A Beam-Search Decoder for Disfluency Detection

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 1457–1467, Dublin, Ireland, August 23-29 2014. A Beam-Search Decoder for Disfluency Detection Xuancong Wang1,3 Hwee Tou Ng1,2 Khe Chai Sim2 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, {nght, simkc}@comp.nus.edu.sg Abstract In this paper1, we present a novel beam-search decoder for disfluency detection. We first pro- pose node-weighted max-margin Markov networks (M3N) to boost the performance on words belonging to specific part-of-speech (POS) classes. Next, we show the importance of measur- ing the quality of cleaned-up sentences and performing multiple passes of disfluency detection. Finally, we propose using the beam-search decoder to combine multiple discriminative models such as M3N and multiple generative models such as language models (LM) and perform multi- ple passes of disfluency detection. The decoder iteratively generates new hypotheses from current hypotheses by making incremental corrections to the current sentence based on certain patterns as well as information provided by existing models. It then rescores each hypothesis based on features of lexical correctness and fluency. Our decoder achieves an edit-word F1 score higher than all previous published scores on the same data set, both with and without using external sources of information. 1 Introduction Disfluency detection is a useful and important task in Natural Language Processing (NLP) because spon- taneous speech contains a significant proportion of disfluency. The disfluencies in speech introduce noise in downstream tasks like machine translation and information extraction. Thus, the task of disfluency detection not only can help improve the readability of automatically transcribed speech, but also the performance of downstream NLP tasks. There are mainly two types of disfluencies: filler words and edit words. Filler words include filled pauses (e.g., ‘uh’, ‘um’) and discourse markers (e.g., “I mean”, “you know”). They are insertions in spontaneous speech that indicate pauses or mark boundaries in discourse. Thus, they do not convey useful content information. Edit words are words that are spoken wrongly and then corrected by the speaker. For example, consider the utterance: I want a flight Edit � �� � to Boston �Filler �� � uh I mean �Repair �� � to Denver The phrase “to Boston” forms the edit region to be replaced by “to Denver”. The words “uh I mean” are filler words that serve to cue the listener about the error and subsequent correction. So, the cleaned-up sentence would be “I want a flight to Denver”, which is what the speaker originally intended to say. In general, edit words are more difficult to detect than filler words, and so edit word prediction accuracy is much lower. Thus, in this work, we mainly focus on edit word detection. In Section 2, we briefly introduce previous work. In Section 3, we describe our improved baseline system that will be integrated into our beam-search decoder. Section 4 presents our beam-search decoder in detail. In Section 5, we describe our experiments and results. …

Content truncated to 500 words