This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!
, with p instructions to the sensori-motor system, l instructions to the language of thought, and g instructions to the computation as such. In this schema the interfaces play a crucial role, since they determine which of the properties of the systems outside CHL are legible for the purpose of computation. What is legible may in fact be only a very impoverished representation of a rich structure. For instance, verbs may have a very rich conceptual structure reXected in their thematic properties. Reinhart (2000b, 2003) 4 Take, for instance, the rules computing entailments. A sentence such as DPplur elected Y does not entail that z elected Y, for z 2 kDPplurk, whereas DPplur voted for Y does entail that z voted for Y, for z 2 kDPplurk. This is reXected in the contrast between we elected me and ??we voted for me. The latter is ill-formed since it entails the instantiation I (lx (x voted for x) ), which is reXexive, but not reXexivemarked (Reinhart and Reuland 1993).
56
The Nature of Gradience
shows that the computational system can read these properties only in so far as they can be encoded as combinations (clusters) of two features: [ + c(ause)] and [ + m(ental)]. Anything else is inaccessible to CHL. The coding of the concept subsequently determines how each cluster is linked to an argument. Note furthermore, that an interface must itself contain operations. That is, it must be a component sui generis. For instance, syntactic constituents do not always coincide with articulatory units. Furthermore, in preparation for instruction to the sensori-motor system, hierarchical structure is broken down and linearized. On the meaning side, it is also well-known that for storage in memory much structural information is obliterated. As Chomsky (1995) notes, at the PF side such operations violate inclusiveness. It seems fair to assume that the same holds for the interpretive system at the C–I interface. Given this schema, more can be ‘wrong’ with an expression than just a violation (crash) of the derivation in CHL. DiVerences in the ‘degree of ungrammaticality’ may well arise by a combination of violations in diVerent components: PF, lexicon, narrow syntax, interpretive system, interpretation itself. Any theory of language with a further articulation into subsystems is in principle well equipped to deal with ‘degrees’ of well- or ill-formedness.
3.3 Issues in binding and co-reference The domain of anaphora is well-suited for an illustration of these issues. The canonical binding theory (Chomsky 1981) represents a strictly categorical approach. However, in many languages anaphoric systems are more complex than the canonical binding theory anticipated. Moreover, not all cases where an anaphoric element is co-valued with an antecedent are really binding relations. Limitations of space prevent me from even remotely doing justice to the discussion over the last decades. Instead I will focus on one basic issue. Going over the literature one will Wnd that sometimes judgements are clearly categorical, in other cases we Wnd judgements that are far more ‘soft’ and ‘variable’. In one and the same domain, languages may in fact vary, even when they are closely related. Below I will discuss some concrete examples. Recently, this property of anaphora has been taken as evidence that a categorical approach to anaphora must be replaced by an approach in which binding principles are replaced by soft, violable constraints along the lines proposed in optimality theoretic approaches to grammar. A useful discussion along these lines is presented in Fischer (2004). Yet, such a conclusion is not warranted. ‘Flexible’ approaches of various kinds may seem attractive at a coarsely grained level of analysis, taking observational entities at face value. The attraction disappears if one sees
Gradedness: Interpretive Dependencies and Beyond
57
that more Wnely grained analyses make it possible to connect the behaviour of anaphoric elements to the mechanisms of the grammar, and to explain variation from details in their feature make-up, or from diVerences in their environment. Current research warrants the conclusion that the computational system underlying binding operates as discretely and categorically as it does in other areas of grammar. However, just as in the cases discussed in Sections 3.1 and 3.2, the computational system does not determine all aspects of interpretation; where it does not, systems of use kick in, evoking the air of Xexibility observed. Much work on binding and anaphora so far is characterized by two assumptions: 1. all binding dependencies are syntactically encoded (by indices or equivalent); 2. all bindable elements are inherently characterized as pronominals or anaphors (simplex, complex, or clitic). These assumptions are equally characteristic of extant OT-approaches to binding. So, one Wnds diVerent rankings of binding constraints on pronominals or subtypes of anaphors, across environments and languages. In Reuland (2001) I argued that both assumptions are false. The argument against (1) rests in part on the inclusiveness condition, and hence is theoryinternal, but should be taken seriously in any theory that strives for parsimony. Clearly, indices are not morphosyntactic objects. No language expresses indices morphologically. Thus, if syntax is, indeed, the component operating on morphosyntactic objects, it has no place for indices. External validation of the claim that syntax has no place for indices rests on the dissociation this predicts between dependencies that can be syntactically encoded and dependencies that cannot be. In so far as such dissociations can be found, they constitute independent evidence (see Reuland 2003 for discussion). The argument against (2) is largely theory-independent. It essentially rests on the observation that there are so many instances of free ‘anaphors’ and locally bound ‘pronominals’, that one would have to resort to massive lexical ambiguity in order to ‘maintain order’. Instead, we can really understand the binding behaviour of pronouns (to use a cover term for anaphors and pronominals) if we consider their inherent features, the way these features are accessed by the computational system, and general principles governing the division of labour between the components of the language system, in particular CHL, logical syntax, and discourse principles as part of the interpretive system.
58
The Nature of Gradience
By way of illustration I will brieXy consider three cases: (a) free anaphors (often referred to as ‘logophors’); (b) reXexives in PPs; and (3) local binding of pronominals. 3.3.1 Free anaphors There are two main issues any theory of binding must address: 1. Why is it that cross-linguistically reXexivity must be licensed? 2. Why is it that certain expressions, ‘anaphors’, must be bound? Question (1) covers a substantial part of Condition B of the canonical binding theory, question (2) reXects the canonical condition A. Consider Wrst question (1). More concretely it asks: Given that pronominals can generally be bound by DP, why cannot we simply use expressions of the form John admires him, or more generally, DP V pronominal, to express that the person who John admires is John? Simply invoking avoidance of ambiguity does not help. There are lots of cases where pronominals are used ambiguously, and no special marking is introduced. An ambiguity story cannot be complete without explaining why in all the other cases there is no marking. Moreover, languages diVer as to where they require special marking (we will discuss one such case below). Again, an avoid-ambiguity story would not lead us to expect such variation. As I argued in Reuland (2001) the essence of the answer to (1) is that reXexivization involves the identiWcation of variables in the logical syntax representation of a predicate.5 Thus, limiting ourselves for simplicity’s sake to binary predicates, reXexivization eVectively reduces a relation to a property, as in (3.7): (3.7) lx ly (P x y) ! lx (P x x) This leads to a theta-violation since the computational system cannot distinguish these two tokens of x as diVerent objects (occurrences). At the C–I interface syntactic structure is broken down and recoverable only in so far as reXected in conceptual structure. This is parallel to what happens at the PF-interface; here syntactic structure is broken down as well, and recoverable in so far as reXected in properties of the signal. SpeciWcally, following the line of Chomsky (1995), a category such as V’, needed to express the speciWer– complement asymmetry in [VP Spec [V’ V Comp] ], is not a term, hence not 5 The notion ‘logical syntax’ is to be distinguished from ‘logical form’. Logical form is the output of the computational system (CHL); the operations yielding logical form are subject to the inclusiveness condition. Logical syntax is the Wrst step in interpretation, with operations that can translate pronominals as variables, can raise a subject leaving a lambda expression, etc., thus not obeying inclusiveness. For a discussion of logical syntax, see Reinhart (2000a) and references cited there.
Gradedness: Interpretive Dependencies and Beyond
59
visible to the interpretive system. Therefore, it is not translatable at the C–I interface, and the hierarchical information it contributes is lost. What about order? Syntax proper only expresses hierarchy, but no order. Order is imposed under realization by spell-out systems. As a consequence, the computational system cannot distinguish the two tokens of x in (3.7) on our mental ‘scratch paper’ by order. Hence, translating DP V pronominal at the C–I interface involves the steps in (3.8): (3.8) [VP x [V’ V x ]] ! ([VP V ‘‘x x’’ ]) ! *[VP V x] 1 2 3 The second step with the two tokens of x in ‘x x’ is virtual, hence it is put in brackets. It is virtual, since with the breakdown of structure, and given the absence of order, it has no status in the computation: eliminate V’ and what you get is the third stage. The transition from (3.8.1) to (3.8.3) does not change the arity of V itself. It is still a 2-place predicate, but in (3.7)/(3.8.3) it sees only one object as its argument. As a consequence, one theta-role cannot be assigned. Under standard assumptions about theta-role discharge a theta-violation ensues. (Alternatively, two roles are assigned to the same argument with the same result.)6 Languages employ a variety of means to obviate the problem of (3.8) when expressing reXexive relations (Schladt 2000). They may mark the verb, they may put the pronominal inside a PP, they may double the pronominal, add a body-part, a focus-marker, etc. Here I will limit discussion to SELF-marking in languages such as English and Dutch. BrieXy, the minimum contribution 6 I am grateful to an anonymous reviewer for pointing out a problem in the original formulation. In the present version I state more clearly the empirical assumptions on which the explanation rests. Note that the problem of keeping arguments apart for purposes of theta-role assignment does not come up in the case of two diVerent arguments, as in (i) (abstracting away from QR/l-abstraction): (i) [VP j [V’ V m ]] ! [VP V m j ]) (–/! [VP V m/j]) 1 2 3 The objects remain distinct. There is no reason that theta-relations established in the conWguration in (i.1) would not be preserved in (i.2). Hence, the issue of (i.3) does not arise. Note, that theta-roles are not syntactic features that modify the representation of the element they are assigned to. That is, they are not comparable to case features. This leaves no room for the alternative in (ii) where xu1 and xu2 are distinguishable objects by virtue of having a diVerent ‘u-feature’ composition. (ii) [VP xu1 [V’ V xu2 ]] ! [VP V xu1 xu2 ] –/! [VP V xu1/xu2] 1 2 3 Rather, with theta-assignment spelled out, but reading xu1 correctly as ‘as x is being assigned the role u1’ we get (iii), which reXects the problem discussed in the main text: (iii) [VP xu1 [V’ V xu2 ]] ! ([VP V ‘‘x x’’u1/u2 ]) ! *[VP V xu1/u2] 1 2 3
60
The Nature of Gradience
that SELF must make in order to make a reXexive interpretation possible is to induce suYcient structure. That is, in a structure such as (3.9) the two arguments of P are distinct. (3.9) lx (P x [x SELF]) This is independent of other eVects SELF may have on interpretation. The semantics of SELF only has to meet one condition: [x SELF] should get an interpretation that is compatible with the fact that the whole predicate has to be used reXexively. That implies that whatever interpretation [x SELF] gets should be suYciently close to whatever will be the interpretation of x. This is expressed in (3.10): (3.10) lx (P x f(x)), with f a function that for each x yields a value that can stand proxy for x It is this property of SELF that gives rise to the statue-reading discussed by JackendoV (1992). A statue reading also shows up in Dutch, as is illustrated in (3.11b), where zichzelf despite being ‘reXexive’ refers to the Queen’s statue rather than to the Queen herself (Reuland 2001). Importantly, the statuereading depends on the presence of SELF. If we replace zichzelf by zich, which is allowed in this environment, the statue reading disappears. So, zich expresses identity, zichzelf stands proxy for its antecedent. (3.11) ‘Madame Tussaud’s’-context: Consider the following example in Dutch: De koningin liep bij Madame Tussaud’s binnen. Ze keek in een spiegel en a. ze zag zich in een griezelige hoek staan b. ze zag zichzelf in een griezelige hoek staan Translation: The queen walked into Madame Tussaud’s. She looked in a mirror and a. she saw SE in a creepy corner stand b. she saw herself in a creepy corner stand Interpretations: (a) zich ¼ the queen: the queen saw herself; (b) zichzelf ¼ the queen’s statue: the queen saw her statue. The diYculty for our computational system to deal with diVerent tokens of indiscernibles is in a nutshell the reason why reXexivity must be licensed. In other languages a ‘protecting element’ can behave quite diVerently from English himself. In Malayalam, for instance, the licensing anaphor does not need to be locally bound at all (Jayaseelan 1997). Compare the sentences in (3.12):
Gradedness: Interpretive Dependencies and Beyond
61
(3.12) a.
raamani tan-nei *(tanne) sneehikunnu ‘Raman SE-acc self loves’ ‘Raman loves himself ’ b. raamani wicaariccu [penkuttikal tan-nei tanne sneehikkunnu enn@] ‘Raman thought [girls SE-acc self love Comp]’ ‘Raman thought that the girls loved him(self).’
Example (3.12a) is a simple transitive clause. A simple pronominal in object position cannot be bound by the subject. A reXexive reading requires the complex anaphor tan-nei tanne. In (3.12b) tan-nei tanne is put in an embedded clause with a plural subject, which is not a possible binder for the anaphor. In English the result would be ill-formed. In Malayalam, however, the matrix subject raaman is available as a binder for the anaphor. Thus, the presence of tanne licenses but does not enforce reXexivity. Note, that tan-ne tanne is essentially a doubled pronoun, and tanne need not be identiWed with SELF. Hence a diVerence in behaviour is not unexpected. As illustrated in (3.5a), in English himself need not always be locally bound either. Reinhart and Reuland (1991, 1993) and Pollard and Sag (1992) present extensive discussion of such cases. Example (3.13) presents an illustrative minimal pair. (3.13) a.
Max boasted that the queen invited Mary and him/himself for a drink b. Max boasted that the queen invited him/*himself for a drink
In (3.13a) the reXexive can have a long-distance antecedent, in (3.13b) it cannot. In their Dutch counterparts we see a diVerent pattern: in both cases, long-distance binding of zichzelf is impossible. (3.14) a.
Max pochte dat de koningin Marie en hem/*zichzelf voor een drankje had uitgenodigd b. Max pochte dat de koningin hem/*zichzelf voor een drankje had uitgenodigd
In Dutch the canonical counterpart of himself is zichzelf. The question is why an LD antecedent is possible in (3.13a) and not in (3.13b), and why it is never possible with zichzelf. If we only consider superWcial properties of himself and zichzelf, no nonstipulative answer will be found. However, Reinhart and Reuland (1991) propose a more Wne-grained analysis (also assumed in Reinhart and Reuland 1993). As we saw, a reXexive predicate must be licensed. In Dutch and English, the licenser is SELF. Unlike Malayalam tan-ne tanne a SELF-anaphor
62
The Nature of Gradience
obligatorily marks a predicate reXexive if it is a syntactic argument of the latter (Reinhart and Reuland 1993: a reXexive-marked syntactic predicate is reXexive). In (3.15b) this requirement causes a clash: the predicate cannot be reXexive due to the feature mismatch between the queen and himself. Hence (3.13b) is ill-formed. In (3.13a) himself is properly contained in a syntactic argument of invite, hence it does not impose a reXexive interpretation. But why does the analysis have to refer to the notion syntactic argument/ predicate? This falls into place if we understand reXexive marking to be a real operation in the computational system. We know independently that coordinate structures or adjuncts resist certain operations that complements easily allow. Movement is a case in point, as illustrated by the coordinate structure constraint. Assume now, that there is a dependency between SELF and the verb that is sensitive to the same constraints as movement. For short, let us assume it is movement and that SELF-movement is really triggered by a property of the verb. Its eVect in canonical cases is represented in (3.15): (3.15) a. John admires himself b. John SELF-admires [him [-]] SELF now marks the predicate as reXexive, and indeed nothing resists a reXexive interpretation of (3.15). Consider now (3.13b), repeated as (3.16): (3.16) a. *Max boasted that [the queen invited himself for a drink] b. *Max boasted that [the queen SELF-invited [him[-]] for a drink] SELF attaches to the verb, marks it as reXexive, but due to the feature mismatch between the queen and him it cannot be interpreted as reXexive, and the sentence is ruled out. Consider next (3.13a), repeated as (3.17): (3.17) a. Max boasted that [the queen invited [Mary and himself ] for a drink] b. Max boasted that [the queen SELF-invited [Mary and [him[-]]] for a drink] X
In this case SELF-movement is ruled out by the coordinate structure constraint. As a consequence, the syntactic predicate of invite is not reXexivemarked, and no requirement for a reXexive interpretation is imposed. Hence, syntax says nothing about the way in which himself is interpreted, leaving its interpretation open to other components of the language system.
Gradedness: Interpretive Dependencies and Beyond
63
SELF also has another use, namely as an ‘intensiWer’. Hence, its interpretation reXects that property precisely in those cases where its use is not regulated by the structural requirements of grammar (as we saw in the case of (3.5) and (3.6)). Note, that we do not Wnd the same pattern in Dutch. This is because himself diVers from zichzelf in features. The zich in zichzelf is not speciWed for number and gender. We know descriptively that zich must be bound in the domain of the Wrst Wnite clause. So, any theory assigning zich a minimal Wnite clause as its binding domain will predict that Max is inaccessible as an antecedent to zich in both (3.14a) and (3.14b) (see Reinhart and Reuland 1991 for references, and Reuland 2001 for an execution in current theory without indices). So, the contrast between Dutch and English follows from independent diVerences in feature content between himself and zichzelf. Why must anaphors be bound? The fact that himself must be bound only where it is a syntactic argument of the predicate, and is exempted where it is not, already shows that this is too crude a question. Also Icelandic sig may be exempted from a syntactic binding requirement, be it in a diVerent environment, namely subjunctive.7 Similar eVects show up in many other languages. Hence, there can be no absolute necessity for anaphors to be bound. Reuland (2001) shows that anaphors have to be bound only if they can get a free rider on a syntactic process enabling the dependency between the anaphor and its antecedent to be syntactically encoded. In the case of anaphors such as sig, or zich a syntactic dependency can be formed as in (3.18): (3.18) DP
I R1
V R2
pronoun R3
R3 is the dependency between object pronoun and V, that is realized by structural case. R1 reXects the agreement between DP and inXection. R2 represents the dependency between verb and inXection (assumed to also present if I is on an auxiliary). All three dependencies are syntactic and independent from binding. Dependencies can be composed. Composition of R1-R2-R3 yields a potential dependency between pronoun and DP. For reasons discussed in Reuland (2001) this dependency can only be formed 7 See the example in (i): (i) Marı´a var alltaf svo andstyggileg. Þegar Olafurj kæmi segd¯i hu´n se´ri/*j a´reid¯anlega ad¯ fara . . . (Thra´insson 1991) Mary was always so nasty. When Olaf would come, she would certainly tell himself [the person whose thoughts are being presented—not Olaf] to leave
64
The Nature of Gradience
eVectively if the pronoun lacks a speciWcation for grammatical number. Thus, zich allows it (lacking a number speciWcation), its pronominal counterpart hem ‘him’ does not. Complementarity between pronominals and anaphors follows not from any local prohibition on binding pronominals, but from a very general principle concerning the division of labour between components of the language system: in order to interpret zich as a variable bound by the antecedent, syntactic processes suYce. The dependency between a pronominal and its antecedent cannot be established in the syntax (grammatical number prevents this), but requires that part of the computation is carried out via the CI Interface. Switching between components incurs a cost, so using zich is less costly than using a pronominal instead. In the case of SELF the trigger must be diVerent. Dutch zelf does indeed have a relevant property which so far has gone unnoticed. Consider the following contrast between zich and zichzelf in Dutch, which shows up with verbs allowing either. One such verb is verdedigen ‘defend’. Suppose a group of soldiers has been given the assignment to occupy a hill and subsequently the enemy attacks. After the battle a number of situations can obtain, two of which can be described as follows: (a) the soldiers kept the hill, but at the cost of most of their lives; (b) the soldiers lost the hill, they all stayed alive. In the Wrst case one can properly say (3.19a), but not (3.19b). In the second case one can say either: (3.19) a.
De The b. De The
soldaten soldiers soldaten soldiers
verdedigden defended verdedigden defended
zich ‘them’ zichzelf themselves
met succes successfully met succes successfully
Zichzelf has a distributive reading (each of the soldiers must have defended himself successfully), whereas zich is collective. If the verbal projection has a position to mark distributivity (Stowell and Beghelli 1997), this is suYcient to warrant attraction of SELF. If so, we have on the one hand an independent account of the fact that SELF is attracted to the verb, we have an account for the meaning contrast in (3.19), and the account accommodates the fact that not all licensers are attracted and that reXexivity is not always enforced. This approach predicts correlations that go beyond being an anaphor or being a pronominal. For instance, given that in Malayalam reXexivity is not enforced by tan-ne tanne, one predicts that it does not mark any special property of the verb. Whether this prediction is in fact correct is a matter of further research.
Gradedness: Interpretive Dependencies and Beyond
65
3.3.2 ReXexives in PPs The following contrast between French (Zribi-Hertz 1989) and Dutch illustrates how a small variation in grammar independent from binding may have a signiWcant eVect on binding. (3.20) a. Jean Jean b. Jean Jean c. Jean Jean d. Jean Jean
est is est is
Wer proud jaloux jealous bavarde avec mocks (of) parle talks
lui/lui-meˆme him/himself *lui/lui-meˆme him/himself *lui/lui-meˆme him/himself de lui/lui-meˆme (of) him/himself de of de of
This pattern is sometimes taken to be a problem for a conWgurational approach to binding, and condition B in particular. ConWgurationally (3.20a) and (3.20b) are identical, and the same holds for (3.20c) and (3.20d). If so, how can a pronominal be allowed in the one case, and not in the other? Alternatively, it is argued, these examples show that the selection of anaphors is sensitive to semantic-pragmatic conditions. How ‘expected’, or ‘normal’, is the reXexivity of the relation expressed by the predicate? Let us assume the relevant factor in French is indeed semantic-pragmatic. Yet, this cannot be all there is to it, since in the corresponding paradigm in Dutch all have the same status and in all cases a complex anaphor is required. (3.21) a.
Jan Jan b. Jan Jan c. Jan Jan d. Jan Jan
is is is is spot mocks praat talks
trots proud jaloers jealous met (of) over (of)
op of op of
zichzelf/*zich him/himself zichzelf/*zich him/himself zichzelf/*zich him/himself zichzelf/*zich him/himself
One, surely, could not seriously claim that Dutch speakers have a diVerent pragmatics, or that trots means something really diVerent than Wer. A clearly syntactic factor can be identiWed, however. As is well-known, Dutch has preposition stranding, but French does not. Whatever the precise implementation of preposition stranding, it must involve some relation between P and the selecting predicate head that obtains in stranding languages like Dutch and does not in French. Let us assume for concreteness sake that this relation
66
The Nature of Gradience
is ‘allows reanalysis’. Thus, P reanalyses with the selecting head in Dutch, not in French (following Kayne 1981). We will be assuming that reanalysis is represented in logical syntax. If so, in all cases in (3.21) we have (3.22) as a logical syntax representation: (3.22) DP [V [P pro]] ! . . . .[ V-P] . . . ! DP (lx ([V-P] x x)) We can see now, that at logical syntax level we have a formally reXexive predicate. Such a predicate must be licensed, which explains the presence of SELF in all cases. In French there is no V-P reanalysis. Hence, we obtain (3.23): (3.23) DP [V [P pro]] ! DP (lx (V x [P x])) Here, translating into logical syntax does not result in a formally reXexive predicate. This entails that no formal licensing is required.8 Hence it is indeed expectations, or other non-grammatical factors that may determine whether a focus marker like meˆme is required. In a nutshell, we see how in one language a grammatical computation may cover an interpretive contrast that shows up in another. 3.3.3 Locally bound pronominals Certain languages allow pronominals to be locally bound. Frisian is a case in point. Frisian is instructive, since the literature occasionally reports that Frisian has no specialized anaphors (Schladt 2000). This is in fact incorrect, since in Frisian as in Dutch, reXexivity must be licensed by adding sels to the pronominal, yielding himsels.9 A rule of thumb is that Frisian has the bare pronominal him where Dutch has the simplex anaphor zich.10 So, in Frisian one Wnds the pattern in (3.24): (3.24) a.
Jan John b. Jan John
wasket washes Welde felt
him himself [him himself
fuortglieden] slip away
8 Note, that by itself meˆme is quite diVerent from SELF (for instance in French there is no *meˆmeadmiration along the lines of English self-admiration, etc.). 9 The occasional claim that Frisian does not have specialized anaphors is reminiscent of the claim that Old English does not have them. For Frisian the claim is incorrect. For Old English I consider the claim as inconclusive. Going over Old English texts, it is striking that all the cases that are usually adduced in support of the claim are in fact cases that would make perfectly acceptable Frisian. What one would need are clear cases of sentences with predicates such as hate or admire to settle the point. 10 Note that him is indeed a fully-Xedged pronominal that can be used as freely as English him or Dutch hem.
