Joint Parsing and Disfluency Detection in Linear Time

Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 124–129, Seattle, Washington, USA, 18-21 October 2013. c⃝2013 Association for Computational Linguistics Joint Parsing and Disfluency Detection in Linear Time Mohammad Sadegh Rasooli∗ Department of Computer Science Columbia University, New York, NY rasooli@cs.columbia.edu Joel Tetreault Nuance Communications, Inc. Sunnyvale, CA joel.tetreault@nuance.com Abstract We introduce a novel method to jointly parse and detect disfluencies in spoken utterances. Our model can use arbitrary features for pars- ing sentences and adapt itself with out-of- domain data. We show that our method, based on transition-based parsing, performs at a high level of accuracy for both the parsing and disfluency detection tasks. Additionally, our method is the fastest for the joint task, running in linear time. 1 Introduction Detecting disfluencies in spontaneous speech has been widely studied by researchers in different com- munities including natural language processing (e.g. Qian and Liu (2013)), speech processing (e.g. Wang et al. (2013)) and psycholinguistics (e.g. Finlayson and Corley (2012)). While the percentage of spo- ken words which are disfluent is typically not more than ten percent (Bortfeld et al., 2001), this addi- tional “noise” makes it much harder for spoken lan- guage systems to predict the correct structure of the sentence. Disfluencies can be filled pauses (e.g. “uh”, “um”, “huh”), discourse markers (e.g. “you know”, “I mean”) or edited words which are repeated or cor- rected by the speaker. For example, in the follow- ing sentence, an edited phrase or reparandum inter- val (“to Boston”) occurs with its repair (“to Den- ver”), a filled pause (“uh”) and discourse marker (“I ∗ The first author worked on this project while he was a research intern in CoreNL research group, NLU lab, Nuance Communications, Sunnyvale, CA. mean”).1 I want a flight to Boston � �� � Reparandum �Interregnum �� � uh ���� FP I mean � �� � DM �to Denver �� � Repair Filled pauses and discourse markers are to some extent a fixed and closed set. The main challenge in finding disfluencies is the case where the edited phrase is neither a rough copy of its repair or has any repair phrase (i.e. discarded edited phrase). Hence, in previous work, researchers report their method performance on detecting edited phrases (reparan- dum) (Johnson and Charniak, 2004). In contrast to most previous work which focuses solely on either detection or on parsing, we intro- duce a novel framework for jointly parsing sentences with disfluencies. To our knowledge, our work is the first model that is based on joint dependency and disfluency detection. We show that our model is ro- bust enough to detect disfluencies with high accu- racy, while still maintaining a high level of depen- dency parsing accuracy that approaches the upper bound. Additionally, our model outperforms prior work on joint parsing and disfluency detection on the disfluency detection task, and improves upon this prior work by running in linear time complexity. The remainder of this paper is as follows. In §2, we overview some the …

Content truncated to 500 words