Probabilistic Parse Scoring Based on Prosodic Phrasing

Probabilistic Parse Scoring Based on Prosodic Phrasing N. M. Veilleuz and M. Ostendorf Boston University 44 Cummington St. Boston, MA 02215 ABSTRACT The relative size and location of prosodic phrase boundaries provides an important cue for resolving syntactic ambiguity. In previous work, we have introduced an analysis/synthesis formalism for scoring parses in terms of the similarity be- tween prosodic patterns recognized from a given utterance and synthesized for the hypothesized parse. This paper de- scribes a new approach to the synthesis problem, as well as an alternative scoring criterion. Specifically, a decision tree is designed to predict prosodic phrase structure for a given syntactic parse, and the tree is used to compute a parse score, which now is the probability of the recognized break sequence. Unlike the rule-based synthesis algorithm used in the previous work, the decision tree can be automatically trained and can therefore be designed specifically for differ- ent speaking styles or task domains. In experiments with a corpus of ambiguous sentences spoken by FM radio announc- ers, we have achieved disambiguation performance similar to the rule-based algorithm, which is close to the performance of human subjects in perceptual experiments using the scoring algorithm with hand labeled breaks. 1. Introduction Spoken language understanding is a difficult problem, in part because of the many ambiguities inherent in natu- ral language. Syntactic ambiguity arises when a given expression can be described by more than one syntactic structure, and contributes substantially to the difficulty of the natural language processing problem. Several fac- tors may be involved in resolving such ambiguities, in- cluding semantics, discourse and bias toward a specific syntactic structure. In spoken language, prosody, or the suprasegmental information in an utterance, is an im- portant cue. Prosody is especially useful in automatic speech understanding, since computer representations of semantics and discourse are not as sophisticated as hu- man knowledge. Experiments have shown that listeners can resolve sev- eral types of syntactic ambiguities by using prosodic in- formation [3, 6]. The results of Price el al. [6] indicated that human listeners could reliably select the intended meaning of two target syntactic structures (86% cor- rect identification for six out of seven types of structural ambiguity). Of the prosodic patterns studied in that work, the relative size and location of phrase boundaries seemed to provide the principal cue for resolving ambigu- ities. Thus, it seems likely that automatically detected prosodic phrase breaks could be used in speech under- standing systems to reduce syntactic ambiguity. Assuming that prosodic cues can be detected automat- ically, there are many different ways in which prosody might be used in syntactic disambiguation for speech un- derstanding. In earlier work [9], we proposed a scoring algorithm to rank candidate parses based on an analy- sis/synthesis method. In this approach prosodic patterns recognized from a given utterance are compared to those synthesized using a hypothesized syntactic parse from a set of possible parses. Specifically the method involves: (1) automatically predicting prosodic break locations for each candidate syntactic structure (synthesis); …

Content truncated to 500 words