Gradedness: Interpretive Dependencies and Beyond c.
67
Jan bewuˆndere him*(sels) John admired himself
This is one of the cases where re-ranking of constraints might descriptively work. However, Reuland and Reinhart (1995) present independent evidence that pronominals such as him in Frisian are under-speciWed for structural case. Fully explaining the consequences of this diVerence in feature speciWcation would lead us beyond the scope of this article. For current purposes a fairly crude suggestion suYces: whereas zich’s under-speciWcation for number makes it possible for Dutch zich to form a syntactic dependency with its antecedent along the lines indicated in (3.18), the case system in Frisian does not enable an element in the position of the pronoun in (3.18) to enter a syntactic dependency with the antecedent, that is the link R3 is not eVective for encoding in Frisian. This goes beyond just saying that Frisian happens to lack a zich-type anaphor. Even if Frisian had a zich-type anaphor, it would not be more economical than using him. In any case, the case correlation shows the importance of considering the Wne grain of grammar to understand the syntactic processes of binding. As discussed in Reuland and Reinhart (1995) German dialects also exhibit interesting variation in the local binding of pronominals. This variation is case related as well, although so far a detailed analysis has not been given. Yet, any theory that fails to take into account that it is case that links DPs, and therefore also anaphors and pronominals, to the fabric of the sentential structure is at risk of missing the source of the variation. This paves the way for discussing an instance of grammar-based gradedness in Dutch. Since Dutch has a three-way system, with a simplex anaphor, a complex anaphor, and a pronominal, the Dutch counterpart of (3.24) has a couple more options than Frisian: (3.25) a.
Jan John b. Jan John c. Jan John
wast washes voelde felt bewonderde admired
zich/1hem himself [zich/1hem wegglijden] himself slip away zichzelf/1zich/ 2hem himself
In (3.25a) and (3.25b), using hem instead of zich only violates the principle that it is more economical to have an anaphor than a pronominal. In (3.25c), however, hem violates two principles, economy and the principle that reXexivity be licensed. And indeed, we do Wnd diVerent degrees of ill-formedness reXecting the number of violations.
68
The Nature of Gradience
3.3.4 Variability and gradedness in binding relations It is time for a summary of how real or apparent variation in binding arises. The collective–distributive contrast in the interpretation of zich versus zichzelf is a potential source of variation in judgements. Unless collectivity versus distributivity is systematically controlled for, an investigator may have the impression of Wnding arbitrary intra- and inter-subject variation where in fact there is underlying systematicity. Assessing binding into PPs, one must distinguish between categorical, grammatical factors, and non-categorical extra-grammatical factors. In Dutch PPs of the type Jan praat over zichzelf ‘Jan talks of himself ’ grammatical factors overrule possible sources of variation from discourse: no variation between complex and simplex anaphors obtains. However, in French Jean parle de lui/lui-meˆme ‘Jean talks of him/himself ’ there are two options, since grammar leaves the choice of strategy open. Judgements will, therefore, be open to variation governed by intra-subject shifts in perspective, like Necker cube eVects in visual perception. Something similar obtains in Dutch locative PPs, where the grammar leaves open the choice between zich and hem, and consequently, grammatically they are in free variation. Yet, the choice is not arbitrary: informally speaking, Jan keek onder zich ‘Jan looked under SE’ is presented from Jan’s perspective, whereas Jan keek onder hem ‘Jan looked under him’ is presented from the speaker’s perspective. Subtle though this is, this is another factor to be controlled for. As the contrasts in (3.5) and (3.6) showed, perspective with its concomitant variability is also involved in the licensing of exempt anaphors. Extragrammatical factors become visible where the categorical distinctions of grammar leave interpretive options open. I presented case as a factor in the local binding of pronominals. In languages without a strong morphology, case by itself is a low-level variable factor in the grammar, and a typical area in which we Wnd dialectal variation and variation across registers in language. Slight variations in the case system may have high level eVects on binding possibilities of pronominals, as may other types of variation in feature composition. In addition we saw that wherever binding phenomena are determined by the interaction between several conditions, gradedness phenomena also can be explained from the satisfaction of some, but not all conditions.
Gradedness: Interpretive Dependencies and Beyond
69
3.4 By way of conclusion I realize that the discussion in this article implies that no easy explanations are forthcoming in linguistics. Being a pronominal or being an anaphor are no longer characterizations that can be taken at their face value. What are the components of a pronominal or anaphoric expression? What grammatical features do they have? What are the properties of the case system? How is agreement eVected? How are prepositions and verbs related? What lexical operations does a language have? What morphological operations are available to manipulate the argument structure of verbs? The answers to such questions are needed in order to assess whether putative instances of tendencies, or grading, are really in the purview of grammar. If certain issues of language have to remain outside our grammars, this simply reXects the fact that, whatever grammar does, it certainly does not tell us all of how to say what we want to say.
4 Linguistic and Metalinguistic Tasks in Phonology: Methods and Findings S T E FA N A . F R I S C H A N D A D R I E N N E M . S T E A R N S
4.1 Introduction Phonologists have begun to consider the importance of gradient phonological patterns to theories of phonological competence (e.g. Anttila 1997; Frisch 1996; Hayes and MacEachern 1998; Ringen and Heinamaki 1999; Pierrehumbert 1994; Zuraw 2000; and for broader applications in linguistics see the chapters in Bod et al. 2003 and Bybee and Hopper 2001). Gradient patterns have been discovered in both internal evidence (the set of existing forms in a language) and external evidence (the use and processing of existing and novel forms in linguistic and metalinguistic tasks). The study of gradient phonological patterns is a new and promising frontier for research in linguistics. Grammatical theories that incorporate gradient patterns provide a means to bridge the divide between competence and performance, most directly in the case of the interface between phonetics and phonology (Pierrehumbert 2002, 2003) and between theoretical linguistics and sociolinguistics (e.g. Mendoza-Denton et al. 2003). In addition, linguistic theories that incorporate gradient patterns are needed to unify theoretical linguistics with other areas of cognition where probabilistic patterns are the rule rather than the exception. A variety of methodologies have been used in these studies, and an overview of these methodologies is the primary focus of this paper. The methodologies that are reviewed either attempt to directly assess phonological knowledge (well-formedness tasks and similarity tasks) or indirectly reXect phonological knowledge through performance (elicitation of novel forms, corpus studies of language use, errors in production and perception). A case
Linguistic and Metalinguistic Tasks in Phonology
71
study demonstrating an approach that combines both internal and external evidence is also presented.
4.2 Approaches to phonological judgements (data sources) The evidence for gradient patterns as part of the phonological knowledge of a language comes from a variety of sources, including psycholinguistic experiments using metalinguistic and language processing tasks, and studies of language corpora. In this section, these methods are brieXy reviewed. While the focus of this chapter is on tasks involving metalinguistic judgements, additional sources of evidence that indirectly reXect linguistic knowledge, such as corpus studies, are also reviewed. When these diVerent methodological approaches are used together, strong evidence for the relevance of gradient phonological patterns can be obtained. For example, lexical corpus studies have revealed numerous systematic statistical patterns in the phonotactics of a variety of languages. If metalinguistic tasks where participants make judgements about novel forms that share those same patterns demonstrate that participants make probabilistic generalizations, then it must be the case that those statistical patterns are part of the phonological knowledge of the participants. 4.2.1 Direct measures Direct measures of phonological judgements are tasks that explicitly ask participants to make a judgement about a lexical item or phonological form. These tasks probe the explicitly available linguistic knowledge of participants. 4.2.1.1 Well-formedness judgements The most commonly used judgement tasks are well-formedness judgement tasks. Well-formedness judgement tasks attempt to directly probe the grammaticality of phonological forms. Two variants of the well-formedness judgement task are the acceptability task and the wordlikeness task. In the acceptability task, the participant is presented with a novel phonological form and given two choices, acceptable/possible or unacceptable/ impossible. This task is equivalent to the prototypical grammaticality judgement task used elsewhere in linguistics. While the acceptability task might, at Wrst glance, appear to reduce the likelihood of collecting gradient data, gradience is still commonly found in acceptability judgement tasks in variation in the response to a stimulus item across participants. For example, Coleman and Pierrehumbert (1997) collected acceptability judgements for
72
The Nature of Gradience
novel multisyllabic words containing illegal onset consonant clusters. They found considerable variability across participants in the acceptability of these clusters, and this variability was predicted in part by the probability of the rest of the novel word. An illegal cluster in a high probability context was accepted by more participants than an illegal cluster in a low probability context. The wordlikeness task attempts to measure gradient well-formedness within individual participants. Participants are asked to judge the degree to which a novel word could be a word in their language. In Frisch et al. (2000) and Frisch and Zawaydeh (2001), these judgements were on a one to seven scale, where ‘one’ represents a word that could not possibly be a word in the language and ‘seven’ represents a word that could easily be a word in the language. Frisch et al. (2000) and Frisch and Zawaydeh (2001) collected wordlikeness judgements for novel multisyllabic words in English and Arabic respectively. As in the Coleman and Pierrehumbert (1997) study, one factor that was shown to strongly aVect wordlikeness is the aggregate probability of the constituents in the novel word. An example of this task applied to novel monosyllabic words is given in the case study section of this chapter. Frisch et al. (2000) collected both acceptability judgements and wordlikeness judgements for the same novel word stimuli in English. In a comparison of the two sets of data, they found that participants appeared to perform a similar judgement task whether they were allowed just two options (acceptable/unacceptable) or seven (1–7). In particular, Frisch et al. (2000) found that the ‘unacceptable’ option was used in the acceptability task more frequently than the rating of ‘one’ in the wordlikeness task. Thus it appeared that participants applied the acceptable/unacceptable rating like a wordlikeness task on a scale from ‘one’ to ‘two’. Overall, they found that phonotactic probability predicted participant judgements in both tasks, and that participant judgements were similar across the two tasks. Given that acceptability and wordlikeness tasks generated similar results, but the wordlikeness task allows participants to express more subtle distinctions between forms, researchers interested in gradient phonological patterns would likely beneWt from using the wordlikeness task. Well-formedness judgement tasks have also been used to probe morphophonological knowledge. Zuraw (2000) investigated nasal substitution in Tagalog. Nasal substitution in Tagalog is a phonological process where a stem initial obstruent consonant is replaced by a homorganic nasal. For example, the stem /kamkam/ ‘usurpation’ surfaces with initial /˛/ in /ma3pa3˛amkam/ ‘rapacious’. This process is lexically gradient, occurring with some stems but not others. The likelihood that a stem participates in nasal substitution depends on the obstruent at the onset of the stem. Substitution is
Linguistic and Metalinguistic Tasks in Phonology
73
more likely for /p t s/ and less likely for /d g/. Zuraw (2000) used elicitation tasks to explore the productivity of this process, and she also used a wellformedness task where participants judged the acceptability of nasal substitution constructions using novel stems in combination with common preWxes. She compared wordlikeness judgements for the same novel stem in forms with and without nasal substitution, and found that ratings were higher for nasal substitution in cases where nasal substitution was more common in the lexicon (e.g. /p t s/ onsets to the novel words). Hay et al. (2004) examined the inXuence of transitional probabilities on wordlikeness judgements for novel words containing medial nasal-obstruent clusters (e.g. strimpy, strinsy, strinpy). Overall, they found that wordlikeness judgements reXected the probability of the nasal-obstruent consonant cluster. However, they found surprisingly high wordlikeness for novel words when the consonant cluster was unattested (zero probability) in monomorphemic words. They hypothesized that these high judgements resulted from participants’ analyses of these novel words as multimorphemic. This hypothesis was supported in additional experiments where subjects were asked to make a forced choice decision between two novel words as to which was more morphologically complex. Participants were more likely to judge words with low probability internal transitions as morphologically complex, demonstrating that participants considered multiple analyses of the forms they were given. Hay et al. (2004) proposed that participants would assign the most probable parse to forms they encountered (see also Hay 2003). The Wndings of Hay et al. (2004) highlight the importance of careful design and analysis of stimuli in experiments using metalinguistic tasks. Although these tasks are meant to tap ‘directly’ into phonological knowledge, the judgements that are given are only as good as the stimuli that are presented to participants and there is no guarantee that the strategy employed by participants in the task matches the expectations of the experimenter. Another example of this problem appears in the case study in this chapter, where perceptual limitations of the participants resulted in unexpected judgements for nonword forms containing presumably illegal onset consonant clusters. 4.2.1.2 Distance from English Greenberg and Jenkins (1964) is the earliest study known to the authors that examined explicit phonological judgements for novel word forms in a psycholinguistic experiment. They created novel CCVC words that varied in their ‘distance from English’ as measured by a phoneme substitution score. For each novel word, one point was added to its phoneme substitution score for every position or combination of positions for which phoneme substitution could make a word. For example, for the novel
74
The Nature of Gradience
word /druk/ there is no phoneme substitution that changes the Wrst phoneme to create a word. However, if the Wrst and second phonemes are replaced, a word can be created (e.g. /Xuk/). For every novel word, substitution of all four phonemes can make a word, so each novel word had a minimum score of one. Greenberg and Jenkins (1964) compared the cumulative edit distance against participants’ judgements when asked to rate the novel words for their ‘distance from English’ using an unbounded magnitude estimation task. In this task, participants are asked for a number for ‘distance from English’ based on their intuition, with no constraint given to them on the range of numbers to be used. Greenberg and Jenkins (1964) found a strong correlation between the edit distance measure and participants’ judgements of ‘distance from English’. They also found similar results when they used the same stimuli in a wordlikeness judgement task where participants rated the words on an 11-point scale. Given that the data from a wordlikeness task is simpler to collect and analyse, there appears to be no particular advantage to collecting distance judgements in the study of phonology. 4.2.1.3 Similarity between words Another phonological judgement task is to ask participants to judge the similarity between pairs of real words or novel words (Vitz and Winkler 1973; Sendlmeier 1987; Hahn and Bailey 2003). This task has been used to investigate the dimensions of word structure that are most salient to participants when comparing two words. Presumably, these same dimensions of word structure would be the most salient to participants in making well-formedness judgements which, in some sense, involve comparing a word to the entire lexicon. These tasks have found that word initial constituents and stressed constituents have greater impact on similarity judgements than other constituents. However, it is well-known from the literature on similarity judgements in cognitive science that similarity comparisons are context dependent (Goldstone et al. 1991). Thus the most salient factors for a pairwise comparison of words might not be the most salient factors when a novel word is compared to the language as a whole. This is a research area where very little direct work has so far been done (see Hahn and Bailey 2003). Frisch et al. (2000) analysed a variety of predictors for their well-formedness judgements and found some evidence that word initial constituents and stressed constituents had a greater role in predicting well-formedness, at least in the case of longer novel words. Thus, it appears that similarity judgements are a tool that can investigate the same dimensions of lexical structure and organization from a diVerent perspective. If a well-deWned relationship between well-formedness and similarity judgements can be found, then that would suggest more explicitly that the
Linguistic and Metalinguistic Tasks in Phonology
75
well-formedness task is somehow grounded in a similarity judgement to one or more existing lexical items. 4.2.2 Indirect measures Indirect measures of well-formedness reXect grammatical linguistic knowledge through linguistic performance. Indirect measures have provided evidence for the psychological reality of gradient phonological patterns, as several studies have shown participant performance to be inXuenced by the phonological probability of novel words or sub-word constituents. 4.2.2.1 Elicitation of novel forms The traditional linguistic approach to exploring competence experimentally is to elicit examples that demonstrate the linguistic process of interest. This procedure has been extended to elicitations that involve nonsense forms, demonstrating unequivocally that a process is productive. This type of experiment is sometimes referred to as ‘wug-testing’, in honour of one of the Wrst uses of this technique by Berko (1958) to explore children’s knowledge of plural word formation using creatures with novel names (e.g. ‘this is a wug, and here comes another one, so now there are two of them; there are now two ____’). Elicitation of novel forms can be used to examine gradient phonological patterns. Variation across participants in their performance in an elicitation task provides evidence for an underlying gradient constraint. Alternatively, if participants are given a number of stimuli of roughly the same phonological form to process, variation within an individual can be observed (Zuraw 2000). 4.2.2.2 Analysis of lexical distributions Linguistic corpora have also been used as a source of external evidence. Generally, these corpora have resulted from large, organized research projects such as CELEX (Burnage 1990) or the Switchboard corpus (Godfrey et al. 1992). Recently, however, web searches have been used to generate corpora from the large amount of material posted on the internet (e.g. Zuraw 2000). Written corpora may be more or less useful for phonological study, depending on the phenomenon to be examined and the transparency of the writing system for the language involved. Written corpora have shown the greatest potential in the study of morphophonology. For example, Anttila (1997) and Ringen and Heinamaki (1999) examined the quantitative pattern of vowel harmony in Finnish suYxes using large written corpora. They found that harmony is variable, with the likelihood of harmony depending on the distance between the harmony trigger and target, and also the prosodic prominence of the trigger. In a subsequent psycholinguistic task, Ringen and Heinamaki (1999) found that type frequencies from their corpus predicted well-formedness judgements of native speakers.
76
The Nature of Gradience
4.2.2.3 Confusability in perception and production Gradient wellformedness of phonological forms is also connected to speech perception and production processes. It is well documented that word perception is inXuenced by the number of phonologically similar words in the lexicon, known as the size of the word’s lexical neighbourhood (e.g. Luce and Pisoni 1998). However, lexical neighbourhood sizes are correlated with phonological probability and the relationship between these two measures has yet to be clearly established (Bailey and Hahn 2001; Luce and Large 2001). Overall, words that share sub-word constituents such as onsets, rimes, or nuclei with many other words are subject to competition during spoken word recognition. In addition, generalizations about more and less frequent subword sequences bias phonetic perception in favour of more frequent sequences (Pitt and McQueen 1998; Moreton 2002). Together, these Wndings suggest that lexical organization for language processing utilizes those subword phonotactic constituents. Similar eVects of phonotactic probability have been demonstrated in speech production. For example, Munson (2001) probed the relationship between phonological pattern frequency and repetition of nonwords by adults and children. The nonwords created for Munson’s study used consonant clusters embedded in a two-syllable construct (CVCCVC). The stimuli were Wrst tested to determine if the frequency of constituents aVected wordlikeness ratings by participants. The wordlikeness ratings were highly correlated with the phonological pattern frequency of the sound sequences within the novel words. Munson then used the same stimuli to study the ability of children and adults to accurately repeat nonwords and found that subjects were less accurate in their repetition of nonwords that contained less frequent phoneme sequences. In addition, Munson (2001) found there was greater variability within participants’ productions for low probability items versus high probability items. Similar speech production eVects have been found elsewhere for monolingual adults (e.g. Vitevitch and Luce 1998) and second language learners (e.g. Smith et al. 1969), although with additional complications that will not be discussed here. These results further support theories that information about phonological pattern frequency is encoded at the processing and production levels of linguistic representation. 4.2.3 Summary Studies of language patterns, language processing, and metalinguistic judgements have found substantial evidence that probabilistic phonological patterns are part of the knowledge of language possessed by speakers. These
Linguistic and Metalinguistic Tasks in Phonology
77
patterns are reXected in the frequency of usage of phonological forms, the ease of processing of phonological forms, and gradience in metalinguistic judgements about phonological forms. In the next section, a case study is presented that demonstrates probabilistic patterns in the cross-linguistic use of consonant clusters and gradience in metalinguistic judgements about novel words with consonant clusters that reXects the cluster’s probability.
4.3 Case study The example case study presented here focuses on initial consonant clusters. In a linguistic corpus study, the frequency of word onset consonant clusters across and within languages is examined. Cross-linguistic patterns of consonant clustering are reXected in the likelihood of use of a consonant cluster type within a language. The pattern appears to be gradient, and based on the scalar property of sonority. In an experimental study, wordlikeness for novel words based on the type frequency of the onset cluster is examined. Wordlikeness judgements are aVected by consonant cluster frequency, showing that native speakers are sensitive to consonant cluster frequency, and thus are capable of encoding a gradient onset consonant cluster constraint. 4.3.1 Cross-linguistic patterns in onset clusters It has been claimed that consonant cluster combinations are restricted within and across languages by sonority sequencing. Sonority is a property of segments roughly corresponding to the degree of vocal tract opening or acoustic intensity (Hooper 1976). Stop consonants have the lowest sonority and liquids and semi-vowels have the highest sonority. Analyses of sonority sequencing constraints have generally assumed that there is a parametric restriction on the degree of sonority diVerence required in a language, where larger sonority diVerences are allowed and smaller sonority diVerences are not allowed (e.g. Kenstowicz 1994). The current study shows that sonority restrictions are gradient, rather than absolute. A closer examination of sonority sequencing for onsets presented in this section shows that the scalar property of sonority is reXected quantitatively in the frequency of occurrence of consonant cluster types across languages. In addition, the data support an analysis of sonority constraints based on ‘sonority modulation’ (Ohala 1992; Wright 1996). The cross-linguistic preference is for onset consonant clusters that have a large sonority diVerence, whether rising or falling. Consonant clusters with large sonority diVerences occur more frequently than consonant clusters with small sonority diVerences across a wide range of languages.
