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

TL;DR

This paper introduces a context-aware OOD intent detection (Caro) framework to model multi-turn contexts in OOD intent detection tasks and follows the information bottleneck principle to extract robust representations from multi-turn dialogue contexts.

Abstract

Out-of-Domain (OOD) intent detection is vital for practical dialogue systems, and it usually requires considering multi-turn dialogue contexts. However, most previous OOD intent detection approaches are limited to single dialogue turns. In this paper, we introduce a context-aware OOD intent detection (Caro) framework to model multi-turn contexts in OOD intent detection tasks. Specifically, we follow the information bottleneck principle to extract robust representations from multi-turn dialogue contexts. Two different views are constructed for each input sample and the superfluous information not related to intent detection is removed using a multi-view information bottleneck loss. Moreover, we also explore utilizing unlabeled data in Caro. A two-stage training process is introduced to mine OOD samples from these unlabeled data, and these OOD samples are used to train the resulting model with a bootstrapping approach. Comprehensive experiments demonstrate that Caro establishes state-of-the-art performances on multi-turn OOD detection tasks by improving the F1-OOD score of over $29\%$ compared to the previous best method.

Figures (4)

• Figure 1: Framework of Caro. For each input sample $\boldsymbol{x}= \langle \boldsymbol{h}, \boldsymbol{u} \rangle$, two views $v_1(\boldsymbol{x})$ and $v_2(\boldsymbol{x})$ are obtained and a multi-view information bottleneck loss $\mathcal{L}_{IB}$ is optimized to learn robust representations. A two-stage training process is introduced to mine OOD samples $\mathcal{D}_O$ from unlabeled data $\mathcal{D}_U$, and optimize the cross entropy loss $\mathcal{L}_{CE}$ with $\mathcal{D}_O \cup \mathcal{D}_I$
• Figure 2: Comparing representations obtained by different objectives on the STAR-Full dataset. A lower score means that the learned representation discards more superficial information. See Appendix \ref{['append:graph']} for measurements used to produce the graph.
• Figure 3: Difference of averaged weight score at each token index for testing samples from STAR-Full.
• Figure 4: Difference of averaged aggregation weights at each dimension for testing samples in STAR-Full.