Nonlinear Rescoring Based on a Dialogue Context in Discourse Understanding

摘要

This paper investigates on a nonlinear rescoring approach for discourse understanding with an application to a Thai spoken dialogue system. Discourse understanding aims to find the most likely meaning of a user utterance given both the current user utterance and dialogue contextual information. While a normal approach is to rescore understanding hypotheses based on a linear combination of dialogue-context scores, this paper shows that these various scores are better combined using a nonlinear estimator. The nonlinear rescoring model is not only applied to improve understanding performance, but also to detect unreliable utterances. By adding appropriate confidence measures to the nonlinear estimator, it is capable to detect understanding-errors with no effect to the rescoring task.