KNN-Contrastive Learning for Out-of-Domain Intent Classification

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 5129 - 5141 May 22-27, 2022 c⃝2022 Association for Computational Linguistics KNN-Contrastive Learning for Out-of-Domain Intent Classification Yunhua Zhou1,3, Peiju Liu1,3, Xipeng Qiu1,2,∗ 1School of Computer Science, Fudan University 2Peng Cheng Laboratory 3Shanghai Collaborative Innovation Center of Intelligent Visual Computing {zhouyh20,xpqiu}@fudan.edu.cn pjliu21@m.fudan.edu.cn Abstract The Out-of-Domain (OOD) intent classifica- tion is a basic and challenging task for dia- logue systems. Previous methods commonly re- strict the region (in feature space) of In-domain (IND) intent features to be compact or simply- connected implicitly, which assumes no OOD intents reside, to learn discriminative seman- tic features. Then the distribution of the IND intent features is often assumed to obey a hy- pothetical distribution (Gaussian mostly) and samples outside this distribution are regarded as OOD samples. In this paper, we start from the nature of OOD intent classification and ex- plore its optimization objective. We further propose a simple yet effective method, named KNN-contrastive learning. Our approach uti- lizes K-Nearest Neighbors (KNN) of IND in- tents to learn discriminative semantic features that are more conducive to OOD detection. No- tably, the density-based novelty detection al- gorithm is so well-grounded in the essence of our method that it is reasonable to use it as the OOD detection algorithm without making any requirements for the feature distribution. Extensive experiments on four public datasets show that our approach can not only enhance the OOD detection performance substantially but also improve the IND intent classification while requiring no restrictions on feature distri- bution. Code is available.1 1 Introduction People are getting accustomed to talking with or sending some instructions to task-oriented dialog system in natural language to assist them in fulfill- ing work. As the environment facing the dialogue systems becomes more open, there are more ut- terances with unknown or Out-of-Domain (OOD) intents that the dialog system does not know how to handle. As shown in Figure 1, the chatbot encoun- ters an unsupported intent (OOD intent) utterance ∗Corresponding author. 1https://github.com/zyh190507/KnnContrastiveForOOD. Schedule me a table at Hilton Hotel OK, what is the specific time? Book a flight from London to Paris I found the following flight information for you. When will the COVID-19 pandemic end ? Book_hotel Book_flight Open Figure 1: An example about interactions between a user (green) and a chatbot (grey). The chatbot correctly understands the intents (In-domain intents) of the user in the first two rounds. Thirdly, the chatbot encounters Out-of-Domain intent and does not know how to do it. in the third interaction. It is significant to distin- guish these utterances for the dialogue system be- cause identifying the intents of the user determines whether subsequent actions can be carried out cor- rectly. To solve this problem, the existing methods roughly can be summarized into two categories according to whether extensive labeled OOD intent samples are used during training. The first kind of method (use OOD samples during training) is rep- resented by (Zheng et al., …

Content truncated to 500 words