The goal of the present study is to use prosodic information to improve automatic syntactic parsing of conversational speech in the Switchboard Corpus. To achieve this, an ensemble classifier, based on a Recurrent Neural Network, is developed to predict the parse with the highest Unlabelled Attachment Score (UAS) from the outputs of multiple dependency parsers, based on syntactic and prosodic features. The main syntactic features proposed, which we refer to as “dependency configurations,” represent the relative dependency location of each of a pair of consecutive words. Empirical analysis indicates that configurations with a direct dependency between consecutive words are less likely to be associated with major prosodic breaks. Using syntactic features alone, the system achieved an improvement of 1.1% of UAS on the test set, above the best parser in the ensemble, while using syntactic features combined with prosodic features (pauses and normalized duration) led to a further improvement of 0.4%. Both empirical analysis of dependency configurations and parsing improvement suggest a relationship between prosody and direct dependency relationships between consecutive words.
Cite as: Ghaly, H., Mandel, M. (2020) Using Prosody to Improve Dependency Parsing. Proc. Speech Prosody 2020, 1014-1018, doi: 10.21437/SpeechProsody.2020-207
@inproceedings{ghaly20_speechprosody,
author={Hussein Ghaly and Michael Mandel},
title={{Using Prosody to Improve Dependency Parsing}},
year=2020,
booktitle={Proc. Speech Prosody 2020},
pages={1014--1018},
doi={10.21437/SpeechProsody.2020-207}
}