Headword Percolation in a Multi-Parser Architecture for Natural Language Understanding

摘要

This paper describes an approach for identifying the information goal(s) of information-seeking queries, such as those in the ATIS domain (Price, 1990). The approach is based the Multi-Parser Architecture (MPA) extended with a headword percolation procedure. MPA can generate full and robust parses for input natural language queries. Headword percolation identifies the prime constituent(s) in the parse trees. These constituents are incorporated in goal identification rules derived from training parse trees using an automatic, data-driven technique. The rules are then evaluated with a test set in terms of in single goal identification, multiple goal identification and out-of-domain rejection. We compare the current approach with a previous one based on Belief Networks (BN) (Meng et al. 1999). Results show that the former gave statistically significant improvements. This is indicative of the merits of rich semantic/syntactic analysis in the parse structures for goal identification.