Running Head: Argument Structure Constructions

Argument Structure Constructions 30

Running Head: Argument Structure Constructions

The Contribution of Argument Structure Constructions to Sentence Meaning

Giulia M. L. Bencini

Adele E. Goldberg

Department of Linguistics

University of Illinois

Urbana IL 61801

Abstract

What types of linguistic information do people use to construct the meaning of a sentence? Most linguistic theories and psycholinguistic models of sentence comprehension assume that the main determinant of sentence meaning is the verb. This idea was argued explicitly in Healy and Miller (1970). When asked to sort sentences according to their meaning, Healy and Miller found that participants were more likely to sort sentences according to the main verb in the sentence than according to the subject argument. On the basis of these results, the authors concluded that the verb was the main determinant of sentence meaning. In this study we used the same sorting paradigm to explore the possibility that there is another strong influence on sentence interpretation: the configuration of complements (the argument structure construction). Our results showed that participants did produce sorts by construction, despite a well-documented tendency for subjects to sort on the basis of a single dimension, which would favor sorts by verb.

The relationship between verb, sentence form and sentence meaning has been at the center of linguistic theory and psycholinguistic models of sentence processing for several decades. Within linguistic theory, the predominant view since Chomsky (1965) has been that the lexical representation of a verb specifies (or projects) the number and types of arguments corresponding to the participants in the event described by the verb (its subcategorization frame or argument structure1). This view suggests that the verb is the best predictor of general sentence interpretation. For example the lexical representation for give would specify that it requires three arguments: a subject, a direct object and an indirect object as illustrated by the sentence: Pat gave a cookie to Kim. We will refer to this view as the verb-centered view. The relationships between sentences with the same lexical items but different argument structures (alternations) have been captured by positing lexical rules or transformations; the two sentences have often been assumed not to differ importantly in meaning (following Katz, 1964). For example, the ditransitive in (1a) has been argued to be derived from (1b) by lexical rule or transformation with little change in meaning (e.g., Larson, 1988).

(1)

a. Pat gave Kim a cookie

b. Pat gave a cookie to Kim

Most psycholinguistic models of sentence comprehension and production also assume that argument structure information is encoded in the verb. Experimental evidence has demonstrated that the main verb is in fact a critical factor in sentence comprehension. Verb-specific statistical preferences for particular argument structures have been shown to affect the on-line processing of temporarily ambiguous sentences (Garnsey, Pearlmutter, Myers & Lotocky, 1997; Trueswell, Tanenhaus & Kello, 1993). Lexicalist models of sentence comprehension (Boland & Boehm-Jernigan, 1998; Juliano & Tanenhaus, 1994; MacDonald, Pearlmutter & Seidenberg, 1994) capture the effects of argument structure biases by assuming that argument structures are probabilistically associated with the lexical representation of verbs, and are available as soon as the verb is recognized. Building on the sentence production model of Levelt (1989) and Levelt et al. (1999), in which syntactic information about the verb is represented in a separate lemma stratum, Pickering and Branigan (1998) proposed that the argument structure possibilities for a verb are also part of the lemma information for that verb along with category type and syntactic feature information (number, person, gender). Although these models allow for one verb to be probabilistically associated with more than one argument structure, they still assume that the verb is the main determinant of syntactic and semantic information about the sentence.

An early study that specifically looked at the contribution of the main verb to the overall meaning of the sentence is one by Healy and Miller (1970). The authors compared the relative contribution of verbs and subject arguments to overall sentence meaning. They constructed 25 sentences by crossing 5 subject arguments (the salesman, the writer, the critic, the student, the publisher), 5 verbs (sold, wrote, criticized, studied, published) and one patient (the book). Participants sorted the sentences according to similarity in meaning. Results showed that participants sorted sentences together that had the same verb much more often than sentences that had the same subject argument. Healy and Miller concluded that the verb is the main determinant of sentence meaning.

