Effective Use of Prosody in Parsing Conversational Speech

Proceedings of Human Language Technology Conference and Conference on Empirical Methods in Natural Language Processing (HLT/EMNLP), pages 233–240, Vancouver, October 2005. c⃝2005 Association for Computational Linguistics Effective Use of Prosody in Parsing Conversational Speech Jeremy G. Kahn† Matthew Lease∗ Eugene Charniak∗ Mark Johnson∗ Mari Ostendorf† University of Washington, SSLI† Brown University∗ {jgk,mo}@ssli.ee.washington.edu {mlease,ec,mj}@cs.brown.edu Abstract We identify a set of prosodic cues for parsing con- versational speech and show how such features can be effectively incorporated into a statistical parsing model. On the Switchboard corpus of conversa- tional speech, the system achieves improved parse accuracy over a state-of-the-art system which uses only lexical and syntactic features. Since removal of edit regions is known to improve downstream parse accuracy, we explore alternatives for edit de- tection and show that PCFGs are not competitive with more specialized techniques. 1 Introduction For more than a decade, the Penn Treebank’s Wall Street Journal corpus has served as a benchmark for developing and evaluating statistical parsing tech- niques (Collins, 2000; Charniak and Johnson, 2005). While this common benchmark has served as a valu- able shared task for focusing community effort, it has unfortunately led to the relative neglect of other genres, particularly speech. Parsed speech stands to benefit from practically every application envisioned for parsed text, including machine translation, infor- mation extraction, and language modeling. In con- trast to text, however, speech (in particular, conver- sational speech) presents a distinct set of opportu- nities and challenges. While new obstacles arise from the presence of speech repairs, the possibility of word errors, and the absence of punctuation and sentence boundaries, speech also presents a tremen- dous opportunity to leverage multi-modal input, in the form of acoustic or even visual cues. As a step in this direction, this paper identifies a set of useful prosodic features and describes how they can be effectively incorporated into a statisti- cal parsing model, ignoring for now the problem of word errors. Evaluated on the Switchboard cor- pus of conversational telephone speech (Graff and Bird, 2000), our prosody-aware parser out-performs a state-of-the-art system that uses lexical and syntac- tic features only. While we are not the first to employ prosodic cues in a statistical parsing model, previous efforts (Gregory et al., 2004; Kahn et al., 2004) in- corporated these features as word tokens and thereby suffered from the side-effect of displacing words in the n-gram models by the parser. To avoid this prob- lem, we generate a set of candidate parses using an off-the-shelf, k-best parser, and use prosodic (and other) features to rescore the candidate parses. Our system architecture combines earlier models proposed for parse reranking (Collins, 2000) and filtering out edit regions (Charniak and Johnson, 2001). Detecting and removing edits prior to parsing is motivated by the claim that probabilistic context- free grammars (PCFGs) perform poorly at detect- ing edit regions. We validate this claim empirically: two state-of-the-art PCFGs (Bikel, 2004; Charniak and Johnson, 2005) are both shown to perform sig- nificantly below a state-of-the-art edit detection sys- tem (Johnson …

Content truncated to 500 words