| This thesis deals with the problem of building fast, accurate and portable parsers for natural language understanding. Our focus is a multi-domain dialogue system in which we need a deep linguistically-motivated parser to produce the representations of the input suitable for reasoning. In this dissertation, we are concerned with building parsers which have the wide coverage and portability offered by a general syntactic grammar without sacrificing parsing speed and accuracy.; Our approach relies on a domain-independent deep parser and grammar which uses selectional restrictions to control parsing speed and accuracy. We develop a feature list representation as the basis for selectional restrictions, and a formal model for using selectional restrictions in a unification based framework.; We then develop a lexicon design for multi-domain parsing and semantic interpretation. We show how the restrictions based on feature sets can be integrated with a traditional frame-based semantics, and extend our formal model to cover inheritance and defaults in the lexicon. We show that none of the existing large-scale ontologies and lexicons provide all the information necessary for parsing and semantic disambiguation, and we develop a parsing lexicon suitable for use with a wide-coverage grammar in multiple domains.; Our domain-independent lexicon provides coverage and portability over four different application domains. To customize the representations produced by the parser for domain reasoning, we designed an architecture with mappings between our domain-independent ontology and a domain model. This architecture allows us to produce semantic representations optimally suited for different application domains. In addition, we use the mappings to specialize the lexicon for improved parsing speed and accuracy, and show that our specialization method significantly improves parsing performance.; Finally, we develop a statistical model to learn selectional restrictions from corpora and show how it can be used to distinguish between acceptable and unacceptable verb-object pairs in data sets derived from our lexicon and from the World Street Journal corpus. |