78
The Nature of Gradience
Evidence for gradient sonority modulation constraints can be found by looking at the distribution of consonant clusters within a language. Although a particular degree of sonority modulation may be permitted in a language, it is usually the case that not every conceivable cluster for a particular sonority diVerence is attested. For example, consider word Wnal liquid+C clusters in English, as in pert, purse, fern, furl. Table 4.1 shows the number of attested word Wnal liquid+C cluster types, the number that are theoretically possible given the segment inventory of English, and the percentage of possible clusters that are attested, for the four diVerent levels of sonority diVerence. For example, attested cluster types include /rt/, /rs/, /rn/, and /rl/ as demonstrated in the examples above, as well as /lt/ (e.g. pelt, wilt, salt), /lb/ (bulb), /lv/ (delve, valve), /lm/ (elm, psalm). Unattested clusters include /lg/, /r˛/, and /lr/. The number of possible consonant clusters is determined by the size of the segment inventory. For example, English has three nasal consonants and two liquids, so there are six theoretically possible nasal-liquid combinations. Notice in Table 4.1 that as the sonority diVerence between C and the preceding liquid decreases, the relative number of attested clusters decreases quantitatively, even when the possible number of cluster combinations is taken into account. Similar quantitative analyses of permissible consonant clusters were conducted for onset clusters in a sample of thirty-seven languages from a variety of language families. In this analysis, attested and possible cluster types were determined for all combinations of two stop, fricative, nasal, and liquid consonants as word onsets. The languages analysed were Abun, Aguatec (Mixtec), Albanian, Amuesha, Chatino, Chinantec (Quioptepec), Chinantec (Usila), Chontal (Hokan), Chukchee, Couer D’Alene, Cuicatec, Dakota (Yankton), English, Greek, Huichol, Hungarian, Ioway-Oto, Italian, Keresan, Khasi, Koryak, Kutenai, Mazatec, Norwegian, Osage, Otomi (Mazahua), Otomi (Temoayan), Pame, Portuguese, Romanian, Takelma, Telugu, Terena, Thai, Totonaco, Tsou, and Wichita. For each language, a four-by-four table similar to Table 4.1 was created with the percentage of attested clusters for
Table 4.1. Word Wnal liquid+C clusters in English C
Stop
Fricative
Attested Possible % Attested
11 12 92%
11 16 69%
Nasal 4 6 67%
Liquid 1 2 50%
Linguistic and Metalinguistic Tasks in Phonology
79
each combination of consonant type (e.g. stop-stop, stop-fricative, fricativestop, etc.). Each four-by-four table was then analysed to determine whether larger sonority diVerences were reXected in greater numbers of attested clusters while smaller sonority diVerences were reXected in fewer numbers of attested clusters. For each four-by-four table, there are twenty-four such comparisons that could be made (e.g. stop-stop to stop-fricative, stop-stop to stop-nasal, stop-stop to fricative-stop, etc.). The mean across languages was 74 per cent of comparisons supporting sonority modulation (e.g. fricative-stop and stopfricative more frequent than stop-stop and fricative-fricative). In other words, possible clusters with large sonority diVerences are more frequently attested than possible clusters with small sonority diVerences cross-linguistically. For these onset clusters, sonority modulation appeared to be equally robust for rises toward the syllable nucleus (75 per cent) as for falls toward the nucleus (73 per cent). The minimum for any particular language was for 50 per cent of comparisons supporting sonority modulation (and so 50 per cent not supporting it) and the maximum was 100 per cent of comparisons supporting sonority modulation. Interestingly, non-modulation was never more common than sonority modulation for any of the thirty-seven languages, suggesting that sonority modulation is a universal gradient cross-linguistic constraint. 4.3.2 Wordlikeness judgements Given the apparent presence of quantitative sonority modulation constraints, the next question to be examined is whether these quantitative constraints are part of the synchronic grammar of a native speaker. In this section, an experiment is presented that demonstrates that onset consonant cluster frequencies inXuence well-formedness judgements of native speakers of English. In this experiment, native speakers of English rated novel monosyllabic words with onset consonant clusters for wordlikeness, as in previous studies (e.g. Frisch et al. 2000). 4.3.2.1 Stimuli The stimulus list was created by randomly matching twenty diVerent onset consonant clusters with varying frequencies of occurrence in the lexicon (type frequency) to rimes selected from the mid-range of rime frequencies. The occurrence frequency of each constituent was obtained from a computerized American English dictionary (Nusbaum et al. 1984). Among the consonant clusters were four clusters that do not occur in English (/sr, tl, dl, ul/) and the remaining sixteen ranged in frequency from clusters that occur very rarely (/gw, sf, dw/) or somewhat rarely (/tw, sm, ur, sn/) to clusters
80
The Nature of Gradience
that occur moderately frequently (/kw, dr, sl, fr/) or very frequently (/pl, sp, X, gr, pr/) in the lexicon. The nonword list was analysed to avoid potential confounding factors for wordlikeness judgements such as violating a phonotactic constraint somewhere other than in the onset. Rime statistics were also compiled to examine the eVects of rime frequency on judgements (as in Kessler and Treiman 1997). After discarding items that were not suitable, 115 novel words remained to be used in the experiment. The nonwords were recorded as spoken by the Wrst author using digital recording equipment. 4.3.2.2 Participants Thirty-Wve undergraduate students in an introductory communication sciences and disorders course participated in the experiment. Subjects were between 19 and 45 years of age, and three males and thirty-two females participated. All participants were monolingual native speakers of American English and reported no past speech or hearing disorders. 4.3.2.3 Procedure The experiment was conducted using ECOS/Win experiment software. Participants were seated at individual carrels for the experiment in groups of one to four. Subjects listened to each of the stimuli presented one at a time through headphones at a comfortable listening level. The computer screen displayed a rating range between 1 (not at all like English) and 7 (very much like English) and the participants gave their responses by clicking with a mouse on the button that corresponded with their rating. The total experiment required approximately 15 minutes of the participants’ time to complete. 4.3.2.4 Results The data were analysed based on the mean rating given to each stimulus word across subjects. Three instances where a participant gave no response to the stimulus were discarded; otherwise, all data were analysed. Correlations were examined for the type frequency of onset CC, C1, C2, rime, nucleus, and coda. As expected, mean wordlikeness judgements correlated signiWcantly with the type frequency of the CC sequences contained in the novel words. The CC frequency was the strongest predictor of how the participants judged wordlikeness (r ¼ 0.38). The frequency of the nucleus was also correlated with the participant’s judgement of wordlikeness (r ¼ 0.19). In a regression model of the wordlikeness data using these two factors, both factors were found to be signiWcant (CC: t(112) ¼ 4.1; p < .001; Nucleus: t(112) ¼ 2.0; p < .05). The amount of variance explained by the CC frequency and the nucleus frequency is relatively small (cf. Munson 2001; Hay et al. 2004). Unexpectedly, we found that participant judgements of the novel words containing CC that do not occur in English were fairly high. Possible explanations for this Wnding
Linguistic and Metalinguistic Tasks in Phonology
81
are being investigated. Based on preliminary data collected to date, it appears that participants do not consistently perceive the illegal CC that is presented, but instead regularly perceive a similar sounding CC that is allowed in English (e.g. /tl/ ! /pl/). As mentioned in the discussion of Hay et al. (2004), experiment participants assigned a more probable parse to the unattested sequences present in the stimuli. Setting aside the unattested clusters, the pattern for the remaining clusters is as expected. Mean average ratings for the attested consonant clusters used in the experiment are shown in Figure 4.1 with a best Wt line for the eVect of log CC type frequency in these clusters. The data clearly show that the CC frequency was a strong indicator of the subjects’ ratings of stimulus words. Overall, then, this experiment demonstrates that English speakers are sensitive to the frequency of occurrence of onset consonant clusters, and thus have learned the patterns in the lexicon that would reXect a universal sonority modulation constraint. Thus it seems possible that English speakers (and presumably speakers of other languages) could learn a sonority modulation constraint.
7
6
fr
Mean Rating
5 θr psn 4 dw 3
fl
gr kw sp snpldr sl
sf
2
gw
1 1
10
100
1000
Log CC Type Frequency
Figure 4.1. Mean wordlikeness rating for occurring consonant clusters in English (averaged across stimuli and subjects) by consonant cluster type frequency
82
The Nature of Gradience
4.4 Implications Phonology provides an ideal domain in which to examine gradient patterns because there is a rich natural database of phonological forms: the mental lexicon. A growing number of studies using linguistic and metalinguistic evidence have shown that phonological structure can be derived, at least in part, from an analysis of patterns in the lexicon (e.g. Bybee 2001; Coleman and Pierrehumbert 1997; Frisch et al. 2004; Kessler and Treiman 1997; and see especially Hay 2003). In addition, it has been proposed that the organization and processing of phonological forms in the lexicon is a functional inXuence on the phonology, and that gradient phonological patterns quantitatively reXect the diYculty or ease of processing of a phonological form (Berg 1998; Frisch 1996, 2000, 2004). For example, the storage of a lexical item must include temporal information, as distinct orderings of phonemes create distinct lexical items (e.g. /tAp/ is diVerent from /pAt/ in English). The temporal order within a lexical item is reXected in some models of speech perception, such as the Cohort model and its descendents (e.g. Marslen-Wilson 1987), where lexical items compete Wrst on the basis of their initial phonemes, and then later on the basis of later phonemes. Sevald and Dell (1994) found inXuences of temporal order on speech production. They had participants produce nonsense sequences of words and found that participants had more diYculty producing sequences where words shared initial phonemes. Production was facilitated for words that shared Wnal phonemes (in comparison to words without shared phonemes). In general, models of lexical activation and access predict that lexical access is most vulnerable to competition between words for initial phonemes and less vulnerable for later phonemes, as the initial portions of the accessed word severely restrict the remaining possibilities. This temporal asymmetry can be reXected quantitatively in functionally grounded phonological constraints. For example, phonotactic consonant co-occurrence constraints in Arabic are stronger word initially than later within the word (Frisch 2000). This asymmetry is compatible with the claimed grounding of this co-occurrence constraint in lexical access (Frisch 2004). It has also been demonstrated in a wide range of studies that the phonological lexicon is organized as a similarity space (Treisman 1978; Luce and Pisoni 1998). The organization of the lexicon as a similarity space is reXected in processing diVerences based on activation and competition of words that share sub-word phonotactic constituents. The impact of similarity-based organization is most clearly reXected in cases of analogical processes between
Linguistic and Metalinguistic Tasks in Phonology
83
phonologically similar words (Bybee 2001; Skousen et al. 2002). In related work, similarity has also been used to explain limitations on phonological processes. For example, Steriade (2000, 2001) claims that grammar is constrained to create maximally similar underlying and derived forms where similarity is deWned over a perceptual lexical space. Given the grounding of gradient phonological patterns in lexical distributions, it is unclear whether a distinct phonological grammar (phonological competence) is required above and beyond what is necessary to explain patterns of phonological processing (phonological performance). Ultimately, this is an empirical question. Arguments for and against a distinct phonological grammar have been made (e.g. Pierrehumbert 2003; Moreton 2002). However, diVerent models of phonological processing and the lexicon make diVerent predictions about the nature of phonological patterns. Traditionally, models of language processing hypothesize that abstract linguistic constituents, such as phonemes, onsets, rimes, and syllables, are used in speech perception, spoken word recognition, speech production, and lexical organization. These models assume that, at some stage of processing, an abstract representation is involved that is independent of speciWc instances (e.g. Dell 1986; Berent et al. 1999). This type of model uses generalizations that are naturally akin to the types of representations found in phonological grammars. These types of representations would encode gradient phonological patterns as frequency counts or probabilities for the symbols. Recently, an alternative processing model based on instance-speciWc exemplars has received empirical support in language processing (e.g. Johnson 1997; Pierrehumbert 2002). Exemplar models can still explain generalizationbased behaviour as a reXex of the collective activation of exemplars that are similar along some phonological dimension. Thus, abstract phonological categories may not be needed, and consequently a grammar containing abstract phonological rules may be an illusion resulting from the stable behaviour that follows from the activation of a large number of overlapping exemplars. Exemplar models of language represent the frequency information in gradient phonological constraints in a completely straightforward and transparent way. The frequency information is a direct consequence of frequency of exposure and use. The symbolic and exemplar representations systems are not necessarily mutually exclusive. It is also possible that both abstract categories and instance-speciWc exemplars are part of a person’s phonological knowledge, creating a phonological system with redundancy and with competing analyses of the same phonological form (cf. Bod 1998; Pierrehumbert 2003).
84
The Nature of Gradience
4.5 Summary Studies of gradience in phonology using linguistic and metalinguistic data have revealed a much closer connection between phonological grammar and the mental lexicon. These new dimensions of phonological variation could not have been discovered without corpus methods and data from groups of participants in psycholinguistic experiments. While the range of patterns that have been studied is still quite limited, the presence of gradient phonological constraints demonstrates that phonological knowledge goes beyond a categorical symbolic representation of possible forms in a language. In order to accommodate the broader scope of phonological generalizations, models of grammar will have to become more like models of other cognitive domains, which have long recognized and debated the nature of frequency and similarity eVects for mental representation and processing. The study of phonology may provide a unique contribution in addressing these more general questions in cognition for two reasons. The scope of phonological variability is bounded by relatively well-understood mechanisms of speech perception and speech production. Also, phonological categories and phonological patterns provide a suYciently rich and intricate variety of alternatives that the full complexity of cognitive processes can be explored.
5 Intermediate Syntactic Variants in a Dialect-Standard Speech Repertoire and Relative Acceptability LEONIE CORNIPS
5.1 Introduction Non-standard varieties, such as dialects throughout Europe, which are under investigation challenge research about the phenomenon of micro-variation in two ways.1 Within the framework of generative grammar, the linguist studies the universal properties of the human language in order to Wnd out the patterns, loci, and limits of syntactic variation. Language is viewed essentially as an abstraction, more speciWcally, as a psychological construct (I-language) that refers primarily to diVerences between individual grammars within a homogeneous speech environment, that is to say, without consideration of stylistic, geographic, and social variation. Given this objective, a suitable investigative tool is the use of intuitions or native-speaker introspection, an abstraction that is quite normal within the scientiWc enterprise. Frequently, however, there are no suYciently detailed descriptions available of syntactic phenomena that are of theoretical interest for investigating micro-variation in and between closely related non-standard varieties in a large geographical area (cf. Barbiers et al. 2002). Subsequently, a complication emerges in that the linguist has to collect his own relevant data from speakers of local dialects who are non-linguists. The elicitation of speaker introspection often calls for a design of experiments in the form of acceptability judgements when the linguist has to elicit intuitions from these speakers (Cornips and Poletto 2005). 1 I like to thank two anonymous reviewers for their valuable comments. Of course, all usual disclaimers apply.
86
The Nature of Gradience
Moreover, the standard variety may strongly interfere with local dialect varieties in some parts of Europe so that there is no clear-cut distinction between the standard and the local dialect. In this contact setting—a so-called intermediate speech repertoire (cf. Auer 2000)—the speakers of local dialects may assess all possible syntactic variants, that is dialect, standard, and emerging intermediate variants to their local dialect. Subsequently, clear-cut judgements between the local dialect and the standard variety are not attainable at all. This is, among other factors, of crucial importance for understanding the phenomenon of gradedness in acceptability judgements. This chapter is organized as follows. In the second part it is proposed that acceptability judgements do not oVer a direct window into an individual’s competence. The third part discusses an intermediate speech repertoire that is present in the province of Limburg in the Netherlands. In this repertoire some constructions are diYcult to elicit. Finally, acceptability judgements that are given by local dialect speakers in the same area are discussed. Using data from reXexive impersonal passives, reXexive ergatives, and inalienable possession constructions, it is argued that the occurrence of intermediate variants and the variation at the level of the individual speaker is not brought about by speciWc task-eVects but is due to the induced language contact eVects between the standard variety and the local dialects.
5.2 Relative acceptability Bard et al. (1996: 33) discuss the important three-way distinction among grammaticality, a characteristic of the linguistic stimulus itself, acceptability, a characteristic of the stimulus as perceived by a speaker, and an acceptability judgement which is the speaker’s response to the linguists’ enquiries. These authors note that relative grammaticality is an inherent feature of the grammar whereas relative acceptability reXects gradience in acceptability judgements. The former has a controversial status since it is not entirely clear how to deal with relative grammaticality in formal theory. According to Schu¨tze, the best-known proponents of the view that grammaticality occurs on a continuum are Ross, LakoV and their followers in the late 1960s and early 1970s (see Schu¨tze 1996: 62, 63 for more detail). With respect to acceptability judgements, every elicitation situation is artiWcial: the speaker is being asked for a sort of behaviour that, at least on the face of it, is entirely diVerent from everyday conversation (cf. Schu¨tze 1996: 3). Moreover, Chomsky (1986: 36) argues that: ‘In general, informant judgments do not reXect the structure of the language directly; judgments of acceptability, for example, may fail to provide direct evidence as to grammatical status
Intermediate Syntactic Variants
87
because of the intrusions of numerous other factors’ (cf. Gervain 2002). ‘The intrusion of numerous other factors’ may lead to a crucial mismatch between the acceptability judgements of a construction and its use in everyday speech. One of these factors is that in giving acceptability judgements people tend to go by prescriptive grammar (what they learned at school, for instance) rather than by what they actually say (cf. Altenberg and Vago 2002). This is consistent with sociolinguistic research that prescriptive grammars usually equal standard varieties that are considered more ‘correct’ or have more prestige than the vernacular forms speakers actually use. Moreover, strong sociolinguistic evidence shows that a speaker may judge a certain form to be completely unacceptable but can, nevertheless, be recorded using it freely in everyday conversation (Labov 1972, 1994, 1996: 78). One way to diminish the prescriptive knowledge eVect is to ask for indirect comparative acceptability judgements. Rather than eliciting direct intuitions by the formula: ‘Do you judge X a grammatical/better sentence than Y?’, speakers can be asked the more indirect: ‘Which variant Y or X do you consider to be the most or the least common one in your local dialect?’ Relative judgements can be administered by asking the speakers to indicate how uncommon or how common (for example represented by the highest/lowest value on a several point scale, respectively) the variant is in their local dialect. Psychometric research shows that subjects are thus much more reliable on comparative, as opposed to independent ratings (cf. Schu¨tze 1996: 79 and references cited there). These Wndings indicate that relative acceptability is an inevitable part of the speaker’s judgements. Relative acceptability is without doubt brought about by the complex relationship between I-language and E-language phenomena. The opposition between these two types of phenomena is not necessarily watertight as is often claimed in the literature. Muysken (2000: 41–3) argues that the cognitive abilities which shape the I-language determine the constraints on forms found in the E-language and that it is the norms created within E-language as a social construct that make the I-language coherent. One example of this complex relationship may be that in a large geographical area two or more dialects may share almost all of their grammar (more objective perspective) but are perceived as diVerent language varieties by their speakers (more subjective perspective). This can be due to the fact that dialects may diVer rather strongly in their vocalism and that non-linguists are very sensitive to the quality of vowels. The perceived diVerences between dialects may be associated with diVerent identities and vice versa. Another consequence may be that speakers actually believe that the norm created within their speech community (E-language; more subjective perspective) reXects their grammar
88
The Nature of Gradience
(I-language; more objective perspective). For example, in the Dutch Syntactic Atlas project (acronym SAND) we asked dialect speakers to ‘translate’ standard Dutch verbal clusters containing three verbs, as exempliWed in (5.1), into their local dialect (cf. Cornips and Jongenburger 2001).2 (5.1) Ik weet dat Jan hard moet kunnen werken I know that Jan hard must can work Quite a number of speakers told us that their dialects are simpler or more informal than standard Dutch. Therefore, sentences such as (5.1) are judged as ungrammatical or less grammatical than sentences containing two-verb clusters. However, there is not one Dutch dialect attested that excludes constructions as in (5.1). So, it is important to realize that the use of the dialect and the standard variety in a speciWc setting may be triggered by stylistic or social factors (in a speciWc setting, with speciWc interlocutors) if these varieties constitute a continuum. This information minimizes the risk that we are obtaining information about the prescriptive norms of the standard (or prestigious or formal) variety while our intention is rather to question speakers about their dialect (or vernacular non-standard) forms.
5.3 The intermediate speech repertoire The so-called intermediate speech repertoire (cf. Auer 2000) as exempliWed in the southeastern part of the Netherlands (province of Limburg) is presumably the most widespread in Europe today. In this repertoire, there is a structural or genetic relationship between the standard variety and the dialects (cf. Auer 2005). The inXuence of the standard variety on the dialects is quite manifest. There is no longer a clear-cut separation between the varieties, that is to say, speakers can change their way of speaking without a clear and abrupt point of transition between these varieties. This is of crucial importance to understanding relative acceptability. In general, syntactic elicitation provides no diYculties if structures are grammatical in the standard variety and ungrammatical in the dialects (Cornips and Poletto 2005). The speakers usually refuse these constructions, either by providing a grammatical alternative or by simply not being able to translate the sentence and, hence, show non-response. For instance, in the local dialect of Heerlen (a city in the south of the province of Limburg in the Netherlands, see Figure 5.1) negation agreement in (5.2a) is ungrammatical.
2 More information about the SAND-project can be found at: http://www.meertens.nl/projecten/ sand/sandeng.html
Intermediate Syntactic Variants
89
The local dialect speaker easily provides an alternative, as shown in (5.2b) (taken from the Dutch Syntactic Atlas-project): (5.2) Instruction: ‘Translate into your local dialect’ a. Er wil niemand niet dansen expl will no one not dance ‘No one wants to dance.’/ ‘Everyone wants to dance.’ Translation: b. Gene wilt danse no one wants dance ‘No one wants to dance.’ Second, speakers may provide syntactic features that are obligatory in the local dialect even if the same phenomenon is banned from the standard variety. In this case speakers seem to be able to distinguish whether a given construction is grammatical without interference from prescriptive norms. In the dialect of Heerlen and standard Dutch, there is a very sharp contrast between the grammaticality of the impersonal passive with and without a reXexive, respectively. The local dialect speakers in Heerlen were asked whether they encounter the variant in (5.3a) in their local dialect. This variant is fully ungrammatical in the standard variety. The majority of the subjects (16 out of 24, 67 per cent) provide an aYrmative answer, which is conWrmed by their translation, as exempliWed in (5.3b):3 (5.3) Local dialect of Heerlen Instruction: ‘Do you encounter this variant’ a. Er wordt zich gewassen refl washed thereexpl is Answer: ‘Yes’ and translation b. 1t wee¨d zich gewessje refl washed thereexpl is ‘One is washing himself.’ Finally, in the case that the phenomenon is optional, speakers tend to reproduce the standard variety, because this is nonetheless grammatical in their dialect (Cornips and Poletto 2005). This issue is nicely illustrated by responses to constructions in which an aspectual reXexive zich (an optional element) is oVered to the Limburg dialect speakers, as illustrated in (5.4) (cf. Cornips 1998):
3 However, the plain impersonal passive is also grammatical in the local dialect of Heerlen. In that case it has no reXexive interpretation: ‘One is washing (clothes).’
90
The Nature of Gradience
(5.4) d’r Jan had zich in twiee¨ minute e beeke gedro`nke the Jan had reX in two minutes a small beer drunk The construction with the reXexive zich is fully ungrammatical in the standard variety but optional in the local dialect. The written questionnaire of the Dutch Syntactic Atlas project shows that in only two out of thirty-Wve possible locations in the province of Limburg and its immediate surroundings, an answer with the reXexive is given. Obviously, the interference with the standard variety is so strong that the reXexive is not presented in the answers. It seems that only a very good subject can provide optional structures or all the possibilities that come to his mind. 5.3.1 Heerlen Dutch as an emerging intermediate regional standard variety Without any doubt, every emerging intermediate speech repertoire is a result of an induced language contact situation and/or processes of standard–dialect and/or dialect–dialect convergence, that is to say, vertical and horizontal levelling, respectively (cf. Cornips and Corrigan 2005). The emergence of intermediate variants, which may result in a regional variety as a second standard in the area, is crucial to understanding the phenomenon of syntactic variation within the speech community and at the individual speaker level. A good example of an intermediate variety due to language contact eVects is Heerlen Dutch (Cornips 1998). Heerlen Dutch is a regional standard Dutch variety in the Netherlands. Heerlen is a town of 90,000 inhabitants, situated in Limburg, a province in the southeast of the Netherlands near the Belgian and German borders (see Figure 5.1). As already discussed above, in the local dialect of Heerlen reXexives may occur in a much wider range of constructions than in standard Dutch. An example is the appearance of the reXexive zich in inchoative constructions. In standard Dutch (SD) the appearance of zich in inchoative verb constructions is far from regular. Zich is required in (5.5a), is optional in (5.5b), and obligatorily absent in (5.5c) (cf. Everaert 1986: 83, SD ¼ standard Dutch): (5.5) a
SD Het the b. SD De the c. SD De the
gerucht rumour suiker sugar boter butter
verspreidt spreads lost dissolves smelt *(zich) melts refl
(*zich) refl (zich) op refl part.
Intermediate Syntactic Variants
91
N
Amsterdam
et he nd
Germany
rla s Belgium (Flemish part)
=
Heerlen
Figure 5.1. The location of Heerlen in the province of Limburg
In the local dialect of Heerlen, the reXexive zich has to be present in (5.5b) and it may be present in (5.5c). Further, it also arises in some more inchoative constructions based on transitive verbs as veranderen ‘change’, krullen ‘curl’, and buigen ‘bend’. All these inchoative constructions are fully ungrammatical with a reXexive in the standard variety. Heerlen Dutch, however, as a second standard variety in the area has regularized the presence of the reXexive throughout the whole verb class. It has an optional reXexive zich in the construction in (5.5c) and also in (5.6) which are ungrammatical in the local dialect (and in the standard variety) (HD ¼ Heerlen Dutch):4 4 In Cornips and Hulk (1996), it is argued that the Heerlen Dutch constructions in (5.6) are ergative intransitive counterparts of transitive change of state verbs with a causer as external argument. Further, it is shown that in constructions such as (5.4) and (5.6), the reXexive marker zich acts as an aspectual marker, that is: the aspectual focus is on the end-point of the event.
92
The Nature of Gradience
(5.6) a.
HD Dit this b. HD De the c. HD Het the
vlees meat jurk dress glas glass
bederft spoils sleept drags breekt breaks
zich refl zich over de vloer refl over the Xoor zich refl
It is important to note that optionality arises as an induced contact outcome (see also later in Section 5.3.2). Another example of an emerging variant in Heerlen Dutch concerns the socalled dative inalienable possession construction in which the referent of the dative object is the possessor of the inalienable body-part(s) denoted by the direct object. Importantly, all (old) monographs of the Heerlen dialect and written dialect literature (see Blancquaert et al. 1962; Jongeneel 1884; Kessels 1883) show that in the local dialect the DP referring to the body-part(s) such as handen ‘hands’, illustrated in (5.7), is headed by the deWnite determiner de ‘the’. The possessive dative construction expressing inalienable possession is abundantly used in the eastern dialect varieties of Dutch, although extremely rare in standard Dutch (cf. Broekhuis and Cornips 1994; Cornips 1998). The inalienable possession construction, as far as possible, has an idiomatic reading in standard Dutch that is completely absent in this regional Dutch variety and in the dialects of Limburg (Hdial ¼ Heerlen dialect): (5.7) Hdial /?*SD Ik was Jandat./hemdat. de handen. I wash Jan/him the hands ‘I am washing Jan’s/his hands.’ Hence, in the standard variety the inalienable possession relation must be expressed by means of a possessive pronoun, namely zijn ‘his’, as illustrated in (5.8). The construction in (5.8) is in turn rare in the local dialect of Heerlen: (5.8) SD/?*Hdial Ik was zijn/Jans handen. I wash his/Jan’s hands ‘I am washing his/Jan’s hands.’ Nowadays, Heerlen Dutch involves a large spectrum of intermediate lects varying between the local dialect and standard Dutch. As a result, in Heerlen Dutch we Wnd both the inalienable dative construction in (5.7) and the possessive pronoun construction in (5.8). Thus, the syntactic variation within a regional variety as Heerlen Dutch corresponds to cross-linguistic diVerences between English and French, as in (5.9) and (5.10), respectively (Cornips 1998; Vergnaud and Zubizarreta 1992):
Intermediate Syntactic Variants (5.9)
93
a. Eng *I am washing himdat. the hands. b. Eng I am washing his hands.
