Disfluency Detection with a Semi-Markov Model and Prosodic Features

Human Language Technologies: The 2015 Annual Conference of the North American Chapter of the ACL, pages 257–262, Denver, Colorado, May 31 – June 5, 2015. c⃝2015 Association for Computational Linguistics Disfluency Detection with a Semi-Markov Model and Prosodic Features James Ferguson, Greg Durrett and Dan Klein Computer Science Division University of California, Berkeley {jferguson,gdurrett,klein}@berkeley.edu Abstract We present a discriminative model for de- tecting disfluencies in spoken language tran- scripts. Structurally, our model is a semi- Markov conditional random field with features targeting characteristics unique to speech re- pairs. This gives a significant performance improvement over standard chain-structured CRFs that have been employed in past work. We then incorporate prosodic features over silences and relative word duration into our semi-CRF model, resulting in further perfor- mance gains; moreover, these features are not easily replaced by discrete prosodic indica- tors such as ToBI breaks. Our final sys- tem, the semi-CRF with prosodic information, achieves an F-score of 85.4, which is 1.3 F1 better than the best prior reported F-score on this dataset. 1 Introduction Spoken language is fundamentally different from written language in that it contains frequent disflu- encies, or parts of an utterance that are corrected by the speaker. Removing these disfluencies is de- sirable in order to clean the input for use in down- stream NLP tasks. However, automatically identify- ing disfluencies is challenging for a number of rea- sons. First, disfluencies are a syntactic phenomenon, but defy standard context-free parsing models due to their parallel substructures (Johnson and Char- niak, 2004), causing researchers to employ other approaches such as pipelines of sequence models (Qian and Liu, 2013) or incremental syntactic sys- tems (Honnibal and Johnson, 2014). Second, hu- man processing of spoken language is complex and mixes acoustic and syntactic indicators (Cutler et al., 1997), so an automatic system must employ fea- tures targeting all levels of the perceptual stack to in the upper school upper four grades Fluent Fluent Disfluent Reparandum Repair Figure 1: Example of a disfluency where the speaker corrected upper school. Our model considers both tran- scribed text and the acoustic signal and predicts disfluen- cies as complete chunks using a semi-Markov conditional random field. achieve high performance. In spite of this, the pri- mary thread of work in the NLP community has fo- cused on identifying disfluencies based only on lex- icosyntactic cues (Heeman and Allen, 1994; Char- niak and Johnson, 2001; Snover et al., 2004; Ra- sooli and Tetreault, 2013). A separate line of work has therefore attempted to build systems that lever- age prosody as well as lexical information (Shriberg et al., 1997; Liu et al., 2003; Kim et al., 2004; Liu et al., 2006), though often with mixed success. In this work, we present a model for disfluency detection that improves upon model structures used in past work and leverages additional prosodic in- formation. Our model is a semi-Markov conditional random field that distinguishes disfluent chunks (to be deleted) from fluent chunks (everything else), as shown in Figure 1. By making …

Content truncated to 500 words