Learning Argument Structure Generalizations

Initial interpretations of sentences

Non-linguistic context alone must be sufficient for the child to learn the meanings of his or her very first words. Gillette et al. (1998) have shown experimentally that this is plausible for nouns. Adult subjects were shown video clips of mother-child interactions; the sound was turned off except for a beep at the point in the clip when the mother uttered a particular word. Subjects were asked to guess the particular word uttered after witnessing six contexts in which the word was uttered (masked by the beep). They performed fairly well when guessing nouns, averaging 45% correct, but performed much more poorly when guessing verbs, with only 15% guessed correctly.

Following Gillette et al. (1998) and Fisher (1996), we would like to suggest that verb-centered categories can be learned without necessarily having prior knowledge of verb meaning. We need only assume some partial understanding of a number of nouns or pronouns, and an ability to make hypotheses about what is intended in certain contexts.

We assume the following principle, directly motivated by Grice (1975)’s Maxim of Relevance (“make your contribution relevant”) and Horn’s (1984) R-maxim (“say no more than you must’’):

If a word w is expressed, then the meaning of the word, “w,” is relevant to the situation being conveyed.

From this assumption alone, the child can infer that each known word expressed that refers to an entity indicates a relevant participant in the scene. This is a type of conversational implicature: an inference drawn in context on the basis of what is said and assumptions about the speaker’s communicative goals.¹ If we assume the child only understands you and bed, upon hearing the expression, you mop to bed, in the relevant context, the child might well be able to infer the relevant meaning of the unknown word mop as possibly “go” or “return.”²

The viability of this route to interpretation specifically for verbs has been demonstrated experimentally. As noted above, Gillette et al. (1998) found that subjects performed poorly at guessing verbs in the presence of only non-linguistic context. At the same time, when subjects were also provided a written list of co-occurring nouns (in alphabetical order—subjects were not provided with syntactic frames or even word order), subjects’ performance in correctly guessing the verbs was significantly increased over performance where only scenes or only nouns were available (29% vs 15% and 16.5%, respectively). Further evidence comes from Fisher (1996), who taught three and five year old children novel verbs in different sentence frames and informative non-linguistic contexts. The identity of subject and object referents was obscured by using pronouns that were ambiguous in context (She’s daxing her over there). Children demonstrated a significant sensitivity to the number and type of arguments in interpreting the novel verbs. Finally, Akhtar (1999) taught 2-4 year old children novel verbs in non-English word orders, and found that children had no trouble correctly interpreting sentences even though they only had knowledge of the two noun referents and the non-linguistic context (Nameera Akhtar, personal communication, May 5, 1998).

These findings predict that early on, children will produce errors in interpreting intransitive utterances that contain two recognizable nouns, if context allows for a transitive interpretation. As Fisher (2000) observes, based on data from Hirsh-Pasek and Golinkoff (1996), this prediction appears to be borne out as children younger than 25 months do misinterpret intransitives with conjoined subjects (Watch Big Bird and Cookie Monster flexing) as transitives.

Thus it seems that using nouns to guide the attentional process is a viable route to interpreting early utterances without prior knowledge of verb meaning. We can view this kind of verb learning as using a type of pragmatic bootstrapping to distinguish it from both syntactic bootstrapping (Landau & Gleitman, 1985) and semantic bootstrapping (Grimshaw, 1981; Pinker, 1984). Classic syntactic bootstrapping assumes that the child uses information about the structural positions of the nouns, e.g., which one is subject, in determining the verb’s semantic properties (Fisher, Gleitman, & Gleitman, 1991, ). While we concur that once such structural knowledge is acquired, it is used to infer properties of unknown verbs (seeSethuraman, Goldberg, & Goodman, 1997), we do not assume that the child must have abstract syntactic knowledge or knowledge of the rules of composition at the point when the first argument structure expressions are interpreted.

Pragmatic bootstrapping is also intended to be distinct from semantic bootstrapping, as the latter is traditionally understood. Classic semantic bootstrapping (Pinker, 1989, ) has been suggested as a way for the child to learn a phrase structure grammar. It assumes that the child is able to initially learn the meanings of both nouns and verbs based solely on the non-linguistic context. Pragmatic bootstrapping suggests a way in which verbs and other grammatical categories can be learned once a handful of nouns are learned. This process can be likened to semantic bootstrapping insofar as any known words that are heard when a sentence is uttered become part of the context, and can therefore be used to help direct a child’s attention to what is being conveyed. We now turn out attention to the question of how generalizations over individual utterances are made.

Verb-Centered Categories

Many studies have demonstrated that the initial acquisition of argument structure patterns is very conservative in that children stick closely to the forms they have heard used with particular verbs (Akhtar & Tomasello, 1997; Baker, 1979; Bates & MacWhinney, 1987; Bowerman, 1982; Braine, 1976; Brooks & Tomasello, 1999; Gropen et al., 1989; Ingram & Thompson, 1996; Lieven et al., 1997; MacWhinney, 1982; Olguin & Tomasello, 1993; Schlesinger, 1982; Tomasello, 1992). For example, Olguin and Tomasello (1993) taught 25 month old children four novel transitive verbs, each in a different syntactic pattern: both participants expressed, agent only, patient only or neither argument expressed. Children almost always reproduced the same exact pattern they had heard. Tomasello (1992) observed that by far the best predictor of his child’s use of a given verb on a particular day was her use of the same verb on the previous few days, not, as might be expected, her use of other verbs on the same day. Tomasello and his colleagues have discussed this verb-centered conservatism under the rubric of verb islands, since children readily substitute new nominals into the frames (Akhtar & Tomasello, 1997; Clark, 1996; Gropen, Epstein, & Schumacher, 1997; Tomasello, 1992; Tomasello, Aktar, Dodson, & Rekau, 1997).

