two unpublished replication studies run by the author; interestingly, the failure to replicate concerns a sample of high school students with a firmer background in mathematics, suggesting (tentatively) a greater difficulty of this pair, although insufficient to affect participants with higher logical abilities. In terms of difficulty, it is noteworthy that performance was, by far, worse for particular than for universal quantifiers, both in terms of certainty ratings and of entailing and nonentailing correct answers: universals were the four most often correctly identified as non-entailing, and they ranked first, third, fourth and fifth among the most often correctly identified as entailing. This result may be understood in set-theoretical terms as reflecting that the universal

Guy Politzer 339

2 This can be illustrated with numerical trees which show that all and no occupy just one edge of the tree, whereas their counterparts occupy the whole complement of the domain. 3 Left entailment is a property attached to the subject, which makes a partition between universal and particular sentences (odd arguments), whereas right entailment is a property attached to the predicate, which makes a partition between affirmative and negative sentences (even arguments).

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

quantifiers all and no are simpler than their particular counterparts:2 the greater formal complexity results in a greater processing load. The other exception, the O2 argument, is also a particular argument. It contradicts the hypothesis more directly as it is entailing; in other words, the right downward entailment was not recognized for the O sentence. If we notice that the associated O4 argument showed only a weak majority against upward entailment, it would seem that a number of individuals were confused with regard to the direction (upward or downward) of right entailment for this specific quantifier. This conclusion is reinforced by the results of the two replication studies, both of which yielded percentages similar to the present experiment. The result for O2 supports Geurts’s claim that negation adds cognitive load to the processing of particular quantifiers (which has been documented for a long time, see, e.g. Anderson 1981). That no similar effect was observed for the other tests of the some . . . not sentence, namely the O1 and O3 arguments, which concern left (subject) entailment, is consistent with the negation hypothesis, as negation in two-place relational quantifiers is linked to the predicate, not to the subject.3 Also, the fact that the certainty ratings for O2 and O4 are among the lowest of the entailing and the non-entailing arguments, respectively, attests to the difficulty of this quantifier, while showing the coherence of the results. However, it could be objected that if the origin of the difficulty of O2 were due to negation, a similar difficulty for the other right-entailing negative quantifier, namely no tested by E2 should have been observed. But this objection is not compelling because considering the greater ease of processing of universals mentioned above, an interaction between quantity and quality can be expected; that is, it is reasonable to assume that the additional processing load due to negation has a compounding effect for particulars but not for universals. It was mentioned earlier that, of the 16 arguments, there was only one significant difference between the are and the have formulations of the sentences. Interestingly, this concerns the O2 argument, for which the rate of correct responses was about twice as high with have as it was with are. That failure on the O2 argument occurred with one verb but not with the other, while the same manipulation did not affect the other arguments, speaks in favour of the idea that this argument has some peculiarities. One could think of the effect of a scalar implicature, for

340 Psychological Reality of Classical Quantifier Entailment Properties

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

example. What such an effect could consist of has to be worked out on a case-by-case basis for each particular premise. It is easy to see that, should a scalar implicature be triggered, this could have no effect for six of the arguments, including O2. For O2, given, for example, ‘Some [ ] are not animals’, one can expect that participants who rightly infer ‘Some [ ] are not cats’ will carry out the same inference in case they add the scalar implicature ‘Some [ ] are animals’ to the premise, because this implicated premise does not interfere with the entailment property of the explicit premise. In brief, a scalar implicature cannot justify the failure to produce the right response for O2. It can be verified that this applies also, mutatis mutandis, to O3, I3 and I4 (which are entailing) and to O1 and I1 (which are not). There are, however, two non-entailing cases, I2 and O4, that in principle could be affected by an implicated premise. Consider I2: participants who correctly fail to draw the conclusion ‘Some [ ] are cats’ from ‘Some [ ] are animals’ are justified to conclude ‘Some [ ] are not cats’ if they have added the implicated premise ‘Some [ ] are not animals’, because this turns I2 into the downward-entailing argument O2. The same applies to the non-entailing O4 whose implicated premise turns it into the upward-entailing argument I4. Now, the raw data indicate that such responses are very rare (6% of the responses to I2 and O4), from which it can be concluded that implicatures do not affect the performance in the present task; this rules out, in particular, a possible explanation for the deviant performance on O2. One cannot underestimate the theoretical importance of the entailments investigated here, even if they might appear easy enough to carry out. In fact, that these inferences seem to be so easy is part of the explanandum: the fact that specific inferences are executed so readily highlights that they are critical for people’s understanding of quantifiers as predicted by the theory. Also, the entailingness task was designed as a precise test and, in view of the number of argument forms tested, a severe test of generalized quantification theory. The test can also be considered as severe from the point of view of the task demands, as the format of response required that participants produce their own conclusion rather than evaluate a correct conclusion offered to them, which might have been too transparent. That the theory has stood up well (and independently of any expectations one might have about the results of the test based on intuition) lends strong support to it. Of course, the conclusion that people recognize most entailment properties of the quantifiers considered here obviously holds to the extent that participants have been submitted to a valid test, that is, to a task containing arguments that faithfully operationalize the formal definition of quantifier entailment. It does not seem questionable that this has been

Guy Politzer 341

4

Of course, a theory based on the entailing properties of quantifiers faces the same challenge: some logical principle would have to be added to such a theory to account for the detailed responses offered by participants.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

done in the present experiment, as (i) the premises presented the quantified sentences of interest; (ii) the inclusion relations could not fail to be understood as such, as they were chosen from highly familiar categorizations and (iii) the conclusions disclosed directly whether or not the expected entailed quantified sentences were produced. This suggests that the task could be exploited to investigate non-classical quantifiers as well. Finally, the present argument forms may be described as syllogisms in disguise; indeed, one can associate a syllogism containing an implicit A minor premise with each argument form (e.g. EA1 to E1, etc.). Consequently, the claim that the performance on the task demonstrates people’s mastery of the entailment properties of classical quantifiers justifies another claim, namely that people can solve this kind of syllogism by whatever means they possess to solve syllogistic tasks in general. This objection must be considered seriously; but to substantiate it, supporters of the relevant theories would have to apply them to explain the present data, including data that have not been reported in section 2.3.4 This concerns a few striking differences in performance on several arguments when compared with performance on standard syllogisms. To take one example, the A3 argument yielded A conclusions in only 8% of the cases, whereas the usual rate of A conclusions for the AA3 syllogism is about 65%. More important, it seems that all theories of syllogism, whether they are psychological or were developed in classical logic, incorporate some principles that are indistinguishable from one or another entailment property. This claim cannot be developed in the present paper, but two examples will suffice: the dictum de omni et nullo, which is at the basis of the medieval theories of the syllogism, can be shown to be a mere consequence of entailment properties (Hoeksema 1986). Similarly, the concept of distribution used in the formulation of the classic laws of the syllogisms can be reformulated in terms of downward entailingness (Makinson 1969). The upshot of all this is that it is not unreasonable, based on considerations of parsimony and generality, to view the quantifier entailment properties as primitive and reasoning principles as derived. The present investigation concerns itself with classically quantified sentences. Quantifier-entailing properties are highly relevant to the psychology of reasoning. As Geurts’s (2003) study exemplifies, monotonicity is a plausible main component of a theory of reasoning with quantified sentences—more specifically of syllogistic reasoning.

342 Psychological Reality of Classical Quantifier Entailment Properties

APPENDIX Four test items with an entailing quantifier in each of the four types are given below in the original French version (with their English translation in the right column). YA1 Tous les animaux le sont. donc: chien(s) . . . dogs(s) . . . [O3 Quelques tulipes ne le sont pas. donc: fleur(s) . . . flower(s) . . . E2Y Aucun n’est un le´gume. donc: . . . poireau(x) . . . leek(s) I4[ Quelques-uns sont des oranges. donc: . . . fruits . . . fruit

All animals are. therefore:

Some tulips are not. therefore:

None are vegetables. therefore:

Some are oranges. therefore:

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

Now, the relevance of monotonicity for reasoning can be extended far beyond quantification, as the various syntactic categories of sentences have monotonicity properties (van Benthem 1986; Sanchez-Valencia 1995). That is, given two expressions of the same type such that [[A]] 4 [[B]], inference in natural language can be viewed as a matter of substituting A with B (or B with A) once the monotonicity of the syntactic category concerned is determined (which requires a procedure of polarity marking). There follows the idea of a general theory of deductive reasoning in natural language based on monotonicity properties of grammatical components of sentences. Because monotonicity properties are alien to the current theories of deductive reasoning, whether these are based on mental models (Johnson-Laird & Byrne 1999) or on mental rules (Rips 1994; Braine & O’Brien 1998) or on a more recently developed probabilistic point of view (Chater & Oaksford 1999; Oaksford & Chater 2007), for psychologists supporting any one of them, the grammatical approach may appear as a revolutionary perspective, certainly worth considering more closely.

Guy Politzer 343

Acknowledgements The author thanks Bart Geurts for his comments on a first draft of this paper.

GUY POLITZER Institut Jean Nicod–CNRS Ecole Normale Supe´rieure, 29 rue d’Ulm, Paris 75005 France. e-mail: [email protected]

Anderson, John R. (1981), ‘Memory for logical quantifiers’. Journal of Verbal Learning and Verbal Behavior 20: 306–21. Barwise, Jon & Robin Cooper (1981), ‘Generalized quantifiers and natural language’. Linguistics and Philosophy 4:159–219. Chater, N. & Mike Oaksford (1999), ‘The probability heuristics model of syllogistic reasoning’. Cognitive Psychology 38:191–258. Chierchia, Gennaro (2004), ‘Scalar implicatures, polarity phenomena, and the syntax/pragmatics interface’. In A. Belletti (ed.), The Cartography of Syntactic Structures. Vol 3: Structures and Beyond. Oxford University Press. Oxford. Geurts, Bart (2003), ‘Reasoning with quantifiers’. Cognition 86:223–51. Geurts, Bart & Frans van der Slik (2005), ‘Monotonicity and processing load’. Journal of Semantics 22: 97–117. Hoeksema, Jack (1986), ‘Monotonicity phenomena in natural language’. Linguistic Analysis 16:25–40. Horn, Laurence R. (1989), A Natural History of Negation. University of Chicago Press. Chicago, IL.

Johnson-Laird, Philip N. & Ruth M. J. Byrne (1991), Deduction. Lawrence Erlbaum. London. Ladusaw, William (1980), Polarity Sensitivity as Inherent Scope Relations. Garland Press. New York. Makinson, David (1969), ‘Remarks on the concept of distribution in traditional logic’. Nouˆs 3:103–8. Oaksford, Mike & Nick Chater (2007), Bayesian Rationality: The Probabilistic Approach to Human Reasoning. Oxford University Press. Oxford. Politzer, Guy (1991), ‘Immediate deduction between quantifiers’. In Kenneth J. Gilhooly, Mark T. G. Keane, Robert H. Logie, and George Erdos (eds.), Lines of Thinking, vol. 1. Wiley. London. 85–97. Sa´nchez-Valencia, Victor (1995), ‘Parsing-driven inference: natural logic’. Linguistic Analysis 25:258–85. van Benthem, Johan (1986), Essays in Logical Semantics. Reidel. Dordrecht, the Netherlands. Westersta˚hl, Dag (2001), ‘Quantifiers’. In Lou Goble (ed.), Philosophical Logic. Blackwell. Oxford. 437–60. Received: 01.08.2006 Accepted: 05.11.2006

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

REFERENCES

Journal of Semantics 24: 345–371 doi:10.1093/jos/ffm006 Advance Access publication June 16, 2007

Suppression in Metaphor Interpretation: Differences between Meaning Selection and Meaning Construction

Abstract Various accounts of metaphor interpretation propose that it involves constructing an ad hoc concept on the basis of the concept encoded by the metaphor vehicle (i.e. the expression used for conveying the metaphor). This paper discusses some of the differences between these theories and investigates their main empirical prediction: that metaphor interpretation involves enhancing properties of the metaphor vehicle that are relevant for interpretation, while suppressing those that are irrelevant. This hypothesis was tested in a cross-modal lexical priming study adapted from early studies on lexical ambiguity. The different patterns of suppression of irrelevant meanings observed in disambiguation studies and in the experiment on metaphor reported here are discussed in terms of differences between meaning selection and meaning construction.

1 CONCEPT CONSTRUCTION IN METAPHOR INTERPRETATION One of the current directions in metaphor theory is based on the assumption that metaphor comprehension involves both the use of previously acquired concepts and schemas and the creation of new ones (Cacciari & Glucksberg 1994). Glucksberg & Keysar (1990) were among the first researchers to develop this idea with their classinclusion model of metaphor interpretation, moving away from the traditional Aristotelian view that nominal metaphors of the form ‘X is a Y’ are understood as implicit comparisons (i.e. ‘X is like a Y’) (Bowdle & Gentner 2005). In their view, nominal metaphors are ‘exactly what they appear to be: class-inclusion assertions’ (Glucksberg & Keysar 1990: 3). According to the class-inclusion model, the content of a metaphor is assigned to a ‘diagnostic category’ that is labelled by the figurative expression or ‘metaphor vehicle’, whose literal denotation is also Ó The Author 2007. Published by Oxford University Press. All rights reserved. For Permissions, please email: [email protected].

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

PAULA RUBIO FERNA´NDEZ University College London

346 Suppression in Metaphor Interpretation a stereotypical subset of that category. Consider the following example from Glucksberg and Keysar (1990): (1) My job is a jail.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

Such a metaphorical statement would be a true class-inclusion assertion, where ‘jail’ not only has its usual denotation but also refers to a broader ad hoc category including both jails and the speaker’s job. Glucksberg et al. (1997a) refer to this aspect of metaphor interpretation as ‘dual reference’. Thus, ‘jail’, as the metaphor vehicle, would give a name to the new category while its salient properties would become prototypical of that category. Since ‘my job’, as the topic of the metaphor, is asserted to be a member of the ad hoc category, typical properties of jails, such as ‘being unpleasant’, ‘constraining and punishing’, would now be attributed to it. Given their experimental approach, Glucksberg and his colleagues make precise empirical predictions about the interpretation and processing of metaphors. For example, interpreting a nominal metaphor like (1) involves enhancing attributes of the vehicle ‘jail’ that are appropriate for the topic ‘my job’ (e.g. ‘unpleasant’, ‘constraining’) while suppressing those attributes that are inappropriate (e.g. ‘has bars on the windows’). This hypothesis has been tested in a number of experiments (see below for review). Another account of metaphor interpretation based on ad hoc concept construction is that of Relevance Theory (Carston 1996, 2002; Sperber & Wilson 1997, 2002; for an earlier relevance-theoretic account, entirely in terms of implicature, see Sperber & Wilson 1986/ 1995). In this theoretical framework, metaphor is seen as a type of ‘loose use of language’ comparable to various other phenomena usually discussed in other terms (e.g. approximations, category extensions, hyperboles and neologisms; see Wilson 2003 for examples and discussion). This unified account of loose use falls within the relevance-theoretic programme for ‘lexical pragmatics’, understood as the branch of linguistics that investigates the pragmatic processes by which the meaning communicated by the use of a word in context comes to differ from the linguistically encoded meaning of that word. Relevance Theory distinguishes two cases where the concept communicated by the use of a word differs from the concept encoded by that word: concept narrowing and concept loosening, both resulting from a general pragmatic process of ‘concept adjustment’ (Carston 2002; Wilson 2003). In instances of narrowing, a word is used to convey a more specific concept than the one it encodes, whereas in

Paula Rubio Ferna´ndez 347

instances of ‘loosening’, the concept communicated is more general than the lexical concept. The interpretation of most nominal metaphors involves a combination of narrowing and broadening (see Carston 1996, 2002 for a discussion of the various types of ad hoc concept constructed in metaphor interpretation). Consider the following example (adapted from Carston 2002): (2)

Caroline is a princess.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

Suppose that (2) was used to convey that Caroline is spoilt. In interpreting this nominal metaphor, an ad hoc concept PRINCESS* would be constructed, resulting from a process of loosening and narrowing the lexical concept PRINCESS. On the one hand, Caroline would be included in the extension of PRINCESS*, which means that the property ‘female royal of a certain sort’, which is definitional of the lexical concept PRINCESS, would not characterize all members of the new category given that Caroline (and other spoilt women included in the ad hoc category) is not a royal. On the other hand, the entities falling within the denotation of PRINCESS* might be characterized as having the property spoilt. Those princesses who do not behave in this way would not be included in the extension of the ad hoc concept. Therefore, in the process of fine-tuning the interpretation of the word ‘princess’, the extension of the communicated concept PRINCESS* would have been not only broadened but also narrowed in relation to that of the encoded concept. Recanati (2004) offers an account of metaphor interpretation that is similar to Carston’s. According to Recanati, a concept has certain ‘conditions of application’ associated with it. Concept narrowing would consist in restricting the reference of a concept by contextually providing further conditions that are not linguistically encoded. Concept loosening is understood as the converse of the process of narrowing: some condition of application packed into the concept literally expressed by a word is contextually dropped so that the extension of the concept is widened (Recanati 2004). Despite the clear similarities in their accounts, there are some major differences between Recanati’s and the relevance-theoretic view of metaphor interpretation, which will be discussed in section 2. Although both Relevance Theory and Recanati’s framework are open to empirical investigation (Recanati 2004; Wilson & Sperber 2004), they do not make specific empirical predictions about metaphor processing and interpretation. However, Carston (2002) and Recanati (2004) describe narrowing and loosening in terms of property promotion and demotion, which could be understood in terms of

348 Suppression in Metaphor Interpretation

2 FURTHER THEORETICAL ISSUES Although the models of metaphor interpretation discussed above share the view that metaphor comprehension involves enhancing relevant properties of the vehicle while suppressing irrelevant ones, their accounts are different in some fundamental respects. In particular, there are important differences between Recanati’s and the relevancetheoretic views of metaphor interpretation (for a discussion of the differences between the class-inclusion and the Relevance Theory models of metaphor interpretation, see Rubio 2005). Although these differences might not be open to direct empirical investigation, the results of the present study might shed some light on some of these issues.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

degree of activation of the properties in question (Rubio 2005). According to these authors, if an ad hoc concept results from a process of narrowing, some encyclopaedic property of the lexical concept is elevated to the status of a logical or content-constitutive component. In example (2) above, ‘princess’ would express PRINCESS*, an ad hoc concept resulting from strengthening the lexical concept PRINCESS by making the encyclopaedic property ‘spoilt’ content constitutive of the new concept (i.e. a princess who is not spoilt would not fall under the ad hoc category PRINCESS*). On the other hand, if an ad hoc concept results from loosening a lexical concept, one or more of the logical or definitional properties of the lexical concept are discarded (Carston 1996, 2002; Recanati 2004). In example (2), the contentconstitutive property ‘female royal of a certain sort’ would be dropped from the logical entry of the lexical concept PRINCESS in order to include Caroline, who is not a royal, under the ad hoc concept PRINCESS*. Similarly to the proposal made by Glucksberg and his colleagues, it is natural to suppose that those encyclopaedic properties of the lexical concept which are promoted to the status of content constitutive of the resulting ad hoc concept become and remain active during the metaphor interpretation process given their contextual relevance (Rubio 2005). Conversely, the activation of those logical properties that have been demoted in concept loosening may be suppressed during processing (Recanati 2004; Rubio 2005). The study reported in this paper was aimed at testing this twofold empirical prediction that is common to the class-inclusion model, the relevance-theoretic account and Recanati’s view of metaphor interpretation.

