Using Weakly Supervised Learning to Improve Prosodic Event Detection

by
Darby Wong

Prosody is the term used for the intonation, emphasis, and word grouping that people use to aid in spoken language communication. We use the term .prosodic event. to describe instances where there is a significant marker of prosody, such as a prosodic phrase boundary, in speech. Prosodic phrase boundaries are particularly important for allowing a listener to parse speech into syntactic phrases which help decode sentence meaning. The ultimate goal of our work is to enable the use of these cues in automatic syntactic parsing.

Unfortunately, it is currently not possible to detect prosodic events in conversational speech with high reliability, in part because there is little data with prosodic annotation available for training models. To address this problem, we explore the usage of weakly supervised learning strategies. Several methods of weakly supervised learning are explored, including cotraining and bagging. We present an analysis of the results for each of these methods, as well as a discussion concerning the tradeoffs between one another.

Advised by Mari Ostendorf

MGH 251
Wednesday
March 10, 2004
3:30 - 4:20 pm