Prompt Conditioned VAE: Enhancing Generative Replay for Lifelong Learning in Task-Oriented Dialogue

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 11153 - 11169 December 7-11, 2022 ©2022 Association for Computational Linguistics Prompt Conditioned VAE: Enhancing Generative Replay for Lifelong Learning in Task-Oriented Dialogue Yingxiu Zhao1∗, Yinhe Zheng2†, Zhiliang Tian1, Chang Gao3, Jian Sun2, Nevin L. Zhang1 1 The Hong Kong University of Science and Technology, Hong Kong 2 Alibaba Group, China, 3 The Chinese University of Hong Kong, Hong Kong {yzhaocx,ztianac,lzhang}@connect.ust.hk, zhengyinhe1@163.com, gaochang@se.cuhk.edu.hk, jiansun_china@hotmail.com Abstract Lifelong learning (LL) is vital for advanced task-oriented dialogue (ToD) systems. To ad- dress the catastrophic forgetting issue of LL, generative replay methods are widely employed to consolidate past knowledge with generated pseudo samples. However, most existing gen- erative replay methods use only a single task- specific token to control their models. This scheme is usually not strong enough to con- strain the generative model due to insufficient information involved. In this paper, we pro- pose a novel method, prompt conditioned VAE for lifelong learning (PCLL), to enhance gen- erative replay by incorporating tasks’ statistics. PCLL captures task-specific distributions with a conditional variational autoencoder, condi- tioned on natural language prompts to guide the pseudo-sample generation. Moreover, it leverages a distillation process to further con- solidate past knowledge by alleviating the noise in pseudo samples. Experiments on natural language understanding tasks of ToD systems demonstrate that PCLL significantly outper- forms competitive baselines in building life- long learning models. We release the code and data at GitHub. 1 Introduction Task-oriented dialogue (ToD) systems are of great importance in advanced AI applications (Zhang et al., 2020b; Dai et al., 2020, 2021; He et al., 2022a,b,c). However, most existing ToD systems are developed under the assumption that the data distribution remains unchanged (Zhu et al., 2022). Unless the entire system is retrained, this setup may not be realistic when the ToD system deployed in practice needs to support new features and pro- vides more services over time based on user de- mands. Without incurring the high cost of retrain- ing, Lifelong Learning (LL) is able to acquire new ∗ Work done while the author was interning at Alibaba. † Corresponding author. knowledge continuously while preserving previ- ously learned knowledge (Delange et al., 2021). Hence, it’s crucial to equip natural language under- standing (NLU) modules, the vital components of ToD systems, with the lifelong learning ability. The main issue for lifelong learning is catas- trophic forgetting (McClelland et al., 1995; Parisi et al., 2019), which refers to the phenomenon that a model forgets previously learned tasks when learn- ing new tasks. Various approaches have been pro- posed to alleviate this issue (Schwarz et al., 2018; Aljundi et al., 2018; Rusu et al., 2016; Aljundi et al., 2017). The replay-based methods are among the most effective and widely used ones (Rebuffi et al., 2017; Shin et al., 2017; Dai et al., 2022). The main idea of replay-based methods is to re- train samples or representations from already seen tasks when learning new tasks (Mundt et al., 2020). Some …

Content truncated to 500 words