Punctuation prediction is very important for automatic speech recognition (ASR). It greatly improves the readability of transcripts and user experience, and facilitates following natural language processing tasks. In this study, we develop a model combination based approach for the recovery of punctuation for Chinese spoken sentence. Our approach models the relationships between punctuation and sentence by the different ways of sentence representation. And the relationships modeled are combined by multi-layer perception to predict punctuation (period, question mark, and exclamation mark). Different from previous studies, our proposed approach is designed to use global lexical information, not only local information. Results indicate that, compared with the baseline, our proposed method results in an absolute improvement of 10.0 % unweighted accuracy and 4.9 % weighted accuracy, respectively. Our approach finally achieves an unweighted accuracy of 86.9 % and a weighted accuracy of 92.4 %.
展开▼