A Probabilistic Model of Early Argument Structure Acquisition

Developing computational algorithms that capture the complex structure of natural language is an open problem. In particular, learning the abstract properties of language only from usage data remains a challenge. In this dissertation, we present a probabilistic usage-based model of verb argument structure acquisition that can successfully learn abstract knowledge of language from instances of verb usage, and use this knowledge in various language tasks. The model demonstrates the feasibility of a usage-based account of language learning, and provides concrete explanation for the observed patterns in child language acquisition. We propose a novel representation for the general constructions of language as probabilistic associations between syntactic and semantic features of a verb usage; these associations generalize over the syntactic patterns and the fine-grained semantics of both the verb and its arguments. The probabilistic nature of argument structure constructions in the model enables it to capture both statistical effects in language learning, and adaptability in language use. The acquisition of constructions is modeled as detecting similar usages and grouping them together. We use a probabilistic measure of similarity between verb usages, and a Bayesian framework for clustering them. Language use, on the other hand, is modeled as a prediction problem: each language task is viewed as finding the best value for a missing feature in a usage, based on the available features in that same usage and the acquired knowledge of language so far. In formulating prediction, we use the same Bayesian framework as used for learning, a formulation which takes into account both the general knowledge of language (i.e., constructions) and the specific behaviour of each verb. We show through computational simulation that the behaviour of the model mirrors that of young children in some relevant aspects. The model goes through the same learning stages as children do: the conservative use of the more frequent usages for each individual verb at the beginning, followed by a phase when general patterns are grasped and applied overtly, which leads to occasional overgeneralization errors. Such errors cease to be made over time as the model processes more input. We also investigate the learnability of verb semantic roles, a critical aspect of linking the syntax and semantics of verbs. In contrary to many existing linguistic theories and computational models which assume that semantic roles are innate and fixed, we show that general conceptions of semantic roles can be learned from the semantic properties of the verb arguments in the input usages. We represent each role as a semantic profile for an argument position in a general construction, where a profile is a probability distribution over a set of semantic properties that verb arguments can take. We extend this view to model the learning and use of verb selectional preferences, a phenomenon usually viewed as separate from verb semantic roles. Our experimental results show that the model learns intuitive profiles for both semantic roles and selectional preferences. Moreover, the learned profiles are shown to be useful in various language tasks as observed in reported experimental data on human subjects, such as resolving ambiguity in language comprehension and simulating human plausibility judgements.