Paula Rubio Ferna´ndez 349

(3)

The ham sandwich has left without paying.

In Recanati’s view, the expression ‘the ham sandwich’ first receives its literal interpretation by activating the representation of HAM SANDWICH. Then activation spreads from the literal representation to other associated representations, such as that of HAM SANDWICH ORDERER. ‘All these representations activated by the description ‘‘the ham sandwich’’ contribute potential candidates for (. . .) going into the interpretation of the global utterance’ (Recanati 2004: 29). Although the literal representation of HAM SANDWICH might initially have been

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

According to Recanati (2004), narrowing and loosening are ‘primary pragmatic processes’. This type of pragmatic process is characterized by being prepropositional (i.e. they do not require that a proposition has been identified already), unconscious (i.e. normal interpreters are not aware of their local operation) and purely associative (i.e. they operate in a blind mechanical fashion that involves no inferential process on the part of the hearer) (Recanati 2004). Relevance Theory does not make a distinction between primary and secondary pragmatic processes, advocating a unified account of the pragmatic processes involved in interpretation (Carston forthcoming). In particular, according to Relevance Theory, all pragmatic processes are uniformly inferential rather than some being merely associative (i.e. the input and output of pragmatic computations are related as premise and conclusion in an argument), and all are constrained by considerations of relevance (i.e. they tend to maximize cognitive effects while minimizing processing effort). Unlike the relevance-theoretic account, Recanati’s model of lexical interpretation is based solely on accessibility: in processing a word, its literal interpretation is accessed first and triggers the activation of associatively related representations (Recanati 2004). Both the literal concept activated by the linguistic expression and some of the other representations activated by association are possible candidates for the concept that will be selected for interpretation. Although these associated candidates are generated via, hence after, the literal concept, they are all processed in parallel and compete for activation. The representation that is most active or accessible when the interpretation process stabilizes will be selected and undergo semantic composition with the other components of the utterance, while all other candidates for the meaning of the word are suppressed (Recanati 1995, 2004). Consider the following example of metonymy (discussed in Recanati 1995, 2004):

350 Suppression in Metaphor Interpretation

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

a more accessible candidate than HAM SANDWICH ORDERER receives increased activation further down the line because of its suitability as an argument of the predicate ‘leave without paying’, so it ends up being the most accessible of the candidates once the entire utterance has been processed. Therefore, the non-literal candidate that derived from the literal one will be retained as part of the proposition expressed by (3), and the literal interpretation is suppressed (Recanati 2004). Recanati’s model is based on the notion of ‘accessibility shifts’: although a given representation of the meaning of some linguistic expression may be the most accessible one at some point in processing (e.g. the literal interpretation at an early stage in processing), activation levels may change and another representation (e.g. the derived nonliteral interpretation) may become the most accessible one at some later point in processing. Recanati (1995, 2004) seems to distinguish two possible factors that might determine these accessibility shifts. The first one is the processing of further linguistic material, as in the example just discussed, where the initial accessibility ordering is reversed once the predicate expression has been processed. The second factor is world knowledge structures or ‘schemata’: according to Recanati, interpretation is generally driven by schemata, so utterance interpretation is to a large extent a top–down process driven by world knowledge (Recanati 2004). In contrast, in the relevance-theoretic account, speaker’s intentions and not only world knowledge structures are among the factors that can affect the accessibility of a word’s interpretation (Sperber & Wilson 2002). Finally, according to Recanati (2004), a global literal interpretation does not necessarily precede figurative interpretation (cf. Grice 1975, 1989). However, at the local lexical level, the literal interpretation of the constituents of a metaphorical expression is accessed before the figurative interpretation of the utterance is derived (cf. Sperber & Wilson 2002). Nevertheless, Recanati (2001) considers a case where the literal interpretation of a figurative expression may not necessarily be suppressed once the intended interpretation of the utterance has been accessed: when a certain threshold of awareness is reached in interpreting an expression metaphorically, it results from some dimensions of the literal meaning remaining active even if they have been filtered out of the interpretation. Notice, however, that even in this case, the figurative interpretation would still need to be the most accessible one in order to be selected. Although these two theoretical issues are probably not empirically testable in any direct way, the experimental study reported below will

Paula Rubio Ferna´ndez 351

give some hints as to whether a figurative lexical interpretation is derived associatively or inferentially and whether it is necessarily the most accessible interpretation once it is selected. 3 METAPHOR PROCESSING AND LEXICAL AMBIGUITY

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

The predictions of the class-inclusion model of metaphor interpretation were tested in a series of experiments (e.g. Gernsbacher et al. 1995; Glucksberg et al. 1997b, 2001; Gernsbacher et al. 2001). The initial hypothesis was that interpreting a metaphor (e.g. ‘That defence lawyer is a shark’) involves enhancing attributes of the vehicle that are appropriate for the topic (‘aggressive’, ‘vicious’, ‘tenacious’) while suppressing those attributes that are inappropriate (‘swims’, ‘has fins’, ‘lives in the ocean’). Glucksberg, Gernsbacher and their colleagues used a sentence verification task, where subjects had to verify a metaphorical statement and then evaluate a second assertion that could be related to either the appropriate aspects of the metaphor vehicle (‘Sharks are tenacious’) or the inappropriate ones (‘Sharks are good swimmers’). The pattern of results observed in these studies seemed to confirm the predictions made by Glucksberg and colleagues: after reading a metaphorical statement, subjects were faster to verify statements related to the metaphoric interpretation of the vehicle and slower to verify statements related to the literal interpretation of the vehicle, when compared to the time it took them to respond to the same statements after reading a control sentence. These data therefore suggest that interpreting a metaphor does involve enhancing attributes that are relevant to the metaphorical interpretation of the vehicle while suppressing those that are irrelevant. However, the experimental design used in these studies makes it difficult to evaluate the power of their results, especially concerning the question whether suppression reduces the activation of metaphorirrelevant properties below baseline: the type of control sentences used in these experiments included the metaphor vehicle, which was the last word in the sentence in both the critical and the control conditions (e.g. ‘That hammerhead is a shark’). The control sentences were therefore not properly unrelated to the target sentences (‘Sharks are tenacious’/‘Sharks are good swimmers’) so the degree of activation of the target properties (‘tenacious’/‘swims’) after processing the control sentences would not correspond to a ‘zero’ level of activation. Because suppression is defined with reference to the baseline condition, an underestimate of baseline performance would result in an overestimate

352 Suppression in Metaphor Interpretation

1

In the third experiment reported in Gernsbacher et al. (2001), they used properly unrelated control sentences. However, in this experiment, rather than being slower, participants were actually faster at verifying a metaphor-irrelevant statement after reading a metaphor than after reading the unrelated control. Gernsbacher and her colleagues explain this facilitation as a result of the imbalance in the repetition effect between the critical and the control conditions. It is unclear, however, to what extent the standardized scores that had to be computed in order to observe suppression of metaphorirrelevant information accounted exclusively for the uneven repetition of the metaphor vehicle across conditions and offered an accurate measure of suppression. Glucksberg et al. (2001) modified the original design of the study in order to avoid the repetition effect between prime and target sentences, although the control condition still included the metaphor vehicle and would therefore have primed the critical properties (see reference for details).

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

of suppression. It follows that this type of baseline measure would not allow for making an accurate distinction between low activation and below-baseline suppression (see Rubio 2005 for further discussion).1 Nonetheless, McGlone & Manfredi (2001) also obtained results supporting the predictions of the class-inclusion model in a similar study where they controlled for the repetition of the metaphor vehicle between prime and target sentences, but using a baseline condition that did not prime the relevant properties (see reference for details of the experimental design). Overall, the results of previous studies investigating the predictions of the class-inclusion model suggest that both enhancement and suppression are indeed involved in metaphor interpretation, although the type of controls used makes it difficult to evaluate the power of such mechanisms. The control condition in the present study included completely unrelated targets. However, rather than the sentence verification task used by Glucksberg, Gernsbacher and their colleagues, I used a cross-modal lexical priming paradigm, which allows making an on-line measure of property activation across time. This in turn would allow investigating at which point in processing suppression dampens the activation of irrelevant literal properties of the metaphor vehicle. This question is related to the role of conscious, attentional processes in metaphor processing. The priming paradigm used in this study was adapted from early studies of lexical ambiguity (Swinney 1979; Tanenhaus et al. 1979; Onifer & Swinney 1981). In the original experiments, participants were presented with sentences in the acoustic modality which included a homonym (e.g. ‘The man found several bugs in his room’). At the offset of the ambiguous prime, participants had to make a lexical decision on a visual target. Critical targets could be related to either of the two meanings of the homonym (e.g. ‘spy’ or ‘ant’ in the above example). Facilitation relative to an unrelated control was interpreted in terms of meaning activation. The results of these experiments showed an early activation of target words related to both meanings of the

Paula Rubio Ferna´ndez 353

2

Although the superordinate ‘cat’ is not strictly speaking a property of CHEETAH, I am taking the activation of this associate as indicative of the activation of ‘being a cat’, which is a contentconstitutive property of the prime concept.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

homonym, which was interpreted in terms of an automatic, exhaustive process of spreading activation of associates (Schvaneveldt & Meyer 1973). However, the activation of the contextually inappropriate meaning dropped as early as 200–300 ms from the offset of the ambiguous word. This pattern of results was interpreted as showing active suppression of the irrelevant reading of the ambiguity, given that passive decay should take considerably longer (Neely 1976; Tanenhaus et al. 1979). However, since controlled, attentional processes take 400–500 ms to operate (Posner & Snyder 1975; Neill et al. 1995), these authors argue that although the meaning selection process must be context-sensitive (unlike the early spreading activation phase), it operates in an almost automatic way (unlike later conscious processes) (see Shiffrin & Schneider 1977). This would explain why hearers are usually unaware of having encountered a homonym in a disambiguating context (Gernsbacher 1990). In the study reported in this paper, participants were presented with contexts biased in favour of metaphorical interpretations, for example ‘Nobody wanted to run against John at school. John was a cheetah.’ Critical targets for a lexical decision were either metaphor-inconsistent properties of the metaphor vehicle (‘cat’) or metaphor-relevant properties (‘fast’).2 As in the above studies, I took facilitation relative to an unrelated control as indicative of property activation. In order to investigate the time course of activation, targets were presented 0, 400 and 1000 ms from the offset of the metaphoric prime. Assuming that metaphor-relevant properties are enhanced in metaphor interpretation, whereas metaphor-inconsistent properties are suppressed (Glucksberg & Keysar 1990; Carston 2002; Recanati 2004), the question that needs to be addressed is at which point in processing the activation level of these two types of associates diverges. In particular, it would be interesting to see whether metaphor-inconsistent properties would be suppressed as early as the irrelevant meanings of homonyms (i.e. 200– 300 ms from the offset of the metaphor vehicle), or whether their suppression would involve later, more attentional processes on the part of the interpreter. Although suppression would be involved in both disambiguation and metaphor interpretation (Gernsbacher & Faust 1991), these linguistic phenomena involve different pragmatic processes: meaning selection and meaning construction, respectively (see Recanati 1995;

354 Suppression in Metaphor Interpretation

4 AN ON-LINE STUDY OF METAPHOR PROCESSING

4.1 Method 4.1.1 Participants The participants in this experiment were 60 undergraduate students at Cambridge University and University College London who volunteered to take part in the experiment. They all had English as their first language. Each session lasted approximately 15 min. 4.1.2 Materials and design A set of 22 common nouns with predictable superordinates and distinctive properties were selected as primes. Two questionnaires based on the literature on prototypes (Rosch & Mervis 1975; Barsalou 1987) offered a direct account of property dominance. After piloting the questionnaires on 15 participants, the final version was distributed among 65 participants. Having chosen a list of words with predictable superordinates and

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

Carston 2002). It is therefore possible that the mechanism of suppression operates differently in each case. Disambiguating a homonymous word like ‘bank’ would involve selecting one of its two meanings (i.e. financial institution or side of a river). Given that these two meanings are part of the mental lexicon of the hearer, in processing the lexical form ‘bank’ activation would spread to two different lexical entries (for the sake of simplicity, BANK-1 and BANK-2). In contrast, interpreting for the first time a metaphor like ‘Mary is a nightingale’ would involve constructing an ad hoc concept NIGHTINGALE* from the lexically encoded concept NIGHTINGALE (Carston 2002; Glucksberg 2003; Recanati 2004). It would therefore be interesting to see whether these differences in the accessibility of the various concepts involved in disambiguation and metaphor interpretation may have an effect on the operation of the mechanism of suppression. To summarize, the following study was aimed at investigating two questions: first, the empirical prediction of various theories that metaphor interpretation involves enhancing metaphor-relevant properties of the vehicle while suppressing metaphor-irrelevant ones and second, at which point in processing suppression reduces the activation of metaphor-irrelevant information in relation to the suppression of contextually inappropriate meanings in early studies of lexical ambiguity. Also, some secondary theoretical issues about the accessibility of the figurative interpretation of the metaphor vehicle will be discussed.

Paula Rubio Ferna´ndez 355

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

distinctive properties, the results were as expected apart from two terms, ‘tip’ and ‘spring’, which are ambiguous and did not elicit a uniform response. These two terms were discarded. For each of the 20 remaining concepts, the most frequent superordinate term was chosen from the brief definition task in the first questionnaire. Likewise, the most frequent distinctive property was selected from the three different tasks in the two questionnaires (i.e. a brief definition task, a distinctive property listing and a free-association task; for the set of primes and targets, see Appendix A). Both types of target were therefore strong associates of the prime concepts (what I have also called ‘core features’ of the primes; see Rubio 2005, forthcoming). A metaphorically biased context was constructed for each one of the 20 primes so that superordinates were inconsistent with the figurative interpretation, whereas distinctive properties were relevant for interpretation. Each context ended in a nominal metaphor of the form ‘X is a Y’, Y being always the prime concept. Because the metaphors were novel, the preceding context included two sentences on average to make sure that the nominal metaphor would be comprehensible (for the set of metaphoric contexts, see Appendix B). The 20 metaphoric contexts were divided into two equal groups matched for word length and frequency of the corresponding superordinates and distinctive properties (Johansson & Hofland 1989). One group of contexts were paired with related superordinates and distinctive properties. For the other group, targets were scrambled so the sentences were paired with unrelated target words. The unrelated contexts and targets served as controls. Two lists of materials were constructed by pairing one group of contexts with related targets in List A and with unrelated targets in List B, and the other group of contexts with unrelated targets in List A and with related targets in List B. Another set of 20 metaphoric contexts was constructed and paired with English-like non words. Critical and filler sentences were randomized individually for each participant in each list of materials. Participants were randomly assigned to one Target Type, List and Inter Stimulus Interval (ISI), so each participant saw each context and the corresponding target only once. Given that a word is identified at the point in time when acoustic information uniquely specifies it, which may actually occur before the physical ending of the word (Marslen-Wilson 1987), for each of the 20 nouns in our materials a point was selected where the prime would be unequivocally recognized. Targets were presented visually at the end of the acoustic signal 0, 400 or 1000 ms after the word-recognition point selected for each prime. This enabled accurate measuring of initial semantic activation, while controlling for the possibility that an early

356 Suppression in Metaphor Interpretation contextual effect may result from an early word recognition followed by a fast property selection given the length of the primes. Sentences were recorded at a normal rate by a male speaker on an Apple Macintosh computer. The auditory stimuli and the visual targets were synchronized using a specialized computer program. The experimental items were preceded by two sets of practise trials. The first one consisted of a lexical decision task and the second one included both sentential contexts in the acoustic modality and visual targets for a lexical decision. The latter contained six metaphoric contexts similar to the critical ones, although the corresponding visual targets were not related to the primes in any of the practice trials.

4.1.4 Procedure The experiment was presented to the participants as a simple psycholinguistic experiment investigating the interpretation of metaphorical language. Participants were told that they would be listening to a series of short texts through the headphones and that each text would end in a metaphor. To make sure that participants derived the intended interpretation of the metaphor, they were asked to try to visualize the figurative meaning of the metaphor. At the end of each sentence, a string of letters would appear on the computer screen and they should try to indicate as fast and accurately as possible whether the string of letters was a word of English or not by pressing the corresponding key on the response box. Participants were first given standard written instructions, which were then explained individually by the experimenter. It was emphasized that both tasks (i.e. listening carefully to the sentences and making a fast lexical decision) were equally important, although they should be taken as independent tasks. Participants were tested individually. They ran through the two sets of practice trials with the experimenter and got appropriate feedback

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

4.1.3 Apparatus The experiment was conducted on a Toshiba laptop computer. The sentences were presented through a pair of headphones plugged into the laptop. The visual probes were presented in capital letters in the middle of the computer screen on a white background. Responses to the visual targets were made via a response box connected to the laptop. ‘Word’ responses were made with the thumb or the index finger of the right hand and ‘non-word’ responses with the thumb or the index finger of the left hand. Target words remained on the screen until the participant had made a decision. There was a 1000-ms delay between the offset of the visual target and the onset of the following acoustic context.

Paula Rubio Ferna´ndez 357

on their performance. When being tested on the critical materials, participants were left on their own in a closed room or cubicle. In order to make sure that participants paid adequate attention to the contexts, a short memory test was given at the end of the experiment. Participants had been told about this memory test in the instructions.

4.2 Results

3 Because of the small scale of the experiment, the design was not powerful enough to carry out reliable analyses per item. However, List was included as an independent variable to see whether the distribution of the materials had had any significant effect on the ANOVAs. Nonetheless, without an item analysis it is not possible to establish whether the results generalize to other metaphors, which is obviously a limitation of the study.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

The minimum of correct responses required in the memory test was 2.5 standard deviations below the participants’ average of correct responses per ISI. No participant had to be replaced for failing to meet this criterion. The mean response time, standard deviation and proportions of missing data for each Relatedness condition, together with the facilitation (i.e. the difference between the experimental [related] and the control [unrelated] conditions) and its significance level per Target Type and ISI are presented in Table 1. A response time data point was treated as ‘missing’ if it was either from an erroneous response or over 2.5 standard deviations above the participant’s average response time to the word targets in his exercise. The activation curves of superordinates and distinctive properties in metaphoric contexts, understood as the priming effect observed across the three ISIs for each target type, are given in Figure 1. The statistical analysis of the data examined the effects of Target Type (superordinate/distinctive property), Target Relatedness (related/ unrelated), ISI (0/400/1000 ms) and List (A/B). Mean reaction times were entered into four-way analyses of variance (ANOVAs), with participants (F1) as the random variable.3 There was a significant main effect of Relatedness, F1(1,48) ¼ 46.03, MSE ¼ 1102.1, P < 0.001; ISI, F1(2,48) ¼ 4.425, MSE ¼ 59251, P < 0.02 and Target Type, F1(1,48) ¼ 4.968, MSE ¼ 59251, P < 0.04. The fastest reaction times were observed in the related condition (42 ms difference), the 400-ms condition (93 ms difference overall), and the distinctive property condition (99 ms difference). Only the 2 3 2 3 3 3 2 interaction (Relatedness 3 Target Type 3 ISI 3 List) was significant, F1(2,48) ¼ 6.381, MSE ¼ 1102.1, P < 0.004. The effect of List was not systematic,

358 Suppression in Metaphor Interpretation ISI Target

Relatedness

0

400

1000

Superordinates

Related Unrelated Facilitation Related

883 (198, 0.03) 919 (187, 0.05)
36* 740 (243, 0.04)

644 (162, 0.03) 698 (163, 0.07)
55*** 658 (58, 0.04)

791 (167, 0.02) 799 (156, 0.03)
8 598 (104, 0.03)

Unrelated Facilitation

782 (230, 0.1)
42**

710 (79, 0.07)
52***

651 (123, 0.06)
53**

Distinctive properties

given that the highest facilitation for each Target Type was not observed for the same List condition across ISIs. This interaction with List could therefore be related to the different speed of reaction of the different groups of participants tested in each condition. A 2 3 2 3 3 3 2 (Relatedness 3 Target Type 3 ISI 3 List) ANOVA was carried out on the arcsine transformation of the missing data using participants as the random factor. The missing data were arcsine transformed to stabilize variances (Winer 1971). Only a significant main effect of Relatedness was observed, F1(1,48) ¼ 7.031, MSE ¼ 0.029, P < 0.02. The highest missing rate was observed in the unrelated condition (0.179), so the facilitation observed could not have been due to missing data points lowering the average response time in the related condition. No interaction reached significance level (all F1 < 1.9). Given that the greatest difference in the level of priming for superordinates and distinctive properties was observed at the longest delay (45 ms difference), a 2 3 2 3 2 (Relatedness 3 Target Type 3 List) ANOVA was carried out on the reaction time data for the 1000-ms condition. This analysis is particularly relevant for the investigation of attentional processes, which take place around 500 ms from the offset of the prime (e.g. Neely 1976; Yee 1991). There was a significant main effect of Relatedness, F1(1,16) ¼ 8.705, MSE ¼ 1078.9, P < 0.01, and Target Type, F1(1,16) ¼ 7.004, MSE ¼ 41614, P < 0.02, with the fastest reaction times being observed in the related condition (31 ms difference) and the distinctive properties condition (171 ms difference). The Relatedness 3 Target Type interaction, which

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

Table 1 Mean reaction times (in milliseconds), standard deviations, proportions of missing data and facilitation for each condition of the experiment *P < 0.1 **P < 0.05 ***P < 0.01.

