TR-IT-0334 :February 14,2000

Ruiqiang Zhang, Ezra Black, Andrew Finch, Yoshinori Sagisaka

Using Detailed Contextual Information To Build Language Models Of Part-Of-Speech Tagging And Language Models Of Speech Recognition By The Maximum Entropy Approach

Abstract:This report is about us the latest results of part-of-speech tagging and language modeling of speech recognition. Detailed information including local N-gram, long distance constraints and information from sentence structure provided by ATR Parser are integrated in language models by maximum entropy approach. The experimental results prove our models are effective to improve pos tagging accuracy and reduce word error rate of ATR speech recognition system — ATRSPREC.