Word Buffering Models for Improved Speech Repair Parsing

Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 737–745, Singapore, 6-7 August 2009. c⃝2009 ACL and AFNLP Word Buffering Models for Improved Speech Repair Parsing∗ Tim Miller University of Minnesota – Twin Cities tmill@cs.umn.edu Abstract This paper describes a time-series model for parsing transcribed speech containing disfluencies. This model differs from pre- vious parsers in its explicit modeling of a buffer of recent words, which allows it to recognize repairs more easily due to the frequent overlap in words between errors and their repairs. The parser implement- ing this model is evaluated on the stan- dard Switchboard transcribed speech pars- ing task for overall parsing accuracy and edited word detection. 1 Introduction Speech repair is a phenomenon in spontaneous speech where a speaker interrupts the flow of speech (at what’s called the interruption point), backtracks some number of words (the reparan- dum), and continues the utterance with material meant to replace the reparandum (the alteration).1 The utterance can be rendered syntactically cor- rect by excising all the words that the speaker skipped over when backtracking. Speech with re- pair is difficult for machines to process because in addition to detecting repair, a system must know what words are meant to be excised, and parsing systems must determine how to form a grammat- ical structure out of the set of words comprising both the error speech and the correct speech. Recent approaches to syntactic modeling of speech with repairs have shown that significant gains in parsing accuracy can be achieved by mod- eling the syntax of repairs (Hale et al., 2006; Core and Schubert, 1999). In addition, others have shown that a parser based on a time-series model that explicitly represents the incomplete ∗This research was supported by NSF CAREER award 0447685. The views expressed are not necessarily endorsed by the sponsors . 1This terminology follows Shriberg (1994). constituents in fluent and disfluent speech can also improve parsing accuracy (Miller and Schuler, 2008). However, these parsing approaches are still not as accurate at detecting reparanda as classifica- tion systems which use a variety of features to de- tect repairs (Charniak and Johnson, 2001; Johnson and Charniak, 2004; Heeman and Allen, 1999). One highly salient feature which classification systems use to detect repair is the repetition of words between the error and the repair. Johnson and Charniak report that 60% of words in the al- terations are copies of words in reparanda in the Switchboard corpus. Typically, this information is not available to a parser trained on context-free grammars. Meanwhile, psycholinguistic models suggest that the human language system makes use of buffers both to keep track of recent input (Bad- deley et al., 1998) and to smooth out generation (Levelt, 1989). These buffers are hypothesized to contain representations of recent phonological events, suggesting that there is a short window where new input might be compared to recent in- put. This could be represented as a buffer which predicts or detects repeated input in certain con- strained circumstances. …

Content truncated to 500 words