Error recovery and sentence verification using statistical partial pattern tree for conversational speech

Chung-Hsien Wu, Yeou-Jiunn Chen, Cher-Yao Yang

Abstract: In this paper, in order to deal with the problems of disfluencies in conversational speech, partial pattern tree (PPT) and a PPT-based statistical language model are proposed. A partial pattern is defined to represent a sub-sentence with a key-phrase and some optional/functional phrases. The PPT is an integrated tree structure of the partial patterns generated from the training sentences and used to model the n-gram and grammatical constraints. In addition, a PPT merging algorithm is also proposed to reduce the number of partial patterns with similar syntactic structure by minimizing an objective cost function. Using the PPT, the undetected/misdetected errors due to disfluencies can be recovered. Finally, a sentence verification approach is proposed to re-rank the recovered sentences generated from the PPT. In order to assess the performance, a faculty name inquiry system with 2583 names has been implemented. The recognition acculacy of the system using the proposed PPT achieved 77.2%. We also contrasted this method with previous conventional approaches to show its superior performance.