It does seem to be true that of all the words in a sentence, verbs are the ones that carry the most information about the syntax and the semantics of the sentence. Because of the high predictive value of verbs, it is reasonable to assume that people use this information during comprehension to predict other lexical items in the sentence and the overall meaning of the sentence. Conversely, given a message to be expressed, speakers can use information about the verb to activate lexical items and syntactic forms. Two additional observations, however, suggest that the predictive value of verbs with respect to the overall meaning of the sentence may not be as strong as assumed by traditional linguistic theories, nor as suggested by Healy and Miller (1970).

The first observation is that verbs occur in many more argument structure configurations than is generally assumed (Goldberg, 1995; Rappaport Hovav & Levin, 1998). For example kick, which is traditionally considered to be a prototypical transitive verb, can occur in at least eight argument structure frames:

1. Pat kicked the wall

2. Pat kicked Bob black and blue

3. Pat kicked the football into the stadium

4. Pat kicked at the football

5. Pat kicked her foot against the chair

6. Pat kicked Bob the football

7. Horses kick

1. Pat kicked his way out of the operating room

The sentences in 1-8 designate a variety of event types including simple transitive action (1), caused change of state (2), caused motion (3), attempted action (4), transfer (6), and motion of the subject referent (8).

Second, even when the predominant treatment of argument structure alternations was the transformational one, many linguists recognized that argument structure configurations are associated with systematic variations in meaning (Anderson, 1971; Borkin, 1974; Fillmore, 1968; Partee, 1965; Wierzbicka, 1988). With respect to the dative alternation in (2) and (3), Partee (1965), for example, noted that the ditransitive argument structure requires that the goal argument be animate, while this is not true of the paraphrase with to:

(2)

a. I brought a glass of water to Pat

b. I brought Pat a glass of water (ditransitive)

(3)

a. I brought a glass of water to the table

b. *I brought the table a glass of water (ditransitive)

One way to account for these variations in meaning is to posit a different verb sense for each possible argument structure configuration (Levin & Rappaport Hovav, 1995; Pinker, 1989), hereafter the multiple-verb-sense approach. In the multiple-sense view, bring in I brought a glass of water to Pat is argued to be a different sense than bring in I brought Pat a glass of water, and ran in Chris ran home is argued to be a different than ran in Chris ran. Psychologically, the multiple-sense view translates into assuming different long-term representations for each verb sense which are stored in the mental lexicon. In production, then, the speaker would select the appropriate sense from the mental lexicon, and this in turn would be associated with a unique argument structure pattern. In comprehension, the task is more complicated, because in English there are no overt markings on the verb that could cue the comprehender as to which sense is intended. Therefore, the comprehender would have to observe the argument structure configuration, reason backwards to the semantic interpretation that is assumed to have produced that configuration, and select the corresponding sense.

An alternative way to account for alternations and the fact that verbs occur in many argument structure patterns is to assign meaning directly to various abstract argument structure types, thereby recognizing the argument structure patterns as linguistic units in their own right. We refer to this approach as the constructional approach (Fillmore & Kay, 1999; Goldberg, 1995; Jackendoff, 1997; Michaelis & Lambrecht, 1996; Rappaport Hovav & Levin, 1998). Examples of English argument structure constructions with their forms and proposed meanings are shown in Table 1.

********************************

* Insert Table 1 about here. *

********************************

On this view, argument structure patterns contribute directly to the overall meaning of a sentence, and a division of labor can be posited between the meaning of the construction and the meaning of the verb in a sentence. While the constructional meaning may, perhaps prototypically, be redundant with that of the main verb, the verb and construction may contribute distinct aspects of meaning to the overall interpretation. For example, the ditransitive construction has been argued to be associated with the meaning of transfer or “giving” (Goldberg, 1995; Green, 1974; Pinker, 1989). When this construction is used with give, as in Kim gave Pat a book, the contribution of the construction is wholly redundant with the meaning of the verb. The same is true when the construction is used with send, mail, and hand. As is clear from these latter verbs, lexical items typically have a richer core meaning than the meanings of abstract constructions.

