Abstract: Many people are now routinely building grammar-based language models for interactive spoken language applications; these language models are typically ad hoc semantic grammars which ignore many standard linguistic constraints, in particular grammatical agreement. We describe a series of experiments in which we took three CFG-based language models from non-trivial implemented systems, and in each case contrasted the performance of a version which included agreement constraints against a version which ignored them. Our findings suggest that inclusion of agreement constraints significantly improves performance in terms of both word error rate and semantic error rate.
0 Replies
Loading