Appropriately Handled Prosodic Breaks Help PCFG Parsing

Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the ACL, pages 37–45, Los Angeles, California, June 2010. c⃝2010 Association for Computational Linguistics Appropriately Handled Prosodic Breaks Help PCFG Parsing Zhongqiang Huang1, Mary Harper1,2 1Laboratory for Computational Linguistics and Information Processing Institute for Advanced Computer Studies University of Maryland, College Park, MD USA 2Human Language Technology Center of Excellence Johns Hopkins University, Baltimore, MD USA {zqhuang,mharper}@umiacs.umd.edu Abstract This paper investigates using prosodic infor- mation in the form of ToBI break indexes for parsing spontaneous speech. We revisit two previously studied approaches, one that hurt parsing performance and one that achieved minor improvements, and propose a new method that aims to better integrate prosodic breaks into parsing. Although these ap- proaches can improve the performance of ba- sic probabilistic context free grammar (PCFG) parsers, they all fail to produce fine-grained PCFG models with latent annotations (PCFG- LA) (Matsuzaki et al., 2005; Petrov and Klein, 2007) that perform significantly better than the baseline PCFG-LA model that does not use break indexes, partially due to mis-alignments between automatic prosodic breaks and true phrase boundaries. We propose two alterna- tive ways to restrict the search space of the prosodically enriched parser models to the n- best parses from the baseline PCFG-LA parser to avoid egregious parses caused by incor- rect breaks. Our experiments show that all of the prosodically enriched parser models can then achieve significant improvement over the baseline PCFG-LA parser. 1 Introduction Speech conveys more than a sequence of words to a listener. An important additional type of informa- tion that phoneticians investigate is called prosody, which includes phenomena such as pauses, pitch, energy, duration, grouping, and emphasis. For a review of the role of prosody in processing spo- ken language, see (Cutler et al., 1997). Prosody can help with the disambiguation of lexical meaning (via accents and tones) and sentence type (e.g., yes- no question versus statement), provide discourse- level information like focus, prominence, and dis- course segment, and help a listener to discern a speaker’s emotion or hesitancy, etc. Prosody often draws a listener’s attention to important information through contrastive pitch or duration patterns associ- ated words or phrases. In addition, prosodic cues can help one to segment speech into chunks that are hy- pothesized to have a hierarchical structure, although not necessarily identical to that of syntax. This sug- gests that prosodic cues may help in the parsing of speech inputs, the topic of this paper. Prosodic information such as pause length, du- ration of words and phones, pitch contours, en- ergy contours, and their normalized values have been used in speech processing tasks like sentence boundary detection (Liu et al., 2005). In contrast, other researchers use linguistic encoding schemes like ToBI (Silverman et al., 1992), which encodes tones, the degree of juncture between words, and prominence symbolically. For example, a simplified ToBI encoding scheme uses the symbol 4 for ma- jor intonational breaks, p for hesitation, and 1 for all other breaks (Dreyer and …

Content truncated to 500 words