(5.10) a. Fr b. Fr
*Je lave ses mains. (*inalienable reading) Je luidat. lave les mains.
Moreover, in spontaneous speech data—although rarely—we even Wnd intermediate forms such as the dative object in combination with the possessive pronoun, as in (5.11b) (cf. Cornips 1998, HD ¼ Heerlen Dutch): (5.11) a.
HD
Ik I b. HD Ik intermediate form I c. HD Ik I
was wash was wash was wash
hemdat. him hemdat. him
de the zijn his zijn his
buik. stomach buik. stomach buik. stomach
It is important to point out that this intermediate variant was already present in 1962 in the neighbouring community of Heerlen, that is Kerkrade.5 Obviously, the inalienable possessive construction is not a binary variable since it allows more than two variants, as illustrated in (5.11). Spontaneous speech data of intermediate variants is presented in (5.12) (see also (5.19) in Section 5.3.2, Cornips 1998). In (5.12) the inalienable possession relation is expressed both by means of a possessive dative je ‘you’ and by means of the possessive pronoun je ‘your’. Note, however, that in all these examples the DP referring to an inalienable body-part is the complement of a PP and not a direct object:6 naar je (5.12) a. HD want ze zeuren jedat van alles because they nag you everything to your hoofd toe (19: Cor) head part ‘They are nagging you about everything.’ bij je neus staan b. HD die had jedat zo they had you right away with your nose stand (35: dhr Berk) ‘They stand in front of you right away.’ 5 Blancquaert et al. (1962) denote the following translation in the local dialect of Kerkrade (in Dutch orthography): (i) De schipper likte zich zijn lippen af The skipper licked refl his lips part. ‘The skipper licked his lips.’ 6 The constructions in (5.12) with a PP are more accepted in standard Dutch than the double object constructions (see Broekhuis and Cornips 1997 for more details about this type of construction).
94
The Nature of Gradience
In (5.13) the semi-copula krijgen ‘get’ cannot assign dative case to the possessor hij ‘he’ which is therefore nominative (cf. Broekhuis and Cornips 1994). However, the spontaneous speech data example in (5.13b) shows that the possessor is also realized by means of the reXexive zich, which is fully ungrammatical in the local dialect (like standard Dutch): (5.13) a. HD Hij krijgt de he gets the ‘His hands are dirty.’ b. HD Die heeft zich (33: dhr Quint) he has refl ‘He gets hell.’
handen hands
vies dirty
enorm
op z’n donder gekregen
enormously a beating
got
5.3.2 The lack of characteristic properties of the dative inalienable possession construction Interestingly, it is not only the case that intermediate variants emerge in an intermediate speech repertoire but they also lack characteristic properties (Vergnaud and Zubizarreta 1992: 598) of the ‘original’ dative inalienable possession construction. Two important properties are the strictly distributive interpretation and grammatical number. The former refers to the presence of a plural possessor combined with a singular inalienable argument as in (5.14a). If that is the case, the referent of the inalienable argument is nevertheless interpreted as referring to more than one bodypart. The latter property refers to the fact that inalienable arguments are obligatorily singular when referring to body-parts of which the number per individual is limited to one such as ‘head’, regardless of whether they have a plural possessor or not. This is illustrated by the grammaticality contrast between (5.14b) and (5.14c) (note that there is no idiomatic reading involved): Ik was hundat./3pl het hoofd (i.e. more ‘heads’ involved) I wash them the head ‘I am washing their heads.’ b. *Ik was hundat./3pl de hoofdenpl I wash them the heads ‘I am washing their heads.’ c. Ik was hundat./3pl de handenpl I wash them the hands ‘I am washing (both) their hands.’
(5.14) a.
Intermediate Syntactic Variants
95
Let us now compare (5.14) with the spontaneous speech data example in (5.15) that occurs very infrequently in the corpus:7 (5.15) HD Ze slaan mekaarpl niet meteen de koppenpl in (5: Stef) they hit each other not at once the heads in ‘They don’t hit each other immediately.’ Apparently, both a distributive interpretation and the property of grammatical number may no longer be characterizing properties of the intermediate forms in Heerlen Dutch, that is: the inalienable argument koppen ‘heads’ is plural although the number of the body-parts kop ‘head’ per individual is limited to one. Finally, the dative construction cannot be modiWed by just any attributive adjective, whereas there is no such restriction in the possessive pronoun constructions, as exempliWed in (5.16a) and (5.16b), respectively (Vergnaud and Zubizarreta 1992): vieze buik HD *Ik was hemdat. de I wash him the dirty stomach ‘I am washing his dirty stomach.’ b. HD Ik was zijn vieze buik. I wash his dirty stomach ‘I am washing his dirty stomach.’
(5.16) a.
I asked speakers of Heerlen Dutch to tell a short story containing the elements vuil ‘dirty’ and handen ‘hands’. In addition to (5.16b), they realize the intermediate variant in (5.17) which lacks, due to the presence of the possessive pronoun, any restriction on the presence of the adjective: (5.17) HD Hij wast hemdat. zijn vuile handen. He washes him his dirty hands ‘He is washing his dirty hands.’ Further, a major important aspect of emerging intermediate variants such as the ones described above is that optionality arises. Hence, a major characteristic of the dative inalienable possession construction is that the [spec TP]subject or the agent, cannot enter into a possessive relation with the direct
7 One reviewer points out that the German counterpart of (5.15) with a plural inalienable argument is grammatical, whereas the German counterpart of (5.14a) with a singular inalienable argument is ungrammatical. However, (5.14a) is fully grammatical both in the local dialect of Heerlen and Heerlen Dutch.
96
The Nature of Gradience
object (or prepositional complement), not even if the indirect object is absent, as illustrated in (5.18a). Thus, a possessive relation between the subject and the direct object can only be expressed indirectly, namely by inserting a dative NP or a reXexive zich, as in (5.18b), respectively (see Broekhuis and Cornips 1994; Vergnaud and Zubizarreta 1992). de handeni (5.18) a. Hdial /SD *Hiji wast b. Hdial /?*SD Hiji wast zichi de handen. he washes reX the hands ‘He is washing his hands.’ However, in all the intermediate variants described so far in which the possessive relation is expressed by the possessive pronoun, the dative object or reXexive referring to a possessor is optional. Importantly, all constructions in (5.19), in contrast to the double object constructions, involve idiomatic readings:8 naar je hoofd toe (19: Cor) (5.19) a. HD ze zeuren (jedat) van alles they nag you everything to your head part ‘They are nagging you about everything.’ bij je neus staan b. HD die had (jedat) zo they had you right away with your nose stand (35: dhr Berk) ‘They stand in front of you right away.’ c. HD Die heeft (zich) enorm op z’n donder gekregen (33: dhr Quint) he has refl enormously a beating got ‘He gets hell.’ Taken together, these intermediate variants are of extreme importance with respect to the locus of syntactic variation, that is: whether the primitive of variation is located outside or inside the grammatical system. It becomes obvious that the facts in Heerlen Dutch indicate that syntactic variation can no longer exhaustively be described by binary settings or diVerent values of a parameter (Cornips 1998). In an intermediate speech repertoire, this concept is a very problematic one and must be open to discussion. DiVerent alternatives are possible but there are no satisfying answers yet. A more recent alternative is to argue that from a minimalist point of view, lexical elements
8 It might be the case that, eventually, the optional dative object will disappear or that it will gain emphatic meaning.
Intermediate Syntactic Variants
97
show minimal morphosyntactic diVerences, more speciWcally, whether they bear un- or interpretable features interacting with general principles.9 Another alternative is to place the notion of choice between syntactic variants into the grammatical system. Henry (1995) shows that imperatives in Belfast allow optional raising of the verb to C and inversion in embedded questions may or may not occur. She accounts for this optionality by arguing that a functional category such as C has both strong and weak features instead of diVerent settings of a parameter. In sum, analyses diVer with respect to the locus of syntactic variation (grammar versus lexicon) and whether individuals may have one or two (or more) grammars responsible for the various syntactic variants.
5.4 Acceptability judgements in an intermediate speech repertoire The coming into existence of intermediate variants is of crucial importance for understanding the phenomena of relative acceptability and perhaps relative grammaticality. Hence, variationist studies have convincingly shown that individual speakers do not show all possible alternatives that exist at their level of their community (Henry 2002; Cornips 1998). So, the behaviour of individual speakers with respect to acceptability judgments cannot be interpreted without knowledge of the community pattern. An individual speaker thus has a passive knowledge of more possible syntactic alternatives than he actually uses due to the fact that these possible alternatives, that is standard, dialect, and emerging intermediate variants, can be heard daily in his community. The intermediate variants form a continuum with the standard and local dialect varieties. This continuum arises not only from a geographic perspective but also from a stylistic (for example the use of dialect and standard features in a more informal and formal setting, respectively) and social perspective (age, gender, ethnicity, levels of education, and occupation of the speaker) as well. I propose that a speaker may no longer be able to judge syntactic features as fully grammatical or ungrammatical. Instead, it is very likely that due to the eVects of the standard–dialect contact situation the speaker can only make relative judgements by comparing those variants. Further, it might be the case that this gradience in acceptability judgements partly arises due to the relative
9 However, even if the syntactic variants are analysed as coming into existence as a result of two competing grammars (cf. Kroch 1989), then some lexical elements must still bear un- and interpretable features as well in order to account for the syntactic alternatives.
98
The Nature of Gradience
grammaticality of the intermediate variants in the community, namely the fact that intermediate variants no longer possess characterizing properties as discussed in 5.3.2. In the former section, a regional standard variety was discussed. Let us now consider the local dialects in the same area. These dialects were investigated in the Dutch Syntactic Atlas project. The design of methodology in the Dutch Syntactic Atlas project consisted of two phases, written elicitation and oral elicitation. The oral acceptability judgement tasks were administered in dialect rather than in the standard variety or some regiolect, in order to avoid accommodation, that is adjustment from the dialect in the direction of the standard-like varieties (cf. Cornips and Poletto 2005). In the phase of oral elicitation, 250 locations were selected throughout the Netherlands. We had a major problem in doing the Weldwork since the large majority of the Weldworkers and Ph.D. students speak only the standard variety. It is for this reason that we had to ask for the assistance of another dialect speaker from the same community speaking the same variety in order to be able to interview the subject in his own dialect. The Weldworker (speaking only standard Dutch) trained a local dialect speaker as an ‘assistant interviewer’. This ‘assistant interviewer’ was asked to translate a standard Dutch structured elicitation task into his or her local dialect. These translations were recorded. In a second session these recordings were played to the second local dialect speaker. In this session, the entire conversation was restricted to the two dialect speakers and the Weldworker did not interfere. 5.4.1 Oral elicitation: the local dialects Two small case studies convincingly show how easily speakers switch between the (base) dialect and the standard variety in an oral task in the southern part of the province of Limburg where an intermediate speech repertoire exists. One of the locations involved in the project was Nieuwenhagen (Landgraaf) a very small ‘rural’ village in the environs of Heerlen. In the local dialect of Nieuwenhagen proper names are obligatorily preceded by the deWnite determiner et or der ‘the’ depending on whether the proper name refers to a female or male, respectively. The presence of the deWnite determiner preceding a proper name, as in (5.20), is fully ungrammatical in standard Dutch: (5.20) et Marie / der Jan is krank det Mary /det Jan is ill
Intermediate Syntactic Variants
99
The recording of the Wrst session between the standard Dutch speaking Weldworker and the local ‘assistant interviewer’ translating standard Dutch into his own dialect shows that the deWnite article in his translation is absent: that is, the proper names Wim and Els show up without it, as illustrated in (5.21). These sentences were elicited in order to investigate the order in the verbal cluster (right periphery): (5.21) 1st session (dialect–standard) Ø Wim dach dat ich Ø Els han geprobeerd e kado te geve Wim thought that I Els have tried a present to give ‘Wim thought I tried to give a present to Els.’ In the same interview session, the ‘assistant interviewer’ shows in another sentence that he may or may not use the deWnite article resulting in der Wim and Ø Els respectively in his local dialect. Note that the deWnite determiner precedes the subject DP whereas it is absent in front of the object DP: (5.22) 1st session (dialect–standard) Der Wim dach dat ich Ø Els e boek han will geve det Wim thought that I Els a book have will give ‘Wim thought I wanted to give a book to Els.’ In the second session, however, in which the ‘assistant interviewer’ exclusively interviews the other dialect speaker in the local dialect, the latter utters the deWnite article both with the subject and object DP as ‘required’: (5.23) 2nd session (dialect–dialect) Der Wim menet dat ich et Els e boek probeerd ha kado te geve. ‘Wim thought I tried to give a book to Els.’ Other indications for easily switching between the base dialect and standard Dutch can be found in (5.24). The inWnitival complementizer has the form om and voor in standard Dutch and the local dialect, respectively (see Cornips 1996 for more details). In the Wrst session, the ‘assistant interviewer’ in interaction with the standard Dutch speaking Weldworker uses the standard Dutch complementizer om whereas in the second session the dialect speaker utters voor, as illustrated in (5.24a) and (5.24b), respectively. Moreover, note that in the Wrst session the proper name Wendy lacks the deWnite article again whereas it is present in the second session, as presented in (5.24a) and (5.24b), respectively: (5.24) a.
Ø Wendy probeerdet om ginne pien te doe.
1st session (dialect– standard)
100
The Nature of Gradience b. Et Wendy hat geprobeerd voor ginne pien te doe. 2nd session (dialect– dialect) ‘Wendy tried not to hurt anyone.’
It is important to note that these functional elements were not explicitly mentioned to the dialect speakers as features we were interested in. From the above, it is obvious that in this linguistic repertoire, speakers can adjust to the standard variety (and surrounding varieties) without a noticeable eVort. This might be due to the fact that speakers are sensitive to their (un)conscious awareness of social diagnosticity of syntactic features, namely that features belonging to the domain of standard Dutch are the prestige variants (Cornips 1996). It is for this reason that training interviewers who are native speakers of the local dialect is necessary although every design has to take into account that standard, non-standard, and intermediate variants represent the daily speech situation, that is: syntactic features from the local dialects and standard Dutch appear in a continuum and have become continuous (cf. Paolillo 1997). The emergence of new intermediate syntactic variants, too, points towards a direction in which it is no longer possible to make a clear-cut distinction between the standard variety and the local dialects from a syntactic point of view. In contrast, as already noted in the introduction, the experiences during Weldwork are that the Limburgian speakers perceive the local dialect and the standard variety as two diVerent varieties and associate them with diVerent identities although they share almost all of their grammar. 5.4.2 Written and oral acceptability judgements in the local dialect In this section, written acceptability judgements about the inalienable possession construction are discussed. The Wrst step in the design of the Dutch Syntactic Atlas project was an extensive written questionnaire containing 424 questions (including sub-questions and remarks to be made by the respondents) that were sent out to 850 respondents and a number of additional informants in Belgium (Cornips and Jongenburger 2001). The grid of the written questionnaires of the Dutch Syntactic Atlas project contains, among others, ten neighbouring villages of Heerlen. In this questionnaire, local dialect speakers were oVered the possessive pronoun construction, as in (5.9), repeated here for convenience as (5.25): (5.25) instruction ‘Translate into your local dialect’ Ik was zijn handen. I wash his hands ‘I am washing his hands.’
Intermediate Syntactic Variants
101
Example (5.26) and Figure 5.2 (on page 102) reveal the translations of (5.25) into the local dialect: (5.26)
‘translations’ location Beek Eijgelshoven Maastricht Vaals Waubach
he ha¨t zien hanj geweschen He had sieng heng gewesche heer heet zien han gewasse Hae hat zieng heng jewaesje Hee¨ hat zieng heng gewessje he has his hands washed f. intermediate variant Eijgelshoven Hee¨ had zieg sieng heng gewesje g. intermediate variant Valkenburg Hae haet zich zien heng gewesje h. intermediate variant Spekholzerhei hea hat ziech zien heng jewesche he has REFL his hands washed i. dialect variant Simpelveld hea hat zich de heng gewesche j. dialect variant Waubach Hea haa zich de heng gewe`sje he has REFL the hands washed a. standard variant b. standard variant c. standard variant d. standard variant e. standard variant
To begin with, the responses show that standard variants, dialect variants, and intermediate variants are among the answers. Further, all deviations of the input, for example intermediate and dialect variants as in (5.26f,g,h) and (5.26i,j) respectively, provide strong evidence that these variants are in the grammar of the speaker (Carden 1976). Moreover, variation arises within a local dialect, as is the case in the spontaneous speech data of Heerlen Dutch, which is a regional standard variety. Thus, two respondents in the location of Eijgelshoven and Waubach reveal diVerent responses. The former displays both the standard and the intermediate variant in (5.26b), and (5.26f), respectively, whereas the latter yields the standard and the dialect variant in (5.26e) and (5.26j), respectively. Finally, the majority of the respondents copy the standard Dutch variant into their local dialect, as illustrated in (5.26a–e).10 In order to control for this task eVect, we also administered this type of construction in the oral acceptability task (see below). Taken together, the translations provide evidence that (a) the standard variety strongly interferes with the local dialect variety, (b) intermediate variants arise, and (c) in this part of the province of Limburg syntactic features from the local dialects and standard Dutch exist in a continuum both in a regional standard variety and in the local dialects (see also (5.11)). 10 Maastricht, in the western part of Limburg, denoted the standard variant in 1962 (cf. Blancquaert et al. 1962). The translations in the atlas of Blancquart seem to suggest that the dative inalienable possessive construction is more spoken in the eastern part of Limburg, i.e. Heerlen and surroundings.
102
The Nature of Gradience
Possessive pronoun
(5)
Intermediate forms
(3)
dative inalienable possession (2)
Figure 5.2. Possible inalienable possession constructions as revealed by the responses to the written questionnaire in ten surrounding locations of Heerlen
Similar to the written translation task, the standard Dutch possessive pronoun construction in (5.25) above was oVered in the oral elicitation task, which was the second step in the Dutch Syntactic Atlas project. The locations in the neighbouring villages of Heerlen where oral Weldwork was conducted are presented in Figure 5.3. In the Wrst section, that is the standard–dialect interaction (see 5.3.2) the assistant interviewers have been asked to translate (5.25) into their local dialect. Only 4 out of 12 respondents (33 per cent) immediately translate (5.25) into the dialect variant which is the dative inalienable possession construction. On the other hand, the majority of the assistant interviewers (8 out of 12, 67 per cent) just copy the possessive pronoun construction into their local dialect. Hence, these respondents show interference from the
Intermediate Syntactic Variants
103
c Meertens Inst
Figure 5.3. Grid of the oral interviews in Heerlen and neighbouring locations
standard variety in shifting towards the more prestigious variety in their response, as is the case in the written elicitation task. What is more, the majority of the respondents (6 out of 8, 75 per cent) reveals an implicational pattern revealing that they copy the standard Dutch variant in the Wrst session (dialect–standard repertoire) whereas they use the intermediate or the local dialect variant in the second session which is the dialect–dialect repertoire, as illustrated in (5.27). (5.27) Location of Vaals: Assistant interviewer: a. Oversetz: ‘Hij Instruction: he ‘Translate’ b. 1st session: ‘Her standard–dialect he c. 2nd session: ‘Her dialect–dialect he Assistant interviewer: d. ‘Komt disse Her satz ook veur?’ ‘Do you also He encounter this variant?’ e. Answer: ‘ja’ ‘yes’
heeft zijn has his
handen gewassen.’ hands washed
had has had has
zien his sich refl
heng hands sien his
gewasse’ washed heng hands
gewasse.’ washed
had
zich
de
heng
gewasse.
has
refl the
hands
washed
104
The Nature of Gradience
Again, the interaction in (5.27) reveals that the speech repertoire in Limburg is a continuous one in which the distinction between standard and dialect varieties is blurred. Consequently, the dialect speaker judges all possible variants, that is to say, the standard possessive pronoun (5.27b), the dialect dative construction (5.27d,e), and the intermediate variant (5.27c) as acceptable. More evidence is presented by the fact that 9 out of 12 speakers (75 per cent) accept both the dative possessive construction and the intermediate form. Strikingly, 6 out of 12 speakers (50 per cent) argue that all forms in (5.27) are acceptable in the local dialect. Two of them give relative judgements without being asked: one considered (5.27c) as slightly more acceptable than (5.27d), the other speaker just considered (5.27d) slightly more acceptable than (5.27c). This small case study indicates (a) extensive variation at the level of the individual speaker such that half of the speakers show all possible syntactic alternatives that exist on the level of their community and (b) the existence of intermediate variants to such an extent that it blurs the distinction between the local dialect and the standard variety. This result is attested in spontaneous speech, and in both the written and oral elicitation data, so we can exclude the possibility that it is primarily due to task eVects. In this intermediate speech repertoire the occurrence of intermediate variants is inevitable in the process of vertical standard–dialect and horizontal dialect–dialect convergence. These Wndings put a question mark on the central sociolinguistic proposal that only phonology is a marker of local identity whereas syntax is a marker of cohesion in large geographical areas. Further, syntactic elicitation shows that speakers of local dialects are no longer able to refuse syntactic variants as fully ungrammatical even if (a) these concern emerging intermediate variants and (b) they did not originally belong to their local dialect variety. Consequently, relative acceptability is the result.
5.5 Conclusion In this paper, I have discussed a so-called intermediate speech repertoire, that results from a contact situation between standard Dutch, a regional Dutch variety (Heerlen Dutch), and local dialects in the southern part of the province of Limburg in the south of the Netherlands. This speech repertoire reveals syntactic diVerences along a continuum to such an extent that it blurs the distinction between the local dialect and the standard variety. It is demonstrated that in this speech repertoire clear-cut judgements are not attainable at all. Using case studies, it has been shown that speakers in this area are not able to judge syntactic features as fully grammatical or ungrammatical. Instead, all variants heard in the community, for example standard,
Intermediate Syntactic Variants
105
dialect, and intermediate variants are considered as acceptable. Moreover, it may be argued that in this speech repertoire dialect and standard varieties form a continuum also beyond the geographic level, that is to say, a continuum from a stylistic and social variation perspective. Subsequently, the Wndings in these case studies can be generalized beyond geographical variation. Moreover, these case studies show that syntax may also be a marker of local identity.
6 Gradedness and Optionality in Mature and Developing Grammars A N TO N E L L A S O R AC E
6.1 Introduction This paper focuses on speciWc patterns of gradedness and optionality in the grammar of three types of speakers: monolingual native speakers, speakers whose native language (L1) has been inXuenced by a second language (L2), and very Xuent non-native speakers. It is shown that in all these cases gradedness is manifested in areas of grammar that are at the interface between the syntax and other cognitive domains. First, evidence is reviewed on the split intransitivity hierarchy (Sorace 2000b, 2003a), indicating that not only auxiliary selection but also a number of syntactic manifestations of split intransitivity in Italian and other languages are lexically constrained by aspectual properties. These constructions tend to show gradedness in native intuitions that cannot easily be accommodated by current models of the lexicon–syntax interface. Moreover, the mappings between lexical properties and unaccusative/unergative syntax are developmentally unstable, whereas the unaccusative/unergative distinction itself is robust and unproblematic in acquisition. Second, it is shown that residual optionality, with its entailed gradedness eVects, occurs only in interface areas of the competence of near-native speakers, and not in purely syntactic domains. Sorace (2003b) indicates that the interpretable discourse features responsible for the distribution of overt and null subject pronouns are problematic in the L2 steady state of L1 English learners of Italian, whereas the non-interpretable syntactic features related to the null subject parameter are completely acquired. Third, it is argued that the diVerentiation between narrow computational syntactic properties and interface properties is also relevant in other domains of language development, such as L1 attrition due to prolonged exposure to a
Gradedness and Optionality in Mature and Developing Grammars
107
second language (Sorace 2000b; Tsimpli et al. 2004; Montrul 2004). A clear parallelism exists between the end-state knowledge of English near-native speakers of Italian and the native knowledge of Italian advanced speakers of English under attrition with respect to null/overt subjects and pre/postverbal subjects. In both cases, the speakers’ grammar is/remains a null-subject language: for example null subjects, when they are used, are used in the appropriate contexts, that is when there is a topic shift. The purely syntactic features of grammar responsible for the licensing of null subjects are not aVected by attrition. The generalization seems to be that constructions that belong to the syntax proper are resilient to gradedness in native grammars; are fully acquired in L2 acquisition; and are retained in L1 attrition. In contrast, constructions that require the interface of syntactic knowledge with knowledge from other domains are subject to gradedness eVects; present residual optionality in L2; and exhibit emergent optionality to L1 attrition. The question of the interpretation of this generalization, however, is still open. There are (at least) two issues for further research. First, there is a lack of clarity about the nature of diVerence interfaces. Are all interfaces equally susceptible to gradedness and optionality? Second, is gradedness inside or outside the speakers’ grammatical representations? The available evidence is in fact compatible both with the hypothesis that the gradedness is at the level of knowledge representations and with the alternative hypothesis that it arises at the level of processing. Possible approaches to these open issues are outlined.
