Continual Learning in Task-Oriented Dialogue Systems

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7452–7467 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 7452 Continual Learning in Task-Oriented Dialogue Systems Andrea Madotto1∗, Zhaojiang Lin1∗, Zhenpeng Zhou2, Seungwhan Moon2 Paul Crook2, Bing Liu2, Zhou Yu3, Eunjoon Cho2, Pascale Fung1, Zhiguang Wang2 1The Hong Kong University of Science and Technology 2Facebook, 3Columbia University amadotto@connect.ust.hk, zgwang@fb.coms Abstract Continual learning in task-oriented dialogue systems allows the system to add new do- mains and functionalities over time after de- ployment, without incurring the high cost of re- training the whole system each time. In this pa- per, we propose a first-ever continual learning benchmark for task-oriented dialogue systems with 37 domains to be learned continuously in both modularized and end-to-end learning settings. In addition, we implement and com- pare multiple existing continual learning base- lines, and we propose a simple yet effective ar- chitectural method based on residual adapters. We also suggest that the upper bound perfor- mance of continual learning should be equiv- alent to multitask learning when data from all domain is available at once. Our experiments demonstrate that the proposed architectural method and a simple replay-based strategy per- form better, by a large margin, compared to other continuous learning techniques, and only slightly worse than the multitask learning up- per bound while being 20X faster in learning new domains. We also report several trade-offs in terms of parameter usage, memory size and training time, which are important in the de- sign of a task-oriented dialogue system. The proposed benchmark is released to promote more research in this direction1. 1 Introduction Task-oriented dialogue systems (ToDs) are the core technology of the current state-of-the-art smart as- sistants (e.g. Alexa, Siri, Portal, etc.). These sys- tems are either modularized as a pipeline of multi- ple components, namely, natural language under- standing (NLU), dialogue state tracking (DST), dia- logue policy (DP) and natural language generation (NLG), or end-to-end, where a single model implic- itly learns how to issue APIs and system responses. ∗ Work done during internship at Facebook 1https://github.com/andreamad8/ToDCL Figure 1: In continual learning, the model is trained one dataset at the time. In this instance, the model is first trained on data from the hotel domain DHotel then on the Taxi domain DT axi and so on. The parameters of the model are updated sequentially based on the loss function L. These systems are continuously updated with new features based on the user’s needs, e.g., adding new slots and intents, or even completely new do- mains. However, existing dialogue models are trained with the assumption of having a fixed dataset at the beginning of the training, and they are not designed to add new domains and function- alities through time without incurring the high cost of retraining the whole system. The ability to ac- quire new knowledge continuously, a.k.a. continual learning (CL) (Thrun and Pratt, 2012), represents a common challenge to many production and on- device dialogue systems where there is a continual growth of …

Content truncated to 500 words