Dual Class Knowledge Propagation Network for Multi-label Few-shot Intent Detection

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 8605–8618 July 9-14, 2023 ©2023 Association for Computational Linguistics Dual Class Knowledge Propagation Network for Multi-label Few-shot Intent Detection Feng Zhang1 Wei Chen1,2,3∗ Fei Ding4,5 Tengjiao Wang1,2,3 1Key Lab of High Confidence Software Technologies (MOE), School of Computer Science, Peking University, Beijing, China 2Research Center for Computational Social Science, Peking University 3Institute of Computational Social Science, Peking University (Qingdao) 4School of Intelligence Science and Technology, Peking University 5Institute for Artificial Intelligence, Peking University {zhangfeng,dingfei}@stu.pku.edu.cn, {pekingchenwei,tjwang}@pku.edu.cn Abstract Multi-label intent detection aims to assign mul- tiple labels to utterances and attracts increasing attention as a practical task in task-oriented di- alogue systems. As dialogue domains change rapidly and new intents emerge fast, the lack of annotated data motivates multi-label few-shot intent detection. However, previous studies are confused by the identical representation of the utterance with multiple labels and overlook the intrinsic intra-class and inter-class interac- tions. To address these two limitations, we propose a novel dual class knowledge propa- gation network in this paper. In order to learn well-separated representations for utterances with multiple intents, we first introduce a label- semantic augmentation module incorporating class name information. For better considera- tion of the inherent intra-class and inter-class relations, an instance-level and a class-level graph neural network are constructed, which not only propagate label information but also propagate feature structure. And we use a sim- ple yet effective method to predict the intent count of each utterance. Extensive experimen- tal results on two multi-label intent datasets have demonstrated that our proposed method outperforms strong baselines by a large margin. 1 Introduction Multi-label intent detection aims to assign multi- ple labels to a single utterance rather than only one label since the user utterance often carries multiple different intents. As a practical task in real-world scenarios, multi-label intent detection attracts more attention with the rapid development of conversational AI (Louvan and Magnini, 2020). Previous works have achieved promising perfor- mance in multi-label intent detection, but they rely on large amounts of labeled training samples (Xu ∗ Corresponding author. 𝒙!: What'day'and'time'did'I'schedule'my'test'for?' 𝒚!={request_date, request_time} 𝒚"={request_location} 𝒙": Where'is'it'located?' 𝒙: What'time'is'my'meeting'and'at'what'location? 𝒚=? 𝑄𝑢𝑒𝑟𝑦 𝐼𝑛𝑠𝑡𝑎𝑛𝑐𝑒 (𝑡𝑒𝑠𝑡 𝑑𝑎𝑡𝑎) 𝑆𝑢𝑝𝑝𝑜𝑟𝑡 𝐸𝑥𝑎𝑚𝑝𝑙𝑒𝑠 (𝑡𝑟𝑎𝑖𝑛𝑖𝑛𝑔 𝑑𝑎𝑡𝑎) Label-semantic Augmentation Dual Class Knowledge Propagation Network Class'Embs. Instance'Embs. request_time request_location Figure 1: An example of 3-way 1-shot multi-label intent detection. and Sarikaya, 2013; Qin et al., 2020a,b, 2021). However, due to fast-emerging intents and rapidly changing domains, the lack of adequate labeled data for each intent is inevitable, thus motivating the multi-label few-shot intent detection (Hou et al., 2021). Few-shot learning, a promising way to tackle data scarcity challenges, can recognize new classes from only a few labeled samples by exploit- ing the prior knowledge from source domain. Recent works for multi-label intent detection mainly focus on threshold-based strategy which consists of relevance scoring and threshold estima- tion (Gangadharaiah and Narayanaswamy, 2019; Hou et al., 2021) to improve performance. Nev- ertheless, they suffer from the following two lim- itations. First, …

Content truncated to 500 words