Decoupling Pseudo Label Disambiguation and Representation Learning for Generalized Intent Discovery

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 9661–9675 July 9-14, 2023 ©2023 Association for Computational Linguistics Decoupling Pseudo Label Disambiguation and Representation Learning for Generalized Intent Discovery Yutao Mou1⇤, Xiaoshuai Song1⇤, Keqing He2⇤,Chen Zeng1,Pei Wang1 Jingang Wang2, Yunsen Xian2, Weiran Xu1⇤ 1Beijing University of Posts and Telecommunications, Beijing, China 2Meituan, Beijing, China {myt,songxiaoshuai,chenzeng,wangpei,xuweiran}@bupt.edu.cn {hekeqing,wangjingang,xianyunsen}@meituan.com Abstract Generalized intent discovery aims to extend a closed-set in-domain intent classifier to an open-world intent set including in-domain and out-of-domain intents. The key challenges lie in pseudo label disambiguation and represen- tation learning. Previous methods suffer from a coupling of pseudo label disambiguation and representation learning, that is, the reliability of pseudo labels relies on representation learn- ing, and representation learning is restricted by pseudo labels in turn. In this paper, we pro- pose a decoupled prototype learning framework (DPL) to decouple pseudo label disambiguation and representation learning. Specifically, we firstly introduce prototypical contrastive repre- sentation learning (PCL) to get discriminative representations. And then we adopt a prototype- based label disambiguation method (PLD) to obtain pseudo labels. We theoretically prove that PCL and PLD work in a collaborative fash- ion and facilitate pseudo label disambiguation. Experiments and analysis on three benchmark datasets show the effectiveness of our method.1 1 Introduction Intent classification (IC) is an important component of task-oriented dialogue (TOD) systems. Tradi- tional intent classification models are based on a closed-set hypothesis (Chen et al., 2019; Yang et al., 2021). That is, they rely on a pre-defined intent set provided by domain experts and can only rec- ognize limited in-domain (IND) intent categories. However, users may input out-of-domain (OOD) queries in the real open world. OOD intent detec- tion (Lin and Xu, 2019; Xu et al., 2020; Zeng et al., 2021; Wu et al., 2022a,b) identifies whether a user query falls outside the range of pre-defined IND intent set. Further, OOD intent discovery task (Lin ⇤The first three authors contribute equally. Weiran Xu is the corresponding author. 1We release our code at https://github.com/ songxiaoshuai/DPL (Labeled) IND Training Data Utterances Labels Train (Unlabeled) OOD Utterances ...... IND Intent Classifier User Utterance Model Collecting OOD Utterances After Deployed IND&OOD Joint Intent Classifier User Utterance Model GID Training IND Classes Logits Logits Extended OOD Classes Figure 1: The illustration of GID task. et al., 2020; Zhang et al., 2022; Mou et al., 2022c,a) (also known as new intent discovery) groups unla- beled OOD intents into different clusters. However, all these work cannot expand the recognition scope of the existing IND intent classifier incrementally. To solve the problem, Mou et al. (2022b) pro- poses the Generalized Intent Discovery (GID) task, which aims to simultaneously classify a set of la- beled IND intents while discovering and recogniz- ing new unlabeled OOD types incrementally. As shown in Fig 1, GID extends a closed-set IND clas- sifier to an open-world intent set including IND and OOD intents and enables the dialogue sys- tem to continuously learn from the open world. Previous …

Content truncated to 500 words