AiM: Taking Answers in Mind to Correct Chinese Cloze Tests in Educational Applications

Proceedings of the 29th International Conference on Computational Linguistics, pages 333–343 October 12–17, 2022. 333 Continual Few-shot Intent Detection Guodun Li1∗, Yuchen Zhai1, QiangLong Chen1, Xing Gao2, Ji Zhang2, Yin Zhang1† 1College of Computer Science and Technology, Zhejiang University 2DAMO Academy, Alibaba Group {guodun.li,zhaiyuchen,chenqianglong,zhangyin98}@zju.edu.cn {gaoxing.gx,zj122146}@alibaba-inc.com Abstract Intent detection is at the core of task-oriented dialogue systems. Existing intent detection sys- tems are typically trained with a large amount of data over a predefined set of intent classes. However, newly emerged intents in multiple do- mains are commonplace in the real world. And it is time-consuming and impractical for dia- logue systems to re-collect enough annotated data and re-train the model. These limitations call for an intent detection system that could continually recognize new intents with very few labeled examples. In this work, we study the Continual Few-shot Intent Detection (CFID) problem and construct a benchmark consist- ing of nine tasks with multiple domains and imbalanced classes. To address the key chal- lenges of (a) catastrophic forgetting during con- tinuous learning and (b) negative knowledge transfer across tasks, we propose the Prefix- guided Lightweight Encoder (PLE) with three auxiliary strategies, namely Pseudo Samples Replay (PSR), Teacher Knowledge Transfer (TKT) and Dynamic Weighting Replay (DWR). Extensive experiments demonstrate the effec- tiveness and efficiency of our method in pre- venting catastrophic forgetting and encourag- ing positive knowledge transfer across tasks. 1 Introduction Intent Detection (ID) is at the core of task-oriented dialogue systems. It aims at understanding the goals underlying user utterances and classifying them into different intents accurately (Zhang et al., 2020; Qin et al., 2021). Traditionally, the ID system is trained with plenty of labeled data to identify a predefined set of intent classes (Larson et al., 2019). However, newly emerged intents in multiple domains are commonplace in the real scenario. A naive approach to detecting new intents is to re- collect annotated data and re-train the model, which * Work done during internship at Alibaba. † Corresponding author: Yin Zhang. … … 𝜯𝒊"𝟏 Task Sequence 𝜯𝒊 Setting Few-shot Continual Multi-Domain #Classes FSID (Zhang et al., 2020) ✓ ✗ ✓ - CID (Liu et al., 2021) ✗ ✓ ✓ - FSCIL-ID (Xia et al., 2021) ✓ ✓ ✗ Balanced CFID (Ours) ✓ ✓ ✓ Imblanced Figure 1: Illustration of Continual Few-shot Intent De- tection (CFID). Compared with existing works, CFID aims to recognize continually new intents from multiple domains with very few labeled examples and imbal- anced number of classes across tasks. is time-consuming and impractical for dialogue systems in deployment. Thus, an intent detection system that could continually recognize new intents with very few labeled examples is called for. Many efforts have been made in existing works to achieve this goal. Zhang et al. (2020) pro- pose a discriminative nearest neighbor classifica- tion method to solve the Few-Shot Intent Detection (FSID) problem. They mainly focus on the data scarcity and ignore the ability to learn consecutive tasks. Liu et al. (2021) and Wang et al. (2021) pro- pose promising solutions …

Content truncated to 500 words