Phonetically aided syntactic parsing of spoken language

The paper presents a technique for parsing a speech utterance from its phonetic representation. The technique is different from a conventional spoken language parsing techniques where a speech utterance is first transcribed at word-level and a syntactic structure is produced from the transcribed words. In a word-level parsing approach, an error caused by a speech recognizer propagates through the parser into the resultant syntactic structure. Furthermore, sometimes transcribed speech utterances are not parse-able even though lattices or confusion networks are used. These problems are addressed by the proposed phonetically aided parser. In the phonetically aided parsing approach, the parsing is performed from a phonetic representation (phone sequence) of the recognized utterance using a joint modeling of probabilistic context free grammars and a n-gram language model. The technique results in better parsing accuracy then word-level parsing when evaluated on spoken dialog parsing task in this paper.