Paula Rubio Ferna´ndez 359 Metaphoric Contexts 60

Priming (msec)

50 40 Superordinates Distinctive Properties

30 20 10 0 400

1000

ISI (msec)

Figure 1 Activation curves of superordinates and distinctive properties in metaphoric contexts.

was critical for the present investigation, was significant, F1(1,16) ¼ 4.559, MSE ¼ 1078.9, P < 0.05, with the highest level of priming being observed in the distinctive property condition. The Relatedness 3 Target Type 3 List interaction was also significant, F1(1,16) ¼ 11.507, MSE ¼ 1078.9, P < 0.005. The highest level of priming was observed on the List A condition for superordinates (
30 ms) and on the List B condition for distinctive properties (
101 ms). Since this pattern of results was not consistent across ISIs in previous analyses and List did not show a significant main effect, this interaction with List could be related to the different performance of the different groups tested in each condition rather than to the particular distribution of the materials. The corresponding ANOVA of the arcsine transformation of the missing data did not show any significant results (all F1 < 1.5).

4.3 Discussion Despite their different roles in metaphor interpretation, both superordinates and distinctive properties were active up to 400 ms from the offset of the metaphor vehicle, their levels of activation deviating only at the 1000-ms delay. However, according to the results of the ANOVAs, the activation patterns of these associates were different, especially at the longest delay, where the critical Relatedness 3 Target Type interaction was significant.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

0

360 Suppression in Metaphor Interpretation

5 GENERAL DISCUSSION The results of the present study offer support to the twofold hypothesis that metaphor interpretation involves enhancing relevant properties of the metaphor vehicle while suppressing irrelevant ones (Glucksberg & Keysar 1990; Carston 2002; Recanati 2004). Both superordinates and distinctive properties, which were, respectively, metaphor-inconsistent and metaphor-relevant properties in their contexts, were active up to 400 ms from the offset of the prime. However, at the longest delay, the level of activation of these associates was significantly different, with only the metaphor-relevant properties remaining active. In another lexical priming study using literal neutral contexts, we observed the opposite pattern of results, with superordinates remaining active up to 1000 ms but distinctive properties decaying between 400 and 1000 ms (Rubio et al. 2003). These different patterns of activation resulted in a crossover effect (Rubio 2005), which supports the view that distinctive properties were enhanced in metaphoric contexts where they were relevant for interpretation, whereas superordinates were actively suppressed given their inconsistency with the figurative interpretation of the nominal metaphor. Therefore, the present study offers support to previous studies of property activation in metaphor

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

It looks as if the loss of activation of superordinates between 400 and 1000 ms must be the result of active suppression of metaphorinconsistent information. In a previous study, we observed that superordinates remain active up to 1000 ms in literal neutral contexts, where no property of the prime was particularly relevant or irrelevant for interpretation (Rubio et al. 2003). Their loss of activation in metaphoric contexts could not, therefore, have been the result of passive decay but has to be due to active suppression. The suppression of superordinates between 400 and 1000 ms in metaphoric contexts would have been due to attentional processes different from the processes involved in lexical disambiguation, which occur as early as 200 ms from the offset of the homonym. The main differences are that attentional processes (i) involve some level of awareness, (ii) need time to develop and (iii) are of limited capacity (Keele & Neill 1978; Neill & Westberry 1987). Therefore, although the results support the theoretical prediction that metaphor interpretation involves the suppression of metaphor-inconsistent information, this mechanism seems to operate faster in the resolution of lexical ambiguity than in metaphor interpretation.

Paula Rubio Ferna´ndez 361

(4)

After getting himself some new clothes and going to the barber, John was a lion.

Unlike in the previous example, the interpretation of ‘John was a lion’ in (4) would be based on the stereotypical view of lions as epitomizing pride and courage. Since there does not seem to be a possible chain of automatic associations between the schemata for buying clothes and going to the barber and John feeling self-confident, some inferential process must take place in interpreting the nominal metaphor along these lines (e.g. some backward inference about John being pleased

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

interpretation (Gernsbacher et al. 2001; Glucksberg et al. 2001; McGlone & Manfredi 2001). Regarding the issue of whether or not the processes involved in metaphor interpretation are inferential, it does seem that some of the nominal metaphors used in this study could be interpreted by purely local associative processes (Recanati 2004). For example, in interpreting ‘Compared to the rest of the boys in the basketball team, John was a minnow’, the activation of the information associated with BASKETBALL TEAM would include information regarding the size of the players, which in turn would make more accessible the metaphor-relevant information of the lexical concept MINNOW. However, not all the metaphoric contexts could be interpreted in this way, some apparently requiring fully inferential processes in the construction of the ad hoc concept (Carston 2002; Wilson 2003). For example, in interpreting ‘After six months without going to the barber, John was a lion’, the hearer would draw the conclusion that John’s hair was long on the basis of processing the first clause, which would increase the accessibility of the contextual assumption lions have manes in interpreting the nominal metaphor ‘John was a lion’. Therefore, the implicature that John’s hair was long, which is inferred from the first clause, would be used in deriving the figurative implication that John had a thick mass of hair, which would follow from John’s belonging to the category LION* characterized by the property ‘having a thick mass of hair’. In order to support his claim that narrowing and loosening involve purely automatic, associative processes, Recanati could argue that in processing the latter example, activation would spread from BARBER to HAIR and from HAIR to MANE, making accessible the figurative interpretation of the nominal metaphor without any inferential process being involved. However, this would imply that any similar metaphoric context including the word ‘barber’ and using ‘lion’ as the metaphor vehicle would result in a similar interpretation, which is not necessarily the case. Consider the following example:

362 Suppression in Metaphor Interpretation

4 The long activation of metaphor-inconsistent properties in the present study could be explained along the lines of ‘the graded salient hypothesis’ (Giora 2002 and references therein), according to which salient (coded, conventional) meanings are processed initially regardless of their contextual relevance. Although this hypothesis would predict the long activation of superordinates in metaphoric contexts as strong associates of the primes, it does not distinguish between a meaning component being highly accessible during processing and it being selected as part of the meaning communicated by use of a word (see Rubio 2005, forthcoming for discussion).

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

with his new look). Therefore, the type of metaphoric contexts used in this experiment support the relevance-theoretic view of narrowing and loosening as inferential pragmatic processes rather than merely associative (Carston 2002; Wilson 2003). On the other hand, the present results support Recanati’s claim about the accessibility of the lexical concept in constructing the derived ad hoc concept (Recanati 1995, 2004; cf. Sperber & Wilson 2002). The superordinates used in this experiment, which were inconsistent with the figurative interpretation of the metaphor vehicle, were highly accessible up to 400 ms from its offset. However, I would not argue that the nominal metaphors were not interpreted until that point in processing, but rather that, because of their strong association with the prime, superordinates remained active past the point where they were discarded from interpretation. Given that the nominal metaphors in this experiment were novel, one may assume that they would have been perceived as figurative uses. The long activation of superordinates could therefore be understood along the lines put forward by Recanati (2001): although certain literal contextual assumptions would have been abandoned in interpreting the utterance figuratively, the corresponding conceptual properties would still be active, giving rise to a certain level of metaphor awareness.4 Although the activation pattern of superordinates can be explained following Recanati (2001, 2004), the results of the present experiment do not generally support his model of lexical processing in terms of accessibility. According to Recanati (1995, 2004), the selection of a conceptual representation is determined exclusively by its accessibility, so that the most highly accessible one will be selected for interpretation. In this study, the activation of metaphor-inconsistent and metaphor-relevant properties was not significantly different at 400 ms, with superordinates showing a slightly higher level of priming at that point in processing. Therefore, according to Recanati’s model of lexical interpretation, metaphor-inconsistent properties would have been selected for interpretation at the intermediate delay, to be suppressed later on in processing. This interpretation of the results would also follow from the standard pragmatic view of metaphor,

Paula Rubio Ferna´ndez 363

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

according to which the literal interpretation of a metaphoric expression is derived first, before looking for an alternative, figurative interpretation that is satisfactory in the context (Grice 1975, 1989; Searle 1979; for empirical evidence against this view see Glucksberg et al. 1982; Gildea & Glucksberg 1983; Keysar 1989). For the present results to have fully supported Recanati’s account of lexical interpretation (Recanati 1995, 2001, 2004), superordinates should have been active at the intermediate delay but at a lower level than distinctive properties, so that the latter would have been selected for interpretation while the former would have stood in conflict with interpretation. However, if the selection of meaning components was directed by considerations of relevance and not by accessibility alone (Carston 2002; Wilson & Sperber 2004), it would still be possible to argue that superordinates had been discarded from interpretation although still highly active at 400 ms. In this view, the selection of the most accessible properties would minimize processing costs, but the maximization of cognitive effects (especially the derivation of implications of the metaphoric meaning) would be the determining factor in selecting the relevant properties for interpretation. Thus, although superordinates would have been active at the intermediate delay because of their strong association with the prime concepts, only distinctive properties would have been selected for interpretation given that those properties would have given rise to greater cognitive effects when promoted to the status of content constitutive of the ad hoc concept which figured in the proposition expressed. Regarding the operation of the mechanism of suppression in metaphor interpretation and disambiguation, the sustained activation of superordinates in metaphoric contexts seems to indicate that the suppression of these metaphor-inconsistent properties is different from the suppression of the irrelevant meanings of homonyms, which takes place 200–300 ms from the offset of the ambiguous prime (e.g. Tanenhaus et al. 1979, Seidenberg et al. 1982). It is therefore reasonable to conclude that the differential availability of the various meanings accessed in lexical ambiguity and novel metaphor interpretation has an effect on the operation of the mechanism of suppression in these two paradigms. Thus, although both disambiguation and novel metaphor interpretation involve dealing with two different concepts, in resolving a lexical ambiguity, various concepts are accessed from the start, although only one is selected in later stages of processing. In contrast, in processing a novel metaphorical expression, a single lexical concept is accessed initially, although an ad hoc concept is constructed on-line later on in processing. The inappropriate meaning of a homonym is

364 Suppression in Metaphor Interpretation

5 Empirical evidence that reading literal and metaphorical expressions takes a comparable amount of time is usually taken to show that processing literal and figurative language takes the same amount of processing effort (see Gibbs 1984, 1989). However, these studies should be interpreted carefully as extra attentional resources can always reduce reading time while still adding to the processing effort account (see Rubio 2005 for discussion).

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

therefore suppressed in a context-sensitive, albeit effectively automatic, way given that the various meanings are available to the processor from the start. On the other hand, suppressing the literal meaning of a novel metaphor would require the operation of later, attentional processes since the alternative figurative meaning becomes available only during the process of understanding the utterance. The idea that the inhibitory processes involved in metaphor interpretation are more demanding of attentional resources than those involved in disambiguation ties in with the idea that metaphorical language is ‘special’ (see Gibbs 1989 for discussion). A recurrent point in the literature on lexical processing is how, despite the initial activation of the various meanings of ambiguous words, people are not usually aware of having encountered an ambiguous word in a sentence (Tanenhaus et al. 1979; Gernsbacher 1990). On the other hand, we are usually aware of figurative language use, especially in the case of novel metaphors like the ones used in the present experiment. It is therefore possible that metaphor awareness is related to the fact that more attentional resources may be involved in metaphor interpretation as opposed to disambiguation, the processing of which is virtually automatic.5 In this view, metaphor interpretation would be special, but not necessarily more so than lexical pragmatic processes such as disambiguation or reference assignment, the difference lying merely in the degree of automatization of the cognitive processes involved. Lexicalized metaphors or metaphors that after frequent use have given rise to a second meaning of an expression (e.g. ‘bulldozer’ as in heavy machine and overbearing person) would be a case in point. If the association between a metaphor vehicle and one of its possible figurative interpretations is strengthened enough by frequent processing, the metaphor vehicle becomes polysemous. Therefore, it is possible that diachronically, the process of meaning construction involved in metaphor interpretation evolves through automatization into a process of meaning selection and so disambiguation (see Neely 1977 and Barsalou 1982 for a discussion of the automatization of controlled processes). It seems reasonable to assume that metaphor awareness would be lost at some point in that diachronic process of automatization of meanings, so the more familiar a metaphor, the less figurative (and probably evocative) it may seem.

Paula Rubio Ferna´ndez 365

6 The career of metaphor model (Bowdle & Gentner 2005) posits that, as a metaphor vehicle becomes conventional, its processing changes from comparison to categorization (cf. Glucksberg & Haught 2006; Jones & Estes 2006). Unlike Bowdle and Gentner, I am not proposing that novel and familiar metaphors involve different modes of processing, but rather that constructing an ad hoc concept when interpreting a novel metaphor may be more demanding of attentional resources than retrieving a figurative meaning that has been lexicalized after frequent use. Another difference with the above studies is that I consider context to play a fundamental role in making a figurative interpretation more or less accessible and therefore I do not see metaphor aptness as a fixed value in the absence of a particular context.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

Regarding the role of suppression in lexical ambiguity and metaphor interpretation, it is possible that the literal interpretation of a metaphor vehicle does not need to be suppressed if the figurative one is made highly accessible by the preceding linguistic context, just as suppression does not need to operate on the inappropriate meaning of a homonym in a biasing context where the appropriate meaning has been lexically primed (Tabossi 1988; Simpson and Krueger 1991). However, this is not necessarily the case in all contexts. The results of the present study suggest that, in contexts where metaphor-relevant properties are not facilitated, suppression reduces the activation of metaphor-inconsistent properties of the vehicle concept. I would therefore like to propose two intersecting continua that may help to delimit the extent to which extra attentional resources are required in processing a metaphorical expression. First, metaphor vehicles may range from novel to familiar to lexicalized, depending on how frequently they have been encountered previously (note that a lexicalized metaphor would be a polysemous word, with the two paradigms being compared touching at that point). This continuum would be related to the availability of the metaphorical meaning in the mental lexicon, with metaphorical meanings not being strongly associated to a novel metaphor but being accessed automatically if the vehicle is a lexicalized metaphor. Second, linguistic contexts may range from nonsensical to metaphoric to priming, depending on how much they facilitate the metaphorical interpretation of the expression. The first type of context would be comparable to ambiguous contexts (e.g. ‘John is a banana’ or ‘John went to the bank’ without any contextual cues that may help interpretation). Metaphoric contexts would be similar to disambiguating contexts in that the appropriate interpretation would be made accessible (e.g. ‘In his flashy coat, John is a banana’), but without going to the extreme of priming it intralexically, in which case the context would be priming (e.g. ‘In his long yellow coat, John is a banana’).6

366 Suppression in Metaphor Interpretation

APPENDICES: EXPERIMENTAL MATERIALS

Appendix A: Primes and Targets Primes

Superordinates

Distinctive Features

Cactus Lion Slippers Skyscraper Lullaby Dalmatian Mercedes Chair Champagne Breakfast Cheetah Sapling Woodpecker Pacific Rugby Steel Minnow Banana Encyclopaedia Norway

Plant Animal Shoe Building Song Dog Car Seat Drink Meal Cat Tree Bird Ocean Sport Metal Fish Fruit Book Country

Spike Mane Comfortable Tall Sleep Spot Expensive Back Bubble Morning Fast Young Noise Large Tough Strong Small Yellow Knowledge Cold

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

The amount of attentional resources involved in interpreting a metaphorical expression would therefore be determined by the combination of these two factors: the degree of familiarity of the metaphorical interpretation and the strength of the contextual bias. It is obvious that the degree of familiarity of a literal expression and the bias of a literal context also combine to determine to some extent the amount of processing effort involved in understanding a literally used expression. However, I would still maintain that, at the lexical level, the availability of a lexically encoded concept as compared to an ad hoc concept constructed on-line makes metaphorical interpretation generally more dependent on context, and sometimes on attentional resources, than literal language interpretation.

Paula Rubio Ferna´ndez 367

Appendix B: Metaphoric Contexts

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

John doesn’t like physical contact. Even his girlfriend finds it difficult to come close to him. John is a cactus. After six months without going to the barber, John was a lion. Mary is very materialistic. She is only interested in men who are rich. Her latest boyfriend is a Mercedes. John knew many people, but there were only a few friends he could lean on. A good friend is a comfy chair. John loved paddling his canoe through the steep canyon. He especially enjoyed rolling over in the white water of the rapids. The river was champagne. On the dunes someone had planted a few pine trees among the local plants. The pine trees were skyscrapers. Mary loved maths but this year the teacher was very boring. Every lesson was a lullaby. It was impossible to study at college during the maintenance work. The carpenter next door was a woodpecker. John spends four hours a day in the gym. His muscles are steel. Compared to the other boys in the basketball team, John was a minnow. Nobody wanted to run against John at school. John was a cheetah. John likes to wear clothes that really stand out in the crowd. In his new coat, John is a banana. Even though she had never been to school, Mary was an encyclopaedia. Mary had been sharing a flat with John for a long time. With him she felt at ease even in silence. John was a pair of old slippers. When he was a kid, John wasn’t allowed to do many things on his own. His mother used to tell him that he was only a sapling. John was making a chocolate milkshake when the lid came off the blender. When his mother saw him, she said John was a Dalmatian. When Maria first came to England, she was very surprised that pubs closed at 11 pm. In Spain, closing time is breakfast. John and Mary have a new house with an amazing garden. Actually, one cannot see the end of it from the back door. Their garden is the Pacific. Things weren’t going well for Mary. Her boyfriend had broken up with her the same week she had lost her job. Sometimes life can be a game of rugby.

