Christoph Scheepers

Barbara Hemforth

Lars Konieczny

a. grüßen soll;

b. Frau Müller vorstellen soll

b. (to) Mrs Müller introduce shall

that the neighbour (for) Mrs Meier Mrs Müller greet should

b. daÁEder Nachbar Frau Meier vorstellen sollte

As the different subcategorisation frames are weighted differently for each type of verb, we expected that the ease of integrating the predicate at the end of a clause is dependent on a) the syntactic commitments made prior to the verb and b) the lexical preferences of the verb itself: after having processed a subject and two objects (i.e. a dative and an accusative object), for instance, the parser should predict a ditransitive verb (satisfying the minimum requirement for the number of argument slots to be provided by the predicate). Thus, after two objects, the integration of a ditransitive bias verb like “vorstellen E(to introduce, cf. 1b) should be fairly easy, whereas the integration of a monotransitive bias verb like “grüßen E(to greet, cf. 2a) should show processing disruption, as it requires the activation of a non-preferred subcategorisation frame.

After having processed a subject and only one object (cf. sentence fragment 1), on the other hand, the ease of integrating an immediately following verb may or may not vary as a function of the verb’s lexical preferences: if the processor commits itself not only to the minimum, but also to the maximum number of potential arguments (i.e. if it expects the most common monotransitive structure, as predicted by the frequency account), integration of a mono-transitive bias verb like “grüßen E(cf. 1a) should be easy, but integration of a ditransitive bias verb like “vorstellen E(cf. 2b) should be hard; if the maximum number of potential arguments remains under-specified until the verb is read, then, after having read only one object NP, both types of verbs should be fairly easy to integrate.

Method

We will report results from an acceptability questionnaire, a self-paced reading study, and an eye-tracking experiment (the former two have been carried out at the University of Freiburg, Germany, and the latter is currently being run with native German participants at the University of Glasgow, Scotland). All of these experiments employ the same two-factor (2x2) design in which the verbs Elexical preferences (monotransitive vs. ditransitive bias) and the respective fit between the number of ar-guments specified in the sentence and the verb’s lexical preferences (matching vs. mismatching) were orthogonally specified (3a-d).

That the neighbour_[nom] (to) the friend_[dat] the brother in law_[acc] introduce should, …

c. DaÁEder Nachbar den Schwager grüßen sollte, …

That the neighbour_[nom] the brother in law_[acc] greet should, …

d. DaÁEder Nachbar dem Freund den Schwager grüßen sollte, …

That the neighbour_[nom] (for) the friend_[dat] the brother in law_[acc] greet should, …

Two sets of verbs were selected from the Wahrig German Dictionary: verbs like “vorstellen E(to introduce, cf. 3a,b) whose first entry specifies a di-transitive verb frame (subcategorising for a subject, a dative object and an accusative object), but which also allow an alternative monotransitive reading; and verbs like “grüßen E(to greet, cf. 3c,d) whose first entry specifies a monotransitive verb frame (sub-categorising for a subject and an accusative object), but which also allow for the alternative ditransitive reading. The verbs were matched for lexical token frequency, and the respective argument structure preferences were confirmed by a sentence completion study run on a different sample of participants than the other experiments.

The verbs were included in VP-head-final con-structions, such that the verbs Elexical preferences either matched the given syntactic context (3a,c) or not (3b,d). The experiments also included a control condition (4) comprising dative-object monotran-sitive verbs like “helfen E(to help) in a matching syntactic environment. The purpose of this control condition was to test whether dative case com-plements can elicit a ditransitive verb expectation.

That the neighbour_[nom] the friend_[dat] help should, …

To summarise the predictions outlined in the introduction, a frequency-based model predicts a main effect of a match vs. mismatch between preferred verb valency and the number of arguments specified in the sentence (processing difficulty in case of a mismatch, regardless of the type of verb involved, cf. 3b,d), whereas the minimal commitment account predicts an interaction: the integration of monotransitive bias verbs like “grüßen E(to greet, cf. 3c,d) should be hard in case of a mismatch (cf. 3d), whereas the integration of ditransitive bias verbs like “vorstellen E(to introduce, cf. 3a,b) should be un-affected by prior syntactic context.

