Learning Syntax-Semantics Mappings to Bootstrap Word Learning

Learning Syntax-Semantics Mappings to Bootstrap Word Learning Chen Yu (chenyu@indiana.edu) Department of Psychological and Brain Sciences, and Program in Cognitive Science Indiana University, Bloomington, 47405, IN USA Abstract to deduce the word argument structures. In contrast, Gleit- man (1990) proposed an alternative account called syntactic bootstrapping. She argued that children use syntactic knowl- edge they have developed to learn what words mean. More specifically, the semantically relevant syntactic structures sur- rounding a verb, such as the subcategorization frames around a verb, provide contextual cues for its meaning. These two hypotheses focus on different aspects of potential interactions between syntax and semantic learning, and both of them have been supported by empirical studies. The present paper proposes a computational model of how syntax-semantics mappings can be learned and emerged from two language learning processes – syntax learning and word learning, and how these mappings can then facilitate word learning. Our study is quite different from previous work in several important ways. First, we propose and implement a general statistical-learning mechanism in which syntactic cues can be seamlessly integrated with already learned se- mantic knowledge to help the learning of new words. We suggest that syntax can act as a linguistic spotlight that fa- cilities word learning by selecting, grouping and highlighting those words that are likely to have the same type of refer- ents. Using the proposed learning mechanism, we demon- strate how syntactic learning could help object name learning and how the development of grammatical abilities continues to be highly linked to lexical development. Second, both the proposed learning mechanism of syntax-semantics mappings and the mechanism of utilizing the mapping knowledge in word learning are general that can be applied not only to a specific syntactic category (verb, etc.) but also to other cat- egories. Thus, we suggest that the acquisition of syntax and the integration of syntactic cues in word learning might par- tially account for the explosive expansion of vocabulary as primary syntactic structures are gradually acquired. Third, we apply the model to raw data collected from everyday parent- children interaction but not to some artificial or synthesized data, and show a dynamic picture of how the learning mech- anism works with realistic input. This paper addresses possible interactive effects between word learning and syntax learning at an early stage of develop- ment. We present a computational model that simulates how the results from a syntax learning process and a word learn- ing process can be integrated to build syntax-semantics map- pings, and how the emergence of links between syntax and word learning could facilitate subsequent word learning. The central idea of our statistical model is to categorize words into groups based on their syntactic roles and then estimate seman- tic meanings of those syntactic categories using lexical knowl- edge acquired from a concurrent word learning process. Once built, those syntax-semantics mappings can be further utilized as a syntactic constraint in statistical word learning. We ap- plied the model to realistic data collected from child-mother picture book reading interaction. A comparative study be- tween a statistical model and the model based on both statisti- cal and syntactic information shows that syntactic cues can be seamlessly integrated in statistical learning and significantly improve word learning performance. Introduction One of the most complex learning tasks young children are faced with is to learn their native language. Language ac- quisition, of course, consists of several distinct tasks, such as speech perception, speech segmentation, word learning and syntax learning. Among others, word learning involves how to map a phonological form to a conceptual representation, such as associating the sound “dog” to the concept of dog. Thus, the crucial issue in word learning is to build word- to-world mappings from language and extralinguistic con- texts. Syntax learning, on the other hand, is mainly about how to categorize words into grammatical categories (e.g. noun, verb, etc.) which are basic building blocks of gram- mar, and then how to acquire the hierarchical and context- sensitive structures that are represented by those syntactic cat- egories. Therefore, syntax learning uses sequential symbolic data (sentences in a language, etc.) to construct a grammar. Although acquisition of the lexicon and acquisition of the grammar seem to address totally different issues, these two learning processes might be closely related due to universal correspondences between syntax and semantics. For instance, Bloom (1994) pointed out bidirectional mappings between syntax and semantics, such as count nouns to kinds of in- dividuals, mass nouns to kinds of portions, and Noun Phrases (NPs) to individuals. Such mappings suggest a possible boot- strapping procedure between these two learning processes – the progresses in one learning process could facilitate the other learning process. In fact, two compelling hypotheses have been proposed by theorists. The semantic bootstrapping hypothesis (Pinker, 1989) argued that word meanings can be antecedently acquired from the observation of events and then used to determine the syntactic category of each word and Related Work There are a number of existing models that account for dif- ferent aspects of word learning. Plunkett, Sinha, Miller, and Strandsby (1992) built a connectionist model of word learning in which a process termed autoassociation mapped preprocessed images with linguistic labels. The linguis- tic behavior of the network exhibited non-linear vocabu- lary growth (vocabulary spurt) that was similar to the pat- tern observed in young children. Colunga and Smith (2005) presented a connectionist model showing that regularities among object and substances categories were learnable and generalizable enabling the system to become, after train- ing, a more rapid learner of new object and substance