Better Punctuation Prediction with Dynamic Conditional Random Fields

Proceedings of the 2010 Conference on Empirical Methods in Natural Language Processing, pages 177–186, MIT, Massachusetts, USA, 9-11 October 2010. c⃝2010 Association for Computational Linguistics Better Punctuation Prediction with Dynamic Conditional Random Fields Wei Lu and Hwee Tou Ng Department of Computer Science National University of Singapore 13 Computing Drive, Singapore 117417 {luwei,nght}@comp.nus.edu.sg Abstract This paper focuses on the task of insert- ing punctuation symbols into transcribed con- versational speech texts, without relying on prosodic cues. We investigate limitations as- sociated with previous methods, and propose a novel approach based on dynamic conditional random fields. Different from previous work, our proposed approach is designed to jointly perform both sentence boundary and sentence type prediction, and punctuation prediction on speech utterances. We performed evaluations on a transcribed conversational speech domain consisting of both English and Chinese texts. Empirical re- sults show that our method outperforms an ap- proach based on linear-chain conditional ran- dom fields and other previous approaches. 1 Introduction Outputs of standard automatic speech recognition (ASR) systems typically consist of utterances where important linguistic and structural information (e.g., true case, sentence boundaries, punctuation sym- bols, etc) is not available. Such information is cru- cial in improving the readability of the transcribed speech texts, and plays an important role when fur- ther processing is required, such as in part-of-speech (POS) tagging, parsing, information extraction, and machine translation. We focus on the punctuation prediction task in this work. Most previous punctuation prediction techniques, developed mostly by the speech process- ing community, exploit both lexical and prosodic cues. However, in order to fully exploit prosodic fea- tures such as pitch and pause duration, it is necessary to have access to the original raw speech waveforms. In some scenarios where further natural language processing (NLP) tasks on the transcribed speech texts become the main concern, speech prosody in- formation may not be readily available. For exam- ple, in the recent evaluation campaign of the Inter- national Workshop on Spoken Language Translation (IWSLT) (Paul, 2009), only manually transcribed or automatically recognized speech texts are provided but the original raw speech waveforms are not avail- able. In this paper, we tackle the task of predicting punctuation symbols from a standard text processing perspective, where only the speech texts are avail- able, without relying on additional prosodic fea- tures such as pitch and pause duration. Specifi- cally, we perform the punctuation prediction task on transcribed conversational speech texts, using the IWSLT corpus (Paul, 2009) as the evaluation data. Different from many other corpora such as broad- cast news corpora, a conversational speech corpus consists of dialogs where informal and short sen- tences frequently appear. In addition, due to the nature of conversation, it also contains more ques- tion sentences compared to other corpora. An ex- ample English utterance randomly selected from the IWSLT corpus, along with its punctuated and cased version, are shown below: you are quite welcome and by the way we may get other reservations so could you please call us as soon as …

Content truncated to 500 words