Improved Syntactic Models for Parsing Speech with Repairs

Human Language Technologies: The 2009 Annual Conference of the North American Chapter of the ACL, pages 656–664, Boulder, Colorado, June 2009. c⃝2009 Association for Computational Linguistics Improved Syntactic Models for Parsing Speech with Repairs∗ Tim Miller Department of Computer Science and Engineering University of Minnesota, Twin Cities tmill@cs.umn.edu Abstract This paper introduces three new syntactic models for representing speech with repairs. These models are developed to test the intu- ition that the erroneous parts of speech repairs (reparanda) are not generated or recognized as such while occurring, but only after they have been corrected. Thus, they are designed to minimize the differences in grammar rule ap- plications between fluent and disfluent speech containing similar structure. The three models considered in this paper are also designed to isolate the mechanism of impact, by systemat- ically exploring different variables. 1 Introduction Recent work in recognition of speech with repairs has shown that syntactic cues to speech repair can improve both overall parsing accuracy and detection of repaired sections (Hale et al., 2006; Miller and Schuler, 2008; Johnson and Charniak, 2004). These techniques work by explictly modeling the structure of speech repair, specifically the tendency of repairs to follow unfinished constituents of the same cate- gory. This is the essence of what was termed the well-formedness rule by Willem Levelt (1983) in his psycholinguistic studies of repair. The work presented here uses the same motiva- tions as those cited above (to be described in more detail below), in that it attempts to model the syn- tactic structure relating unfinished erroneous con- ∗This research was supported by NSF CAREER award 0447685. The views expressed are not necessarily endorsed by the sponsors. stituents to the repair of those constituents. How- ever, this work attempts to improve on those mod- els by focusing on the generative process used by a speaker in creating the repair. This is done first by eschewing any labels representing the presence of an erroneous constituent while processing the text. This modeling representation reflects the intuition that speakers do not intend to generate erroneous speech – they intend their speech to be fluent, or a correction to an error, and can stop very quickly when an error is noticed. This corresponds to Lev- elt’s Main Interruption Rule, which states that a speaker will “Stop the flow of speech immediately upon detecting the occasion of repair.” Rather than attempting to recognize a special syntactic category called EDITED during the processing phase, this work introduces the REPAIRED category to signal the ending of a repaired section only. The second part of the modeling framework is the use of a right-corner transform on training data, which converts phrase-structure trees into heavily left-branching structures. This transformation has been shown to represent the structure of unfinished constituents like those seen in speech repair in a nat- ural way, leading to improved detection of speech repair (Miller and Schuler, 2008). Combining these two modeling techniques in a bottom-up parsing framework results in a parsing architecture that is a …

Content truncated to 500 words