A Syntactic Time-Series Model for Parsing Fluent and Disfluent Speech

Proceedings of the 22nd International Conference on Computational Linguistics (Coling 2008), pages 569–576 Manchester, August 2008 A Syntactic Time-Series Model for Parsing Fluent and Disfluent Speech ∗ Tim Miller Department of Computer Science and Engineering University of Minnesota tmill@cs.umn.edu William Schuler Department of Computer Science and Engineering University of Minnesota schuler@cs.umn.edu Abstract This paper describes an incremental ap- proach to parsing transcribed spontaneous speech containing disfluencies with a Hier- archical Hidden Markov Model (HHMM). This model makes use of the right-corner transform, which has been shown to in- crease non-incremental parsing accuracy on transcribed spontaneous speech (Miller and Schuler, 2008), using trees trans- formed in this manner to train the HHMM parser. Not only do the representations used in this model align with structure in speech repairs, but as an HMM-like time- series model, it can be directly integrated into conventional speech recognition sys- tems run on continuous streams of audio. A system implementing this model is eval- uated on the standard task of parsing the Switchboard corpus, and achieves an im- provement over the standard baseline prob- abilistic CYK parser. 1 Introduction Disfluency is one obstacle preventing speech recognition systems from being able to recog- nize spontaneous speech. Perhaps the most chal- lenging aspect of disfluency recognition is the phenomenon of speech repair, which involves a speaker realizing a mistake, cutting off the flow of speech, and then continuing on, possibly re- tracing and replacing part of the utterance to that point. This paper will describe a system which ap- plies a syntactic model of speech repair to a time- ∗The authors would like to thank the anonymous review- ers for their input. This research was supported by National Science Foundation CAREER/PECASE award 0447685. The views expressed are not necessarily endorsed by the sponsors. ∗c⃝ 2008. Licensed under the Creative Commons Attribution-Noncommercial-Share Alike 3.0 Unported li- cense (http://creativecommons.org/licenses/by-nc-sa/3.0/). Some rights reserved. series parsing model, and evaluate that system on the standard Switchboard corpus parsing task. The speech repair terminology used here fol- lows that of Shriberg (1994). A speech repair con- sists of a reparandum, an interruption point, and the alteration. The reparandum contains the words that the speaker means to replace, including both words that are in error and words that will be re- traced. The interruption point is the point in time where the stream of speech is actually stopped, and the repairing of the mistake can begin. The alter- ation contains the words that are meant to replace the words in the reparandum. 2 Background Historically, research in speech repair has focused on acoustic cues such as pauses and prosodic con- tours for detecting repairs, which could then be ex- cised from the text for improved transcription. Re- cent work has also looked at the possible contribu- tion of higher-level cues, including syntactic struc- ture, in the detection of speech repair. Some of this work is inspired by Levelt’s (1983) investigation of the syntactic and semantic variables in speech re- pairs, particularly his well-formedness rule, which …

Content truncated to 500 words