Distributional cues to grammatical categorization : acquiring categories in a miniature artificial grammar

Thesis (Ph. D.)--University of Rochester. Dept. of Brain & Cognitive Sciences, 2010.A crucial component of language acquisition involves organizing words into grammatical categories and discovering relations between them. The organization of words into categories, and the generalization of patterns from some seen word combinations to novel ones, account for important aspects of the expansion of linguistic knowledge in the early stages of language acquisition. One hypothesis of how learners handle the problem of categorization is that they exploit distributional information in the input to discover the category structure of natural languages (e.g., Braine, 1987). However, given the information processing limitations of young children and the complexity of the computational processes that would be entailed, this hypothesis has often been thought to be an unlikely contributor to categorization. Indeed, many studies have argued that in order for a learner to successfully utilize distributional information for category learning, there must be multiple correlated cues to category structure in the input (e.g., Gomez & Gerken, 2000). This, however, has been somewhat of a puzzle: grammatical categories and subcategories in natural languages do not always have reliable phonological, morphological, or semantic cues (Maratsos & Chalkley, 1980). Rather, learners must utilize distributional cues about the linguistic contexts in which words occur to acquire such categories. In this thesis we hypothesize that the patterning of tokens in a corpus of linguistic input is sufficient, along with a small set of learning biases, to extract the underlying structural categories in a language. We present a series of artificial grammar learning studies that examine how distributional variables will shift learners from forming a category of lexical items to maintaining lexical specificity. We begin with a series of experiments testing whether learners can acquire a single category, generalizing from some instances of the distributional contexts of individual words in the exposure set (but some withheld) to the full range of contexts for all the individual words in the set. To do this, we vary a number of distributional variables to category structure and test how adult learners use this information to inform their hypotheses about categorization. Our results show that learners are sensitive to the contexts across words, the non-overlap of contexts (or systematic gaps in information), and the size of the exposure. These variables taken together determine whether learners fully generalize or preserve lexical specificity. We also explore whether subcategories are learnable from distributional information, if the learner is given adequate overlap inside each subcategory and adequate non-overlap between subcategories. Contrary to much of the existing literature, our results demonstrate that learners can use distributional information alone to find subcategory boundaries. Furthermore, we demonstrate that learners are able to learn lexical exceptions in their input while maintaining subcategory boundaries. Lastly, we explore the role of frequency variation for individual items, and we show that learners can overcome variations in input frequencies in order to maintain a rational generalization strategy based on their exposure to the language. Overall, our results show that learners are able to take account of a rich set of variables to aid them in grammatical categorization, including degree of overlap among category members, amount of input, consistency of gaps and overlaps in the input, and conflicts or consistency among distributional cues