368 Suppression in Metaphor Interpretation Mary didn‘t like spending the night at her grandmother’s. No matter how many blankets she would put on the bed, that attic room was Norway. Acknowledgements

PAULA RUBIO FERNA´NDEZ Department of Phonetics and Linguistics University College London Gower Street London WC1E 6BT e-mail: [email protected]

REFERENCES Barsalou, L. W. (1982), ‘Contextindependent and context-dependent information in concepts’. Memory and Cognition 10:82–93. Barsalou, L. W. (1987), ‘The instability of graded structure: Implications for the nature of concepts’. In U. Neisser (ed.), Concepts and Conceptual Development: Ecological and Intellectual Factors in Categorization. pp. 101–40. Cambridge University Press. Cambridge. Bowdle, B. & D. Gentner (2005), ‘The career of metaphor’. Psychological Review 112:193–216.

Cacciari, C. & S. Glucksberg (1994), ‘Understanding figurative language’. In M. Gernsbacher (ed.), Handbook of Psycholinguistics. Academic Press. pp. 447–77. San Diego, CA. Carston, R. (1996), ‘Enrichment and loosening: Complementary processes in deriving the proposition expressed?’ UCL Working Papers in Linguistics 8:61–88. Reprinted in (1997) Linguistische Berichte 8:103–27. Carston, R. (2002), Thoughts and Utterances: The Pragmatics of Explicit Communication. Blackwell. Oxford.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

The study reported in this paper is part of the experiments completed by the author towards her Ph.D. degree at Cambridge University. This research was initially supported by an Arts and Humanities Research Board Postgraduate Award, and currently by a British Academy Postdoctoral Fellowship and a Marie Curie Outgoing International Fellowship (Project 022149—Inferential Processes in Language Interpretation). This paper was written at University College London and reviewed at Princeton University, where I am currently working on a postdoctoral project. I would like to thank Bart Geurts, Zachary Estes, Lara Jones and an anonymous reviewer for their comments and suggestions, which I think have improved the paper greatly. Thanks as well to Richard Breheny and John Williams for supervising the setting and analysis of my experiments. I would also like to thank Ira Noveck and Nausicaa Pouscoulous for their comments on an earlier version of this paper, as well as Robyn Carston and Sam Glucksberg for the valuable insight and support in all their comments.

Paula Rubio Ferna´ndez 369 metaphors?’ Journal of Verbal Learning and Verbal Behavior 21:85–98. Glucksberg, S. & C. Haught (2005), ‘On the relation between metaphor and simile: When comparison fails’. Mind and Language 21:360–78. Glucksberg, S. & B. Keysar (1990), ‘Understanding metaphorical comparisons: Beyond similarity’. Psychological Review 97:3–18. Glucksberg, S., D. A. Manfredi, & M. S. McGlone (1997a), ‘Metaphor comprehension: How metaphors create new categories’. In T. B. Ward, S. M. Smith, and J. Vaid (eds.), Creative Thought: An Investigation of Conceptual Structures and Processes. American Psychological Association. Washington, DC. Glucksberg, S., M. R. Newsome, & Y. Goldvarg (1997b), Filtering out irrelevant material during metaphor comprehension. Paper presented at the Meeting of the Psychonomic Society, Philadelphia, PA. Glucksberg, S., Newsome, M. R., & Y. Goldvarg (2001), ‘Inhibition of the literal: Filtering metaphor-irrelevant information during metaphor comprehension’. Metaphor and Symbol 16:277–98. Grice, H. P. (1975), ‘Logic and conversation’. In P. Cole and J. Morgan (eds.), Syntax and Semantics III: Speech Acts. Academic Press. London. Grice, H. P. (1989), Studies in the Way of Words. Harvard University Press. Cambridge, MA. Johansson, S. & K. Hofland (1989), Frequency Analysis of English Vocabulary and Grammar based on the Lob Corpus: Vol. 1, Tag Frequencies and Word Frequencies. Clarendon Press. Oxford. Jones, L. L. & Z. Estes (2005), ‘Roosters, robins and alarm clocks: Aptness and conventionality in metaphor

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

Carston, R. (2007), ‘How many pragmatic systems are there?’ In M.-J. Frapolli (ed.), Saying, Meaning, Referring: Essays on the Philosophy of Francois Recanati. Palgrave. London, pp. 18–48. Gernsbacher, M. A. (1990), Language Comprehension as Structure Building. Lawrence Erlbaum. Hillsdale, NJ. Gernsbacher, M. A. & M. Faust (1991), ‘The role of suppression in sentence comprehension’. In G. B. Simpson (ed.), Understanding Word and Sentence. Elsevier Science Publishers. North Holland. Gernsbacher, M. A., B. Keysar, & R. W. Robertson (1995), The role of suppression in metaphor interpretation. Paper presented at the annual meeting of the Psychonomic Society, Los Angeles, CA. Gernsbacher, M. A., B. Keysar, R. W. Robertson, & N. K. Werner (2001), ‘The role of suppression and enhancement in understanding metaphors’. Journal of Memory and Language 45:433–50. Gibbs, R. W. (1984), ‘Literal meaning and psychological theory’. Cognitive Science 8:275–304. Gibbs, R. W. (1989), ‘Understanding and literal meaning’. Cognitive Science 13:243–51. Gildea, P. & S. Glucksberg (1983), ‘On understanding metaphor: The role of context’. Journal of Verbal Learning and Verbal Behavior 22:577–90. Giora, R. (2002), ‘Literal vs. figurative language: Different or equal?’ Journal of Pragmatics 34:487–506. Glucksberg, S. (2003), ‘The psycholinguistics of metaphor’. Trends in Cognitive Sciences 7:92–6. Glucksberg, S., P. Gildea, & H. B. Bookin, (1982), ‘On understanding nonliteral speech: Can people ignore

370 Suppression in Metaphor Interpretation and Cognition: The Loyola Symposium. Lawrence Erlbaum. Hillsdale, NJ. Recanati, F. (1995), ‘The alleged priority of literal interpretation’. Cognitive Science 19:207–32. Recanati, F. (2001), ‘Literal/nonliteral’. Midwest Studies in Philosophy 25: 264–74. Recanati, F. (2004), Literal Meaning. Cambridge University Press. Cambridge. Rosch, E. & C. B. Mervis (1975), ‘Family resemblances: Studies in the internal structure of categories’. Cognitive Psychology 4:573–605. Rubio Ferna´ndez, P. (2005), ‘Pragmatic processes and cognitive mechanisms in lexical interpretation: The on-line construction of concepts’. Ph.D. thesis, University of Cambridge. Rubio Ferna´ndez, P. ‘Concept narrowing: The role of context-independent information’. Journal of Semantics, in press. Rubio Ferna´ndez, P., R. Breheny & M. Wei Lee (2003), ‘Contextindependent information in concepts: An investigation of the notion of ‘‘core features’’’. In F. Schmalhofer, R. M. Young, and G. Katz (eds.), Proceedings of EuroCogSci 03: The European Cognitive Science Conference 2003. Lawrence Erlbaum. London. Schvaneveldt, R. W. & D. E. Meyer (1973), ‘Retrieval and comparison processes in semantic memory’. In S. Kornblum (ed.), Attention and Performance IV. Academic Press. New York. Searle, J. (1979), Expression and Meaning: Cambridge University Press. Cambridge, England. Seidenberg, M. S., M. K. Tanenhaus, J. L. Leiman, & M. Bienkowski (1982), ‘Automatic access of the meanings of ambiguous words in context: Some limitations of

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

comprehension’. Journal of Memory and Language 55:18–32. Keele, S. W. & W. T. Neill (1978), ‘Mechanisms of attention’. In E. C. Carterette and M. P. Friedman (eds.), Handbook of Perception IX. Academic Press. New York. Keysar, B. (1989), ‘On the functional equivalence of literal and metaphorical interpretations in discourse’. Journal of Memory and Language 28:375–85. Marslen-Wilson, W. D. (1987), ‘Functional parallelism in spoken wordrecognition’. Cognition 25:71–102. McGlone, M. S. & D. Manfredi (2001), ‘Topic-vehicle interaction in metaphor comprehension’. Memory and Cognition 29:1209–19. Neely, J. H. (1976), ‘Semantic priming and retrieval from lexical memory: Evidence for facilitatory and inhibitory processes’. Memory and Cognition 4:648–54. Neely, J. H. (1977), ‘Semantic priming and retrieval from lexical memory: Roles of inhibitionless spreading activation and limited capacity attention’. Journal of Experimental Psychology 106:226–54. Neill, W. T., L. A. Valdes, & K. M. Terry (1995), ‘Selective attention and the inhibitory control of attention’. In F. N. Dempster and C. J. Brainerd (eds.), Interference and Inhibition in Cognition. Academic Press. San Diego, CA. Neill, W. T. & R. L. Westberry (1987), ‘Selective attention and the suppression of cognitive noise’. Journal of Experimental Psychology: Learning, Memory, and Cognition 13:327–34. Onifer, W. & D. A. Swinney (1981), ‘Accessing lexical ambiguities during sentence comprehension: Effects of frequency of meaning and contextual bias’. Memory and Cognition 9:225–6. Posner, M. I. & C. R. R. Snyder (1975), ‘Attention and cognitive control’. In R. L. Solso (ed.), Information Processing

Paula Rubio Ferna´ndez 371 Tanenhaus, M. K., Leiman, J. M., & M. S. Seidenberg (1979), ‘Evidence for multiple stages in the processing of ambiguous words in syntactic contexts’. Journal of Verbal Learning and Verbal Behavior 18: 427–40. Wilson, D. (2003), ‘Relevance theory and lexical pragmatics’. Italian Journal of Linguistics/Rivista di Linguistica 15:273–91 (special issue on pragmatics and the lexicon). Wilson, D. & D. Sperber (2004), ‘Relevance theory’. In G. Ward and L. Horn (eds.), Handbook of Pragmatics. Blackwell. Oxford. Longer earlier version published in (2002) UCL Working Papers in Linguistics, 14: 249–87. Winer, B. J. (1971), Statistical Principles in Experimental Design. McGraw-Hill. New York. Yee, P. L. (1991), ‘Semantic inhibition of ignored words during a figure classification task’. Quarterly Journal of Experimental Psychology 43A: 127–53. First version received: 06.08.2006 Second version received: 29.01.2007 Accepted: 02.02.2007

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

knowledge-based processing’. Cognitive Psychology 14:470–519. Shiffrin, R. M. & W. Schneider (1977), ‘Controlled and automatic human information processing: II. Perceptual learning, automatic attending, and a general theory’. Psychological Review 84:127–90. Simpson, G. B. & M. A. Krueger (1991), ‘Selective access of homograph meanings in sentence context’. Journal of Memory and Language 30:627–43. Sperber, D. & D. Wilson (1986/1995), Relevance: Communication and Cognition. Blackwell. Oxford. Sperber, D. & D. Wilson (1997), ‘The mapping between the mental and the public lexicon’. UCL Working Papers in Linguistics 9:107–25. Sperber, D. & D. Wilson (2002), ‘Pragmatics, modularity and mindreading’. Mind and Language 17:3–23. Swinney, D. A. (1979), ‘Lexical access during sentence comprehension: (Re)consideration of context effects’. Journal of Verbal Learning and Verbal Behavior 18:645–59. Tabossi, P. (1988), ‘Accessing lexical ambiguity in different types of sentential context’. Journal of Memory and Language 27:324–40.

Journal of Semantics 24: 373–416 doi:10.1093/jos/ffm011 Advance Access publication September 3, 2007

Processing Presupposed Content FLORIAN SCHWARZ University of Massachusetts Amherst

Abstract

1 INTRODUCTION Presuppositions have been an important topic in both the philosophy of language and in linguistic semantics and pragmatics, but only more recently have they been investigated with psycholinguistic methods. However, a lot can be gained from such investigations, both with respect to theoretical issues in presupposition theory and with respect to our understanding of semantic processing. In the following, I present three experimental studies, two of which employ the German additive particle auch ‘too’, while the last one uses English also. The results of these studies reveal processing effects of presuppositions. These support analyses of additive particles that account for the fact that their presuppositions cannot easily be accommodated. They furthermore suggest that presuppositions play a role in online sentence processing, that is that they can affect the way incoming linguistic input is analysed by the parser. In theoretical terms, they can be seen as constraining the Ó The Author 2007. Published by Oxford University Press. All rights reserved. For Permissions, please email: [email protected].

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

This paper presents three experimental studies investigating the processing of presupposed content. The first two experiments employ the German additive particle auch ‘too’, and the third uses English also. In experiment 1, participants were given a questionnaire containing biclausal, ambiguous sentences containing auch. The presupposition introduced by auch was only satisfied on one of the two readings, which corresponded to a syntactically dispreferred parse of the sentence. The prospect of having the auch presupposition satisfied made participants choose this syntactically dispreferred reading more frequently than in a control condition. Experiment 2 used the self-paced reading paradigm and compared the reading times on clauses containing auch, which differed in whether the presupposition of auch was satisfied or not. Participants read the clause more slowly when the presupposition was not satisfied. Experiment 3 followed up a number of issues that arose from experiment 2 and confirmed the results found there. These studies show that presuppositions play an important role in online sentence comprehension and affect the choice of syntactic analysis. Some theoretical implications of these findings for the semantic analysis of auch/also and dynamic accounts of presuppositions as well as for theories of semantic processing are discussed.

374 Processing Presupposed Content

2 BACKGROUND One might start out the enterprise of investigating presuppositions in processing by wondering about how we can capture their effects in online sentence comprehension studies at all. After all, they are most commonly thought of as crucially relating to the context, and in the experimental settings typically used in psycholinguistic work, there is no realistic context. So, it is at least possible that participants in experiments more or less ignore such context-related information. This would be especially likely if presuppositions were dealt with in very late pragmatic processes that are more like conscious reasoning. If, on the other hand, the processor automatically makes use of presupposed content, we would expect participants to be unable to ignore it. The question then becomes in what ways presuppositions affect the parsing of incoming strings of linguistic expressions, and how quickly their content is accessible to the parser: does it occur online, that is, during the process of parsing the linguistic input or is it part of later, more general inferencing processes that may take place after the parser has decided on a structure and an interpretation of the linguistic input. It is generally assumed that interpretation proceeds incrementally, but the details of how incremental interpretation of specific semantic phenomena takes

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

possible analyses of presuppositional phenomena in general. I argue that the specific results presented here suggest that the processor carries out something like updates of the representations of contexts (in the sense of Dynamic Semantics or Discourse Representation Theory, DRT) below the sentence level in actual processing, namely, at the level of noun phrases. Assuming that the processor does this by using the available grammatical mechanisms, this, in turn, requires that our theory of semantic interpretation in context allows for updates at such a lower level. In addition to these theoretical conclusions, some questions arising for a theory of semantic processing are also discussed. The paper is organized as follows: in section 2, I provide some background on the main issues relevant to the experiments, including my theoretical assumptions about presuppositions and a few remarks about existing work on semantic processing. Section 3 presents the three experimental studies that were carried out. Section 4 discusses implications of the experimental results for the analysis of additive particles and for issues relating semantic theory and semantic processing, as well as some perspectives on future research. Section 5 concludes the paper.

Florian Schwarz 375

1 DRT is usually assumed to be equivalent to File Change Semantics in most respects. One difference is that the former makes reference to formal properties of Discourse Representation Structures (DRSs). File Change Semantics is more neutral in that it does not formulate its generalizations in terms of any specific representation, but rather at the level of content. This is not incompatible with talking about mental representations in connection with processing, however, which can be characterized as having a particular content. It simply does not say anything about the format of those representations.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

place are only partially understood. Looking at processing effects of presuppositions, which have been studied in depth in formal semantics and pragmatics, can provide us insights into how and when a specific aspect of meaning enters the picture in sentence comprehension processes. A further related point of interest is whether presuppositions interact with other factors known to be relevant in parsing, and if so in what ways. From a theoretical viewpoint, we are, of course, especially interested in what implications experimental results might have for semantic and pragmatic theory. First of all, we should consider how our semantic analyses of the specific presupposition triggers relate to them. The relevant issues for auch and also will be introduced in section 3, in connection with the experimental design. Furthermore, we want to relate them to the bigger picture of possible semantic frameworks for presupposition theory. In connection with this it is interesting to note that most of these share a procedural view of some sort which determines how presupposed content is integrated with contextual information (although typically they do not make any explicit claims about actual processing). The family of approaches to semantic theory going under the labels of Dynamic Semantics (Heim 1982; Heim 1983a,b) or DRT (Kamp 1981; Kamp & Reyle 1993), which formalize earlier insights by Stalnaker and Karttunen (Karttunen 1973, 1974; Stalnaker 1973, 1974), has been particularly important in presupposition theory (van der Sandt 1988, 1992; Geurts 1999; Beaver 2001). For concreteness and simplicity, I will frame the discussion in this paper in terms of one specific proposal in this family, namely Heim’s original File Change Semantics. However, it should be clear that the results could just as well be described in other dynamic systems, for example, DRT.1 In approaches to presupposition theory in the tradition of Stalnaker and Karttunen, presuppositions are assumed to have two crucial properties. First, they are something that is taken for granted by the discourse participants. Secondly, they behave differently from asserted content in most embedded contexts. This is at the heart of the projection problem (for an overview, see Beaver 1997; von Fintel 2004). In File Change Semantics, which can be viewed as a formal

376 Processing Presupposed Content

2

But recent work is becoming more diverse in terms of the presupposition triggers covered. See, for example, Chambers & Juan (2005) on again and for new work on pragmatic processing more generally the volume edited by Noveck & Sperber (2004).

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

implementation of the accounts for presuppositional phenomena by Stalnaker and Karttunen, being taken for granted is modelled by the common ground, which is the set of worlds in which all of the beliefs that the discourse participants knowingly share are true. A sentence can only be felicitously uttered when the presuppositions that come with uttering the sentence are entailed by the common ground. The behaviour of presuppositions in embedded contexts is accounted for by the way that the common ground is updated when a new utterance is made in the discourse. Under certain circumstances, presupposition failure can be remedied by a process of accommodation (Lewis 1979), in which the common ground is adjusted in such a way that it does entail the presupposition at issue prior to the update. File Change Semantics represents the meanings of sentences as context change potentials. More concretely, sentence meanings are understood as functions from contexts to contexts (where contexts are modelled either as sets of worlds or as sets of pairs of worlds and assignment functions). One of the crucial issues in this type of theory is where or when context updates take place, and this is where the procedural viewpoint becomes relevant: the issue of when the adjustments to the context are made is determined by the procedural steps that the theory assumes. Quite frequently the discussion in the literature focuses on the sentence or clause level as the locus of updates, which seems intuitively plausible. However, in the full version of Heim’s system, which includes assignment functions, updates also take place at the level of noun phrases (which are viewed as denoting atomic propositions). Furthermore, in order to account for certain facts concerning the behaviour of presuppositions in embedded contexts, Heim (1983a) introduces the notions of local and global accommodation. As I will discuss in some detail below, the issue of where updates take place is crucial for semantic processing viewed from the perspective of File Change Semantics: if the processor is to make use of compositional semantic information, the way in which it can be used depends on the time at which it has access to it. Before turning to the discussion of the experiments, let me briefly review some existing work on presuppositions in processing. Much related work focuses on the presupposition of the definite article and follows the approach taken in the seminal study of Crain & Steedman (1985).2 Looking at locally ambiguous sentences like the one in (1),

Florian Schwarz 377

their experiment 2 showed that varying the discourse context [as in (2)] affects the way that the sentence is parsed. The psychologist told the wife that he was having trouble with . . . a. . . . her husband. b. . . . to leave her husband. (2) a. Complement Inducing Context A psychologist was counseling a married couple. One member of the pair was fighting with him but the other one was nice to him. b. Relative Inducing Context A psychologist was counseling two married couples. One of the couples was fighting with him but the other one was nice to him. (1)

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

In (1a) the that clause is interpreted as the complement of told, while in (1b), it is a relative clause modifying wife. The latter reading is much harder to see due to a typical garden-path effect (especially out of context). The preceding contexts were varied in introducing either one or two couples, the idea being that if two couples are introduced, the definite description consisting of the noun only (the wife) cannot refer successfully, while the complex description consisting of the noun and the following that clause analysed as a relative clause does have a unique referent. The sentences were judged to be ungrammatical 54% of the time in a grammaticality judgment task when (1a) was presented in the relative inducing context, but they were judged to be grammatical 78% (1a) and 88% (1b) of the time when the contexts matched the target sentence. Crucially, even the garden-path in (1b) was ameliorated by putting it in a matching context. This finding motivated Crain and Steedman to propose a principle of parsimony, which guides the selection between different syntactic parses in their parallel parsing architecture, so that the reading carrying the fewest unsatisfied presuppositions will be the preferred one. Similar designs are used in more recent work by van Berkum and colleagues (van Berkum et al. 1999, 2003), which shows that there are ERP effects related to whether the definite description can refer successfully or not. These studies focus on definite descriptions and show effects of presuppositions relative to preceding discourse. The studies presented here aim to broaden the range of triggers being studied and to look at effects of presuppositions in relation to material within the same sentence. The experimental techniques used here contribute a new type of evidence to presupposition theory, where many hotly debated issues involve subtle intuitions. Furthermore, an attempt is made to integrate the experimental results into the theoretical discussion, in order to contribute to a theory of semantic processing informed by linguistic semantics.

