Discriminative learning using linguistic features to rescore n-best speech hypotheses

We describe how we were able to improve the accuracy of a medium-vocabulary spoken dialog system by rescoring the list of n-best recognition hypotheses using a combination of acoustic, syntactic, semantic and discourse information. The non-acoustic features are extracted from different intermediate processing results produced by the natural language processing module, and automatically filtered. We apply discriminative support vector learning designed for re-ranking, using both word error rate and semantic error rate as ranking target value, and evaluating using five-fold cross-validation; to show robustness of our method, confidence intervals for word and semantic error rates are computed via bootstrap sampling. The reduction in semantic error rate, from 19% to 11%, is statistically significant at 0.01 level.