1Submitted. (Feb 2003)Learning Argument Structure Generalizations*Adele E. GoldbergUniversity of [email protected] M. CasenhiserUniversity of [email protected] SethuramanUniversity of California, San [email protected] correlations between form and meaning at the level of argument structurepatterns have often been assumed to be innate. Claims of innateness typically rest on theidea that the input is not rich enough for general learning strategies to yield the requiredrepresentations. The present work demonstrates that the semantics associated withargument structure generalizations can indeed be learned, given the nature of the inputand an understanding of general categorization strategies. Examination of an extensivecorpus study of children's and mother's speech shows that one particular verb is found toaccount for the lion’s share of instances of each argument frame considered.Experimental results are described which demonstrate that high token frequency of asingle prototypical exemplar do indeed facilitate the learning of constructional meaning.1 IntroductionFor some time, linguists have observed that within a given language, there exist certainformal patterns that correlate strongly with the meaning of the utterance in which theyappear. Such correlations between form and meaning have been variously described aslinking rules projected from the main verb’s specifications (e.g., Bresnan & Kanerva,1989; Davis, 1996; Dowty, 1991; Grimshaw, 1990; Jackendoff, 1983) as lexicaltemplates overlain on specific verbs (Hovav & Levin, 1998), or as phrasal form andmeaning correspondences (constructions) that exist independently of particular verbs(Goldberg, 1995; Jackendoff, to appear).One way to account for the association of meanings with particular forms is to claimthat the association is innate (Gleitman, 1994; Grimshaw, 1990; Pinker, 1984, 1989,1994). This claim generally rests on the idea that the input is not rich enough for the3relevant generalizations to be learned; this is the well-known “poverty-of-the-stimulus”argument (Chomsky, 1980, 1988; Pinker, 1994). This view of learning a grammar can belikened to customizing a software package: everything is there, and the learner simplyselects the parameters that are appropriate for his environment (Jackendoff, to appear,Chapter 7). Many have criticized this approach for its biological implausability (Bates &Goodman, 1998; Deacon, 1997; Elman et al., 1996; Sampson, 1997). Moreoever, therehave been virtually no successful proposals for what any specific aspect of the parametersmight look like (Ackerman & Webelhuth, 1988; Culicover, 1999; Jackendoff 2002;Newmeyer, 1998).This paper joins the growing body of literature that detracts from the poverty-of-the-stimulus argument by presenting evidence that the nature and properties of at least certainpatterns in language are learnable on the basis of general categorization strategies (seealso e.g., Bybee & Moder, 1983; Bybee & Slobin, 1982; Jackendoff, Chapter 7 2002Lakoff, 1987; Taylor, 1995, Croft 2001, Tomasello to appear). In this paper we argue thatlanguage input provides more than adequate means by which learners can induce theassociation of meaning with certain argument structure patterns insofar as well-established categorization principles apply straightforwardly to this domain. Throughoutthis paper, we adopt constructional terminology, but the ideas we present are notexclusive to a constructionist account. Those who favor one of the other terminologiesmentioned above need only construe this account as a proposal for how children can learnlinking rules or learn the semantics associated with various lexical templates on the basisof the input. What is crucial is the uncontroversial notion that there do in fact existcorrelations between formal linguistic patterns and meaning.In Table 1, we provide a partial list of such form and meaning correspondences (lexicaltemplates, combination of linking rules, constructions) along with the labels we use asmnemonics throughout the paper to refer to them.Form/Example Meaning Construction Label1. Subj V Oblpath/locX moves Ypath/locIntr. motion (VL)e.g., The fly buzzed into the room.42. Subj V ObjOblpath/locX causes Y to move Zpath/locCaused-Motion (VOL)e.g., Pat sneezed the foam off the cappuccino.3. Subj V Obj RP X causes Y to become ZstateResultative (VOR)e.g., She kissed him unconscious.4. Subj V Obj Obj2 X causes Y to receive ZDouble Object (VOO)e.g., She faxed him a letter.Table 1. Examples of Correlations Between Form and MeaningBefore we discuss our corpus and experimental findings, we review evidence thatchildren store the associations of meanings with forms on two levels. The first involvesthe acquisition of verb-centered categories whereby children conservatively producesyntactic patterns on a verb-by-verb basis (Akhtar & Tomasello, 1997; Baker, 1979;Bates & MacWhinney, 1987; Bowerman, 1982; Braine, 1976; Brooks & Tomasello,1999; Gropen, Pinker, Hollander, Goldberg, & Wilson, 1989; Ingram & Thompson,1996; Lieven, Pine, & Baldwin, 1997; MacWhinney, 1982a; Olguin & Tomasello, 1993;Schlesinger, 1982; Tomasello, 1992, to appear). Ultimately, however, generalizationsover specific verbs are made, forming speakers’ knowledge of argument structurepatterns (Akhtar, 1999; Bowerman, 1982; Brooks & Tomasello, 1999).We do not address the question of exactly when generalizations emerge in this study;that is, no specific time line is suggested (see Fisher, 2002; Tomasello, 2000 for a rangeof views on this issue). Instead we address the more general question of how it ispossible for children to go from specific knowledge of individual verb usage toknowledge of more general linking patterns, using general inductive strategies. Thatchildren do have both levels of generalization is fairly uncontroversial (evidence isbriefly reviewed in sections 2 and 3).2 Verb-Centered Categories ExistMany studies have demonstrated that the initial production of argument structurepatterns is very conservative in that children stick closely to the forms they have heard5used with particular verbs (Akhtar & Tomasello, 1997; Baker, 1979; Bates &MacWhinney, 1987; Bowerman, 1982; Braine, 1976; Brooks & Tomasello, 1999; Gropenet al., 1989; Ingram & Thompson, 1996; Lieven et al., 1997; MacWhinney, 1982b;Olguin & Tomasello, 1993; Schlesinger, 1982; Tomasello, 1992, to appear). Forexample, Olguin and Tomasello (1993) taught 25