6.2 The syntax–lexicon interface in native grammars: split intransitivity According to the unaccusative hypothesis (Perlmutter 1978; Burzio 1986), there are two types of intransitive verbs, unaccusative and unergative, with distinct syntactic properties. The essential insight (variously expressed by diVerent syntactic theories) is that the subject of unaccusative verbs is syntactically comparable to the object of a transitive verb, while the subject of an unergative verb is a true subject. Evidence for the distinction is both syntactic and semantic: in several European languages unaccusative verbs generally select BE as a perfective auxiliary while unergative verbs select have; semantically, the subject of unaccusative verbs tends to be a patient while that of unergative verbs is an agent. However, it has proved diYcult to Wt many verbs unambiguously into one class or the other. On the one hand, there are verbs
108
The Nature of Gradience
that do not satisfy unaccusativity diagnostics in consistent ways, both within and across languages: so blush is unaccusative in Italian (arrossire, selecting be) but unergative in Dutch (blozen, selecting have); Worire ‘blossom’ can take either have or be. On the other hand, there are verbs that can display either unaccusative or unergative syntax depending on the characteristics of the predicate: for example, all verbs of manner of motion (e.g. swim) select have in Dutch and German when they denote a process but take be in the presence of a prepositional phrase denoting a destination; verbs of emission (e.g. rumble; see Levin and Rappaport Hovav 1995 for extensive discussion) are unergative in their default case but in some languages may exhibit unaccusative behaviour when they receive a telic interpretation. One of the main challenges opened up by the unaccusative hypothesis is therefore how to account for the variable behaviour of verbs. A great deal of research in the last ten years has been devoted to explaining the complex mappings between a lexical-semantic level of representation and the level of syntactic structure. This eVort has taken two broad and seemingly incompatible directions. Theories of argument structure (which, following Levin and Rappaport Hovav (1996) may be termed ‘projectionist’) assume that the verb’s lexical entry contains the necessary speciWcation for the mapping of arguments onto syntactic positions. This approach posits Wne-grained distinctions in lexical-semantic representations, singles out the syntactically relevant lexical-semantic components in diVerent languages, and identiWes a set of linking rules that deterministically project lexical-semantic features onto syntactic positions, hence determining the unaccusative or unergative status of verbs. The second direction—named ‘constructional’ by Levin and Rappaport Hovav (1996)—empties the lexical entries of verbs of any syntactic speciWcation and makes semantic interpretation directly dependent on the syntactic conWgurations in which the verb can appear. If verbs are thus not tied to deterministic linking rules but have freedom of mapping, unaccusativity or unergativity become by-products of the verb’s compatibility with particular syntactic conWgurations, instead of inherent lexical properties of verbs (Borer 1994). However, both the projectionist and the constructional solutions to the lexicon–syntax puzzle have limitations; the most relevant is that the former allows for too little variation, because of the deterministic nature of its linking rules, whereas the latter allows too much variation, because of the lack of a mechanism that rules out impossible mappings. These limitations have been highlighted in particular by Sorace (1992, 1993a, 1993b, 1995, 2000b, 2003a), who has shown that there is systematic variation that cannot be explained by
Gradedness and Optionality in Mature and Developing Grammars CHANGE OF LOCATION >
109
Categorical unaccusative syntax
CHANGE OF STATE > CONTINUATION OF STATE > EXISTENCE OF STATE > UNCONTROLLED PROCESS > MOTIONAL PROCESS > NON-MOTIONAL PROCESS
Categorical unergative syntax
Figure 6.1. The split intransitivity hierarchy
either approach.1 Instead, she proposes that intransitive verbs are organized along a hierarchy (the split intransitivity hierarchy (SIH), originally called the auxiliary selection hierarchy (ASH)) deWned primarily by aspectual notions (telicity/atelicity), and secondarily by the degree of agentivity of the verb (Figure 6.1). The SIH is therefore an empirical generalization that identiWes the notion of ‘telic change’ at the core of unaccusativity and that of ‘atelic non-motional
1 The study of optionality and gradedness at interfaces can be measured experimentally with behavioural techniques that are able to capture subtle diVerences in speakers’ performance. The informal elicitation techniques traditionally used in linguistics and language development research (such as binary or n-point acceptability judgement tests) are unlikely to be reliable for such data, because they can measure only very broad distinctions and typically yield ordinal scales, in which the distance between points cannot be evaluated (Sorace 1996). A suitable experimental paradigm that has gained ground in recent years is magnitude estimation (ME), a technique standardly applied in psychophysics to measure judgements of sensory stimuli (Stevens 1975). The magnitude estimation procedure requires subjects to estimate the perceived magnitude of physical stimuli by assigning values on an interval scale (e.g. numbers or line lengths) proportional to stimulus magnitude. Highly reliable judgements can be achieved in this way for a whole range of sensory modalities, such as brightness, loudness, or tactile stimulation (see Stevens 1975 for an overview). The ME paradigm has been extended successfully to the psychosocial domain (see Lodge 1981 for a survey) and recently Bard et al. (1996), Cowart (1997), and Sorace (1992) showed that it may be applied to judgements of linguistic acceptability. Unlike the n-point scales conventionally employed in the study of psychological intuition, ME allows us to treat linguistic acceptability as a continuum and directly measures acceptability diVerences between stimuli. Because ME is based on the concept of proportionality, the resulting data are on an interval scale, which can therefore be analysed by means of parametric statistical tests. ME has been shown to provide Wne-grained measurements of linguistic acceptability, which are robust enough to yield statistically signiWcant results, while being highly replicable both within and across speakers. ME has been applied successfully to phenomena such as auxiliary selection (Bard et al. 1996; Sorace, 1992, 1993a, 1993b; Keller and Sorace 2003), binding (Cowart 1997; Keller and Asudeh 2001), resumptive pronouns (Alexopoulou and Keller 2003; McDaniel and Cowart 1999), that-trace eVects (Cowart 1997), extraction (Cowart 1997), and word order (Keller and Alexopoulou 2001; Keller 2000b).
110
The Nature of Gradience
activity’ at the core of unergativity. The closer to the core a verb is, the more determinate its syntactic status as either unaccusative or unergative. Verbs that are stative and non-agentive are the most indeterminate. Sensitivity to contextual or compositional factors correlates with the distance of a verb from the core. Thus, the ASH helps to account both for variability and for consistency in the behaviour of intransitive verbs. In contrast to the constructionist view, where context is always critical, the ASH account prescribes that core verbs have syntactic behaviour that is insensitive to non-lexical properties contributed by the sentence predicate. On the other hand, peripheral verbs, which are neither telic nor agentive, do seem to behave according to the constructionist observation, with syntactic behaviour depending on the properties of the predicate in which they appear. The SIH substantiates the intuition that, within their respective classes, some verbs are ‘more unaccusative’ and ‘more unergative’ than others (Legendre, Miyata, and Smolensky 1991). Crucially, however, this does not mean that unaccusativity or unergativity are inherently gradient notions, or that the distinction is exclusively semantic, but rather that some verbs allow only one type of syntactic projection whereas other verbs are compatible with diVerent projections to variable degrees. This is the reason why any approach that focuses exclusively on the syntactic or on the semantic side of split intransitivity is ultimately bound to provide only a very partial picture of the phenomena in this domain. While no formal model yet exists that can comprehensively account for the SIH, the SIH has given a new impetus to the search for such a model. Theoretical research inspired by the SIH has in fact been developed within diVerent frameworks and for diVerent languages (e.g. Bentley and Eyrtho´rsson 2004; Cennamo and Sorace in press; Keller and Sorace 2003; Legendre in press; Legendre and Sorace 2003; Randall in press; Mateu 2003; among others). Developmental evidence for the SIH comes from research on second language acquisition (Montrul 2004, in press; DuYeld 2003) and on Wrst language attrition (Montrul in press). Core verbs are the Wrst ones to be acquired with the correct auxiliary both in Wrst and second language acquisition. Data from the acquisition of Italian as a non-native language show that the syntactic properties of auxiliary selection are acquired earlier with core verbs and then gradually extended to more peripheral verb types (Sorace 1993a, 1995), although L2 learners do not attain the same gradient intuitions as those displayed by native Italians. Moreover, Italian learners of French Wnd it more diYcult to acquire avoir as the auxiliary for verbs closer to the core than for peripheral verbs (Sorace
Gradedness and Optionality in Mature and Developing Grammars
111
1993b, 1995), and do not completely overcome this diYculty even at the advanced level. A study by Montrul (in press) conWrms this pattern for L2 learners of Spanish, who have determinate intuitions on the syntactic correlates of split intransitivity in this language, but only with core verbs. These developmental regularities suggest two things. First, the acquisition of the syntax of unaccusatives crucially depends on the internalization of both the hierarchical ordering of meaning components, and the lexicon– syntax mapping system instantiated by the target language. The pattern uncovered by these data is consistent with an enriched constructional model, equipped with a checking mechanism that is sensitive to the degree of lexical speciWcation of verbs and rules out impossible mappings (see Mateu 2003). As it is the position of verbs on the ASH, rather than their frequency, which determines the order of acquisition, it seems that L2 learners do rather more than engaging in the kind of statistical learning envisaged by a basic constructional model. Second, and more generally, there are two sides to the split intransitivity question: a syntactic side (the structural conWguration that determines unaccusativity or unergativity) and a lexicon–syntax interface side (the mapping system that decides the syntactic behaviour of any given verb). Gradedness and indeterminacy in native grammars, as well as learning diYculties and residual problems in non-native grammars, tend to be situated on the interface side: the syntactic distinction itself is categorically stable.
6.3 The syntax–discourse interface in language development Developmental research points to the conclusion that the same areas of grammar appear to be unstable in other domains of language development and change, regardless of the circumstances in which development takes place. Areas of grammar that have been found to be particularly vulnerable to variable crosslinguistic inXuence in diVerent bilingual populations are those that involve the coordination of syntax and discourse. The obvious questions are why this should be so, whether crosslinguistic inXuence is the only cause of these phenomena, and whether the source of crosslinguistic inXuence is the same for each bilingual group. Before addressing these questions, the convergence between two developmental domains—L2 acquisition and L1 attrition—will be brieXy illustrated. 6.3.1 Endstate grammars One of the characteristics of L2 advanced grammars that has received attention recently is residual optionality, that is unsystematic L1 eVects surfacing in
112
The Nature of Gradience
L2 speakers’ production.2 A much-discussed case is subject realization in null-subject languages spoken by non-native speakers. It is well established that null subjects in these languages are syntactically licensed but their distribution is governed by discourse-pragmatic features (Rizzi 1982; Cardinaletti and Starke 1999). In Italian, a typical agreement-licensed nullsubject language, sentences such as (6.1a) are possible, whereas the equivalent sentence in English, (6.1b), is not: E` partito is-3rd SG gone. b. *Is gone.
(6.1) a.
Moreover, in Italian the option of a null or overt subject is conditioned by pragmatic factors, such as the [topic-shift] and the [focus] feature (Grimshaw and Samek-Lodovici 1998). Thus in (6.2), an overt pronoun lui in the subordinate clause can be co-referential with the complement Pietro, or with an extralinguistic referent, but not with the matrix subject Gianni. In contrast, a null pronoun in the same context signals co-referentiality with the topic Gianni. (6.2) Giannii ha salutato Pietrok quando proi / lui*i/k/j l’ha visto. Gianni has greeted Pietro when pro / he him-saw ‘Gianni greeted Pietro when he saw him.’ A characteristic feature of English near-native speakers of Italian (Sorace 2000a, 2003b; Filiaci 2003; Belletti et al. 2005) is that they optionally produce (6.3b), where a monolingual Italian speaker would produce (6.3c). Perche` why b. (perche`) (because) c. (perche`) (because)
(6.3) a.
Maria Maria lei she ___ has
e` is ha has ha found
andata gone trovato found trovato another
via? away? un altro lavoro. another job un altro lavoro. job
In contrast, the same speakers never produce a null pronoun when there is a shift of topic, as in (6.4b), or when the subject is contrastive, as in (6.5b). (6.4) a. Perche` Maria non ha parlato con nessuno? b. Perche *Ø (¼ Gianni) non l’ha neanche guardata because Ø (¼ Gianni) didn’t even look at her 2 Optionality is regarded here as the pre-condition for gradedness: the term refers to the co-existence in the same grammar of two alternative ways of expressing the same semantic content, of which one appears to be preferred over the other by the speaker in production and comprehension, creating gradedness eVects (Sorace 2000b).
Gradedness and Optionality in Mature and Developing Grammars (6.5) a. Maria Maria b. *No, Ø (¼ Paolo) No, Ø
ha has ha
detto che said that detto che
113
andava da Paolo? was going to Paolo’s? andava da lei
said that
he was her. going to A similar pattern obtains for the position of subjects with respect to the verb. In answer to an all-focus question, such as ‘what happened’, L1 English nearnative speakers of Italian optionally place the subject in preverbal position (6.6b), whereas native Italians would naturally place it after the verb (6.6c). This also happens in a narrow-focus context (6.7b), in which Italian requires the topic to be in postverbal position (6.7c). (6.6) a. Che cosa e` successo? b. Gianni Gianni c. E` is-3s
‘What happened ?’ e` is-3s partito left
partito. left Gianni Gianni
(6.7)
‘‘Who sneezed?’’ ha has-3s starnutito) sneezed
starnutito sneezed Gianni Gianni
a. Chi ha starnutito? b. Gianni Gianni c. (Ha Has-3s
These patterns are noticeably asymmetric: near-native speakers of Italian overgeneralize overt subject pronouns and preverbal subjects to contexts which would require null subjects and postverbal subjects in native Italian, but they do not do the reverse, namely they do not extend null and postverbal subjects to inappropriate contexts. In fact, when they use null pronouns and postverbal subjects, they use them correctly. These speakers therefore have acquired a nullsubject grammar. The optionality in their grammar does not aVect the syntactic licensing of null subjects, but is at the level of the discourse conditions on the distribution of pronominals and on the placement of subjects.3 It is worth noting that although the behaviour of native speakers is statistically diVerent from that of near-natives, it is not categorical. In a small number of cases, native speakers also over-produce overt subjects and postverbal subjects in inappropriate discourse contexts. The signiWcance of 3 The few existing studies on near-native L2 grammars point to a similar split between purely syntactic constraints, which are completely acquired, and interpretive conditions on the syntax, which may or may not be acquired. See Sorace (2003b, in press) for details.
114
The Nature of Gradience
this detail lies in the fact that the options favoured by near-native speakers are (strongly) dispreferred by natives, but they are not illicit in their grammar. 6.3.2 L1 attrition There is evidence that the same pattern of asymmetric optionality is exhibited by native speakers of null subject languages who have had prolonged exposure to English. Research on changes due to attrition from another language (Sorace 2000a, Tsimpli et al. 2004) indicates that native Italians who are near-native speakers of English exhibit an identical pattern of optionality as the English near-native speakers of Italian described above: these speakers overgeneralize overt subjects and preverbal subjects to contexts which require a null subject or a postverbal subject. The reverse pattern is not found. It is worth noting that the phenomenon is found both in production and in comprehension. For example, in the forward anaphora sentences in (6.8b), speakers under attrition are signiWcantly more likely than monolingual Italians to judge the overt pronoun as coreferential with the matrix subject ‘Maria’; however, the null pronoun in (6.8a) is correctly interpreted as referring to the matrix subject. These speakers are also more likely to produce sentences such as (6.9a) regardless of whether the subject is deWnite or indeWnite, whereas monolingual speakers would prefer a postverbal subject, as in (6.9b), particularly when the subject is indeWnite. (6.8)
a. b.
(6.9)
a.
b.
Mentre while Mentre while
pro LEI she
attraversa is crossing attraversa is crossing
la strada, the street, la strada, the street,
Maria Maria Maria Maria
saluta greets saluta greets
la sua her la sua her
amica friend amica friend
Hai sentito che un palazzo/il palazzo dell’ONU e` crollato? Have heard that a building /the UN building collapsed? you Hai sentito che e` crollato un palazzo/il palazzo dell’ONU? Have heard that is collapsed a building/the UN building? you
Thus, there is a parallelism between the end-state knowledge of English nearnative speakers of Italian and the native knowledge of Italian near-native speakers of English under attrition with respect to null/overt subjects and
Gradedness and Optionality in Mature and Developing Grammars
115
pre/postverbal subjects: the speakers’ grammar is and remains a null-subject language. The computational features of grammar responsible for the licensing of null subjects are acquired completely, and are not aVected by attrition.4 More generally, changes induced by attrition in individual speakers primarily aVect morphosyntactic features that are interpretable at the interface with conceptual systems. The aVected features may become unspeciWed, giving rise to optionality. Thus, attrition is expected to aVect the use of overt subjects in L2 Italian (given that this is regulated by the interpretable [topic-shift] and [focus] features). If these features become unspeciWed, overt subjects in Italian under attrition are not necessarily being used or interpreted as shifted topics or foci.5 The lexicon–syntax interface conditions governing the syntactic behaviour of intransitive verbs are also vulnerable to attrition. Montrul’s (in press) study on generational attrition in Spanish heritage speakers found that attrition aVects the mappings of individual verbs onto unaccusative/unergative syntax: these speakers do not show a sensitivity to diVerent subclasses represented on the SIH, and their determinate intuitions are restricted to core verbs. Montrul’s study shows that this is the same pattern obtained for L2 learners of Spanish. Once again, both bilingual groups have a robust syntactic representation of split intransitivity, but exhibit instability with respect to the lexicon– syntax interface conditions regulating the distribution of verbs into one syntactic class or the other.
6.4 Interpreting gradedness and optionality At this point we need a generalization that describes these converging patterns of results. A Wrst approximation might be the following:
4 Other cases of selective attrition at interfaces are discussed in Montrul (2002) with respect to the tense/aspect domain in Spanish; Polinsky (1995) with respect to the distinction between reXexive and possessive pronouns in Russian; Gu¨rel (2004) on pronominals in Turkish. 5 Studies on bilingual Wrst language acquisition converge with the results of research on L2 acquisition and L1 attrition. The syntax–pragmatics interface has been identiWed as a locus of crosslinguistic inXuence between the bilingual child’s syntactic systems (Mu¨ller and Hulk 2001). Bilingual children who simultaneously acquire a null-subject language and a non-null-subject language overproduce overt subjects in the null-subject language (see Serratrice 2004 on Italian–English bilinguals; Paradis and Navarro 2003 on Spanish–English bilinguals; Schmitz (2003) on Italian–German bilinguals). Thus, crosslinguistic eVects obtain only from the non-null-subject language to the null-subject language and never in the other direction, regardless of dominance.
116
The Nature of Gradience
(6.10) ‘Narrow’ versus ‘Interface’ syntax: . Non-interpretable features that are internal to the computational system of syntax proper and drive syntactic derivations are categorical in native grammars; are acquired successfully by adult L2 learners; and are retained in the initial stages of individual attrition. . Interpretable features that ‘exploit’ syntactic options and belong to the interface between syntax and other domains, such as the lexicon, discourse, or pragmatics, may exhibit gradedness in native grammars; may present residual optionality in near-native grammars, due to the inXuence of the native language even at the most advanced competence stage; and are vulnerable to change in individual attrition.6 This generalization, which is compatible with theoretical assumptions in the minimalist programme (Chomsky 1995), assumes the existence of diVerent ‘layers’ of syntactic knowledge and places these phenomena at the level of syntactic representations: hence, within the speaker’s grammatical competence. Interpretable features at interfaces are more vulnerable to underspeciWcation, in both native and non-native grammars, and are therefore more prone to gradedness and optionality (Sorace and Keller 2005). This is the analysis adopted by Tsimpli et al. (2004) for L1 attrition: no evidence of attrition is found in the parameterization of purely formal syntactic features, whereas attrition is evident with respect to the distribution of subjects in appropriate pragmatic contexts, which is regulated by interpretable interface features. A similar diVerentiation between narrow syntax and interface properties may be found in the domain of split intransitivity. The unaccusative–unergative distinction is a syntactically represented, potentially universal property, and it belongs to narrow syntax. As argued in much recent research, the syntactic conWguration that determines the unaccusativity of verbs contains a telicity aspectual projection (Borer 1994; van Hout 2000). Core unaccusative verbs are inherently lexically speciWed for telicity and categorically project onto the unaccusative conWguration: they are determinate, acquirable, and 6 At Wrst sight, it may appear as if the generalization just presented contradicts decades of L2 acquisition research. In particular, early research showed that semantically more transparent properties are easier to learn than more abstract syntactic properties that do not correspond in any clear way to semantic notions (see e.g. Kellerman 1987). Moreover, studies of the ‘basic variety’ argued that early interlanguage grammars favour semantic and pragmatic principles of utterance organization (Klein and Perdue 1997). However, the argument here is NOT that syntactic aspects are easier than semantic aspects, but that aspects of grammar that require not only syntactic knowledge, but also the integration of syntactic knowledge with knowledge from other domains are late acquired, or possibly never completely acquired by L2 learners.
Gradedness and Optionality in Mature and Developing Grammars
117
stable. Knowledge of split intransitivity, however, also involves mastery of the behaviour of non-core verbs and their compositional interpretation in the predicates in which they appear: this is acquired gradually through exposure to particular verbs in speciWc aspectual contexts; gives rise to variable intuitions; and is unstable in a situation of attrition. Recent proposals in syntactic theory further reWne the distinction between narrow syntax and interface properties. The distinction between formal licensing of a null pro and the discourse-related conditions on postverbal subjects, as well as on the use/interpretation of subject pronouns is highlighted by the ‘cartographic’ theoretical framework (Belletti 2004; Rizzi 2004). It is assumed within this theory that the low part of the clause contains discourse-related positions, labelled ‘Topic’ and ‘Focus’, which constitute a clause internal VP periphery. Postverbal subjects Wll one of these dedicated positions according to their interpretation in discourse contexts. For example, a sentence such as (6.11b) is associated with the representation in (6.12), where the subject Wlls the speciWer of the (new information) low Focus projection and is therefore interpreted as conveying new information; the preverbal subject position is occupied by pro: (6.11) a. Chi ha starnutito? b. Ha starnutito Gianni (6.12) [CP . . . [TP pro . . . ha starnutito . . . [TopP [FocP Gianni [TopP [VP . . . .] ] ] ] ] ] According to this position, the formal syntactic licensing of pro is a necessary, but not suYcient condition for VS, since the postverbal subject also requires activation of the VP periphery. The experimental data illustrated earlier indicate that it is precisely this further condition that remains problematic in near native L2 speakers of Italian: these speakers often fail to activate the VP-internal focus position required by focalization in Italian.7 The picture is furthercomplicated by the existence of diVerent phenomena that involve an interface between syntax and discourse. Do they all present gradedness and optionality? At a developmental level, there are theoretical and empirical arguments in favour of a distinction between discourse-related phenomena that are also relevant to LF, and those that are relevant only to the syntax–discourse 7 The result is the use of focus in-situ, namely an L1-based strategy that is more economical because it involves a DP-internal focus position (as the one overtly manifested in a sentence like ‘John himself sneezed’). It is worth noticing that L1 French speakers of L2 Italian often use clefting in the same context (Leonini and Belletti 2004), which is an alternative way of activating the VP-periphery (as shown in the example below) and is widely available in French.
