6 ? ) Automatic Grammar Induction from Semantic Parsing

In this thesis, we investigate an approach for grammar induction that relies on semantics to drive the automatic learning of syntax rules. Specifically, we develop a semantic parser, which parses utterances based on "meaning," rather than syntax, by combining only words and phrases that satisfy a given set of semantic constraints. We subsequently extract a syntactic grammar from the resulting semantic-level phrases, trying various approaches to generalize this grammar. We evaluate the learned grammar utilizing two sets of experiments, restricting our test sets to semantically valid utterances. First, we use the grammar to parse new utterances from the same domain in which it was learned; the learned grammar covers 98% of the utterances handled by the semantic constraints. Second, we parse utterances from a new domain, assessing the portability of the grammar. Here, the grammar covers 85% of the semantically valid utterances. Semantic parsing proves to be a very powerful and useful mechanism, independent of syntax, providing an utterance's "meaning representation" directly. Furthermore, our experiments illustrate that this technique has potential for automatically developing portable grammars, making the task of moving an understanding system to a new domain easier. Company Supervisor: Dr. David Goddeau Title: Research Staff, DIGITAL Cambridge Research Laboratory Thesis Supervisor: Dr. James R. Glass Title: Principal Research Scientist