A Graph-theoretic Model of Lexical Syntactic Acquisition

Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing, pages 917–926, Honolulu, October 2008. c⃝2008 Association for Computational Linguistics A graph-theoretic model of lexical syntactic acquisition Hinrich Sch¨utze and Michael Walsh Institute for Natural Language Processing University of Stuttgart, Germany {hs999,walsh}@ifnlp.org Abstract This paper presents a graph-theoretic model of the acquisition of lexical syntactic representa- tions. The representations the model learns are non-categorical or graded. We propose a new evaluation methodology of syntactic ac- quisition in the framework of exemplar theory. When applied to the CHILDES corpus, the evaluation shows that the model’s graded syn- tactic representations perform better than pre- viously proposed categorical representations. 1 Introduction In recent years, exemplar theory has had great ex- planatory success in phonetics. Exemplar theory posits that linguistic production and perception are not mediated via abstract categories, but that instead each production and perception of a linguistic unit is stored and retained. Linguistic inference then di- rectly operates on these stored exemplars. In this pa- per, we propose a new approach to lexical syntactic acquisition in the framework of exemplar theory. Our approach uses an evaluation measure that is different from previous work. Lexical syntac- tic acquisition is most often evaluated with respect to standard syntactic categories like verb and noun. Our first contribution in this paper is that we instead evaluate learned representations in the context of a syntactic task. This task is the determination of an aspect of grammaticality that we call local syntactic coherence. Our second contribution is a graph-theoretic model of the acquisition of lexical syntactic rep- resentations that is more rigorous than previous heuristic proposals. The graph-theoretic model can learn both categorical and non-categorical (or graded) representations. The model is also a unified framework for syntagmatic and paradigmatic rela- tions (as will be discussed below), and for lower- order syntactic relations (those that can be directly observed from the input) and higher-order syntac- tic relations (those that require some generalization from what is directly observable). Redington et al. (1998) give an influential account of the acquisition of lexical syntactic representations in which a standard syntactic category like verb or noun is assigned to each word. Our third contribu- tion is to show that, in the context of acquisition, graded representations are superior to standard cat- egorical representations in supporting judgments of local syntactic coherence. A graded representation formalism is one that, for any two words, can rep- resent a third word whose syntactic properties are intermediate between the two words (Manning and Sch¨utze, 1999). Clearly exemplar theory is not the only frame- work in which lexical acquisition has been explored. Gleitman (1990) for example argues for syntactic bootstrapping to infer lexical semantics, work not at odds with our own (see discussion on the role of se- mantics below). Our argument for the importance of distributional evidence does not call into question the large body of work in child language acquisition that demonstrates that “part of the capacity to learn languages must be ’innate’ ” (Gleitman …

Content truncated to 500 words