(i) Ce . . . . [Top [Foc [Top [VP eˆtre [sc Jean [CP qui a e´ternue´] ] ] ] ] ]
118
The Nature of Gradience
interface. LF-relevant phenomena may pose developmental problems at intermediate stages, but they are ultimately acquired; LF-irrelevant phenomena raise persistent problems at all stages. Moreover, the former normally determine grammaticality eVects, whereas degrees of preference are associated with the latter. For example, syntactic focusing in languages such as Hungarian and Greek involves the formation of an operator-variable structure at LF (cf. Kiss 1998; Szendroi 2004), which causes verb-raising to C/F and associated grammaticality eVects, as opposed to degrees of preference. Focus movement of nonsubject arguments as well as adverbs and participles is unproblematic both in advanced L2 speakers of Greek and in native Greek speakers in an attrition situation. In contrast, both groups of speakers exhibit more variable performance on overt subject pronouns, a discourse-related phenomenon (Tsimpli and Sorace 2005). The diVerences observed between these structures may be due to the fact that discourse, in the sense of pragmatic conditions on the distribution of subject pronouns, is outside grammar proper, whereas the LF-interface is aVected by modular computations within the language system. Even when L2 speakers attain native-like knowledge of properties relevant to LF representations, optionality and crosslinguistic eVects remain possible at the discourse level where pragmatic and processing constraints aVect L2 use. These theoretical reWnements have begun to unravel the complexity of the factors that determine instability at interfaces. In doing so, however, they have also magniWed the fundamental ambiguity of the notion of interface and thus the diYculty of establishing the origins of interface instability. Do interfaces give rise to indeterminacy at the representational level, or is gradedness a phenomenon external to syntactic representations? The ambiguity is apparent in many recent studies. For example, Jakubowicz (2000) argues for the relevance of the notion of syntactic complexity in research on early normal and SLI child grammars, claiming that: (a) constructions requiring the integration of syntactic knowledge and knowledge from other domains are more complex than constructions requiring syntactic knowledge only; and (b) a syntactic operation is less complex if it is obligatorily required in every sentence; it is more complex if it is present only in some sentences because of semantic or pragmatic choices. The felicitous use of complex constructions, according to this deWnition, demands the simultaneous mastery of both the morphosyntactic properties of given constructions and of the discourse conditions governing their distribution and use.8 But is ‘complexity’ related 8 Avrutin (2004) goes a step further and regards ‘discourse’ as ‘a computational system (my emphasis) that operates on non-syntactic symbols and is responsible for establishing referential dependencies, encoding concepts such as ‘‘old’’ and ‘‘new’’ information, determining topics, introducing discourse presuppositions, etc’. Investigating the interface between syntax and discourse necessarily requires going beyond ‘narrow syntax’.
Gradedness and Optionality in Mature and Developing Grammars
119
to problems internal to the speaker’s representation of syntactic knowledge, or are these problems external to these representations and resulting from processing diYculties in integrating knowledge from diVerent domains? L2 studies on other potentially problematic interfaces (e.g. the syntax– morphology interface (Lardiere 1998; Pre´vost and White 2000; White 2003)) point towards the latter explanation, suggesting that persistent problems with inXectional morphology in endstate grammars may in fact be ‘surface’ problems related to the retrieval of the correct morphological exponents for abstract syntactic features. The fact that learners’ problems tend to be with missing inXection, as opposed to wrong inXection, suggests the existence of diYculties at the level of access to knowledge, rather than with knowledge itself, which lead to the optional use of ‘default’ underspeciWed forms. The choice of referential pronouns in Italian qualiWes as complex, since it demands the simultaneous mastery of both morphosyntactic properties and discourse conditions. In contrast, referential subject pronouns in English are less complex because there is no choice of diVerent forms that is conditioned by discourse factors.9 Similar diVerentiations in terms of complexity can be made for some manifestations of split intransitivity. Perfective auxiliaries in Italian are more complex than in English, because only Italian requires a choice of auxiliaries governed by lexical-semantic and aspectual features of the verb. Auxiliary choice in Italian is also more complex than in French, because in French the only verbs that take BE are those inherently speciWed for telicity, and therefore there is no auxiliary selection dependent on the evaluation of the properties of the predicate. Within Italian, auxiliary selection with core verbs is less complex than with non-core verbs, because selection with the latter has to take into account both the properties of the verb and other characteristics of the predicate. It is therefore possible to propose an alternative generalization on the nature of interfaces: (6.13) Processing complexity . Structures requiring the integration of syntactic knowledge and knowledge from other domains are more complex than structures requiring syntactic knowledge only. . Complex structures may present gradedness and variation in native grammars; may pose residual diYculties to near-native L2 speakers; may pose emerging diYculties to L1 speakers experiencing attrition 9 As pointed out by a referee, the interface with discourse conditions obviously aVects other aspects of pronominal use in English, such as the distribution of stress.
120
The Nature of Gradience from a second language because of increasingly frequent failure to coordinate/integrate diVerent types of knowledge.
This hypothesis Wnds a fertile testing ground in recent research on L2 processing. . For L2 speakers, recent evidence from on-line psycholinguistic (Felser et al. 2003; Kilborn 1992) and neuroscience experiments (particularly ERPs, see Hahne and Friederici 2001); . Sabourin (2003) indicates that syntactic processing (i.e. access to ‘Narrow Syntax’) continues to be less than optimally eYcient in nonnative speakers even at advanced levels. If syntactic processing is less eYcient in L2 speakers than in L1 speakers, the coordination of syntax with other domains is aVected because speakers have insuYcient processing resources to carry it out. When coordination fails, speakers resort to the most ‘economical’ option. Crosslinguistic inXuence from English may thus not be the only cause of the over-use of overt subject pronouns in Italian– English bilinguals. Rather, this behaviour is favoured by two concomitant factors: on the one hand, the availability of the English option, which is economical in processing terms; on the other, the speakers’ sub-optimal processing resources.10 Factors related to inadequate parsing resources also Wgure prominently in a recent proposal on the nature of language learners’ grammatical processing by Clahsen and Felser (in press). Accounting for L2 speakers’ divergent behaviour, according to this proposal, does not necessarily involve positing ‘representational deWcits’: L2 speakers can, and indeed do, attain target representations of the L2, but may compute incomplete (‘shallow’) syntactic parses in comprehension. Such shallow processing is often accompanied by reliance—or overreliance—on lexical, semantic, and pragmatic information, which can lead to seemingly trouble-free comprehension in ordinary communication. If the notion of shallow processing is extended to production (see Sorace, in press, for discussion), one may plausibly assume that shallow processing may result in non-native speakers’ lack of activation of the VP periphery in narrow focus contexts. Interpreting these phenomena in the light of Clahsen and Felser’s hypothesis allows us to identify their source in the persistence of an L1-based discourse ‘prominent’ strategy, employed to compensate for the failure to compute the required L2 syntactic representation, despite the potential grammatical availability of the latter. In comprehension, shallow 10 A similar argument is developed by Rizzi (2002), who accounts for the presence of null subjects in early child English grammars by assuming that this is an option structurally available to the child, which also happens to be favoured in terms of limited processing resources.
Gradedness and Optionality in Mature and Developing Grammars
121
processing may similarly involve the optional lack of activation of the VP periphery, which is necessary to the reading of the postverbal subject as carrying focus on new information. The result may be, for example, an ‘old information’ interpretation of an indeWnite postverbal DP. Analogous considerations may be extended to the diVerent distribution of overt subject pronouns in near-native Italian. The non-native production and interpretation of overt subject pronouns may be the result of shallow processing of the interface mapping governing the use of overt subjects (e.g. the obligatory presence of the feature ‘topic shift’; see Tsimpli et al. 2004) and the consequent assimilation of strong Italian pronouns to the corresponding weak English pronouns, which—unlike Italian overt pronouns—can refer to topic antecedents. The strategy used in these circumstances would be diVerent from the use of overt pronouns in a default form to relieve processing overload due, for example, to insuYcient knowledge of (or access to) agreement inXection (Bini 1993; Liceras et al. 1999; Sorace in press). This account crucially involves the optionality of shallow processing, that is, the L2 speakers’ ability to perform full processing, at least at the near-native level. Shallow processing, in this sense, would be a relief strategy that is available to all speakers but is relied on especially by bilingual speakers.11 For this reason, native speakers should not be immune from occasional interface coordination diYculties, for example in situations of competing processing demands. Indeed, a study by Serratrice (2004) shows that older monolingual Italian children (aged 8+) overproduce overt referential subjects, although not to the same extent as English–Italian bilingual children. This seems to suggest that the ‘interface’ conditions relating subject pronouns to discourse factors are late acquired because they are more demanding. As already mentioned, even Italian adult monolingual control groups in bilingual studies (Tsimpli et al. 2004; Filiaci 2003; Belletti et al. 2005) do not show categorically correct behaviour with respect to subject pronouns; they (sporadically) make unidirectional mistakes that always involve the inappropriate use of overt subjects in contexts that would favour null subjects.12 11 One should not lose sight of the fact that these diYculties are resolved in ways that betray the inXuence of universal factors. Optionality favours the retention and occasional surfacing of unmarked options which are subject to fewer constraints, consistent with typological trends (see Bresnan 2000). 12 The extension of overt subject pronouns to null subject contexts is attested in another situation in which knowledge of English is not a factor. Bini (1993) shows that Spanish learners of Italian up to an intermediate proWciency level use signiWcantly more overt subjects than monolingual Italians and monolingual Spanish speakers. Since the two languages are essentially identical with respect to both the syntactic licensing of null subjects and the pragmatic conditions on the distribution of pronominal forms, L1 inXuence is not a relevant factor here. This pattern is therefore likely to be due exclusively to coordination diYculties leading to the use of overt subjects as a default option.
122
The Nature of Gradience
6.4.1 The role of external ‘destabilizing’ factors Finally, the quantitative and qualitative characteristics of the input to which speakers are exposed may play a role in accounting for the instability found at interfaces in diVerent speaker populations. The quantitative factor is evident in bilingual use. What L2 near-native speakers and L1 speakers under attrition have in common is the fact that their total exposure to the language is reduced compared to that of monolingual speakers: in the case of L2 speakers, because they started the process of L2 acquisition in adulthood; in the case of L1 speakers under attrition, because they are no longer exposed to the L1 in a continuous way.13 Qualitative diVerences may be less obvious but are equally relevant: both the near-native speakers of Italian and the native Italian experiencing attrition are likely to receive input from native Italians in a situation of attrition and from other non-native Italian speakers; they may be exposed to non-native Italian from their spouse, and to ‘bilingual’ Italian from their children. Thus, these speakers receive qualitatively diVerent input that is consistent with, and reinforces, their own grammar. Gradedness and indeterminacy in split intransitivity is also fed and maintained, both in native and non-native speakers, by the input, which is categorical and uniform (and therefore rich in terms of frequency) for core verbs and variable for non-core verbs.14 It is intriguing to ask exactly what ‘destabilizing’ eVects may be brought about by the quantitative and qualitative diVerences in the input to which speakers are exposed, and grammars are aVected in diVerent ways. One possible hypothesis is that quantitative diVerences are likely to aVect processing abilities, because speakers have fewer opportunities to integrate syntax 13 Clearly the quantitative factor is also a function of age of Wrst exposure: it cannot be considered in absolute terms. Thus, an L2 speaker may have been exposed to the language for many decades and still exhibit non-native behaviour compared to a younger native speaker who has been exposed to the language for a shorter time, but since birth. 14 Variation at interfaces may be regarded as the motor of diachronic change, because it is at this level that ‘imperfect acquisition’ from one generation to the next is likely to begin. Sprouse and Vance’s (1999) study of the loss of null subjects in French indicates that language contact created a situation of prolonged optionality, that is competition between forms that make the same contribution to semantic interpretation, during which the null-subject option became progressively dispreferred in favour of the overt-subject option because it is the less ambiguous in processing terms. An analogous situation is experienced by the native Italian speaker after prolonged exposure to English: this speaker will be exposed both to null pronouns referring to a topic antecedent (in Italian) and to overt pronouns referring to a topic antecedent (in English, and also in the Italian of other native speakers in the same situation). Optionality, and competition of functionally equivalent forms, is therefore as relevant in this situation as in diachronic change. The diachronic loss of auxiliary choice in Romance languages may also be traced as beginning from non-core verbs and gradually extending to core verbs (Sorace 2000b; Legendre and Sorace 2003).
Gradedness and Optionality in Mature and Developing Grammars
123
and other cognitive domains in interpretation and production; qualitative diVerences, on the other hand, may aVect representations, because speakers would receive insuYcient evidence for interface mappings. Generally, it seems that exposure to consistent input up to a certain threshold level is necessary both for acquiring and maintaining an eYcient syntactic system.
6.5 Conclusions To conclude, I have presented evidence of gradedness and optionality in native and non-native grammars whose locus seems to be the interface between syntactic and other cognitive domains. There are two potential explanations for these patterns. One involves underspeciWcation at the level of knowledge representations that involves the interaction of syntax and other cognitive domains, such as lexical-semantics and discourse-pragmatics. The other involves insuYcient processing resources for the coordination of diVerent types of knowledge. Furthermore, there are diVerent kinds of interfaces, not all of which are susceptible to gradedness eVects either in stable or in developing grammars. Behavioural and neuropsychological evidence suggests that syntactic processes are less automatic in L2 speakers than in L1 speakers, which in turn may increase coordination diYculties. L2 speakers may also have inadequate resources to carry out the right amount of grammatical processing required by on-line language use, independently of their syntactic knowledge representations. The processing and the representational explanations, however, do not necessarily exclude each other, and indeed seem to work in tandem, particularly in the case of bilingual speakers. Furthermore, syntactic representations and processing abilities may be diVerentially aVected over time by quantitative and qualitative changes occurring in the input to which speakers are exposed. Future research is needed to ascertain the plausibility, and work out the details, of a uniWed account of gradedness and optionality in native and non-native grammars.
7 Decomposing Gradience: Quantitative versus Qualitative Distinctions M AT T H I A S S C H L E S EWS K Y, I NA B O R N K E S S E L A N D BRIAN MCELREE
7.1 Introduction Psycho- and neurolinguistic research within the last three decades has shown that speaker judgements are subject to a great deal of variability. Thus, speakers do not judge all sentences of a given language equally acceptable that are assumed to be grammatical from a theoretical perspective. Likewise, ungrammatical sentences may also vary in acceptability in rating studies conducted with native speakers. These Wndings stand in stark contrast to the classical perspective that grammaticality is categorical in that a sentence is either fully grammatical or fully ungrammatical with respect to a particular grammar. This apparent contradiction has, essentially, been approached from two diVerent directions. On the one hand, it has been proposed that judgement variability—or gradience—results from extra-grammatical ‘performance factors’ and that it therefore has an origin distinct from linguistic ‘competence’ (Chomsky 1965). Alternatively, the gradience of linguistic intuitions has been described in terms of varying markedness of the structures in question. Rather than appealing to variation in grammaticality, this latter approach introduces and appeals to an additional grammar-internal dimension. The idea that structures can vary in acceptability for grammar-internal reasons has found expression in the use of question marks, hash marks, and the like to describe the perceived deviation from the endpoints of the grammaticality scale. Importantly, it must be kept in mind that judgements of acceptability— whether they are binary judgements or numerical ratings—represent
Quantitative versus Qualitative Distinctions
125
unidimensional assessments of what is inherently a multidimensional signal. In essence, intuitions of acceptability reXect the endpoint of a complex sequence of processes underlying sentence comprehension or production. Consequently, it is possible that two sentence structures judged to be equally unacceptable may be unacceptable for rather diVerent reasons. Consider, for example, the following three German examples: (7.1) a. Dann hat der Lehrer dem Jungen den Brief gegeben. then has [the teacher]NOM [the boy]DAT [the letter]ACC given ‘Then the teacher gave the letter to the boy.’ b. ??Dann hat dem Jungen den Brief der Lehrer gegeben. then has [the boy]DAT [the letter]ACC [the teacher]NOM given ‘Then the teacher gave the letter to the boy.’ c. *Dann hat der Lehrer gegeben dem Jungen den Brief. then has [the teacher]NOM given [the boy]DAT [the letter]ACC
Example (7.1a) illustrates the canonical argument order in German: nominative > dative > accusative. In (7.1b), two argument permutations have resulted in the order dative > accusative > nominative. Argument serializations of this type are typically analysed as grammatical, but are highly marked. Example (7.1c), by contrast, is ungrammatical because of the positioning of the participle, which should be clause-Wnal. Interestingly, structures such as (7.1b) and (7.1c) are consistently judged to be equally (un-)acceptable in rating studies of various types, including questionnaire studies (Pechmann et al. 1996) and speeded acceptability ratings (Ro¨der et al. 2002; Fiebach et al. 2004). Thus, it is not possible to discriminate between the two sentence types by relying on linguistic intuitions alone. Fortunately, however, other measures can eVectively discriminate between the structures in (7.1). In a recent study using functional magnetic resonance imaging (fMRI) to map the brain areas involved in the processing of sentences such as (7.1), Fiebach et al. (2004) showed that the observed pattern of acceptability can be traced back to two distinct sources of neural activation (Figure 7.1). Whereas complex grammatical sentences (e.g. 7.1b) gave rise to an enhanced activation in Broca’s area (the pars opercularis of the left inferior frontal gyrus, BA 44), ungrammatical structures (e.g. 7.1c) engendered activation in the posterior deep frontal operculum. The data reported by Fiebach et al. (2004) thus provide a compelling demonstration that overt judgements of sentence acceptability (or grammaticality) may not provide an adequate means of determining the underlying diVerences in acceptability of various sentence structures.
126
The Nature of Gradience
scrambling effect ungrammaticality effect
Figure 7.1. A schematic illustration of the activations elicited by the complexity (scrambling) and the grammaticality manipulation in Fiebach et al. (2004)
In this chapter, we draw upon studies of word order variation in German to examine how linguistic judgements emerge from the real-time comprehension processes. On the basis of a number of empirical observations, we argue that gradient data need not be interpreted as evidence against categorical grammars. Rather, gradience can arise from a complex interaction between grammar-internal requirements, processing mechanisms, general cognitive constraints1 and the environment within which the judgement task is performed. We begin by describing the critical phenomena from a behavioural perspective (Section 7.2), before turning to experimental methods providing more Wne-grained data (event-related brain potentials, ERPs; speed-accuracy trade-oV, SAT) (Section 7.3).
7.2 The phenomenon: argument order in German Argument order variations in German are typically classiWed along several dimensions. First, a permuted argument may occupy either the sentenceinitial position (the Vorfeld) or reside in a clause-medial position (in the Mittelfeld).2 Secondly, the type of permuted argument (wh-phrase, pronoun, etc.) is also of crucial importance. In this way, four permutation types are distinguished: topicalization (7.2a), wh-movement (7.2b), scrambling (7.2c), and pronoun ‘movement’ (7.2d). 1 It has been shown, for example, that the Wnal interpretation of a sentence may vary interindividually as a function of general cognitive capacity. Thus, researchers have distinguished between fast and slow comprehenders (e.g. Mecklinger et al. 1995), good and poor comprehenders (e.g. King and Kutas 1995), high and low verbal working memory capacity as measured by the reading span test (King and Just 1991) and individual alpha frequency (Bornkessel et al. 2004b). However, a discussion of these factors is beyond the scope of this chapter. 2 The Mittelfeld is the region of the German clause that is delimited to the left by a complementizer (subordinate clauses) or Wnite verb in second position (main clauses) and to the right by a clause-Wnal participle or particle.
Quantitative versus Qualitative Distinctions
127
(7.2) a. Topicalization (Vorfeld, -wh) Den Arzt hat wahrscheinlich der Journalist [the journalist]NOM [the doctor]ACC has probably eingeladen. invited ‘The journalist most likely invited the doctor.’ b. Wh-movement (Vorfeld, +wh) Welchen Arzt hat wahrscheinlich der Journalist [the journalist]NOM [which doctor]ACC has probably eingeladen? invited ‘Which doctor did the journalist most likely invite?’ c. Scrambling (Mittelfeld, non-pronominal) Wahrscheinlich hat den Arzt der Journalist probably has [the doctor]ACC [the journalist]NOM eingeladen. invited ‘The journalist most likely invited the doctor.’ d. Pronoun ‘movement’ (Mittelfeld, pronominal) Wahrscheinlich hat ihn der Journalist eingeladen. probably has [him]ACC [the journalist]NOM invited ‘The journalist most likely invited him.’ In addition to these four theoretically motivated permutation types, psychoand neurolinguistic studies implicate an additional dimension, namely whether the permuted argument is unambiguously case marked (e.g. den ¨ rztin, ‘[the Arzt, ‘[the (male) doctor]NOM’) or case ambiguous (e.g. die A (female) doctor]NOM/ACC’). From a comprehension perspective, the diVerence between unambiguous and ambiguous case marking lies in the fact that the former immediately signals the presence of an argument order variation, while the latter does not. Empirical evidence indicates that, when faced with an ambiguity, German speakers initially pursue a strategy in which the Wrst ambiguous argument is analysed as the subject of the clause (e.g. Frazier and d’Arcais 1989; de Vincenzi 1991; Bader and Meng 1999). Only when information contradicting this analysis is encountered is a reanalysis of the clause initiated. In this way, the processes leading to the recognition of an argument order variation diVer qualitatively in unambiguous and ambiguous situations. The various types of argument order variations in German have been subject to a number of empirical investigations using diVerent types of acceptability measurements. From these Wndings, the four central generalizations in (7.3) emerge.
128
The Nature of Gradience
(7.3) Generalizations with regard to the acceptability of argument order variations in German (i) object-initial sentences are generally less acceptable than their subject-initial counterparts (Krems 1984; Hemforth 1993); (ii) acceptability decreases with an increasing number of permutations (Pechmann et al. 1996; Ro¨der et al. 2000); (iii) the acceptability of object-initial sentences decreases when the permuted object is case ambiguous (Meng and Bader 2000a); (iv) the acceptability diVerence between object- and subject-initial structures varies according to the following hierarchy: scrambling > topicalization > wh-movement > pronoun movement (Bader and Meng 1999). The four generalizations summarized in (7.3) interact to produce the overt acceptability pattern seen in German argument-order variations, giving what appear to be a highly gradient set of linguistic intuitions. To cite just one example, Meng and Bader (2000b) report a 49 percent acceptability rate for the scrambling of ambiguous accusative objects. As participants were asked to provide yes–no judgements, this amounts to chance-level performance.3 However, the generalizations in (7.3) emerged almost exclusively from studies on the permutation of accusative objects in transitive sentences. By contrast, when the relative ordering between dative- and nominative-marked arguments is examined, at least two interesting exceptions to this general pattern become apparent. First, the severe drop in acceptability for scrambled (transitive) objects is attenuated when the object bears dative case marking: Schlesewsky and Bornkessel (2003) report an 86 percent acceptability rate for initially ambiguous dative-initial structures similar to those engendering a 49 percent acceptability rate for accusative-initial structures in the Meng and Bader (2000b) study. Secondly, in sentences with dative object-experiencer verbs—which project an argument hierarchy in which the dative-marked experiencer outranks the nominative-marked stimulus—the acceptability decrease for object-initial orders is neutralized or even tendentiously reversed (Schlesewsky and Bornkessel 2003; see Table 7.1). Dramatic diVerences of this sort call for an explanation. We believe that an adequate explanation requires, as a Wrst step, an accurate, Wne-grained characterizationoftheoutputsignalthatconstitutesanacceptability judgement.
3 Note that this essentially amounts to the same level of performance that a non-human primate might be expected to produce when confronted with the same sentences and two alternative pushbuttons. As such, chance-level acceptability deWes interpretation.
Quantitative versus Qualitative Distinctions
129
Table 7.1. Acceptability ratings for locally ambiguous subjectand object-initial sentences with dative active and dative objectexperiencer verbs Sentence type Subject-Wrst, active verb Subject-Wrst, object-experiencer verb Object-Wrst, active verb Object-Wrst, object-experiencer verb
Mean acceptability in % 92.3 84.6 86.5 91.7
Source : Schlesewsky and Bornkessel 2003
Thus, we must examine how the judgement ‘emerges’ from the on-line comprehension process.
