5th International Conference on Spoken Language Processing

Sydney, Australia
November 30 - December 4, 1998

Automatic Grammar Induction from Semantic Parsing

Debajit Ghosh (1), David Goddeau (2)

(1) Nuance Communications, USA
(2) Compaq Cambridge Research Laboratory, USA

One approach to spoken language understanding converts a transcribed utterance into a semantic representation, which is then interpreted to produce a response. This can be accomplished with conventional parsing technology given a syntactic grammar and semantic composition rules. However, constructing such a grammar can be difficult and time-consuming. An alternative approach is to learn the rules from translated examples. This eliminates the need for knowledge engineering but requires the collection and annotation of the examples, which can be as difficult. This research investigates using semantic information to learn syntax automatically. After describing a semantic parsing mechanism for parsing utterances based on meaning, we illustrate a grammar induction technique which uses semantic parsing's results to create syntactic rules. We also present experiments which use these rules in syntactic parsing experiments in two domains. The learned grammar covers 98% of semantically-valid utterances in its original domain and 85% in a different domain.

Full Paper

Bibliographic reference.  Ghosh, Debajit / Goddeau, David (1998): "Automatic grammar induction from semantic parsing", In ICSLP-1998, paper 0167.