378 Processing Presupposed Content 3 EXPERIMENTAL STUDIES ON AUCH AND ALSO

3 A related question of great interest is to what extent accommodation has measurable effects. Although this is just as important, I would not pursue this question here.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

How should we go about testing the potential effects of presuppositions in sentence processing? One of the standard techniques in psycholinguistics is to compare a normal or unproblematic form to a somehow deviant (or temporarily deviant seeming) form. This basic idea is applied to presuppositions in the studies below in two ways: first, participants were shown ambiguous sentences containing auch, where one reading of the sentence satisfied the presupposition introduced by auch, whereas the other did not. The task, then, was to choose a paraphrase corresponding to the participants’ understanding of the sentence. The second approach was to show unambiguous sentences with auch (experiment 2) and also (experiment 3) that varied in whether the presupposition was satisfied or not. These studies employed the selfpaced reading method, and participants simply had to read the sentences region by region. In experiment 2, they also had to answer simple questions about the sentences. Let us now turn to the question of what the presupposition of also is, and how it relates to the experimental design. It is well known that the presuppositions introduced by many triggers can easily be accommodated. It certainly is a possibility to be considered that in an experimental setting participants are willing to accommodate just about any content, since the situation they are in is obviously artificial. Just compare this situation to reading an example sentence in a linguistics article. It might very well contain, say, a definite description. There is nothing odd about reading such a sentence, even if it is completely unclear and left open whether the relevant presuppositions are satisfied or not. The danger for an experimental inquiry into presuppositions in processing might be that their effects cannot be measured at all, at least to the extent to which they can be accommodated without a problem.3 There are, however, a few presupposition triggers that have been argued to either strongly resist accommodation or be unaccommodable altogether (Beaver and Zeevat forthcoming). One case in point is additive particles like too and also. According to the early analysis of too by Karttunen & Peters (1979), this type of additive particle introduces an existential presupposition, requiring that there is another individual that has the property attributed to the focus of the sentence with too (e.g. ‘BILL likes Mary too’ would presuppose that there is someone other than Bill who likes

Florian Schwarz 379

Mary). However, Kripke has argued that this type of analysis is inadequate, based on examples like the one in (3a) (from Kripke 1991). (3)

a. JOHN is having dinner in New York tonight too. b. Did you know that Bill is having dinner in New York tonight?

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

In an out of the blue context, the sentence in (3a) is very odd, presumably due to a presupposition failure that cannot be remedied by accommodation. This is unexpected on the existential analysis, since in just about any context, it will be uncontroversial that there are many people having dinner in New York tonight. But it is clear that the utterance of (3a) is only felicitous when there is some individual salient in the discourse that has the relevant property, for example, in the context of (3b). One way to capture this property is by assuming that too is anaphoric, much like a pronoun, and that its presupposition can only be satisfied if there is an antecedent in the discourse context (Heim 1992; van der Sandt & Geurts 2001). It is exactly this property that makes also a useful presupposition trigger for the present purposes. If these additive particles strongly resist accommodation, we can have good hopes of finding processing effects when their presuppositions are not satisfied. The flip side of this is that strong processing effects of presupposition failure with these triggers support the idea that they are impossible (or at least very hard) to accommodate and analyses that account for this property, such as the anaphoric accounts mentioned above. In connection with this, it is also worth noting the work by Spenader (2002), who provides solid empirical evidence that the presupposition of too is hardly ever accommodated. In a corpus study of the London-Lund Corpus, she finds that too lacks an antecedent only 4% of the time, whereas many other presupposition triggers (e.g. definite descriptions and factives) lack an antecedent much more often (40% and 80% of the time, respectively) and are apparently easily accommodated in such situations. In summary, too (as well as also and its German counterpart auch) lends itself to experimental investigation. On the one hand, we have more control over whether presupposition failure takes place or not, since it is quite clear intuitively that sentences like (3b) are infelicitous without the right kind of supporting context. And on the other hand, we can hope to find a new form of empirical support for accounts that have an explanation for the difficulty of accommodation, if we find strong effects of presupposition failure.

380 Processing Presupposed Content

3.1 Questionnaire study on auch 3.1.1 Methods and materials The basic strategy for the experimental items for the first study was to construct biclausal, ambiguous sentences consisting of a relative clause and a main clause. One of the readings was preferred based on well-known syntactic parsing preferences. The other reading was the one that satisfied the presupposition of also, which appeared in the second clause. An example is given in (4). N and A stand for nominative and accusative, respectively.

The relative clause is syntactically ambiguous due to the ambiguity in the case marking. In German, there is a strong and extremely wellstudied parsing preference for interpreting such clauses as having a subject-initial, that is, as having subject-object order (Hemforth 1993; Bader & Meng 1999; beim Graben et al. 2000; Schlesewsky et al. 2000; Schlesewsky & Friederici 2003). In the main clause, the unambiguously nominative marked subject appears in final position and is preceded by auch. Assuming that auch is understood as being unstressed (a plausible assumption for function words), it associates with an expression that follows it (for a discussion of stressed v. unstressed auch, see Krifka 1999), here most naturally the subject (der Mann), which yields the presupposition that someone else had seen the woman. This presupposition is not satisfied on the syntactically preferred subject-initial interpretation of the relative clause. However, the syntactically dispreferred object-initial interpretation of the relative clause (that the girl saw the woman) does satisfy this presupposition. The task for the participants then was to choose a paraphrase that best matched their understanding of the sentence. The paraphrases for (4) would have been ‘The man and the girl saw the woman’ and ‘The woman saw the girl and the man saw the woman’. This choice between paraphrases amounted to a choice between the syntactically preferred interpretation and the interpretation on which the presupposition of auch was satisfied. As a control condition, the same sentence was used 4

Here and below, the passive is only used in the English paraphrase to keep the word order similar to the German one. Note that the sentences given here are only used for illustration purposes and were not used in the actual studies. Samples of the experimental materials are provided in Appendix.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

(4) Die Frau, die das Ma¨dchen sah, hatte The womanN/A whoN/A the girlN/A saw had auch der Mann gesehen. also the manN seen ‘The woman that (saw the girl/the girl saw) had also been seen by the man.’4

Florian Schwarz 381

except that auch was replaced by vorher (here best translated as earlier), which does not introduce any presupposition whose satisfaction depends on the interpretation of the relative clause. Two further conditions followed the same basic idea, but had the order of the clauses reversed, with auch appearing in the relative clause. An example is given in (5).

In this case, the matrix clause is ambiguous, and the relative clause contains auch. Note that this time the noun phrase den Mann ‘the man’ in the relative clause is unambiguously marked accusative, so that the clause can only mean that the girl saw the man. As above, the ambiguous clause had a syntactic parsing preference for a subject-initial interpretation, whereas the dispreferred object-initial interpretation satisfied the presupposition introduced by auch (that the girl saw someone else apart from the man). A control condition was again constructed by replacing auch by vorher. Finally, a fifth condition was included, which was identical to the previous one, except that all noun phrases were ambiguously case marked: (6) Die Frau sah das Ma¨dchen, das auch die The womanN/A saw the girlN/A whoN/A also the Lehrerin gesehen hatte. teacherN/A seen had (i) ‘The woman saw the girl that had also seen the teacher.’ (ii) ‘The woman was seen by the girl that had also seen the teacher.’ (iii) ‘The woman saw the girl that had also been seen by the teacher.’ As a result, the sentence was three-way ambiguous.5 Two of the readings satisfied the presupposition of auch [namely (ii) and (iii)], but 5

In principle, there even is a fourth reading on which both clauses are interpreted as being objectinitial. But since that reading does not satisfy the also presupposition, it is unlikely that this reading will come to mind.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

(5) Die Frau sah das Ma¨dchen, das auch den The womanN/A saw the girlN/A whoN/A also the Mann gesehen hatte. had manA seen ‘The woman saw the girl that had also seen the man.’ or ‘The woman was seen by the girl that had also seen the man.’

382 Processing Presupposed Content

3.1.2 Results The results were analysed with the percentage of the type of paraphrase chosen as the dependent variable, where the paraphrases corresponded to either the subject-initial interpretation or the object-initial interpretation. The mean percentages of how often the object-initial paraphrase was chosen are 57% and 28% in the relative clause auch and vorher conditions, respectively, and 17% and 6% in the matrix clause ones. Figure 1 illustrates this. The object-initial interpretation was chosen more frequently in the auch conditions (A and C) than in the corresponding control conditions with vorher (B and D). It was also chosen more frequently in general for the relative clause-first order than for the matrix clause-first order. A 2 3 2 analysis of variance (ANOVA) (auch v. vorher and relative-first v. matrix-first) was performed. There was a main effect of auch [F1(1,89) ¼ 112.3, p < 0.001; F2(1,29) ¼ 277.2, p < 0.001] and a main effect of clause type [F1(1,89) ¼ 183.3, p < 0.001; F2(1,29) ¼ 92.1, p < 0.001]. There was also an interaction between the two factors [F1(1,89) ¼ 30.7, p < 0.001; F2(1,29) ¼ 37.2, p < 0.001]. Two-tailed t-tests were carried out to test for simple effects of auch for the two types of clause orders. Both effects were significant (condition A v. B: t1(89) ¼ 10.3, p < 0.001; t2(29) ¼ 13.2, p < 0.001, condition C v. D: t1(89) ¼ 5.4, p < 0.001; t2(29) ¼ 7.3, p < 0.001). This shows

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

differed in whether the matrix clause or the relative clause was interpreted as being object-initial. Therefore, the results for this condition provide a further perspective on the differences between the first two pairs of conditions. The set-up resulted in a 2 3 2 design (plus the fifth condition, which was treated separately), with the presence or absence of auch as the first factor and clause order as the second factor. For the questionnaire, 30 sentences were constructed with versions for each of the five conditions above. Five versions of the questionnaire were created, varying sentences across conditions, so that each list contained six sentences per condition, resulting in a counterbalanced design. The questionnaire was created in HTML and made available online. The sentences were followed by disambiguated paraphrases and participants were asked to choose the paraphrase that matched their initial understanding of the sentence or their preferred interpretation of the sentence if more than one reading was possible. In addition to the experimental items, there were three items similar to the experimental ones, but preceded by a short text. Also, there were 20 unrelated filler items. Altogether, 90 native speakers of German completed the questionnaire.

Florian Schwarz 383

that the differences between the auch and vorher conditions are significant for each of the clause orders. In the three-way ambiguous fifth condition, the paraphrase corresponding to the object-initial interpretation of the relative clause was chosen 43% of the time and the paraphrase corresponding to the object-initial interpretation of the matrix clause was chosen 8% of the time. The syntactically preferred subject-initial interpretation of both clauses was chosen 49% of the time. During the initial inspection of the data, the percentage of objectinitial interpretations seemed to be higher in the later parts of the questionnaire. To test whether there was a significant increase, post hoc regression analyses with order position as a factor were carried out. Since the two clause orders varied substantially in how often the objectinitial paraphrase was chosen, this was done separately for the two auch conditions. There was no significant correlation between order position and the percentage of B readings for the relative clause condition (r ¼ .065, B ¼ .1%, P ¼ 0.73). For the matrix clause condition, on the other hand, there was a significant correlation between order position and percentage of B readings (r ¼ .544, B ¼ .6%, P < 0.01). The control conditions without the presupposition patterned with the relative clause presupposition condition and did not display any significant correlation between order position and percentage of B readings. To test whether there actually was an interaction between the relative clause and matrix clause auch conditions

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

Figure 1 Percentage of object-initial paraphrases per condition.

384 Processing Presupposed Content with respect to order position, the percentages for the B readings were converted into z scores to control for differences in variability found in the two conditions. Regressing the z scores of the percentage of B readings on order, sentence type, and order position 3 sentence type yielded a marginally significant interaction coefficient (B ¼ .054, P ¼ 0.057). We can thus conclude with fairly high certainty that the relative clause and matrix clause conditions do differ in the way order position affects the percentage of B readings chosen, which indicates that the two differ in the presence of practice effects.

6 This description assumes a non-parallel parsing architecture. I briefly discuss the relevance of the present studies to this issue of parsing architectures in section 4.3

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

3.1.3 Discussion The results from the questionnaire study clearly show that participants’ choice of paraphrase is influenced by the presupposition introduced by also. When it is present, as in conditions A and C, the otherwise dispreferred object-initial paraphrase is chosen more frequently than when it is not, presumably because this order yields the also presupposition satisfied. This effect is present and significant for both clause orders, but stronger in the relative clause-first order. Altogether, the object-initial paraphrase is chosen more frequently in the relative-first order. This, together with the statistical interaction, suggests that the effect of the presupposition interacts with other parsing factors. Such an interaction is highly relevant to the question mentioned in the introduction of how exactly incremental interpretation takes place and what contribution presupposed content might make to the process of assigning meaning to linguistic input as it is being parsed. One way of describing the process that readers might go through in reading these sentences is that they first commit themselves to a subject-initial interpretation of the ambiguous clause and then reanalyse that clause once they see that this renders the presupposition of also satisfied.6 While this reanalysis is fairly unproblematic in the case of the ambiguous relative clause, it is most likely harder and involves at least one additional confounding factor in the matrix clause: interpreting the sentence-initial noun phrase as the object requires a special interpretation (e.g. as a topic), which is not supported by anything in the context. Therefore, it is altogether harder and less likely that participants will end up with the object-initial interpretation for the matrix-first order, and the effect of the presupposition is smaller in the condition with this order of clauses. An interesting further result

Florian Schwarz 385

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

in the statistical analysis is that there was a practice effect reflected in a significant correlation between the percentage of object-initial paraphrases chosen for the matrix-first order and the order position of the sentence within the questionnaire. For the relative-first order, there was only a small numerical increase throughout the questionnaire that was not significant. This supports the conclusion made above that it is harder to get the object-initial interpretation in the matrix-first order. Apparently, participants become more likely to choose the objectinitial interpretation after having been exposed to a number of these constructions and paraphrases for the matrix-first order, whereas they start out at a fairly high level for the other clause order. The results from the three-way ambiguous fifth condition are also important in a number of ways. First they support the point made at the end of the last paragraph, since they show that what is behind the object-initial paraphrases being chosen less often in the matrix clausefirst condition really is that the matrix clause has to be reanalysed. In the three-way ambiguous condition, either clause could have been given the object-initial interpretation in order to satisfy the also presupposition. But again, we find a strong asymmetry between the relative clause and the matrix clause, with 43% object-initial paraphrases chosen for the relative clause and only 8% object-initial paraphrases for the matrix clause. This asymmetry shows that the differences between the matrix-first and the relative clause-first conditions are not due to parallelism, as one might be tempted to hypothesize, since the objectinitial interpretation of the relative-first conditions results in both clauses having the same order, whereas the matrix-first conditions have non-parallel orders on that interpretation. Furthermore, the asymmetry helps to fend off another alternative hypothesis, namely, that the higher percentage in object-initial interpretations for the relative-first order is due to the obligatory object-initial interpretation of the matrix clause. But since the object-initial paraphrase of the relative clause was chosen so frequently in the three-way ambiguous condition, where no such obligatory object-initial interpretation was present, this explanation does not seem promising. In sum, then, we have found that both the presupposition of also and the type of clause that is ambiguous have a great impact on the choice of paraphrase. The interaction seen between the effect of the presupposition and other parsing factors related to the differences between relative clauses and matrix clauses can be taken as a first indication that the evaluation of presuppositions with respect to their context takes place in online processing, although we need to be cautious in drawing any firm conclusions in this regard from an offline

386 Processing Presupposed Content questionnaire study. The experiment reported in section 3.2 attempts to address this issue in a more direct way.

3.2 Self-paced reading study on auch

(7) a. Die Frau,/ die der Junge sah,/ hatte auch der The womanN/A whoN/A the boyN saw had also the Mann gesehen. manN seen ‘The woman that the boy saw had also been seen by the man.’ b. Die Frau,/ die den Jungen sah,/ hatte auch der The womanN/A whoN/A the boyN saw had also the Mann gesehen. manA seen ‘The woman that saw the boy had also been seen by the man.’ In the sentence in (7a), the noun phrase in the relative clause (der Junge ‘the boy’) is unambiguously marked nominative, which results in the clause being object-initial and meaning that the boy saw the woman. The main clause contains auch, which (again assuming that it associates with der Mann ‘the man’) introduces the presupposition that someone else saw the woman. Given the meaning of the relative clause, this presupposition is satisfied. In (7b), on the other hand, the noun 7 As before, this example is only used for illustrative purposes. See Appendix for a sample of the actual materials used in the experiment. The slash indicates the frame breaks between the parts of the sentence that were displayed at one time in the moving-window display (this is described in more detail below).

