Designing Statistical Language Learners: Experiments on Noun Compounds

The goal of this thesis is to advance the exploration of the statistical language learning design space. In pursuit of that goal, the thesis makes two main theoretical contributions: (i) it identifies a new class of designs by specifying an architecture for natural language analysis in which probabilities are given to semantic forms rather than to more superficial linguistic elements; and (ii) it explores the development of a mathematical theory to predict the expected accuracy of statistical language learning systems in terms of the volume of data used to train them. The theoretical work is illustrated by applying statistical language learning designs to the analysis of noun compounds. Both syntactic and semantic analysis of noun compounds are attempted using the proposed architecture. Empirical comparisons demonstrate that the proposed syntactic model is significantly better than those previously suggested, approaching the performance of human judges on the same task, and that the proposed semantic model, the first statistical approach to this problem, exhibits significantly better accuracy than the baseline strategy. These results suggest that the new class of designs identified is a promising one. The experiments also serve to highlight the need for a widely applicable theory of data requirements.