In many cases, however, the meaning of the construction contributes an aspect of meaning to the overall interpretation that is not evident in the verb in isolation. For example the verb kick need not entail or imply transfer (cf. Kim kicked the wall). Yet when kick appears in the ditransitive construction, the notion of transfer is entailed. The ditransitive construction itself appears to contribute this aspect of meaning to the sentence. That is, the sentence Kim kicked Pat the ball can be roughly paraphrased as “Kim caused Pat to receive the ball by kicking it”. The construction contributes the overall meaning of “X causes Y to receive Z”, while the verb specifies the means by which the transfer is achieved, i.e. the act of kicking. On this view, it is a conventional fact of English grammar that the pattern “Subject Verb Object1 Object2” is productively associated with the meaning of transfer; many other languages do not have this pattern, and others have it with a narrower or broader range of meanings. Although this study does not attempt to discriminate between the multiple-verb sense view and the constructional view, we will return to a comparison of the two approaches in the General Discussion.

The two experiments reported here were designed to test whether argument structure constructions play a role in determining sentence meaning. Both experiments used a sorting paradigm (Healy & Miller, 1970) in which participants were asked to sort sentences according to their overall meaning. The stimuli consisted of 16 sentences obtained by crossing four constructions and four verbs. That is, there were four sets of sentences which contained the same verb (four sentences each with throw, take, get and slice), and there were four instances of each of the transitive, ditransitive, caused-motion and resultative constructions.

The use of a sorting paradigm is a particularly stringent test to detect the role of argument structure constructions. Research in the category formation literature has shown that there is a strong domain-independent bias towards performing unidimensional sorts, even with categories that are designed to resist such unidimensional sorts in favor of a sort based on a family resemblance structure.

For example, in a series of experiments, Medin et al. (1987) found that participants persistently performed unidimensional sorts, despite the fact that the category and stimulus structures were designed to induce family resemblance sorting (Rosch & Mervis, 1975). The unidimensional bias was found across a variety of stimulus materials (pictures and verbal descriptions), procedures (on-line and from memory), and instructions. The unidimensional bias was not eliminated by adding complexity to the stimuli nor by adding individuating information to the exemplars. Participants sorted unidimensionally even when they were explicitly told to pay attention to all of the dimensions of the stimuli. Unidimensional sorting has been found even with large numbers of dimensions (Smith, 1981), ternary values on each dimension (Ahn & Medin, 1992), holistic stimuli, and stimuli for which an obvious multidimensional descriptor was available (Regehr & Brooks, 1995).

Lassaline and Murphy (1996) hypothesized that the reason participants do not sort categories according to family resemblance even when the category has a family resemblance structure is that family resemblance sorting is computationally more difficult. It requires attending to information across different dimensions and identifying the relations among them. Given the choice, people will take the easier route and sort unidimensionally.

In the stimuli used in both of our experiments, the verb provides a concrete shared dimension among subsets of sentences. In the light of the studies just cited, this is an invitation to perform a sort based solely on the verb. In contrast, sentences that share the same construction need not (and in this study, did not) have anything concrete in common; their similarity is abstract and relational, requiring the recognition that several grammatical relations co-occur.

It would of course be possible to design stimuli with a great deal of overlapping propositional content such that we could a priori predict either a verb or constructional sort. For example, the sentences Pat shot the duck and Pat shot the duck dead would very likely be grouped together on the basis of overall meaning, despite the fact that the constructions are distinct. Conversely, Pat shot the elephant and Patricia stabbed a pachyderm would likely be grouped together despite the fact that no exact words are shared. Our stimuli were designed to minimize such contentful overlap contributed by anything other than the lexical verb. No other lexical items in the stimuli were identical or near synonyms. Thus, we manipulated the stimuli such that we could compare the contribution of the verbs with the contribution of the constructions.

Verb-centered approaches predict that the verb should be an excellent predictor of overall sentence meaning. Because the verb also provides a route to a simple unidimensional sort, these approaches predict that subjects would sort overwhelmingly on the basis of the main verb. Both the constructional approach and the multiple-verb sense approach, albeit for different reasons, described below, predict that the argument structure configuration should play a critical role in determining meaning. Therefore at least some constructional sorts would be expected to occur.