Appropriately Handled Prosodic Breaks Help PCFG Parsing

This paper investigates using prosodic information 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 approaches can improve the performance of basic probabilistic context free grammar (PCFG) parsers, they all fail to produce fine-grained PCFG models with latent annotations (PCFGLA) (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 alternative ways to restrict the search space of the prosodically enriched parser models to the nbest parses from the baseline PCFG-LA parser to avoid egregious parses caused by incorrect breaks. Our experiments show that all of the prosodically enriched parser models can then achieve significant improvement over the baseline PCFG-LA parser.