Search in a Learnable Spoken Language Parser

Abstract

We describe and experimentally evaluate a system, FeasPar, that learns parsing spontaneous speech. The FeasPar architecture consists of neural networks and a search. The neural networks learns the parsing task, and the search improves performance by finding the most probable and consistent feature structure. This paper focuses on the search component, and shows how the search improves overall performance considerably. N-best lists of feature structure fragments and agendas are used to speed up the search. To train and run FeasPar (Feature Structure Parser), only limited handmodeled knowledge is required. FeasPar with the search component performs better than a hand modeled LR-parser in all six comparisons that are made. FeasPar is trained, tested and evaluated in the Time Scheduling Domain, and compared with the LR-parser. The handmodeling effort for FeasPar is 2 weeks. The handmodeling effort for the LRparser was 4 months.