7.3 Sources of gradience A natural Wrst step in tracing the emergence of a linguistic judgement lies in the examination of the comprehension system’s initial response to the variation under consideration, that is, for present purposes, to the argument order permutation. How does the system react when it encounters an object before a subject and is there evidence for an immediate diVerentiation between the diVerent permutation types? A methodological approach optimally suited to answering this question is the measurement of event-related brain potentials (ERPs; see Appendix 1). Because of their very high temporal resolution and their multi-dimensional characterization of neuronal activity, ERPs allow for an exquisitely precise diVerentiation of various cognitive processes, and, not surprisingly, many researchers have capitalized upon these properties to explore language processing. Most of the argument order variations discussed above have been subjected to examinations using ERPs, but we restrict our discussion to order permutations in dative constructions and how these contrast with those in accusative structures. Moreover, we will focus primarily on Wndings for initially ambiguous structures, as these reveal the inXuence of processing considerations on acceptability ratings most clearly. From the perspective of a strong competence versus performance distinction, these structures might be considered a ‘worst case scenario’ and are as such well-suited to examining the limitations of supposedly time insensitive linguistic judgements. The exploration of ambiguous sentences provides fruitful ground for investigating these issues. Consider, for example, the sentence fragment in (7.4):
130
The Nature of Gradience
(7.4)
. . . dass Dietmar Physiotherapeutinnen . . . . . . that DietmarNOM/ACC/DAT physiotherapistsNOM/ACC/DAT
When confronted with an input such as (7.4), the processing system initially analyses the Wrst argument Dietmar as the subject of the clause (e.g. Hemforth 1993; Schriefers et al. 1995; Bader and Meng 1999; Schlesewsky et al. 2000). Accordingly, the second argument Physiotherapeutinnen—which does not contradict the initial assignment—is analysed as an object. If, however, the clause is completed by a plural verb such as beunruhigen (‘to disquiet’), the supposed subject of the clause no longer agrees with the Wnite verb. Thus, a reanalysis towards an object-initial order must be initiated in order for a correct interpretation to be attained. In terms of ERP measures, reanalyses are typically associated with a late (approximately 600–900 ms) positivity with a posterior distribution (P600; e.g. Osterhout and Holcomb 1992). Indeed, this component has also been observed for the reanalysis of argument order in German, for example in whquestions (beim Graben et al. 2000), topicalizations (Frisch et al. 2002) and scrambled constructions (Friederici and Mecklinger 1996). Note, however, that all of these studies only manipulated the word order of accusative structures. However, a qualitative diVerence emerges when the reanalysis towards a dative-initial order is examined and compared to the reanalysis towards an accusative-initial order in an otherwise identical sentence (i.e. completing sentence fragments as in (7.4) with either an accusative or a dative verb, e.g. besuchen (‘to visit’) versus danken (‘to thank’)). While the reanalysis in (a)
Accusative
(b)
Dative CP3
CP3 N400
0.600 ..0.900 s −4.0
mV
0.600 ..0.600 s
P600
+4.0
−4.0
−5 mV
5
1.0
+4.0
OBJECT−SUBJECT
s 0.5
mV
CP3
SUBJECT−OBJECT
Figure 7.2. Grand average ERPs for object- and subject-initial structures at the position of the disambiguating clause-final verb (onset at the vertical bar) for sentences with accusative (A) and dative verbs (B). Negativity is plotted upwards. The data are from Bornkessel et al. (2004a)
Quantitative versus Qualitative Distinctions
131
accusative structures gives rise to a P600 eVect as discussed above, the revision towards a dative-initial word order elicits a centro-parietal negativity between approximately 300 and 500 ms post onset of the disambiguating element (N400; Bornkessel et al. 2004a; Schlesewsky and Bornkessel 2004). The diVerence between the two eVects is shown in Figure 7.2. In accordance with standard views on the interpretation of ERP component diVerences (e.g. Coles and Rugg 1995), we may conclude from this distinction that reanalysis towards an object-initial order engages qualitatively diVerent processing mechanisms in accusative and dative structures. The processes in question, which may be thought to encompass both conXict detection and conXict resolution, therefore reXect underlyingly diVerent ways of resolving a superWcially similar problem (i.e. the correction of an initially preferred subject-initial analysis). Before turning to the question of whether the ERP diVerence between reanalyses towards accusative and dative-initial orders may be seen as a correlate of the diVerent acceptability patterns for the two types of object cases—and thereby a source of gradience in this respect—we shall Wrst examine a second exception to the generalizations in (7.3), namely the behaviour of dative object-experiencer verbs (e.g. gefallen, ‘to be appealing to’). As brieXy discussed above, this verb class is characterized by an ‘inverse linking’ between the case/grammatical function hierarchy and the thematic hierarchy: the thematically higher-ranking experiencer bears dative case, while the lower-ranking stimulus is marked with nominative case. In the theoretical syntactic literature, it has often been assumed that these verbs are associated with a dative-before-nominative base order, which comes about when the lexical argument hierarchy (experiencer > stimulus) is mapped onto an asymmetric syntactic structure (e.g. Bierwisch 1988; Wunderlich 1997, 2003; Haider 1993; Haider and Rosengren 2003; Fanselow 2000). These properties of the object-experiencer class lead to an interesting constellation for argument order reanalysis, which may again be illustrated using the sentence fragment in (7.4). As with the cases discussed above, the comprehension system initially assigns a subject-initial analysis to the input fragment in (7.4). Again, when this fragment is completed by a dative object-experiencer verb that does not agree with the Wrst argument, reanalysis towards an object-initial order must be initiated. However, in contrast to the structures previously discussed, here the verb provides lexical information in support of the target structure, speciWcally an argument hierarchy in which the dative outranks the nominative. The ERP diVerences between reanalyses initiated by dative object-experiencer verbs and dative active verbs (which project a ‘canonical’ argument hierarchy) are shown in Figure 7.3 (Bornkessel et al. 2004a).
The Nature of Gradience
132
(b) Object-experiencer verbs
(a) Active verbs
N400
0.350 .. 0.550 s
−3.0
mV
N400
P4
P4
0.350 .. 0.550 s −3.0
+3.0
mV
−5 mV
SUBJECT−OBJECT
s 5
0.5
1.0
+3.0
OBJECT−SUBJECT P4
Figure 7.3. Grand average ERPs for object- and subject-initial structures at the position of the disambiguating clause-final verb (onset at the vertical bar) for sentences with dative active (A) and dative object-experiencer verbs (B). Negativity is plotted upwards. The data are from Bornkessel et al. (2004a)
As Figure 7.3 shows, reanalyses initiated by a dative object-experiencer verb also engender an N400 component, rather than a P600. However, the N400 eVect is less pronounced than for the analogous structures with dative active verbs. The diVerence between the two types of dative constructions is therefore quantitative rather than qualitative in nature. This suggests that reanalysis is more eVortful in the case of dative active verbs, but that the same underlying processes may be assumed to take place with both verb classes. To a large extent, the ERP patterns mirror the acceptability judgements described above. On the one hand, there is a general diVerence between dative- and accusative-initial sentences: the former are not only more acceptable than the latter, they also engage qualitatively diVerent processing mechanisms in reanalysis. Secondly, there is also a diVerence within the dative verbs themselves such that reanalysis towards a dative-initial order is less costly when it is triggered by an object-experiencer rather than an active verb. Nonetheless, reanalyses with both dative verb types appear to proceed in a qualitatively similar manner. However, despite this strong convergence of measures, it is unrealistic to expect a one-to-one mapping between the ERP data and the acceptability ratings. Indeed, not all the diVerences found in ERP measures are expressed in overt judgements. For example, the disadvantage for the object-initial
Quantitative versus Qualitative Distinctions
133
structures is measurable in ERPs even with dative object-experiencer verbs, while the diVerence between the two word orders is no longer visible in the acceptability rates. In order to precisely predict the relationship between the two types of measures we must fully understand how an overt judgement ‘emerges’ from the comprehension process. This requires tracing the development of the judgement from the point at which the problem is detected to later points when the system has settled on a Wnal assessment of the acceptability of the structure. The speed–accuracy trade-oV procedure (SAT; see Appendix 2) is one experimental method that allows for an examination of how a linguistic judgement develops over time. This method traces the emergence of an acceptability judgement from its beginnings (i.e. from the point at which the judgement departs from chance-level) up to a terminal point (i.e. a point at which the judgement no longer changes even with functionally unlimited processing time). Under the assumption that ERPs characterize the processing conXict and its resolution, while time-insensitive linguistic judgements reXect the endpoint of a multidimensional set of processing mechanisms, SAT procedures provide a bridge between the two measures. Let us Wrst consider the SAT results for reanalysis towards a dative-initial order in sentences with dative active and dative object-experiencer verbs. The SAT functions for the four critical conditions are shown in Figure 7.4 (Bornkessel et al. 2004a). The SAT data shown in Figure 7.4 were best Wt with an exponential approach to a limit function (Eq. 1), that assigned a distinct asymptote (l)
Subject-Object, Active Object-Subject, Active Subject-Object, Object-Experiencer Object-Subject, Object-Experiencer
Accuracy (d⬘)
4 3 2 1 0 0
1 2 3 4 5 Processing time (lag plus latency) in seconds
6
Figure 7.4. SAT functions for object- and subject-initial structures with (dative) active and (dative) object-experiencer verbs The data are from Bornkessel et al. (2004a)
134
The Nature of Gradience
to each of the four conditions and distinct intercepts (d) to the subject-initial and object-initial conditions, respectively. (Eq. 1) d’ (t) ¼ l (1–e–b(t–d)) for t > d, otherwise 0 The intercept diVerence between subject-initial and object-initial structures, with a longer intercept for object-initial structures, indicates that the Wnal analysis of the object-initial sentences takes longer to compute than the Wnal analysis of their subject-initial counterparts. This is the characteristic pattern predicted for a reanalysis operation: as reanalysis requires additional computational operations, the correct analysis of a structure requiring reanalysis should be reached more slowly than the correct analysis of an analogous structure not requiring reanalysis. The dynamics (intercept) diVerence occurs in exactly the same conditions as the N400 eVect in the ERP experiment. The asymptotic diVerences appear to result from two sources. First, the object-initial structures are generally associated with lower asymptotes than the subject-initial controls. This diVerence likely reXects a decrease in acceptability resulting from the reanalysis operation required to interpret the object-initial sentences. A principled explanation for this pattern, one that is consistent with the concomitant dynamics diVerences, is that, on a certain proportion of trials, the processing system fails to recover from the initial misanalysis, thus engendering lower asymptotic performance for an initially ambiguous object-initial structure as compared to a subject-initial structure. More interesting, perhaps, are the diVerences in asymptote between the two object-initial conditions: here, the sentences with object-experiencer verbs were associated with a reliably higher asymptote than those with active verbs. This diVerence may directly reXect the diVerences in the accessibility of the object-initial structure required for a successful reanalysis. Whereas the active verbs provide no speciWc lexical information in favour of such a structure, the object-experiencer verbs are lexically associated with an argument hierarchy calling for precisely this type of ordering. Thus, while a garden path results for both verb types, the object-experiencer verbs provide a lexical cue that aids the conXict resolution. Again, the correspondence to the ERP data is clear: the higher asymptote for the object-initial structures with object-experiencer verbs—which we have interpreted as arising from the higher accessibility of the object-initial structure in these cases—corresponds to the reduced N400 for this condition, which also reXects a reduction of the reanalysis cost. Two conclusions concerning acceptability patterns follow from this analysis. First, despite the presence of an almost identical processing conXict in both cases, dative-initial structures with object-experiencer verbs are more
Quantitative versus Qualitative Distinctions
135
acceptable than those with active verbs because only the former are lexically associated with an object-initial word order. Secondly, however, even the presence of an object-experiencer verb can never fully compensate the cost of reanalysis, as evidenced by the fact that an initially ambiguous dative-initial structure never outstrips its nominative-initial counterpart in terms of acceptability. From a surface perspective, therefore, the observed acceptability patterns are the result of a complex interaction between diVerent factors. The observed gradience does not result from uncertainty in the judgements, but rather from interactions between the diVerent operations that lead to the Wnal intuition concerning acceptability. Having traced the emergence of the acceptability judgements for the two types of dative structures, a natural next step appears to be to apply the same logic to the diVerence between accusative and dative structures and to thus examine whether similar types of parallels between the on-line and oV-line Wndings are evident in these cases. Recall that, while the reanalysis of a dative structure generally engenders an N400 eVect in ERP terms, reanalysis towards an accusative-initial order has been shown to reliably elicit a P600 eVect. In addition, the surface acceptability is much lower for the accusative than for the dative-initial structures. How might these Wndings be related? Essentially, the diVerent ERP patterns suggest that the two types of reanalysis take place not only in a qualitatively diVerent manner, but also in diVerent phases of processing: while the N400 is observable between approximately 300 and 500 ms after the onset of a critical word, the time range of the P600 eVect is between approximately 600 and 900 ms. Thus, the reanalysis of an accusative structure appears to be a later process than the reanalysis of a dative structure and, in terms of the SAT methodology, we might therefore expect to observe larger dynamics diVerences between subject- and object-initial accusative sentences than in the analogous contrast for dative sentences. As discussed above, dynamics diVerences can subsequently lead to diVerences in terminal (asymptotic) acceptability and the distinction between the dative and the accusative structures might therefore also—at least in part—stem from a dynamic source. The diVerence between subject- and object-initial dative and accusative structures as shown in an SAT paradigm is shown in Figures 7.5.a and 7.5.b (Bornkessel et al. submitted). As the accusative and dative sentences were presented in a between-subjects design, model Wtting was carried out separately for the two sentence types. While the accusative structures were best Wt by a 2l–2b–2d model (adjusted R2 ¼ .994), the best Wt for the dative structures was 1l–1b–2d (adjusted R2 ¼ .990). Estimates of the composite dynamics (intercept + rate) were
136
The Nature of Gradience
Accuracy in d⬘ units
(a) 5
Accusatives
4
S-INITIAL O-INITIAL
3 2 1 0 0
Accuracy in d⬘ units
(b) 5
1
2 3 4 5 Processing time in seconds
6
7
6
7
Datives
4
S-INITIAL O-INITIAL
3 2 1 0 0
1
2 3 4 5 Processing time in seconds
Figure 7.5. SAT functions for object- and subject-initial structures with accusative (a) and dative verbs (b) The data are from Bornkessel et al. (submitted)
computed for each condition using the formula 1/d + b, which provides a measure of the mean time required for the SAT function to reach the asymptote. The dynamics diVerence between the two accusative structures was estimated to be 588 ms (4430 ms for object- versus 3842 ms for subjectinitial sentences), while the diVerence for the dative sentences was estimated at 332 ms (3623 ms versus 3291 ms). Thus, while both dative- and accusativeinitial sentences show slower dynamics than their subject-initial counterparts, the dynamics diVerence between the subject- and object-initial accusative structures is approximately 250 ms larger than that between the dative structures. This Wnding therefore supports the hypothesis that the large
Quantitative versus Qualitative Distinctions
137
acceptability disadvantage for the initially ambiguous accusative-initial structures—and the corresponding asymptote diVerences for the accusative sentences—results to a large extent from the highly pronounced dynamics diVerence between these structures and the corresponding subject-initial sentences. In other words, the likelihood that the correct analysis fails to be computed is much higher in the accusative-initial sentences because the computational operations required to obtain this analysis are much more complex for this sentence type. Consequently, accusative-initial sentences are rejected as unacceptable in a higher proportion of trials, thereby yielding a lower acceptability rate/asymptote. Again, SAT provides a principled means of establishing the correspondence between the ERP data and the acceptability ratings. The reanalysis mechanisms that are reXected in an N400 eVect—those that enable reanalysis towards a dative-initial structure—are also associated with a smaller dynamics increase than those reXected in a P600 component—those that enable reanalysis towards an accusative-initial structure. Therefore, we might speculate that the diVerence in the underlying neural processes, which is reXected in the diVerent ERP components, gives rise to the concomitant diVerences in SAT dynamics and, thereby, to the diVerences in surface acceptability. If both of the SAT studies discussed here are considered together, an interesting diVerence between the two becomes apparent. In the Wrst experiment, in which only dative structures were examined in a design identical to that used to obtain the acceptability rates in Table 7.1, there were reliable asymptote diVerences between dative- and nominative-initial structures with dative active verbs. In the second study, there were comparable dynamics diVerences, but the asymptote diVerence—although apparent in visual inspection—failed to signiWcantly improve the model Wt. How can we account for this variation or inter-experimental gradience? Assuming that the asymptotic acceptability measured using SAT reXects the endpoint of processing and, thereby, the time-independent acceptability of a given structure, one plausible explanation appears to lie in the diVerent experimental environments in which the structures were presented. It is well-known that sentences judgements are inXuenced by various factors including context, Wller sentence type, etc. (Bard et al. 1996; Schu¨tze 1996). One crucial diVerence between the two SAT studies is that dative object-experiencer verbs were only included in the Wrst experiment. It may therefore be the case that the acceptability ratings for the object-initial dative active sentences arise not only from a contrast with the corresponding subject-initial sentences but also with the object-initial sentences with
138
The Nature of Gradience
experiencer verbs. In the face of more acceptable object-initial structures, the acceptability disadvantage for the object-initial active sentences may be ‘overestimated’. If true, this observation suggests that terminal acceptability may result from the interaction of a variety of diVerent factors. Whatever the source of this discrepancy, it serves to highlight the highly variable nature of acceptability judgements, and to contrast these measures with those that more directly reXect intrinsic properties of the underlying processing mechanisms, such as the dynamics diVerence between object- and subject-initial structures, which may be assumed to be more stable and less subject to environmental inXuences.
7.4 Final remarks In this chapter, we have attempted to show how linguistic judgements arise from diVerent facets of language comprehension. In particular, our data suggest the following caveats concerning the interpretation of acceptability judgements: 1. One-dimensional judgements often result from the interaction of a variety of factors, and hence are inherently multi-dimensional. 2. SuperWcially similar judgements may stem from qualitatively diVerent sources, which should be disentangled. DiVerences that may appear quantitative, for example as diVerent ‘strengths’ on a single dimension may nevertheless have qualitatively diVerent origins. 3. Acceptability decreases may be dynamic or non-dynamic in nature. Concerning gradience in linguistic judgements, these Wndings indicate that a considerable amount of variation in judgements may be accounted for by carefully considering factors that interact to produce the end state that constitutes an acceptability judgement. The question thus arises whether gradience should indeed be attributed to linguistic competence, or whether it is better described as a product of the language in use, that is of processing mechanisms—which may or may not be language speciWc—and of general cognitive factors (e.g. working memory, see footnote 1). From our perspective, the burden of the evidence rests with the advocates of gradient grammaticality, for it appears very diYcult to mount a convincing argument in favour of grammar-internal gradience on the basis of acceptability judgements alone. Thus, when all possible alternative sources of surface gradience are considered, a categorical grammar still appears to be the simplest and therefore most appealing means of accounting for the data.
Quantitative versus Qualitative Distinctions
139
7.5 Appendix 1. Event-related brain potentials (ERPs) Event-related brain potentials (ERPs) are small changes in the spontaneous electrical activity of the brain, which occur in response to sensory or cognitive stimuli and which may be measured non-invasively by means of electrodes applied to the scalp. The high temporal resolution of ERP measures is of particular importance for the examination of language comprehension. Furthermore, ERP patterns (‘components’) are characterizable in terms of the following parameters: polarity (negative versus positive); topography (at which electrode sites an eVect is visible); latency (the time at which the eVect is visible relative to the onset of a critical stimulus); and amplitude (the ‘strength’ of an eVect). While a number of language-related ERP components have been identiWed (cf., for example, Friederici 2002), we will not introduce these here for the sake of brevity. For a more detailed description of the ERP methodology and how it has been applied to psycholinguistic domains of investigation, the reader is referred to the overviews presented in Coles and Rugg (1995), Garnsey (1993), and Kutas and Van Petten (1994). Ongoing EEG Amplifier S
S
S
S
one sec Auditory event-related potential −6 mV N400 Signal averager
ELAN
Auditory stimulus (S)
P600 P200 ms
+6 mV 200 Stimulus onset
400
600 800 Time
Figure 7.6. Schematic illustration of the ERP methodology
1000
140
The Nature of Gradience
The ERP methodology only provides relative measures, that is an eVect always results from the comparison of a critical condition with a minimally diVering control condition. For example, at the position of socks in He spread the warm bread with socks in comparison to the position of butter in He spread the warm bread with butter, a negativity with a centro-parietal distribution and a maximum at 400 ms post critical word onset (N400) is observable (Kutas and Hillyard 1980). Thus, in the experiments presented here, we always compare the response to a critical condition with that to a control condition at a particular (critical) position in the sentence. A schematic illustration of the ERP methodology is shown in Figure 7.6.
7.6 Appendix 2. Speed–accuracy trade-oV (SAT) Reading time (eye-movement tracking or self-paced) procedures are often used as a natural and unintrusive measure of processing time. However, these measures do not provide an estimate of the likelihood that readers have successfully processed a sentence and, conversely, do not provide a direct estimate of the time it takes to compute an interpretation. A reading time diVerence can reXect the time needed to compute a particular interpretation, but it can also reXect the likelihood that readers can compute that interpretation or how plausible readers Wnds the resulting interpretation (McElree 1993, 2000; McElree and GriYth 1995, 1998; McElree and Nordlie 1999; McElree et al. 2003). A standard solution to this problem is to derive a full time–course function that measures how the accuracy of processing varies with processing time (Wickelgren 1977). The response-signal, speed–accuracy trade-oV (SAT) procedure provides the required conjoint measures of processing speed and accuracy. The response-signal speed–accuracy trade-oV task requires subjects to make their judgement of acceptability at particular times. This serves to chart the full time–course of processing, measuring when discrimination departs from a chance level, the rate at which discrimination grows as a function of processing, and the asymptotic level of discrimination accuracy reached with (functionally) unlimited processing time. Figure 7.7 presents illustrative SAT functions derived from this procedure. The accuracy of discriminating acceptable from unacceptable sentences is measured in d’ units (the z-transform of the probability of correctly accepting an acceptable sentence minus z-transform of the probability of falsely accepting an unacceptable sentence). Typical SAT functions display three distinct phases: a period of chance performance (d’ ¼ 0), followed by a period of increasing accuracy, followed by an asymptotic period where further processing does not
Quantitative versus Qualitative Distinctions
141
improve performance. In a sentence acceptability task, the SAT asymptote provides a measure of the probability (across trials and materials) that readers arrive at an interpretation suYcient to support an ‘acceptable’ response. If two conditions diVer in asymptote, as illustrated in Panel (a), it indicates that they diVer in the likelihood that a meaningful interpretation can be computed or in overall acceptability/plausibility of the respective interpretation. The point at which accuracy departs from the chance level (the intercept of the function) and the rate at which accuracy grows over processing time are joint measures of the underlying speed of processing. If one type of structure
(a) Probability of computing of acceptable interpretation
2.0 1.5 Proportional dynamics Functions reach a given proportion of their asymptote at the same time.
1.0
Accuracy (d⬘ units)
0.5 Condition A Condition B
0.0 0.0 (b)
1.0
2.0
3.0
4.0
Speed of computing an acceptable interpretation
2.0 1.5 Disproportional dynamics Functions reach a given proportion of their asymptote at different times.
1.0 0.5 0.0 0.0
1.0
2.0
3.0
Processing time (response time) in seconds
Figure 7.7. Illustrative SAT functions
4.0
142
The Nature of Gradience
can be interpreted more quickly than another, the SAT functions will diVer in rate, intercept, or some combination of the two parameters. This follows from the fact that the SAT rate and intercept are determined by the underlying Wnishing time distribution for the processes that are necessary to accomplish the task. The time to compute an interpretation will vary across trials and materials, yielding a distribution of Wnishing times. Intuitively, the SAT intercept corresponds to the minimum of the Wnishing time distribution, and the SAT rate is determined by the variance of the distribution. Panel (b) depicts a case where the functions diVer in rate of approach to asymptote, leading to disproportional dynamics; the functions reach a given proportion of their asymptote at diVerent times. Dynamics (rate and/or intercept) diVerences are independent of potential asymptotic variation. Readers may be less likely to compute an interpretation for one structure or may Wnd that interpretation less acceptable (e.g. less plausible) than another; however, they may not require additional time to compute that interpretation (McElree 1993, 2000; McElree et al. 2003; McElree and Nordlie 1999).
