The MSIIP system for dialog state tracking challenge 4

Abstract

This article presents our approach for the Dialog State Tracking Challenge 4, which focuses on a dialog state tracking task on human-human dialogs. The system works in an turn-taking manner. A probabilistic enhanced frame structure is maintained to represent the dialog state during the conversation. The utterance of each turn is processed by discriminative classification models to generate a similar semantic structure to the dialog state. Then a rule-based strategy is used to update the dialog state based on the understanding results of current utterance. We also introduce a slot-based score averaging method to build an ensemble of four trackers. The DSTC4 results indicate that despite the simple feature set, the proposed method is competitive and outperforms the baseline on all evaluation metrics.