Results

Table 1 shows the average residual reading times (i.e. reading times corrected for length effects) at the main verb and at the following auxiliary in each experimental condition (obtained in a self paced reading study with 32 native German participants). At both regions, there was a significant interaction between Syntactic Fit and Lexical Preference (main verb: F1(1,31) = 8.71; p < .01; F2(1,19) = 8.88; p < .01; auxiliary: F1(1,31) = 11.41; p < .005; F2(1,19) = 13.55; p < .005). Planned comparisons revealed significant processing disruption for monotransitive bias verbs in the mismatch condition as compared to the match condition (main verb: F1(1,31) = 8.85; p < .01; F2(1,19) = 18.73; p < .001; auxiliary: F1(1,31) = 21.75; p < .001; F2(1,19) = 37.74; p < .001). For ditransitive bias verbs, on the other hand, no reliable match vs. mismach contrast was established (all Fs < 1 at both regions). Reading times in the control condition (4) revealed no indication of a ditransitive expectation after having read a dative object.

Mean off-line acceptability ratings (on a scale from 1 = unacceptable to 4 = perfectly acceptable) are shown in Table 2: the first column represents judgements made after each trial in the self paced reading study (i.e. by exactly the same participants), whereas the second column shows the judgements of a different sample of participants (N = 32) whose task was to rate exactly the same materials in a paper-and-pencil questionnaire. In both experiments, participants were instructed to base their accept-ability judgements on grammaticality as well as plausibility.

The results were the same for both samples, and differed considerably from the on-line data: instead of an interaction (all p > .50), there was a significant main effect of Syntactic Fit (p < .001 for both samples by Wilcoxon Signed-Ranks Tests across subjects and items, respectively), indicating that a mismatch between the lexical preference of the verb and its syntactic environment leads to a decrease in acceptability, regardless of the type of verb involved, i.e. regardless of whether the verb subcategorises for one or two objects, respectively.

Since the acceptability judgements almost certainly do not reflect the on-line syntactic integration processes of interest (but rather indicate the results of processes induced after the verb has become available), we will base our conclusions on the reading time data obtained at the main verb itself and at the following auxiliary. Here, the obtained data patterns clearly favour the minimal commitment account over the frequency-based account: although syntactic integration of a monotransitive bias verb into a ditransitive construction leads to considerable processing difficulty, integration of a ditransitive bias verb into a monotransitive construction does not. (Preliminary data from the eye-tracking study indicate a very similar pattern). This suggests that when the parser is generating expectations about the verb’s sub-category during on-line sentence proces-sing, it predicts a minimum number of argument-slots of the potential verb (justifying the number of arguments already available), but leaves the maximum number of potential arguments under-specified.

Several possible counter-arguments to this con-clusion can be ruled out. Firstly, it could be argued that  EI>beneficiary Econstructions like (3d, cf. 2a) are considerably less felicitous, even when compared to  EI>implicit argument Econstructions like (3b, cf. 2b). However, the obtained acceptability judgements do not convincingly support such a claim (cf. Table 2). Secondly, it may be the case that the two sets of verbs differed with respect to the strength of their respective lexical preferences. Not only the results from the sentence completion study (pre-test), but also the off-line acceptability ratings militate against such a hypothesis, as for both types of verbs non-matching syntactic environments lead to about the same decrease in acceptability.

The present evidence suggests that the human parser, in fact, does make predictions about the subcategorisation properties of a verb yet to appear, which may be taken as evidence against strictly head-driven accounts of sentence processing (e.g., Pritchett, 1992). However, in doing so, it restricts itself to predicting only the minimum number of potential arguments of the verb (as specified by how many complements have already been read), rather than relying on frequency information to generate expectations about the maximum number of potential arguments. Such a ‘cautious Estrategy appears opti-mal in the sense that it avoids generating wrong predictions, which may otherwise require lots of syntactic revisions during structure building. Nevertheless, expanding valency classes of potential verbs in a monotonic fashion allows incremental processing of verb-final constructions. Hence, it can explain part of the miracle of why verb-final constructions are easy to comprehend.