Assessing the Use of Prosody in Constituency Parsing of Imperfect Transcripts

This work explores constituency parsing on automatically recognized\ntranscripts of conversational speech. The neural parser is based on a sentence\nencoder that leverages word vectors contextualized with prosodic features,\njointly learning prosodic feature extraction with parsing. We assess the\nutility of the prosody in parsing on imperfect transcripts, i.e. transcripts\nwith automatic speech recognition (ASR) errors, by applying the parser in an\nN-best reranking framework. In experiments on Switchboard, we obtain 13-15% of\nthe oracle N-best gain relative to parsing the 1-best ASR output, with\ninsignificant impact on word recognition error rate. Prosody provides a\nsignificant part of the gain, and analyses suggest that it leads to more\ngrammatical utterances via recovering function words.\n