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

3.2.1 Methods and materials The second experiment used the selfpaced reading method to investigate the effect of presuppositions on the time people spend reading the relevant parts of the experimental sentences. For this study, the basic strategy was to present morphosyntactically unambiguous versions of the materials in the first experiment, which varied in whether the presupposition of also was satisfied or not. To disambiguate the sentences, masculine, rather than feminine or neuter noun phrases were used in the critical positions, so that the case marking on the definite article was unambiguously nominative (der) or accusative (den). Since the effect in the questionnaire was larger for the relative-first order, sentences using this order were used for the online study. An example illustrating the set-up of the experimental items is given in (7).7

Florian Schwarz 387

8

For examples of the actual questions, see the materials in Appendix.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

phrase den Jungen ‘the boy’ is unambiguously marked accusative, so that the relative clause is subject-initial and can only be understood as the woman seeing the boy. The presupposition of the main clause is as in (7a), and is therefore not satisfied by the relative clause. If we found any reading time effects related to whether or not the presupposition is satisfied (and which did not show up in the controls), this would tell us that information about presupposition satisfaction has to be available to the processor at that time, and hence that any semantic processes necessary to determine presupposition satisfaction must have already taken place. As in experiment 1, control conditions were constructed by replacing auch with vorher. As before, this resulted in a 2 3 2 design, with the presence or absence of auch as the first factor and subject-initial v.object-initial structures as the second factor. The experiment included 24 sentences with versions in each of the four conditions. The sentences were counterbalanced across conditions in four lists. Participants only saw each sentence in one condition. The experiment was programmed using E-Prime software. The presentation order of the items was randomized. Sentences were presented using the moving-window technique. On the first screen, all characters were replaced by underscores. Participants had to press the space bar to see the first part of the sentence. When they pressed the space bar again, the first part was replaced by underscores, and the next part of the sentence was displayed. Reading times were recorded for each displayed phrase. After each sentence, a yes–no question about that sentence was presented, and participants had to push ‘s’ to answer ‘yes’ and ‘k’ to answer ‘no’. Half of the questions asked about the relation in the relative clause (‘Did the boy see the woman?’8) and the other half about the relation in the matrix clause (‘Did the man see the boy?’). Overall, half of the questions had ‘yes’ as a correct answer and half of them ‘no’. For the relative clause questions, the correct answer varied across conditions, since the relation depended on the experimental manipulation of subject- v. object-relative clauses. Both the responses and the response times were recorded. Apart from the experimental items, there were 72 items from unrelated experiments and 12 from a related experiment. Furthermore, there were 12 filler items. Subjects were instructed that they were going to read sentences on the screen and that they had to answer short questions about them, which did not necessarily have right or wrong answers. They also were told to answer questions with ‘yes’ only if this

388 Processing Presupposed Content followed directly from the sentence in question and that they had to press the ‘s’ key for ‘yes’ and the ‘k’ key for ‘no’. On average it took about 30 minutes to complete the experiment. In total, 20 native speakers of German participated in the experiment.

Figure 2 Reading time on final clause in milliseconds.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

3.2.2 Results The measure of most interest was the reading times on the clause containing auch (or vorher). Their means were 3555 and 4911 ms in the object-initial conditions with auch and vorher, respectively, and 5469 and 4480 ms in the subject-initial ones. They are illustrated in Figure 2. When auch was present, the reading time in the object-initial condition A (where the presupposition of auch was satisfied) was almost two seconds faster than in the subject-initial condition C (where the presupposition was not satisfied). When auch was replaced by vorher, the subject-initial condition (D) had a small advantage over the objectinitial condition (B). Interestingly, the auch phrase was read almost 1.5 seconds faster than the vorher phrase in the object-initial condition (A v. B), but roughly one second slower in the subject-initial condition (C v. D). A 2 3 2 ANOVA revealed an interaction between the two factors [F1(1,19) ¼ 26.00, P < 0.001; F2(1,23) ¼ 17.81, P < 0.001]. In addition, there was a main effect of order (subject-initial v. objectinitial) [F1(1,19) ¼ 11.58, P < 0.01; F2(1,23) ¼ 7.88, P ¼ 0.01], which

Florian Schwarz 389

9 In light of the fairly low accuracy rates, an anonymous reviewer suggested to also analyse the data by only looking at data points from sentences to which the subjects had responded correctly. The overall pattern of the data looked very similar: condition A: 3582 ms, B: 4916 ms, C: 5079 ms, D: 4648 ms. The interaction was significant [F1(1,19) ¼ 9.85, P < 0.01; F2(1,23) ¼ 10.80, P < 0.01]. The main effect of order was only marginally significant by subjects, but significant by items [F1(1,19) ¼ 3.73, P ¼ 0.07; F2(1,23) ¼ 6.24, P < 0.05]. The simple effect comparing conditions A and C was still present, [t1(19) ¼
3.34, P < 0.01; t2(23) ¼
3.90, P ¼ 0.001], as was the simple effect comparing A and B [t1(19) ¼
3.78, P ¼ 0.001; t2(23) ¼
4.59, P < 0.001]. 10 After removal of eight outliers that were over 3 standard deviations (SDs) from the mean of each condition (RTs over 10 s).

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

was dominated by the interaction. A number of t-tests were also carried out to test for simple effects of auch v. vorher and object-initial v. subject-initial relative clauses separately. The difference between conditions A and C was significant [t1(19) ¼
6.49, P < 0.001; t2(23) ¼
4.58, P < 0.001], which shows that there was a simple effect of subject-initial v. object-initial structures in the auch-conditions. There also was a significant difference between A and B [t1(19) ¼
4.72, P < 0.001; t2(23) ¼
5.03, P < 0.001], that is, a simple effect of auch in the object-initial conditions. The difference between C and D was significant by subject and near significant by items [t1(19) ¼ 3.07, P < 0.01; t2(23) ¼ 1.96, P ¼ 0.06], but the difference between B and D was not significant [t1(19) ¼
1.28, P ¼ 0.22; t2(23) ¼ 1.25, P ¼ 0.23]. In terms of the statistical analysis, then, the main results are the interaction between the two factors and the simple effect of the order of subject and object in the relative clause. The simple effect of auch in the object-initial conditions is of interest as well, but its interpretation is less clear as it could in principle be due to a lexical effect involving auch and vorher.9 Taken together, these results show that the reading times in the auch conditions were strongly influenced by subject-initial v. object-initial order (corresponding to whether the presupposition of auch is satisfied or not), while the reading times in the vorher conditions were only slightly influenced by this factor, and in the opposite direction. The data for the relative clause region were analysed as well to provide a comparison with the effects in the auch region. The reading times by condition were as follows:10 condition A 3615 ms, condition B 3776 ms, condition C 3648 ms and condition D 3429 ms. A 2 3 2 ANOVA did not find any significant effect. As additional measures, the response times and the accuracy rates for the yes–no questions following the display of the sentence were also analysed. In the response times, there was a main effect of order, with the object-initial conditions having roughly an advantage of one second over the subject-initial conditions [object-initial: 3885 ms,

390 Processing Presupposed Content

11 The effect of order on response times was also significant when including only the data for correct responses [object-initial: 3633 ms, subject-initial: 4754 ms; F1(1,19) ¼ 18.45, P < 0.001; F2(1,23) ¼ 12.81, P < 0.01]. 12 Analyses involving this factor as well as the question type factor below were only done by subjects, since the levels of these factors were not varied systematically within items. 13 The numbers and results for the accuracy rates reported here differ from those previously reported in Schwarz (2006), because a error was discovered during a re-examination of the data.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

subject-initial: 4960 ms, F1(1,19) ¼ 16.4, P ¼ 0.001; F2(1,23) ¼ 16.41, P < 0.001].11 There was no significant interaction and no other significant main effect. Response times were faster when the correct response was ‘yes’ (3955 ms) than when it was ‘no’ (4996 ms) [t1(19) ¼ 2.54, P < 0.05]12 and correct answers (4134 ms) were faster than incorrect answers (5878 ms) [t1(1,19) ¼ 4.08, P ¼ 0.001; t2(1,23) ¼ 3.80, P ¼ 0.001]. The overall mean accuracy rate was 81.25%.13 There was a main effect of order [F1(1,19) ¼ 7.69, P < 0.05; F2(1,23) ¼ 5.11, P < 0.05], with means of 86.25% for the object-relative clause conditions and 76.17% for the subject-initial ones. There was no significant interaction and no other significant main effect. Accuracy in the object-relative clause condition with auch (85%) was higher than in the subject-initial one (73%) [t1(19) ¼ 2.67, P < 0.05; t2(23) ¼ 1.94, P ¼ 0.07], which indicates a simple effect of presupposition satisfiction, with higher accuracy rates when the presupposition was satisfied. There was a numerical difference between questions that asked about the relative clause (78%) and those that asked about the matrix clause (85%), which could simply be due to recency of the phrase asked about. Accuracy was the lowest when both of these last two factors were considered in combination, namely in the subject-relative clause condition with auch when the question was about the relative clause (67%). It was highest, on the other hand, in the object-relative clause conditions when the question was about the matrix clause (91.5%), in which case the presupposition is satisfied and the question is about the most recently seen part of the sentence. Looking at order, auch, and question type together in a three-way ANOVA, there was a marginally significant three-way interaction [F(1, 19) ¼ 4.04, P ¼ 0.06]. This suggests that some of the questions were particularly hard in certain conditions and that the relatively low overall accuracy rates were predominantly due to these combinations of questions and conditions. Whether the correct response was ‘yes’ or ‘no’ did not alter accuracy rates significantly (‘yes’: 82%, ‘no’: 80%). In summary, the question response times and accuracy rates did not display the auch 3 order interaction, but were predominantly affected by order. This

Florian Schwarz 391

might well be due to the fact that it was easier to keep the relations in the relative clause and the matrix clause straight when they were parallel (the woman was seen by the boy and the man) than when they were not (the boy was seen by the woman and the woman was seen by the man). In terms of simple effects, both recency of the phrase asked about and the satisfaction of the auch presupposition affected accuracy.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

3.2.3 Discussion The results from the self-paced reading study clearly show that the reading time on the final clause containing also was substantially affected by whether the presupposition of also was satisfied or not. This is not merely an effect of parallel order in the two clauses, as the effect was reversed in the earlier conditions, in which no relevant presupposition interfered. Interestingly, this effect was not reflected in the accuracy rates or the question response times, which only exhibited main effects but no interaction of also and order. Nonetheless, a simple effect of presupposition satisfaction showed up in the accuracy rates. The effect of the presupposition is rather large, at almost two-second difference between conditions A and C. It is very likely that this is due to the relatively demanding task, especially in certain conditions, of answering the yes–no questions that followed the display of the sentence. Almost all subjects reported that it was often quite difficult to keep in mind who did what to whom amongst the three people talked about in each sentence. When the presupposition did not match the content of the relative clause, it must have been even harder to keep this information straight, and this may have caused rather substantial delays when reading the final part of the sentence. The simple effect of presupposition satisfaction on the accuracy rates supports this as well. In connection with this, one particularly telling comment made by a participant after the experiment was that she thought there were some spelling mistakes in the sentences, especially with respect to the case marking on noun phrases (e.g. der Mann rather than den Mann). Apparently, the expectation raised by the presupposition of also was so strong that the mismatch was perceived as a mistake. One thing that is remarkable about this is that when sentences like those in the unsatisfied also condition are seen out of context and without the question, they do not stand out much at all (the reader can make his/her own judgment about the corresponding English examples in experiment 3). It thus seems like the presence and nature of the questions contributed substantially to the slow reading times and large effect sizes. The strong effect on the reading time suggests that the presupposed content is evaluated online. This lends further support to the speculative

392 Processing Presupposed Content

3.3 Self-paced reading study on also 3.3.1 Methods and materials In order to address the issues with the German self-paced reading study raised above, a follow-up study was undertaken in English. The additive particle chosen for this study was also, rather than too, in order to allow for a similar paradigm where other adverbials could replace also in the control conditions. The main new features introduced in this study were that the critical region was 14 Part of the reason for this may be that a large number of object-initial structures were presented throughout the various experiments included in the self-paced reading study.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

conclusion above that the results from the questionnaire study are based on online effects of presuppositions. This finding is consistent with the results from previous studies on the presuppositions of definite descriptions that were mentioned above (e.g. Crain & Steedman 1985; van Berkum et al. 2003). An additional point of interest here is that the reading times for the clause containing also, preceded by the relative clause that satisfied the also presupposition (A), were faster than the reading times for the same clause with earlier preceded by the same relative clause (B). However, it is possible that this is simply a lexical effect of also compared to earlier. If this difference turned out to be real, it could be taken to tell us something interesting about the role of presupposed content in natural language. The advantage of the also condition might be that the presupposed content facilitates the integration of new content into the contextual representation by connecting new and old information. Since this effect was not replicated in experiment 3 discussed below, we should be careful not to overinterpret the effect at this point. While the results in general reinforce the conclusion that there are online effects of presuppositions, a number of questions remain open that might undermine the interpretation of these results to some extent. First, the critical region was the final region, which makes it impossible to distinguish between online effects during the actual reading and potential sentence-final wrap-up effects. Secondly, the rather slow reading times and the large effect size, together with the rather low accuracy rates in some conditions, give rise to the possibility that the effects found are due to the task demands of answering the questions. Another worry in this direction is that the well-documented subject-relative clause advantage did not show up significantly in the relative clause reading times.14 Finally, the possibility of the also v. earlier advantage being a lexical effect keeps us from drawing any strong conclusions in this respect.

Florian Schwarz 393

non-final, that no questions were asked (although there were filler items and items from other studies for which questions had to be answered), and that a range of different adverbs were used for the control conditions (e.g. just, once, almost, recently, now). An example from the materials is provided below.15

In the English set-up, unlike in the German set-up, the matrix clause is subject-initial, which means that it is now the subject-initial relative clause (a) that satisfies the presupposition of also. One additional manipulation was introduced in this experiment: whereas half of the items had the same verb in the relative clause and in the matrix clause, the other half had two different verbs in the two clauses. These two verbs were more or less synonymous and were all chosen in such a way that the verb in the relative clause implied the verb in the matrix clause (at least in the specific usage in the sentence). An example is given in (9). (9) a. The lawyer/ who contacted Allison/ will also get in touch with her neighbors/ to discuss the problems/ with the new zoning law. b. The lawyer/ who contacted Allison/ will later get in touch with her neighbors/ to discuss the problems/ with the new zoning law. c. The lawyer/ who Allison contacted/ will also get in touch with her neighbors/ to discuss the problems/ with the new zoning law. d. The lawyer/ who Allison contacted/ will later get in touch with her neighbors/ to discuss the problems/ with the new zoning law. Apart from making the materials more diverse and more natural, this served as a first attempt to shed light on an important question about the properties of the levels of representation at which the processes studied here take place. The intuitive idea is that seeing that the presupposition of also is satisfied might be easier when the very 15

See Appendix for more examples from the materials.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

(8) a. The congressman/ who wrote to John/ had also written to the mayor/ to schedule a meeting/ for the fundraiser. b. The congressman/ who wrote to John/ had just written to the mayor/ to schedule a meeting/ for the fundraiser. c. The congressman/ who John wrote to/ had also written to the mayor/ to schedule a meeting/ for the fundraiser. d. The congressman/ who John wrote to/ had just written to the mayor/ to schedule a meeting/ for the fundraiser.

394 Processing Presupposed Content

16 But note that this is by no means a necessary feature of any DRT account, and hence a negative or non-confirming outcome would not be evidence against DRT.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

same verb appears in both clauses than when there are two different verbs that are in an implicational relation. In the former case, it can be read off the surface structure (or a representation very close to it) that the presupposition is satisfied, while in the latter case, additional inferences have to be made to the extent that, speaking in terms of the example above, ‘contacting Allison’ will do in order to satisfy the presupposition that the lawyer got in touch with someone else apart from the neighbours. In DRT approaches to presupposition (van der Sandt 1992; Geurts 1999), the process of evaluating presuppositions makes reference to formal properties of DRSs. It might be possible, then, to decide whether or not an anaphoric presupposition is satisfied by only considering a formal level of representation that is close to the surface structure (i.e. a level where the relationship between ‘contact’ and ‘get in touch’ is not transparent).16 In the case of the same-verb condition, it would be possible to determine that the presupposition of also is satisfied based on such a representation, since the predicates would be formally identical [say, contact(x)(y)]. But in the different-verb condition, this would not be possible, because it can only be determined based on the meaning of the two verbs that the also presupposition is satisfied. From a processing perspective, checking whether the presupposition is satisfied should then be computationally cheaper in the same-verb condition than in the different-verb condition. This is the case at least, for the conditions where the presupposition is satisfied. If it is not, the processor presumably would go through the same steps in both cases, trying every possibility to get the presupposition satisfied. This hypothesis would therefore predict an interaction between the presupposition effect familiar from experiment 2 and the same-verb v. different-verb factor. While it would be easy to capture such an effect in DRT, it would not be expected for a theory like File Change Semantics and other dynamic theories that do not make reference to representational structures. Hence, such a result would strongly support an approach to presupposition resolution processes that involve reference to formal properties of representations, like DRT. The procedures for this experiment were the same as in experiment 2, except that no questions were asked after the experimental items. The study included 60 items from three unrelated experiments as well

Florian Schwarz 395

as 40 filler items. Most of these other items were followed by a question. A total of 48 undergraduates from the University of Massachusetts Amherst participated in the experiment.

17 Outliers that were over 3 SDs from the mean of their condition were excluded from the analyses. This removed 2.2% of the data points.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

3.3.2 Results As in experiment 2, the main interest was in the reading times on the region containing also. The means are shown by condition in Table 1.17 A 2 3 2 ANOVA was performed on the data. There was a significant interaction [F1(1,47) ¼ 4.55, P < 0.05; F2(1,23) ¼ 5.57, P < 0.05] and a main effect of order [F1(1,47) ¼ 15.61, P < 0.001, F2(1,23) ¼ 12.19, P < 0.05]. Finally, there was a main effect of also that was significant by items [F1(1,47) ¼ 2.33, P ¼ 0.134; F2(1,23) ¼ 4.92, P < 0.05]. Turning to simple effects, the subject-relative clause condition with also was faster than the object-initial one [t1(47) ¼ 4.16, P < 0.001; t2(23) ¼ 3.97, P ¼ 0.001]. The only other significant simple effect was comparing the object-relative clause condition with also to the one with another adverb [t1(47) ¼ 2.31, P < 0.05; t2(23) ¼ 3.39, P < 0.05]. In summary, we find the same interaction as in experiment 2, as well as main effects that are dominated by the interaction. The reading times for the same- and different-verb conditions are presented in Table 2. In numerical terms, the two-way interaction seems to be present in both conditions, with the advantage of condition A over C being bigger than the advantage of B over D. Including the factor of verb sameness in the analysis by running a three-way ANOVA (order 3 also 3 verb) did not yield a significant three-way interaction [F1(1,46) ¼ .34, P ¼ 0.56; F2(1,22) ¼ .37, P ¼ 0.55]. Reading times were slightly higher in the different-verb conditions, which was reflected in a main effect of verb sameness that was significant by items and marginally significant by subjects [F1(1,46) ¼ 3.44, P ¼ 0.07; F2(1,22) ¼ 4.47, P < 0.05]. The main effects of order and the order 3 also interaction were also significant. There were no other significant effects. In terms of the hypothesized stronger effect in the same-verb condition, the numerical results go in the opposite direction of what we would expect if the representations would facilitate the process of checking whether the presupposition is satisfied. While the difference between the two also conditions is slightly bigger in the same-verb conditions than in the different-verb conditions (227 v. 196 ms), the difference in the corresponding control conditions patterns the other

396 Processing Presupposed Content Condition

A

B

C

D

RTs in ms

1601

1633

1821

1692

Table 1 Reading times on also region

Condition

A

B

C

D

Same verb: RTs in ms Different verb: RTs in ms

1540 1635

1554 1674

1767 1829

1682 1693

way (128 v. 19 ms), so that the advantage of the satisfied also presupposition is actually bigger in the different-verb condition when viewed relative to the control conditions. Turning to the relative clause region, the subject-relative clauses (1382 ms) were read significantly faster than the object-relative clauses (1667 ms), which was reflected in a main effect of order [F1(1,47) ¼ 15.97, P < 0.001; F2(1,23) ¼ 19.18, P < 0.001]. No other effects were significant. This effect illustrates the well-known advantage of subjectrelative clauses over object-relative clauses. Finally, since there were two additional regions following the also region, we should also look at the reading times for the region immediately following the one with also. The mean reading times are shown in Table 3.18 A 2 3 2 ANOVA did not find a significant interaction. There was a main effect of order [F1(1,47) ¼ 3.76, P ¼ 0.06; F2(1,23) ¼ 5.58, P < 0.05]. The only significant simple effect was between the subject-initial and the object-initial also conditions [t1(47) ¼ 2.12, P < 0.05; t2(23) ¼ 2.08, P < 0.05]. Thus, although there seems to be some spillover from the also region yielding this simple effect, the bulk of the effect we are looking at is confined to the region containing also. 3.3.3 Discussion Experiment 3 avoided some of the shortcomings of experiment 2, discussed above. Although no specific task other than reading the sentences was performed, we found the same interaction effect as before. The reading times in experiment 3 were much shorter than in experiment 2, and the effect between the subject-relative clause and the object-relative clause also conditions is in the order of 200 ms. 18

Again, outliers over 3 SDs from the condition means were excluded from the analyses.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

Table 2 Reading times on also region

Florian Schwarz 397 Condition

A

B

C

D

RTs in ms

890

904

969

931

Table 3 Reading times on region after the also region

4 THEORETICAL IMPLICATIONS Ideally, results from psycholinguistic studies can contribute to theory in two directions, which correspond to the following two questions: what do the results tell us about (the relevant part of) linguistic theory, and what can we learn from them with respect to processing theories? I will focus on the implications for semantic theory, which I turn to in

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

This is about the same size as the subject-relative clause advantage found in the relative clause region, which was around 250 ms. Furthermore, the effect in its full form showed up only on the region containing the also, with minimal spillover to the following region. This excludes the possibility that we are only dealing with sentence-final wrap-up processes. Finally, the difference between the subject-relative clause also and adverb conditions (corresponding to the object-relative clause conditions with auch and vorher) was not replicated in experiment 3, which suggests that in experiment 2 this difference reflected a lexical effect or that it was task specific in that it was helpful in answering the questions asked after each sentence. With respect to the newly introduced factor, which varied between having identical and different verbs in the relative and matrix clauses, no relevant interaction effect of verb sameness could be determined. Assuming that presupposition resolution can take place on a representational level close to the surface structure, we would have expected a stronger effect in the same-verb conditions. But there is no evidence of it being harder to see that the presupposition of also was satisfied in the different-verb condition than in the same-verb condition. Numerically the reading times patterned in the opposite way, with a larger advantage of the satisfied also condition in the different-verb conditions (relative to the control conditions). However, these results are inconclusive since they are not significant. A number of theoretical issues arise in connection with the results of the experimental studies reported here, which in turn have the promise of providing new approaches for empirical research on presuppositions. I turn to these points in section 4.

