Out-of-Domain Intent Detection Considering Multi-turn Dialogue Contexts

Hao Lang, Yinhe Zheng, Binyuan Hui, Yongbin Li

Abstract

Out-of-Domain (OOD) intent detection is vital for practical dialogue systems, and it usually requires considering long dialogue histories. However, previous OOD intent detection approaches are limited to single-turn contexts since it is non-trivial to gather or synthesize high-quality OOD samples in multi-turn settings, and the long distance obstacle exhibited in multi-turn contexts hinders us from obtaining robust features for intent detection. In this paper, we introduce a context-aware OOD intent detection (Caro) framework that aims to consider multi-turn contexts in OOD intent detection tasks. Specifically, we follow the information bottleneck principle to extract robust representations from multi-turn dialogue contexts by eliminating superfluous information that is not related to intent detection tasks. We also propose to synthesize pseudo OOD samples with the help of unlabeled data under the constraint of dialogue contexts, i.e., candidate OOD samples are retrieved from unlabeled data based on their context similarities and representations of these candidates are mixed-up to produce pseudo OOD samples. A three stage training process is introduced in Caro to combine above approaches. Empirical results validate the superiority of our method on benchmark datasets.
BibTeX
@inproceedings{lang2024out,
  author = {Hao Lang and Yinhe Zheng and Binyuan Hui and Yongbin Li},
  title = {{Out-of-Domain Intent Detection Considering Multi-turn Dialogue Contexts}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2024},
  url = {https://openreview.net/forum?id=fUoojCBf0v},
}
Figure