A Neural Attention Model for Disfluency Detection

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 278–287, Osaka, Japan, December 11-17 2016. A Neural Attention Model for Disfluency Detection Shaolei Wang, Wanxiang Che, and Ting Liu Center for Social Computing and Information Retrieval, Harbin Institute of Technology, China {slwang, car, tliu}@ir.hit.edu.cn Abstract In this paper, we study the problem of disfluency detection using the encoder-decoder framework. We treat disfluency detection as a sequence-to-sequence problem and propose a neural attention- based model which can efficiently model the long-range dependencies between words and make the resulting sentence more likely to be grammatically correct. Our model firstly encodes the source sentence with a bidirectional Long Short-Term Memory (BI-LSTM) and then uses the neural attention as a pointer to select an ordered subsequence of the input as the output. Experi- ments show that our model achieves the state-of-the-art f-score of 86.7% on the commonly used English Switchboard test set. We also evaluate the performance of our model on the in-house annotated Chinese data and achieve a significantly higher f-score compared to the baseline of CRF-based approach. 1 Introduction Disfluency detection is the task of detecting the infelicities in spoken language transcripts. The task is important for natural language understanding, since most downstream NLU systems are built on the fluent utterances. Disfluency of a sentence can be categorized into five classes (Wu et al., 2015): uncom- pleted words, filled pauses (e.g. “uh”, “um”), editing terms (e.g. “you know”), discourse markers (e.g. “i mean”) and repairs that are discarded, or corrected by its following words. Typically, as shown in Figure 1, a repair type disfluency consists of a filled pause (“um”), a reparandum (“Boston”) and an Interreg- num(“I mean”) followed by its repair (“Denver”). The goal of the repair type disfluency detection is to detect the reparandum. The former four classes of disfluencies are easy to detect as they often consist of fixed phrases (e.g. “uh”, “you know”). However, the repair type disfluency (see Table 1 for more examples) is more difficult to detect, because reparandums are in arbitrary form. Most of the previous disfluency detection works focus on detecting the repair type disfluencies. A flight to um FP� Boston RM ��� I mean IM ��� Denver RP �� � � � Tuesday Figure 1: A sentence with disfluencies annotated in the style of (Shriberg, 1994) and the Switchboard corpus. FP=Filled Pause, RM=Reparandum, IM=Interregnum, RP=Repair. We follow previous works in evaluating the system on the accuracy with which it identifies speech-repairs, marked reparandum above. Modeling long-range dependencies between repair phrases is one of the core problems for disfluency detection. Previous sequence tagging methods (Ferguson et al., 2015; Georgila, 2009; Qian and Liu, 2013) require carefully designed features to capture the information of long distance, but usually suffer from the sparsity problem. Another line of syntax-based disfluency detection works (Honnibal and John- son, 2014; Wu et al., 2015) try to model the repair phrases on a syntax tree by compressing the unrelated * Corresponding author: Wanxiang Che This work …

Content truncated to 500 words