There is evidence that adults retain much verb-specific knowledge as well. Verbs are occasionally quite idiosyncratic in the types of argument structure patterns they appear in (Bresnan, 1982; Chomsky, 1965; Pollard & Sag, 1987). For example, the near synonyms help and aid differ in their distribution:

1. 
   a. Pat helped her grandmother walk up the stairs.
   

b. *Pat aided her grandmother walk up the stairs.

2. 
   a. ??Pat helped her grandmother in walking up the stairs.
   

b. Pat aided her grandmother in walking up the stairs.

Psycholinguistic studies have demonstrated that speakers are influenced by the relative frequencies with which they have heard particular verbs used in various argument structure constructions (Ford, 1982; MacDonald, Pearlmutter, & Seidenberg, 1993). For example, knowledge that believed is more likely to appear with a clausal complement than with an object complement influences speakers’ on-line interpretation of potentially ambiguous sentences (Garnsey, Pearlmutter, Myers, & Lotocky, 1997; Trueswell, Tanenhaus, & Kello, 1993). The relative frequencies play a role despite the fact that both possibilities are fully grammatical:³

3. 
   a. Pat believed the speaker might cause a riot.
   

b. Pat believed the speaker.

We are suggesting that the earliest argument structure knowledge is, for example, roughly of the form given in . The angled brackets are intended to capture the fact that children readily substitute arguments other than the verb early on, subject to semantic constraints.

< actor>put< thing >< location >

Why should verbs be the basis of initial argument structure patterns? That is, why should the verb tend to be the only initially fixed slot in sentences rather than any other single word?

The predictive value of verbs in argument structure patterns

A critical factor in the primacy of verbs in argument structure patterns stems from their relevant predictive value. As many psychologists have emphasized, human categorization is generally driven by some functional pressure, typically the need to predict or infer certain properties on the basis of perceived characteristics (Anderson, 1991; Holland, Holyoak, & Thagard, 1989; Kersten & Billman, 1997; Leake & Ram, 1995; Ross & Makin, 1999; Wisniewski, 1995). That is, cognitive systems do not generalize randomly or completely; Holland et al. (1989), in their monograph on induction, emphasize that generalizations are constrained in that “the inferences drawn by a cognitive system will tend to be...relevant to the system’s goals” (p. 5). In the case of language, the language learner’s goal is to understand and to be understood: to comprehend and produce language. There is ample functional pressure to predict meaning on the basis of given lexical items and grammatical characteristics (comprehension); conversely, there is pressure to predict the choice of lexical items and grammatical characteristics given the message to be conveyed (production). Since the sentences the child is learning to understand and produce form an open-ended set, it is not sufficient to simply memorize the sentences that have been heard. The child must necessarily generalize those patterns at least to some extent in order to understand and produce new utterances.

If we compare verbs with other words (e.g., nouns), verbs are much better predictors of overall sentence meaning. Experimental evidence for this idea is provided by Healy & Miller (1970). Healy and Miller compared the relative contribution of verbs and subject arguments to overall sentence meaning. These two candidates, verb and subject, were presumably chosen because they appear to be the best candidates for representing overall sentence meaning. The subject argument is often referred to as the “topic” argument in a sentence or what the sentence is “about” (Kuno, 1972; Lambrecht, 1994; Reinhart, 1982). At the same time, the verb provides a great deal of information about who did what to whom. Healy and Miller 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 were asked to sort the sentences into five piles 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 (p < .008). That is, for example, all five sentences with the verb criticized were categorized together much more often than five sentences with the subject the critic. Healy and Miller concluded that the verb is the main determinant of sentence meaning.

Another source of evidence for the idea that the verb is a good predictor of sentence meaning comes from work on analogy. It has been richly documented that relational aspects of meaning are fruitful sources of analogy and similarity judgements (Gentner, 1982). Markman and Gentner (1993), for example, found that in making non-linguistic similarity judgments, similarity is judged to be greater when two representations are related in a systematic way via the relations that hold between the entities in each representation. The entities are aligned based on the structure that relates them, rather than on the basis of independent characteristics of the entities. The relevance to language is straightforward. In comparing two sentences, the main relational predicates, the verbs, are more likely to be used than the independent characteristics of the arguments (seeTomasello, 2000, for similar observations). Why should this be so? The purpose of analogies is generally one of drawing inferences and making predictions: what can be predicted on the basis of one situation about another situation (Gentner & Medina, 1998). Thus it is the value of verbs as good cues to sentence meaning that results in the child’s early learning of verb-centered argument structure patterns (verb islands).

Many verbs appear frequently in more than one construction, with concomitant variation in meaning. For example, while get seems to mean “acquire” in She got it, it is better interpreted as “make” in She got him angry; go appears frequently as a future marking auxiliary in addition to being the most frequent verb in the intransitive motion construction. It must be borne in mind that it is not the verb in isolation that is claimed to be predictive of sentence meaning, but rather the use of a verb in a particular construction, that is associated with a particular meaning. We return to this idea in section . As was outlined initially in section , the early verb islands involve verbs in particular argument patterns or constructions, not verbs in isolation.

Given the fact that early learning is based on particular verbs in constructional patterns, we can now turn to the questions of how and why learners ever generalize away from the verb to the more abstract level of argument structure constructions. Before focusing on this issue, we first review evidence that generalizations over particular verbs are in fact eventually made.

2.