398 Processing Presupposed Content sections 4.1 and 4.2. A few brief remarks about related processing issues are made in the final part of this section.

4.1 Implications for the semantics of additive particles

(10) a. Presupposition of also in general (Heim 1992)19 U alsoi [a]F presupposes xi 6¼ a in c & U(xi) b. The congressman who wrote to John had also written to the mayor . . . c. Presupposition of also in (a) (with focus on the mayor) kx. write(congressman, x) also [the mayor]F presupposes xi 6¼ the mayor in c & write(congressman, xi) The presupposition introduced by also, that there is another individual in the discourse context, apart from the mayor, to whom the congressman wrote, is satisfied by the information in the relative 19 For a recent formulation of this idea within DRT and discussion of some further issues, see van der Sandt & Geurts (2001).

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

First, let us turn to the implications that the results presented above have for the analysis of additive particles like also and auch. In experiment 1, the fact that the presupposition of also was not satisified on the syntactically preferred analysis of the sentence resulted in a significant increase of the percentage with which the paraphrase corresponding to the syntactically dispreferred analysis was chosen. And in the self-paced reading studies, there were significant delays in the reading times when the presupposition was not satisfied. These very strong effects of presupposition failure are relevant for the analysis of the presupposition of additive particles like too, auch and also. As was discussed in section 3, Kripke (1991) argued that too differed from many other presupposition triggers in that its presupposition strongly resists accommodation. Following Kripke, current proposals assume that their presupposition is at least partly anaphoric, which means that they require an antecedent of some sort in the discourse context (Heim 1992; van der Sandt & Geurts 2001). Let us take a closer look at the formulation of an anaphoric analysis of also in connection with the experimental materials. I will focus on the English examples from experiment 3 for ease of exposition, but the same points of course apply to the German studies. Heim’s analysis of too (adapted here for also) is provided in (10a). Note that the presupposed x has to correspond to a discourse referent that is already in the context. This is what captures the anaphoricity of also. In (10c), this analysis is applied to the example sentence for condition A.

Florian Schwarz 399

20

In experiment 2, it was the subject-initial relative clause condition, because the matrix clause was object-initial and hence required an object-initial relative clause in order for the also presupposition to be satisfied.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

clause, which states that the congressman wrote to John. Crucially, this requires that the matrix clause is evaluated with respect to a context that contains the information in the relative clause—this will be discussed in detail in section 4.2. For the moment, let us focus on the cases where the order in the relative clause is switched around (‘. . . who John wrote to . . .’), which results in there not being any individual in the discourse that the congressman had written to and the presupposition not being satisfied. These were the syntactically preferred analyses of the ambiguous sentences in the questionnaire study (experiment 1) and the object-initial relative clause condition C in experiment 3.20 In the former case, we found that this led to the paraphrase corresponding to the syntactically dispreferred analysis being chosen more often (which yielded the presupposition satisfied). In the latter, we found substantial delays in the reading times on the region containing also. Since the contrast in presupposition satisfaction between conditions A and C is the only relevant difference that is not also present in the control contrast between B and D, we can conclude that the reading time differences between A and C (that is not present between B and D) are due to this difference in presupposition satisfaction. These results are very much consistent with the claim that the presupposition of auch and also cannot be accommodated, and with accounts that take this aspect into consideration, for example the one above, for which van der Sandt & Geurts (2001) have argued that it is due to the anaphoric aspect of the presupposition of too that it cannot be accommodated. In the case of the ambiguous questionnaire items, the only way to save the sentence from presupposition failure, then, is to override the strong syntactic preference for the initial analysis of the ambiguous clause. And in the case of the unambiguous self-paced reading materials, the inability to accommodate leads to a substantial slow down in reading. On the other hand, the results are unexpected on earlier proposals, for example, the one by Karttunen & Peters (1979), which assumed that additive particles like too merely have an existential presupposition. It is hard to imagine how a merely existential presupposition, which should not be too hard to accommodate, could have such strong effects in processing, especially given that it has proven to be difficult to find effects of accommodation for other presupposition triggers like the definite article, which also is generally assumed to have an existential presupposition (for a recent discussion, see Frazier 2006).

400 Processing Presupposed Content While Kripke’s example in (3a) above already provides a strong argument against such a proposal, the difficulty of accommodating the presupposition of auch and also in the present experiments can be taken as further evidence against an existential account. In summary, the present experiments can be taken as evidence for the impossibility (or at least great difficulty) of accommodating the presupposition of auch and also. Consequently, they also provide support for accounts that can explain this property, such as the anaphoric account.

In this section, I turn to the issue of how the experimental results relate to the bigger issue of how the parser makes use of the semantic part of the grammar and what implications this might have for semantic theory in general. To begin with, let us briefly consider some general aspects of the relevant questions. I take it to be the null hypothesis that the processor makes use of the grammar when parsing linguistic input. It may have additional principles that help to rule out many of the grammatical analyses of the structure that might in principle be possible, but it certainly should make use of the grammatical system to exclude ungrammatical analyses. If we can conclude from experimental results that the processor has access to certain information for a given structure, then we can conclude that the grammatical system must function in a way that allows it to provide this information to the parser at that point and on the basis of the information available at the time. This is the general form of the line of argumentation taken below. To anticipate, I will argue that the processor evaluates the also clauses from the self-paced reading materials with respect to the preceding noun phrase (including the relative clause), which means that the semantic component of the grammar must have made it possible to integrate the content of that noun phrase into the representation of the context when it encounters the also clause. This means that the processing results constrain us in formulating grammatical theories in that they have to be compatible with the incremental steps found in presupposition interpretation in the reported experiments. We have seen in section 4.1 that the processing effects depend on whether or not the content of the relative clause in the experimental materials satisfies the presupposition of also in the matrix clause. As far as the processing perspective is concerned, the process of determining this seems to take place online, since the effect shows up in the reading time on the clause that contains the presupposition trigger.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

4.2 Implications for semantic theory

Florian Schwarz 401

(11) a. Let c be a context (here a set of assignment functions) and let p be an atomic formula, then, if defined : c + p ¼ {g : DOM(g) ¼ [ Dom( f ) s.t. f 2 c [ {i : xi occurs in p} & g is an extension of one of the functions in c & g verifies p} b. The Novelty/Familiarity Condition c + p is only defined if for every NPi that p contains, if NPi is definite, then xi 2 Dom(c), and if NPi is indefinite, then xi ; Dom(c). 21

For simplicity, I restrict the formal characterization of contexts to sets of assignment functions.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

This suggests that as one is reading the part of the sentence containing also, one is aware of the content of the relative clause (of course, that also matches our intuitive sense of what happens when we read). When we look at processing in terms of dynamic semantics, this is rather interesting: to evaluate the presupposition of also is to check whether the context entails it (and in the case of also, it also involves something like checking whether there is an appropriate discourse entity having the relevant property). Since the sentence is not at all problematic in any way (neither intuitively nor in terms of the reading time results), it seems to be the case that the content of the relative clause is already integrated into the representation of the context by the time the part of the sentence containing also is semantically processed. In other words, it looks as if the representation of the context has been updated with the sentence-initial noun phrase, including the relative clause, by the time the next part of the matrix clause is interpreted. Let us now turn to a more detailed analysis of what the grammar has to provide for the processor. If we think of updates of the representation of the context as only taking place at the level of a sentence or a full clause, we cannot explain how the initial noun phrase can satisfy the presupposition: if we try to apply the context change potential of the entire sentence to the neutral context, the update would fail, since the presupposition of also is not satisfied in the initial context (and no repair would work, since the presupposition of also cannot be accommodated). However, as I already mentioned in section 2, in the full version of File Change Semantics of Heim (1983b), contexts consist of sets of pairs of worlds and assignment functions, and noun phrases denote atomic propositions (and hence have complete context change potentials of their own). The meaning of definite and indefinite noun phrases is as in (11), with the difference between definite and indefinite ones being captured with the Novelty Condition in (11b).21

402 Processing Presupposed Content With denotations such as these, the progression of updates of the representations of the context for the sentences of condition A can proceed without a problem. First, the initial noun phrase is interpreted and its presupposition is evaluated with respect to the input context. It is not satisfied, but it can be accommodated without a problem. Next, the rest of the matrix clause is interpreted, and the presupposition of also is evaluated with respect to the local context. In this context it is satisfied, and the update can proceed smoothly. The semantic characterization of these steps is sketched in semiformal terms in (12).

This contrasts with condition C, where the order in the relative clause has been switched around, so that even after the initial noun phrase has been integrated into the representation of the context by the time the rest of the matrix clause is interpreted, the presupposition of also is not satisfied, and there is no chance to accommodate it, since the presupposition of also resists accommodation. The presence of this effect requires that the semantic analyses necessary for recognizing this contrast have been carried out by the time the also phrase is being read and interpreted. More specifically, the initial noun phrase, including its relative clause, must have been syntactically parsed and compositionally interpreted—the relation in the relative clause must have been fully understood by the time the presupposition of also is evaluated, since the A and C conditions only differ in the structure of the relative clause. In addition to these purely semantic steps of analysis, the noun phrase as a whole, being definite, needs to be accommodated prior to the evaluation of the also presupposition. The results from these studies thus provide insight into the timing of compositional semantic processing, including the evaluation and accommodation of presupposed content with respect to the sentenceinternal context. It is worth comparing this aspect of the present studies with previous work on definite descriptions and their presuppositions.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

(12) p : The congressman x that wrote to John q : x also wrote to the mayor a. c + p defined only if there is a unique congressman that wrote to John b. after accommodation: c + p ¼ {g: g verifies congressman(x) & write(x)( john)} ¼ c# c. c# + q defined only if there is a z 6¼ the mayor in c# & write(x)(z) defined, since the congressman wrote to John, hence c# + q ¼ {g: g verifies congressman(x) & write(x)(john) & mayor(z) & write(x)(z)}

Florian Schwarz 403

(13) a. My teacher works as a DJ too. b. Critics of science use it, too. Apparently, the noun teacher suffices to satisfy the presupposition that the relevant individual works as something else than a DJ. And the (admittedly slightly playful) example in (13b) can be understood with focus on critics, which introduces the presupposition that other people use science. The occurrence of science seems to make it salient enough that there are scientists who do science, so that the presupposition is satisfied. In addition to the level of the noun phrase, updates can, of

22 The example in (13b) was a headline in the Valley Advocate on 1 December 2005. Kai von Fintel (personal communication) suggests that this is to be understood with focus on use, with critics as the antecedent for too. While agreeing that that is a possible reading, I and several other people I have consulted find the reading discussed in the text at least as plausible.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

Recall the Crain and Steedman type of experimental design discussed in section 2. The effect found there concerned the evaluation of a definite description with respect to a context, consisting of the preceding sentence, where the uniqueness presupposition of the definite article was or was not satisfied. While this allows conclusions about the timing of the evaluation of the uniqueness presupposition, it does not reveal anything about the sentence-internal dynamics of interpretation. The studies presented here, on the other hand involve a more complex sentence-internal set-up, in which both a noun phrase with its relative clause and the relationship between also and its presupposition and the rest of the structure of the matrix clause have to be fully analysed and interpreted in order for the effects observed here to arise. They therefore contribute new insights to our understanding of how exactly the parser’s incremental interpretation proceeds and more specifically, of the timing of when presupposed content is evaluated and integrated into the representation of the discourse context. The more general picture that is evolving from this discussion is that in processing, the representation of the context is updated as soon as possible. Since noun phrases have context change potentials of their own, the processor can update the representation of the context as soon as it has been given a noun phrase. Further support for updates at this level comes from examples such as the following, where the presupposition of too is satisfied by a noun phrase which does not have any phrasal subpart as in the relative clause cases considered above.22

404 Processing Presupposed Content

23 Quick updates that take place whenever a propositional unit has been parsed might be part of the explanation for the surprising findings in Christianson et al. (2001), where subjects are reported to answer ‘yes’ 60% of the time to the question ‘Did Anna dress the baby’ after reading the sentence ‘While Anna dressed the baby baby spit up on the bed.’ This finding suggests that even though subjects revise their syntactic analysis of the garden-path structure, they hold on to the incorrect interpretation (that Anna dressed the baby) corresponding to the initial syntactic analysis of the first part of the sentence. 24 This is the case, at least, if we assume that the processor makes use of the grammar, as argued above.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

course, also occur at the level of the full clause (or any propositional level, for that matter).23 In summary, I conclude that the results of the experiments presented here contribute a new kind of evidence to the theoretical discussion. They show that the steps the processor goes through in interpreting a sentence, including its meaning and presuppositions, require a grammatical system for handling presupposed content that can be used to evaluate a presupposition with respect to a local context, such as the initial noun phrase in the experiments above.24 Dynamic theories, such as File Change Semantics or DRT, seem suitable to do this job, and thus are compatible with the experimental results. In addition to these considerations about the online studies, we should also note the relevance of the findings of the questionnaire study in this respect. Assuming a model of the syntactic parser that only pursues one structural analysis at the time, and given that there is independent evidence supporting a syntactic parsing preference for subject-initial clauses, we find a remarkable amount of effort put into reanalysis of the first clause in the questionnaire items. If that clause has initially been parsed as subject-initial, it must be revised in order to satisfy the presupposition. The fact that this revision is even considered indicates that the parser can in some sense see that the reversal of the syntactic roles of the subject and the object yields an interpretation that will just be of the right kind to satisfy the presupposition of too. With respect to this point, it seems crucial for the parser to have access to representations very much like the DRSs posited by DRT, since those would provide the parser with representations such as see(x, y), which might suffice to make the inverse of this relation accessible somehow. In connection with this, it is worthwhile noting that it seems to be fairly easy in general for the processor to invert relations when there is enough evidence, as was shown in recent work by Kim & Osterhout (2005). In this respect, we have a first bit of suggestive evidence, then, about what the relevant representations might look like. Needless to say, a lot of work needs to be done to relate more complex theoretical issues to processing results. One interesting

Florian Schwarz 405

question is what happens when a presupposition trigger like auch appears early on in a sentence, with the part that satisfies it following later on, as in the following sentence:25 (14) Auch der Mann sah die Frau, die saw theN/A woman whoN/A Also theN man das Kind gesehen hatte. had theN/A child seen ‘Also the man saw the woman that the child had seen.’ or ‘Also the man saw the woman that had seen the child.’

4.3 Implications for processing theories Let us now turn to some considerations about what the results reported here mean for a theory of semantic processing. At this point, we are not anywhere close to having a realistic idea of how compositional semantic processing takes place online. One central question, of course, is 25 Thanks to Francesca Panzeri and Hans Kamp for independently bringing my attention to this question.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

While it is unclear to me what exactly to expect in connection with this in terms of processing results, it is intuitively clear that there is a certain element of suspense in sentences like this, with a high expectation that the presupposition of auch will be satisfied by something that is coming later on in the sentence. Future work will hopefully be able to address questions related to current issues in the presupposition literature more directly, for example the ever pressing issue of local and global accommodation (see Heim’s work and for a recent critical position on local accommodation, van Rooy 1999). If the general approach pursued in this paper is on the right track, local accommodation becomes a very plausible mechanism from the viewpoint of processing. Another important issue, partly related to this, is the question of whether presuppositions are at heart semantic or pragmatic (Stalnaker 1974; Beaver 2001; Simons 2001; Abusch 2005). One might take the apparent automatic nature of presupposition processing to support a semantic view (at least for the presupposition of also), but that, of course, depends on how we deal with pragmatic phenomena in processing in general. Without being able to go into the details of these issues, I hope that the present findings will inspire further exploration of these topics from a processing perspective.

406 Processing Presupposed Content

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

at what point the processor actually goes through steps of semantic composition and at what point the content of the currently processed linguistic unit is integrated with the information present in the context (which crucially should involve the evaluation of presuppositions with respect to that context). A viable hypothesis can be constructed from what has been said here: apart from the level of full clauses, where we obviously are dealing with propositional units, updates also take place at the level of noun phrases. This amounts to a straightforward extension of dynamic semantics to the theory of processing. Whether or not this can be upheld, it is the simplest assumption that the processor makes use of the system supplied by the grammar, and it has the advantage of making predictions that should, at least in principle, be experimentally testable. Hopefully, this will also enable us to investigate further theoretical issues in presupposition theory in new ways. Apart from these issues related immediately to semantic processing, the studies might also contribute to more general architectural questions in processing theory. Let me just mention one particularly interesting point, namely, that the results from the questionnaire study are most likely problematic for a simple version of a parallel parsing architecture along the lines of the one proposed by Crain & Steedman (1985). The idea in this work is that when the processor deals with an ambiguous structure, it considers all possible structures at the same time, with some structures being filtered out by certain principles. One central principle that they assume to account for the data mentioned above in (1) is the principle of parsimony, which only keeps those interpretations that have the fewest presuppositions violated. One of the more intriguing aspects of the questionnaire study discussed here was the interaction of how often subjects would choose the syntactically dispreferred structure (to have the presupposition of also satisfied) with the order the clauses appeared in (which affected whether the matrix clause or the relative clause was ambiguous). If people were always considering both interpretations of the ambiguous clauses at the same time, and then choosing one of them based on which one had the fewest presupposition violations, we would expect that they would choose the reading on which the also presupposition is satisfied more often than they actually did (in the matrix-first condition with also, they chose it only 17% of the time, and even in the relative-first condition, they chose it only 57% of the time). Furthermore, we would not expect that the two clause orders would differ so drastically in this respect. Of course, we need to be cautious in drawing conclusions about online processing from the results of an

Florian Schwarz 407

offline study. Nonetheless, it is worth considering possible predictions that online accounts make for tasks in offline studies. And unless other factors can be identified that account for the differences between the relative and matrix-first conditions as well as the overall fairly low percentages for the readings where the also presupposition is satisfied, these effects are unexpected from the perspective of the framework assumed by Crain & Steedman (1985). Thus, the questionnaire results introduce an interesting question to be considered in this debate between different parsing architectures.