Part II Gradience in Phonology
This page intentionally left blank
8 Gradient Perception of Intonation C A RO L I N E F E´ RY A N D RU B E N S TO E L
8.1 Introduction Many phonologists associate the term ‘gradience’ with the distinction between phonology—which is supposed to be categorical—and phonetics— which is supposed to be gradient (see Cohn, this volume, for a review of the issues associated with this distinction).1 In recent years, a diVerent role for gradience in phonology has emerged: the well-formedness of phonological structures has been found to be highly gradient in a way that correlates with their frequency. In their chapter, Frisch and Stearns (this volume) show that phonotactic patterns, like consonant clusters and other segment sequences, as well as morphophonology, word-likeness, etc. are gradient in this way. The examination of large corpora is a reliable indicator of relative frequency. Crucially, the less frequent sequences are felt by speakers to be less prototypical exemplars of their category. In grammaticality judgement tasks, wordlikeness tasks, assessment of novel words, etc., less frequent items are likely to get lower grades than more frequent ones. In short, speakers reproduce in their judgements the pattern of relative frequency that they encounter in their linguistic environment. In light of this well-documented (see Frisch and Stearns, this volume and references cited there), but controversial result, the question has arisen for some phonologists as to the need of a grammar operating with abstract phonological categories, like features and phonemes. In their opinion, if phonotactic distribution is learnable by executing probabilistic generalizations over corpora, the only knowledge we need in order to
1 A pilot experiment for this paper was presented at the Potsdam Gradience Conference in October 2002 and some of the results discussed here were presented at the Syntax and Beyond Workshop in Leipzig in August 2003. Thanks are due to two anonymous reviewers, as well as to Gisbert Fanselow and Ede Zimmermann for helpful comments. Thanks are also due to Frank Ku¨gler for speaking the experimental sentences, and to Daniela Berger, Laura Herbst, and Anja Mietz for technical support. Nobody except for the authors can be held responsible for shortcomings.
146
Gradience in Phonology
elaborate ‘grammars’ may turn out to be a stochastic one. But before we can take a stand on this important issue in a competent way, we need to be wellinformed on other aspects of the phonology as well. In this chapter, we take a Wrst step and investigate the question of whether intonational contours are gradient in the same way that segment sequences are. Is it the case that more frequent tonal patterns are more acceptable than less frequent ones? We use the term gradience in the sense of gradient acceptability. Unfortunately, for a number of reasons, large corpora—at least in their present state—are useless for the study of tonal pattern frequencies. One of the reasons relates to the analysis and annotation of tonal patterns. Scholars not only disagree on the kinds of categories entering intonation studies but also on the annotation for ‘rising contour’, ‘fall–rise’, etc. Melodies—like Gussenhoven’s (1984, 2004) nuclear contours or the British school’s ‘heads’ and ‘nuclei’—may well exist as independent linguistic elements, but they are not transcribed uniformly. Even though autosegmental-metrical representations of tonal contours, like ToBI (Beckman and Ayers-Elam 1993; Jun 2005) are evolving to become a standard in intonation studies, they are not suYciently represented in corpora. Most large corpora consist of written material anyway, and those which contain spoken material generally only display segmental transcription rather than tonal. In short, the development of corpora which are annotated in a conventional way for intonation patterns is an aim for the future, but as of now, it is simply not available for German. As a result, we must rely on the intuition of speakers. The questions we address in this chapter are: Which tonal contours are accepted most? Which are less accepted? We will see that the question must be made precise in the following way: given a certain syntactic structure, is there a contour which is accepted in the largest set of contexts? And this is related to the question of pitch accent location. Which constituents are expected to be accented? Which accent structure is the least marked, in the sense of being accepted in the greatest number of contexts? Are some accent patterns (tonal patterns) ‘unmarked’ (more frequent, acquired earlier, but also accepted more easily) in the same sense as consonant clusters or other segment sequences are? Below, we present the results of a perception study bearing on tonal contours. But before we turn to the experiment, we Wrst sum up some relevant issues in the research on prosody and situate our research in this broader context.
Gradient Perception of Intonation
147
8.2 Prosody and intonation Prosody plays a crucial role in communication. To begin with, we partition our utterances in prosodic chunks, like phonological phrases and intonation phrases, which correspond to syntactic constituents (Nespor and Vogel 1986; Truckenbrodt 1999) or information structural blocks (Vallduvı´ 1992). These phrases, which help both speakers and hearers structure the discourse, are signalled phonetically by boundary tones, segmental lengthening, or some other phonological cues. A second factor playing a role in phonological patterning is the distribution and form of pitch accents, associated with prominent syllables. A syllable may be prominent if it is the bearer of the lexical stress of a word or of a larger constituent which is itself prominent. A speaker may decide to speak about some object in her surroundings or an object she knows about, and decide to focus on one property of this object. Or she may answer a question asked by a protagonist because she feels she has to deliver some bit of information. In other words, prominence may be assigned to some linguistic constituents because of contextual or cognitive reasons (Bolinger 1972). The other reason to assign a pitch accent to a syllable is purely grammatical. An internal argument of a German predicate + argument complex, for example, may receive a pitch accent, and the verb may be unaccented. Still, the whole phrase may be prominent (see Bierwisch 1968; Schmerling 1976; Gussenhoven 1983, 1992; von Stechow and Uhmann 1986; Cinque 1993; Fe´ry and Samek-Lodovici 2006, among others). In Standard German, nuclear accents (the Wnal or most prominent accents of an intonation phrase) are either bitonally falling, HL, or rising, LH, whereas prenuclear accents can be rising or falling as well or monotonally high (H) or low (L) (see Fe´ry 1993; Grabe 1998; Grice et al. 2003; Peters 2005 for phonological studies of intonation of standard German). Prosodic phrases may be terminated with a boundary tone, which is written with a subscripted P for a phonological phrase, and a subscripted I for an intonation phrase (following Hayes and Lahiri’s 1991 notation). For the sake of illustration, two pitch tracks of a sentence used in the experiments described below are shown with their phonological tone structure. The Wrst pitch track, in Figure 8.1, is equivalent to a wide-focused realization with two pitch accents, a rising one on the subject Ruderer, and a falling one on the object Boote. The verb, adverb, and particle mit are unstressed. This realization may be dubbed ‘unmarked prosodic structure’ (UPS, see below). It is expected to be the most frequent one, and thus, the most widely accepted pattern for such a declarative sentence. In German, a topic-focus realization, in which the subject is
Gradience in Phonology
148 250
200
Pitch (Hz)
150
100
50 RUDERER
L*H
bringen immer
Hp
BOOTE
H*L
mit
Li 1.67134
0 Time (s)
Figure 8.1. Pitch track of Ruderer bringen immer Boote mit. ‘Oarsmer always bring boats.’
topicalized and the remainder of the sentence is focused, is identical to a wide-focused realization. The second pitch track (Figure 8.2) shows a marked pattern, with just one pitch accent located early in the sentence. This kind of pattern is expected to be conWned to special contexts, in particular those eliciting a narrow focus on the subject. It is not possible to investigate the gradience of tonal patterns out of context. Tonal patterns do not exist as pure melodies: they need to be interpreted as linguistic units, thus as pitch accents or as boundary tones. This can only happen when tonal excursions are associated with text. Moreover, tonal contours are more or less marked only when they are associated with speciWc locations in a sentence, since accent locations are dependent on syntax and information structure. We introduce ‘focus projection’ brieXy in the next section, but have no space to develop all arguments for this phenomenon (see Selkirk 1995; Schwarzschild 1999; Fe´ry and Samek-Lodovici 2006 among others). We propose the concept of ‘Unmarked Prosodic Structure’ (UPS, Fe´ry 2005) as the intonation used when the sentence is realized in a whole-focused environment. It refers to the phrasing and the tonal contour projected when the speakers have no clue about the context. Unmarked Prosodic Structure relies solely on the syntactic structure. A tonal contour
Gradient Perception of Intonation
149
250
200
Pitch (Hz)
150
100
50 RUDERER
bringen immer
H*L
Boote mit
Li 1.73211
0 Time (s)
Figure 8.2. Pitch track of Ruderer bringen immer Boote mit
compatible with unmarked prosody is expected to be acceptable in more environments than other, more marked contours.
8.3 Previous studies on gradient tone perception Few studies, if any, have explicitly addressed the gradience of intonational contours, so we cannot base our work on a rich empirical basis. There are, however, quite a number of studies investigating the question of categories in intonational morphemes, which have found more or less gradient accents or boundaries.2 The most relevant studies for our aim have looked at the adequacy of pitch accent patterns in some speciWc contexts. The issue of the location of pitch accents and their role for the focus structure has been investigated for English by Gussenhoven (1983) and Birch and Clifton (1995), among others, who examine the role of prenuclear accents on the verb in a VP consisting of a verb plus an argument (or an adjunct by Gussenhoven) in English. Gussenhoven’s (1983) sentence accent assignment rules (SAAR) predict that in a focused predicate argument complex, only the argument needs to be stressed, but that a prenuclear accent can be added 2 Some have found categories in the domain of pitch accent realization; for example Pierrehumbert and Steele (1989) or Ladd and Morton (1987).
150
Gradience in Phonology
freely on a verb without impairing processing. In a verbal phrase, by contrast, both the verb and the adjunct need to be stressed. Gussenhoven himself Wnds conWrmation of this prediction in experimental work. In mini-dialogues such as (8.1), there is a diVerence between the focus structure of the sentences answering (8.1a) and (8.1b). In (8.1a), the whole VP share a Xat is focused, whereas in (8.1b) only the direct object is focused, the diVerence being elicited by the preceding question. The same kind of contrast is obtained in the dialogues in (8.2) which contain a verb followed by an adjunct. (8.1) Verb and argument a. C: Do you live by yourself? b. C: I hate sharing things, don’t you? c. U: I share a Xat. (the whole VP or the argument NP is focused) (8.2) Verb and adjunct a. C: Where will you be in January? b. C: Where will you be skiing? c. U: We will be skiing in Scotland. (the whole VP or the adjunct PP is focused) Gussenhoven cross-spliced questions and answers, spoken by native speakers, so as to obtain both answers in both contexts. Subjects then had the task of deciding which of the two answers was the more appropriate response to the preceding question. Gussenhoven found that the presence of an accent on the verb in addition to the expected accent on the object in (8.1) does not change the acceptability of the pitch accent structure, and that this held in both narrow and broad focused contexts. The speakers did not do better than by chance when required to choose between the two contexts on the basis of such an accent pattern. But in (8.2), the absence of a stress on the verb in (8.2a) was an indicator that the verb had to be given (and thus not focused), so that the speakers did better than in the predicate-argument condition in the same task. The reliability of the accent on the verb in deciding for the wide-focus context depended gradiently on the number of unstressed syllables intervening between the two accents. Birch and Clifton (1995) conducted similar experiments, but obtained slightly diVerent results. They also prepared matched and mismatched pairs of questions and answers. An example of a dialogue set is reproduced in (8.3). Only the pairs QA/R1 and QB/R3 match perfectly, all others are predicted to be more or less deviant along the same lines as those just explained, although the authors acknowledge that QA/R2 could be as good as QA/R1 if SAAR make the right predictions.
Gradient Perception of Intonation
151
(8.3) a.
Questions QA: Isn’t Kerry pretty smart? QB: Isn’t Kerry good at math? b. Responses R1: Yes, she teaches math. R2: Yes, she teaches math. R3: Yes, she teaches math.
In judgement and decision tasks, Birch and Clifton found that as an answer to question QA, speakers prefer R1, with two accents, over R2, with just one accent on the argument NP. The diVerence was small but signiWcant. And unsurprisingly, R3 was by far the preferred answer to QB. All other pairs obtained poorer scores. In a second experiment, speakers had to decide how well the pairs made sense. In this case, the results for QA were similar to those of Gussenhoven: there was no diVerence between a sentence with two accents (R1) and a sentence with just one accent on the argument (R2).3 These results, as well as other perception experiments bearing on the location of pitch accents conducted for Dutch (Nooteboom and Kruyt 1987; Krahmer and Swerts 2001) and for German (Hruska et al. 2001) show that, for these three languages at least, a prenuclear accent is readily acceptable, but that a postnuclear one is less easily accepted and that accents on narrowly focused items in an otherwise non-nuclear position are more readily perceived than accents on words accented per default in their unmarked accent pattern. Nooteboom and Kruyt (1987) rightly explain the acceptability of a prenuclear accent in terms of topicalizing or thematicizing the bearer of such an accent, and observe that a sentence with a supplementary prenuclear accent can get an interpretation in which the prenuclear accent is information structurally prominent. In psycholinguistic experiments studying the role of prosody in disambiguating syntactic structures (see for instance Lehiste 1973; Kjelgaard and Speer 1999; Schafer et al. 2000), garden path sentences or sentences with an ambiguous late or early closure/attachment have been tested. These experiments deliver gradient results correlating with the strength and the location of boundaries. Comparing the two realizations of the sentences in (8.4), there is no doubt that intonation can disambiguate the readings. Example (8.4a) is realized as one Intonation Phrase, but in (8.4b), an Intonation Phrase boundary is located after heiratet, which is then understood as an intransitive verb. 3 Birch and Clifton’s results also indicate that a single accent on the verb is readily accepted in a context eliciting broad focus (78 percent of yes). The only situation where speakers accepted a pair less (with 54 per cent and not between 71 and 84 percent as in the other pairs) was when the context was eliciting a narrow focus on the verb, and the answer had a single accent on the argument (QB/R2).
152
Gradience in Phonology
Much more subtle is the question of whether prosody can help with the sentence in (8.5). In one reading, it is the woman who lives in Georgia, and in the other reading, her daughter. The phrasing, in the form of a Phonological Phrase boundary, is roughly the same in both readings. Nevertheless, it is possible to vary the quantity and the excursion of the boundary tone in such a way that the preference for one or the other reading is favoured. L*H L*H H*L LI (8.4) a. [Maria heiratet Martin nicht]I ‘Mary does not marry Martin’ H*L LI L*H H*L LI L*H b. [Maria heiratet]I [Martin nicht]I ‘Mary gets married. Martin does not.’ (8.5) [ [Ich treVe mich heute]P [mit der Tochter der Frau]P]I [ [die in Georgien lebt]P]I ‘I am meeting today with the daughter of the woman who lives in Georgia.’ We are only marginally interested in syntactic disambiguation in this chapter. Rather, our experiment aimed at testing the gradience of German intonational structures. This experiment diVers from the ones conducted by Gussenhoven and by Clifton and Birch in a crucial way: several parameters were systematically varied: sentence type, context, and tonal contours. We were explicitly interested in Wnding out whether some kinds of intonation patterns are more acceptable than others and whether gradience can be observed in the domain of tonal contours.
8.4 Experiment 8.4.1 Background The experiment reported in this section was intended to elucidate the question formulated above: How gradient are tonal contours? We wanted to understand what triggers broad acceptance for intonational patterns. To this aim, we used three diVerent kinds of sentences, which were inserted in diVerent discourse contexts, and cross-spliced. If an eVect was to be found, we expected it to be of the following kind: the unmarked tonal contours should be generally better tolerated than the marked ones. The hypothesis can be formulated as in (8.6).
Gradient Perception of Intonation
153
(8.6) Unmarked Prosodic Structure (UPS) Hypothesis An unmarked prosodic structure, i.e. a prosodic structure adequate in a broad focus environment, is readily accepted. It can be inserted successfully in more environments than a marked prosodic structure, which is appropriate in a restricted number of contexts only. The topic-focus contour that we used in our experiment has the same contour as a broad focus one. Both have a rising pitch accent on the subject, and a falling accent on the focused word (the ‘focus exponent’). We chose a topicfocus environment instead of a broad focus one because of the slightly clearer accent pattern produced with a topic and a focus. Even though we did not include a broad focus context in our experiment, we are conWdent that the pattern we call TF would get high scores in it. 8.4.2 Material Three diVerent kinds of sentences served as our experimental material: six short sentences, six long sentences, and three sentences with ambiguous scope of negation and quantiWer. Every sentence was inserted in three or four matching contexts (see below). In (8.7) to (8.9), an example for each sentence is given along with their contexts. The remaining sentences are listed in the appendix. (8.7) Short sentences Maler bringen immer Bilder mit. Painters bring always pictures with a. Narrow focus on the subject (NFS): Tom hat mir erza¨hlt, dass Fotografen unserer Nachbarin immer Bilder mitbringen. Aber das stimmt nicht: ‘Tom told me that photographers always bring pictures to our neighbour. But this is not true:’ b. Narrow focus on the object (NFO): Angeblich bringen Maler unserer Nachbarin immer Bu¨cher mit. Aber das stimmt nicht: ‘It is said that painters always bring books to our neighbour. But this is not true:’ c. Topic-focus (TF): Meine Nachbarin schmeißt oft große Partys, dafu¨r bekommt sie aber auch viele Geschenke. Regisseure schenken ihr Filme, Schriftsteller Bu¨cher und . . . ‘My neighbour often throws big parties, and therefore she also gets lots of presents. Movie directors give her movies, writers give her books and . . .’
154
Gradience in Phonology
(8.8) Long sentences Passagiere nach Rom nehmen meistens den spa¨ten Flug.4 Passengers to Rome take mostly the late Xight a. Narrow focus on the subject (NFS): Angeblich nehmen die Leute nach Athen meistens den spa¨ten Flug. Aber das stimmt nicht: ‘It is said that the people (Xying) to Athens mostly take the late Xight, but this is not true:’ b. Narrow focus on the object (NFO): Mona sagt, dass Passagiere nach Rom meistens die fru¨he Maschine nehmen. Aber das stimmt nicht: ‘Mona says that passengers to Rome mostly take the early Xight, but this is not true:’ c. Topic-focus (TF): Pendler, die ziemlich weit von zuhause arbeiten, haben oft a¨hnliche Angewohnheiten. Gescha¨ftsleute Richtung Paris fahren oft mit dem Auto, Reisende nach London nehmen den Zug aus Calais und . . . ‘Commuters who work far away from home often have similar habits. Business people who go to Paris often take their car, travellers to London take the train from Calais and . . .’ (8.9) Quantifier-negation sentences Beide Autos sind nicht bescha¨digt worden. Both cars were not damaged a. Two foci (‘two’): Es wa¨re schlimm gewesen, wenn Karl bei dem Unwetter seinen Jaguar und seinen Porsche auf einmal verloren ha¨tte, aber glu¨cklicherweise war es nicht so. ‘It would have been too bad if Charles had lost both his Jaguar and his Porsche because of the bad weather, but fortunately this was not the case.’ b. Narrow focus on the quantiWer (FQ): Ist nur Peters Auto nicht bescha¨digt worden? Nein, . . . ‘Has only Peter’s car not been damaged? No, . . .’ c. Narrow focus on the negation (FN): Ich habe gesehen, dass Deine beiden Autos seit Wochen in der Garage stehen. Sind sie bei dem Unfall bescha¨digt worden?—Nein, ich habe Dir doch schon gesagt: ‘I have seen that both your cars have been sitting in the garage for ages. Were they damaged in the accident?—No, I already told you, . . .’ 4 As Ede Zimmermann (p.c.) observes, it is not undisputed whether there is a structural ambiguity between the temporal and the quantiWcational reading of meistens. We suspect that, even if conWrmed, this ambiguity played no role in the experimental results.
Gradient Perception of Intonation
155
d. Topic-focus (TF): Bei dem Unfall ist verschiedenes passiert. Drei Fahrra¨der sind jetzt Schrott, ein Fußga¨nger ist im Krankenhaus, aber bei den Autos, die dabei involviert waren, war es nicht dramatisch: ‘Several things happened at the accident. Three bikes are now ruined, a pedestrian is at the hospital, but nothing dramatic happened to the cars involved:’ Contexts and stimuli sentences were spoken by a trained speaker and recorded in a sound-proof booth on a DAT recorder. The speaker was instructed to speak naturally, in a normal tempo. He read the context-target pairs at once, Wrst the context and then the stimulus sentences. There were 48 matching pairs for the three experiments altogether (six short sentences, six long sentences, and three quantiWer-negation sentences in their contexts, thus 18 + 18 + 12 pairs). All pitch accents of a speciWc type were realized similarly (see Figures 8.1 to 8.3 for illustrations), and controlled carefully with the help of the speech analysis program PRAAT. Several recording sessions were necessary. The sentences were evaluated by three independent trained phonologists as to their naturalness. Context sentences and stimulus sentences were digitized into individual sound Wles, ready to be cross-spliced. No manipulation whatever was undertaken in order to not endanger the naturalness. We prepared 36 non-matching pairs for the short sentences, 36 for the long sentences, and 32 for the scope sentences, a total of 104 non-matching pairs. The sentences to be evaluated thus consisted in 48 matching and 104 nonmatching pairs, an overall total of 152 pairs. 8.4.3 Subjects Four non-overlapping groups of Wfteen subjects (altogether sixty students at the University of Potsdam) took part in the experiment. They were native speakers of Standard German and had no known hearing or speech deWcit. All were paid or acquired credit points for their participation in the experiment. Two groups judged the sentences on a scale of 1 (very bad) to 8 (perfect), and two groups judged the same sentences in a categorical way: acceptable (yes) or non-acceptable (no). All sixty informants evaluated the scope sentences. In addition, the Wrst and third groups also judged the short sentences, while the second and fourth groups judged the long sentences, thus thirty matching sentences plus sixty-eight non-matching ones each. 8.4.4 Procedure The subjects were in a quiet room with a presentation using the DMDX experiment generator software developed by K. and J. Forster at the
156
Gradience in Phonology
University of Arizona. The experimenter left the subject alone in the room after brief initial instructions as to beginning and ending the session. The subjects worked through the DMDX presentation in a self-paced manner. It led them through a set of worded instructions, practice utterances, and Wnally the experiment itself, consisting of 102 target sentences. No Wllers were inserted, but three practice sentences started the experiment. This experiment was itself included in a set of experiments in which the subjects performed diVerent tasks: production of read material, and dialogues. The instructions made it clear that the aim of the experiment was to test the intonation and stress structure of the sentences, and not their meaning or syntax. The stimuli were presented auditorily only: pairs of context and stimulus sentences were presented sequentially. The subject heard Wrst a context, and after hitting the return key, the test sentence. The task consisted in judging the adequacy of the intonation of the sentence in the given context. Every recorded sentence of the groups of short and long sentences was presented nine times, in three diVerent intonational and stress patterns, and each of these patterns in three diVerent contexts. The scope sentences were presented sixteen times each, in all possible variants. The sentences were presented in a diVerent randomized order for each subject. The set-up and the instructions included the option of repeating the context–stimulus pair for a particular sentence. Most subjects made occasional use of this possibility. Only the last repetition was included in the calculation of the reaction time (see Section 8.4.9). 8.4.5 Short and long sentences There were six short sentences like the one illustrated in (8.7), consisting of a simple subject (an animate noun in plural), a verb (mitbringen ‘bring’), an adverb (immer ‘always’) and a simple object (an inanimate noun in plural). The separable but unstressed particle mit was located at the end of the sentence, resulting in a non-Wnal object. The sentences were inserted in three diVerent contexts inducing the following information structures: narrow corrective focus on the subject (NFS), see Figure 8.2, narrow corrective focus on the object (NFO), see Figure 8.3, and topic-focus (TF), the unmarked prosodic structure, see Figure 8.1. The sentences with narrow focus were elicited by replacing a pre-mentioned element with another one. Our decision to use a corrective narrow focus was driven by the intention to have a very clear accentual structure. A topic-focus was elicited by pre-mentioning some pairs of elements with the same structure as the tested sentence.
Gradient Perception of Intonation
157
250
200
150
100
50 Ruderer
bringen immer
L*H
BOOTE
H*L
mit
Li 1.74512
0 Time (s)
Figure 8.3. Pitch track of Ruderer bringen immer Boote mit
Figure 8.3 displays a narrow focus on the object. The subject Ruderer has a rising prenuclear pitch accent with a much smaller excursion than in the unmarked topic-focus conWguration. The object Boote carries the high-pitched nuclear accent. 8.4.6 Results and discussion Table 8.1 displays the data for the Wrst group of subjects, who had to give scalar judgements. Each cell shows the mean score of the six sentences having the same context-intonation pair. The second group of subjects judged the same sentences in a categorical way, and the mean scores for these subjects are given in Table 8.2. The correlation between the mean scores in Tables