Dialogue state tracking is a key sub-task of dialogue management. The fourth Dialog State Tracking Challenge (DSTC-4) focuses on dialogue state tracking for real world human-to-human dialogues. The task is more challenging than previous challenges because of more complex domain and coreferences, more synonyms and abbreviations, sub-dialogue level labelled utterances, and no spoken language understanding output provided. To deal with these challenges, this paper proposes a novel hybrid dialogue state tracking method, which can take advantage of the strength of both rule-based and statistical methods. Thousands of rules are first automatically generated using a template-based rule generation approach and then combined together with several manually designed rules to yield the output of the rule-based method. In parallel, a statistical method is applied to track the state. The tracker finally takes the union of the outputs of the two methods. In DSTC-4 evaluation, the proposed hybrid tracker obtained state-of-the-art results. It ranked the second and significantly outperformed the baseline system and most submissions.
Cite as: Sun, K., Zhu, S., Chen, L., Yao, S., Wu, X., Yu, K. (2016) Hybrid Dialogue State Tracking for Real World Human-to-Human Dialogues. Proc. Interspeech 2016, 2060-2064, doi: 10.21437/Interspeech.2016-949
@inproceedings{sun16c_interspeech,
author={Kai Sun and Su Zhu and Lu Chen and Siqiu Yao and Xueyang Wu and Kai Yu},
title={{Hybrid Dialogue State Tracking for Real World Human-to-Human Dialogues}},
year=2016,
booktitle={Proc. Interspeech 2016},
pages={2060--2064},
doi={10.21437/Interspeech.2016-949}
}