I have argued that the results from the studies reported here support analyses of additive particles like also and auch that take into consideration the impossibility of accommodating their presupposition. Furthermore, they suggest that the processor has access to and makes use of presupposed content in online processing and must at least be able to update representations of the context at the level of noun phrases. In a sense, this means taking the ‘dynamic’ aspect of dynamic semantics quite literally by claiming that the linguistic processor can update representations of the context in the process of interpreting a sentence compositionally. Crucially, the grammar has to provide the means to the processor for accessing the relevant levels of semantic representation at the right time, that is at a time where only parts of sentences, namely the initial noun phrase in the experimental materials, are available. Bringing our theoretical frameworks and processing theories closer together in this way has the advantage of being temptingly simple. Hopefully, this will lead to interesting new predictions that we can test in further work, and open up the possibility of extending the empirical foundation for work in theoretical semantics and of addressing central issues in presupposition theory that often involve disputes about the intuitive status of presupposed content. Investigating these issues in a more direct empirical way will make an important contribution to the theoretical discussion by providing evidence for the psychological reality of the theoretical notions in question. With a better understanding of what kind of effects related to presuppositions there are in processing, we can hope to address more sophisticated questions in presupposition theory (e.g. the issue of local and global accommodation) by employing psycholinguistic methods to collect empirical evidence.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

5 CONCLUSION

408 Processing Presupposed Content APPENDIX: MATERIALS

A. Sample of the auch-Questionnaire Materials

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

1. a/b. Das Marketingteam, das TheN/A marketing-team RPN/A die Abteilungsleiterin beraten hat, hatte theN/A department-head advised has, had (auch/vorher) der Gescha¨ftsfu¨hrer beraten. also/earlier theN CEO advised. ‘The marketing-team that the department head advised, had also been advised by the CEO.’ or ‘The marketing-team that advised the department-head had also been advised by the CEO.’ Paraphrases to choose from for (a) and (b): i. Die Abteilungsleiterin wurde the department-head was von dem Marketingteam beraten, und das by theD marketing-team advised and the Marketingteam von dem Gescha¨ftsfu¨hrer. marketing-team by theD CEO ‘The department-head was advised by the marketing-team and the marketing-team by the CEO.’ ii. Das Marketingteam wurde von der the marketing-team was by theD Abteilungsleiterin und vom Gescha¨ftsfu¨hrer beraten. department-head and by-the CEO advised ‘The marketing-team was advised by the department-head and by the CEO.’ c/d. Das Marketingteam beriet die marketing-team advised theN/A TheN/A Abteilungsleiterin, die department-head RPN/A (auch/vorher) den Gescha¨ftsfu¨hrer beraten hatte. also/earlier theA CEO advised had. ‘The marketing-team advised the department-head that also had advised the CEO.’ or ‘The marketing-team was advised by the department-head that also had advised the CEO.’ Paraphrases to choose from for (c) and (d): i. Die Abteilungsleiterin wurde von dem the department-head was by theD

Florian Schwarz 409

ii.

i.

ii.

iii.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

e.

Marketingteam beraten und der marketing-team advised and the Gescha¨ftsfu¨hrer von der Abteilungsleiterin. CEO by theD department-headi. ‘The department-head was advised by the marketing-team and the CEO by the department-head’ Das Marketingteam und der Gescha¨ftsfu¨hrer wurden the marketing-team and theN CEO were von der Abteilungsleiterin beraten. by theD department-head advised ‘The marketing-team and the CEO were advised by the department-head’ Das Marketingteam beriet die TheN/A marketing-team advised theN/A Abteilungsleiterin, die auch die department-head RPN/A also, theN/A Gescha¨ftsfu¨hrerin beraten hatte. CEO advised had. Paraphrases to choose from for (e): Die Abteilungsleiterin wurde von the department-head was by dem Marketingteam beraten, und die theD marketing-team advised and the Gescha¨fsfu¨hrerin von der Abteilungsleiterin. department-head CEO by theD ‘The department-head was advised by the marketing-team and the CEO by the department-head.’ Die Abteilungsleiterin wurde von dem the department-head was by theD Marketingteam und von der marketing-team und by theD Gescha¨ftsfu¨hrerin beraten. CEO advised ‘The department-head was advised by the marketing-team and by the CEO.’ Das Marketingteam und die Gescha¨ftsfu¨hrerin the marketing-team and the CEO wurden von der were by theD Abteilungsleiterin beraten. department-head advised

410 Processing Presupposed Content

2. a/b.

c/d/e.

c/d/e.

4. a/b.

c/d/e.

5. a/b.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

3. a/b.

‘The marketing-team and the CEO were advised by the department-head.’ Die Mitarbeiterin, die die Sekreta¨rin The employee RPN/A the secretary auswa¨hlte, hatte (auch/vorher) der chose had also/earlier theN Direktor ausgewa¨hlt. director chosen Die Mitarbeiterin wa¨hlte die Sekreta¨rin The employee chose the secretary aus, die (auch/vorher) out RPN/A also/earlier (den/die) Direktor(in) ausgewa¨hlt hatte. had theA/theN/A director( fem) chosen Die Spionin, die die Kommissarin verfolgt The spy RPN/A the superintendent chased hat, hatte (auch/vorher) der has had also/earlier theN KGB Mann verfolgt. KGB man chased. Die Spionin verfolgte die Kommissarin, die (auch/vorher) (den/die) KGB (Mann/Frau) verfolgt hatte. The spy chased the superintendent RPN/A also/earlier theA/theN/A KGB man/woman chased Die Grenzbeamtin, die die Polizistin the police-officer The border-officer RPN/A kontrollierte, hatte (auch/vorher) examined had also/earlier der Staatsanwalt kontrolliert. theN prosecutor examined. Die Grenzbeamtin kontrollierte die Polizistin, The border-officer examinded the police-officer die (auch/vorher) RPN/A also/earlier (den/die) Staatsanwalt(in) kontrolliert hatte. examinded had theA/theN/A prosecutor(fem) Die Professorengruppe, die das Expertenteam expert-team The group-of-professors RPN/A the begutachtete, hatte reviewed had

Florian Schwarz 411

B. Sample of the auch-Self-Paced Reading Materials 1.

2.

Das Marketingteam, das (der/den) Manager TheN/A marketing-team RPN/A theN/A manager beraten hat, hatte advised has, had (auch/vorher) der Gescha¨ftsfu¨hrer beraten. also theN CEO advised. Question: Had the marketing-team advised the manager? Die Spionin,/ die (der/den) Kommissar super-intendent the spy RPN/A theN/A verfolgte,/ hatte (auch/vorher) der chased had also/earlier theN KGB-Mann verfolgt. KGB-man chased Question: Was the super-intendent chased by the spy?

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

(auch/vorher) der Universita¨tspra¨sident begutachtet. also/earlier theN university-president reviewed. c/d/e. Die Professorengruppe begutachtete das Expertenteam, The group-of-professors reviewed the expert-team das RPN/A (auch/vorher) (den/die) Universita¨tspra¨sidenten(in) also/earlier theA/theN/A university-president( fem) begutachtet hatte. reviewed had 6. a/b. Die Redakteurin, die das Projektmitglied The editor RPN/A the project-member begleitete, hatte accompanied had (auch/vorher) der Computertechniker begleitet. also/earlier theN computer-technician accompanied c/d/e. Die Redakteurin begleitete das Projektmitglied, The editor accompanied the project-member das (auch/vorher) RPN/A also/earlier (den/die) Computertechniker(in) begleitet hatte. theA/theN/A computer-technician( fem) accompanied had

412 Processing Presupposed Content 3.

4.

6.

C. Sample of the Also Self-Paced Reading Materials 1. 2. 3.

The congressman/ who (John) wrote to (John)/ had (also/just) written to the mayor/to schedule a meeting/ for the fundraiser. The electrician/ that (Justin) helps (Justin)/ (also/once) helped the old lawyer/ in setting up/ his new computer. The singer/ who (Josh) met (Josh)/ had (also/once) met Kurt Cobain/ at a benefit concert/ in Boston.

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

5.

Die Grenzbeamtin,/ die (der/den) Polizist(en) police-officer the border-officer RPN/A theN/A kontrollierte,/ hatte examined had (auch/vorher) der Staatsanwalt kontrolliert. also/earlier theN prosecutor examined Question: Had the border officer examined the prosecutor? Die Cellistin,/ die (der/den) Komponist(en) composer the cellist RPN/A theN/A bewunderte,/ hatte(auch/vorher) admired had also/earlier der Dirigent bewundert. theN director admired Question: Was the director admired by the cellist? Die Gruppe,/ die (der/den) the group RPN/A theN/A Naturschutzverein unterstu¨tzte,/ hatte nature-conversancy-orangisation supported had (auch/vorher) der Bu¨rgermeister unterstu¨tzt. supported also/earlier theN mayor Question: Had the group supported the nature conservancy organisation? Die Norwegerin,/ die (der/den) Finne(n) Fin the NorwegianFEM RPN/A theN/A besiegte,/ hatte (auch/vorher) der defeated had also/earlier theN Swede defeated Schwede besiegt. Question: Was the Fin defeated by the Norwegian?

Florian Schwarz 413

4.

Acknowledgements An earlier version of this paper has appeared in the Proceedings of Sinn und Bedeutung 10 (Schwarz 2006). Special thanks are due to Lyn Frazier, Angelika Kratzer and Chuck Clifton for countless helpful discussions throughout the development of this project, and to Chris Potts for very helpful comments on previous versions of this paper. I would also like to thank the following people for helpful comments and discussion: Jan Anderssen, Greg Carlson, Kai von Fintel, Florian Ja¨ger, John Kingston, Paula Menendez-Benito, Barbara Partee, Kristen Syrett, the participants of 2nd Year Seminar and Semantics Reading Group and the audience at Sinn and Bedeutung 10. Ira Noveck and two anonymous reviewers for the Journal of Semantics also provided very helpful comments on an earlier draft. Parts of this work were supported by National Institutes of Health Grant HD-18708 to

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

The coach/ that (Sam) chose (Sam)/ had (also/almost) chosen a Canadian player/ after the Canadian team/ did so well at the Olympics. 5. The police officer/ who (the burglar) noticed (the burglar)/ had (also/once) noticed a suspicious car/ in front of the building/ on the corner. 6. The model/ that (the producer) spoke with (the producer)/ had (also/recently) spoken with Calvin Klein/ at the show/ in Paris last month. 7. The soccer team/ who (Courtney’s school) defeated (Courtney’s school)/ (also/recently) beat the state champion/ in the tournament/ last year in Boston. 8. The actor/ that (Shannon) irritated (Shannon) at the party/ had (also/once) annoyed Shannon’s Dad/ at the dinner/ the night before. 9. The law professor/ who (the committee) advised (the committee)/ (also/often) coun-seled the governor/ about the education program/ for underprivileged youth. 10. The sports reporter/ that (the sponsor) commended (the sponsor)/ (also/once) praised the tennis player/ for showing up/ despite his injury. 11. The poet/ who (the cellist) admired (the cellist)/ (also/still) looked up to Harvey Keitel/ for his performance/ in Blue in the Face. 12. The agent/ that (the Iranian) was watching (the Iranian)/ (also/still) kept an eye on the shop-owner/ who was suspected of/ dealing with illegal weapons.

414 Processing Presupposed Content the University of Massachusetts and by the National Science Foundation under Grant BCS-0418311 to Barbara H. Partee and Vladimir Borschev.

FLORIAN SCHWARZ Department of Linguistics, University of Massachusetts Amherst, 226 South College, 150 Hicks Way, Amherst MA 01003, USA, e-mail: fl[email protected]

Abusch, D. (2005), Triggering from Alternative Sets and Projection of Pragmatic Presuppositions. Unpublished manuscript, Cornell University. Bader, M. & M. Meng (1999), ‘Subjectobject ambiguities in German embedded clauses: an across-the-board comparison’. Journal of Psycholinguistic Research 28:121–43. Beaver, D. (1997), ‘Presupposition’. In J. van Benthem and A. ter Meulen (eds.), The Handbook of Logic and Language. Elsevier. Amsterdam. 939–1008. Beaver, D. (2001), Presupposition and Assertion in Dynamic Semantics. CSLI Publications. Stanford, CA. Beaver, D. & H. Zeevat (2007), ‘Accommodation’. In G. Ramchand and C. Reiss (eds.), Oxford Handbook of Linguistic Interfaces. Oxford University Press. Cambridge, 503–38. beim Graben, P., J. D. Saddy, M. Schlesewsky, & J. Kurths (2000), ‘Symbolic dynamics of event-related brain potentials’. Physical Review 62:5518–41. Chambers, C. & V. S. Juan (2005), ‘Accommodation and the interpretation of presupposition during referential processing’. Poster presented at the 18th CUNY Conference on Human Sentence Processing. The University of Arizona.

Christianson, K., A. Hollingworth, J. F. Halliwell, & F. Ferreira (2001), ‘Thematic-roles assigned along the garden path linger’. Cognitive Psychology 42:368–407. Crain, S. & M. Steedman (1985), ‘On not being led up the garden path: the use of context by the psychological parser’. In D. R. Dowty, L. Karttunen, and A. Zwicky (eds.), Natural Language Parsing: Psychological, Computational, and Theoretical Perspectives. Cambridge University Press. Cambridge. 320–58. Frazier, L. (2006), The Big Fish in a Small Pond: Accommodation and the Processing of Novel Definites. Unpublished manuscript, University of Massachusetts Amherst. Geurts, B. (1999), Presuppositions and Pronouns. Elsevier. Amsterdam, New York. [Includes bibliographical references ([223]–9) and indexes]. Heim, I. (1982), The Semantics of Definite and Indefinite Noun Phrases. Ph.D. thesis, University of Massachusetts. Amherst, MA. Heim, I. (1983a), ‘File change semantics and the familiarity theory of definiteness’. In R. Baeuerle, R. Schwarze, and A. von Stechow (eds.), Meaning,

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

REFERENCES

Florian Schwarz 415 Lewis, D. (1979), ‘Scorekeeping in a language game’. In R. Baeuerle, U. Egli, and A. von Stechow (eds.), Semantics from Different Points of View. Springer. Berlin. 172–87. Noveck, I. & D. Sperber (eds.) (2004), Experimental Pragmatics. Palgrave Macmillan. Basingstoke. Schlesewsky, M. & A. Friederici (2003), ‘Sentence processing mechanisms’. In L. Nadel (ed.), Encyclopedia of Cognitive Science. Macmillan Reference. London. 1149–55. Schlesewsky, M., G. Fanselow, R. Kliegl, & J. Krems (2000), ‘The subjectpreference in the processing of locally ambiguous wh-questions in german’. In B. Hemforth and L. Konieczny (eds.), German Sentence Processing. Kluwer. Dordrecht. 65–93. Schwarz, F. (2006), ‘Presuppositions in Processing—a case study of German auch’. In C. Ebert and C. Endriss (eds.), Proceedings of Sinn und Bedeutung 10. ZAS. Berlin. 301–15. Simons, M. (2001), ‘On the conversational basis of some presuppositions’. In R. Hastings, B. Jackson and Z. Zvolensky (eds.), Proceedings of Semantics and Linguistic Theory 11. CLC Publications. Ithaca, NY. 427–46. Spenader, J. (2002), Presuppositions in Spoken Discourse. Ph.D. thesis, Computational Linguistics, University of Stockholm. Stalnaker, R. (1973), ‘Presuppositions’. Journal of Philosophical Logic 2: 447–57. Stalnaker, R. (1974), ‘Pragmatic presuppositions’. In M. Munitz and P. Unger (eds.), Semantics and Philosophy. New York University Press. New York. 197–214. van Berkum, J. J. A., C. M. Brown, & P. Hagoort (1999), ‘Early referential context effects in sentence processing:

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

Use, and Interpretation of Language. De Gruyter. Berlin. 164–89. Heim, I. (1983b), ‘On the projection problem for presuppositions’, In D. Flickinger, et al. (eds.), Proceedings of the Second West Coast Conference on Formal Linguistics. Stanford University. 114–25. Heim, I. (1992), ‘Presupposition projection and the semantics of attitude verbs’. Journal of Semantics 9:183–221. Hemforth, B. (1993), Kognitives Parsing: Repra¨sentation und Verarbeitung grammatischen Wissens. Infix. Sankt Augustin. Kamp, H. (1981), ‘A theory of truth and semantic representation’. In J. Groenendijk, T. Janssen, and M. Stokhof (eds.), Formal Methods in the Study of Language: Proceedings of the Third Amsterdam Colloquium, vol. I. Mathematical Center. Amsterdam. 227–321. Kamp, H. & U. Reyle (1993), From Discourse to Logic. Kluwer. Dordrecht. Karttunen, L. (1973), ‘Presuppositions of compound sentences’. Linguistic Inquiry 4:169–93. Karttunen, L. (1974), ‘Presupposition and linguistic context’. Theoretical Linguistics 1:181–94. Karttunen, L. & S. Peters (1979), ‘Conventional implicature’. Syntax and Semantics 11: Presupposition. Academic Press. New York. 1–56. Kim, A. & L. Osterhout (2005), ‘The independence of combinatory semantic processing: evidence from eventrelated potentials’. Journal of Memory and Language 52:205–25. Krifka, M. (1999), ‘Additive particles under stress’. In SALT. CLC Publications. Ithaca, NY. 111–28. Kripke, S. (1991), Presupposition and Anaphora: Remarks on the Formulation of the Projection Problem. Unpublished manuscript, Princeton University.

416 Processing Presupposed Content van Rooy, R. (1999), ‘Against local accommodation’. In M. K. Bart Geurts and R. van der Sandt (eds.), Focus and Presupposition in Multispeaker Discourse. Workshop held at the 11th European Summer School in Logic, Language, and Information. ESSLI. Amsterdam. von Fintel, K. (2004), ‘Would you believe it? The king of France is back! (presuppositions and truth-value intuitions)’. In M. Reimer and A. Bezuidenhout (eds.), Descriptions and Beyond. Oxford University Press. Oxford. 315–41. Received: 04.03.2006 Final version received: 09.02.2007 Accepted: 14.06.2007

Downloaded from jos.oxfordjournals.org by guest on January 1, 2011

evidence from event-related brain potentials’. Journal of Memory and Language 41:147–82. van Berkum, J. J. A., C. M. Brown, P. Hagoort, & P. Zwitserlood (2003), ‘Event-related brain potentials reflect discourse-referential ambiguity in spoken language comprehension’. Psychophysiology 40:235–48. van der Sandt, R. (1988), Context and Presupposition. Croom Helm. London. van der Sandt, R. & B. Geurts (2001), ‘Too’, In P. Dekker and R. van Rooy (eds.), 13th Amsterdam Colloquium. 180–5. van der Sandt, R. A. (1992), ‘Presupposition projection as anaphora resolution’. Journal of Semantics 9:333–77.