N-best list rescoring using syntactic trigrams

Abstract

This paper demonstrates the usefulness of syntactic trigrams in improving the performance of a speech recognizer for the Spanish language. This technique is applied as a post-processing stage that uses syntactic information to rescore the N-best hypothesis list in order to increase the score of the most syntactically correct hypothesis. The basic idea is to build a syntactic model from training data, capturing syntactic dependencies between adjoint words in a probabilistic way, rather than resorting to the use of a rule-based system. Syntactic trigrams are used because of their power to express relevant statistics about the short-distance syntactic relationships between the words of a whole sentence. For this work we used a standarized tagging scheme known as the EAGLES tag definition, due of its ease of use and its broad coverage of all grammatical classes for Spanish. Relative improvement for the speech recognizer is 5.16%, which is statistically significant at the level of 10%, for a task of 22,398 words (HUB-4 Spanish Broadcast News).