Combining Punctuation and Disfluency Prediction: An Empirical Study

Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 121–130, October 25-29, 2014, Doha, Qatar. c⃝2014 Association for Computational Linguistics Combining Punctuation and Disfluency Prediction: An Empirical Study Xuancong Wang1,3 Khe Chai Sim2 Hwee Tou Ng1,2 1NUS Graduate School for Integrative Sciences and Engineering 2Department of Computer Science, National University of Singapore 3Human Language Technology, Institute for Infocomm Research, Singapore xuancong84@gmail.com, {simkc, nght}@comp.nus.edu.sg Abstract Punctuation prediction and disfluency pre- diction can improve downstream natural language processing tasks such as ma- chine translation and information extrac- tion. Combining the two tasks can poten- tially improve the efficiency of the over- all pipeline system and reduce error prop- agation. In this work1, we compare var- ious methods for combining punctuation prediction (PU) and disfluency prediction (DF) on the Switchboard corpus. We com- pare an isolated prediction approach with a cascade approach, a rescoring approach, and three joint model approaches. For the cascade approach, we show that the soft cascade method is better than the hard cascade method. We also use the cas- cade models to generate an n-best list, use the bi-directional cascade models to per- form rescoring, and compare that with the results of the cascade models. For the joint model approach, we compare mixed- label Linear-chain Conditional Random Field (LCRF), cross-product LCRF and 2- layer Factorial Conditional Random Field (FCRF) with soft-cascade LCRF. Our re- sults show that the various methods link- ing the two tasks are not significantly dif- ferent from one another, although they perform better than the isolated predic- tion method by 0.5–1.5% in the F1 score. Moreover, the clique order of features also shows a marked difference. 1 Introduction The raw output from automatic speech recogni- tion (ASR) systems does not have sentence bound- 1The research reported in this paper was carried out as part of the PhD thesis research of Xuancong Wang at the NUS Graduate School for Integrated Sciences and Engineering. aries or punctuation symbols. Spontaneous speech also contains a significant proportion of disflu- ency. Researchers have shown that splitting input sequences into sentences and adding in punctua- tion symbols improve machine translation (Favre et al., 2008; Lu and Ng, 2010). Moreover, dis- fluencies in speech also introduce noise in down- stream tasks like machine translation and informa- tion extraction (Wang et al., 2010). Thus, punc- tuation prediction (PU) and disfluency prediction (DF) are two important post-processing tasks for automatic speech recognition because they im- prove not only the readability of ASR output, but also the performance of downstream Natural Lan- guage Processing (NLP) tasks. The task of punctuation prediction is to insert punctuation symbols into conversational speech texts. Punctuation prediction on long, unseg- mented texts also achieves the purpose of sentence boundary prediction, because sentence boundaries are identified by sentence-end punctuation sym- bols: periods, question marks, and exclamation marks. Consider the following example, How do you feel about the Viet Nam War ? Yeah , I saw that as well . The question mark splits the sequence into …

Content truncated to 500 words