Atila 2011 (antwerp, 1-2 December)

Developing computational algorithms that capture the complex structure of natural languages is an open problem. In particular, learning the abstract properties of language only from usage data without built-in knowledge of language structure remains a challenge. We have developed a Bayesian model of the acquisition and use of verb argument structure from child-directed data. In our model, the general constructions of language (such as transitive and intransitive) are viewed as a probability distribution over the syntactic and semantic features, e.g., the semantic properties of the verb and its arguments, and their relative word order in an utterance. Constructions are learned through clustering similar verb usages. Language use, on the other hand, is modeled as a Bayesian prediction problem, where the missing features in a usage are predicted based on the available parts and the acquired constructions (e.g., in sentence production, the best syntactic pattern for an utterance is predicted from the available semantic information). The model can successfully learn the common constructions of language, and its behaviour shows similarities to actual child data, both in sentence production and comprehension. Moreover, the acquired knowledge of language in this model is robust yet flexible, and many general patterns of behaviour that are observed in children can be simulated and explained by this approach.