Hierarchical Inductive Transfer for Continual Dialogue Learning

Findings of the Association for Computational Linguistics: ACL 2022, pages 693 - 699 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Hierarchical Inductive Transfer for Continual Dialogue Learning Shaoxiong Feng1,2 Xuancheng Ren3 Kan Li1 Xu Sun3,4 1Beijing Institute of Technology 2University of Technology Sydney 3MOE Key Laboratory of Computational Linguistics, School of CS, Peking University 4Beijing Academy of Artificial Intelligence {shaoxiongfeng, likan}@bit.edu.cn {renxc, xusun}@pku.edu.cn Abstract Pre-trained models have achieved excellent per- formance on the dialogue task. However, for the continual increase of online chit-chat sce- narios, directly fine-tuning these models for each of the new tasks not only explodes the capacity of the dialogue system on the embed- ded devices but also causes knowledge forget- ting on pre-trained models and knowledge inter- ference among diverse dialogue tasks. In this work, we propose a hierarchical inductive trans- fer framework to learn and deploy the dialogue skills continually and efficiently. First, we intro- duce the adapter module into pre-trained mod- els for learning new dialogue tasks. As the only trainable module, it is beneficial for the dialogue system on the embedded devices to ac- quire new dialogue skills with negligible addi- tional parameters. Then, for alleviating knowl- edge interference between tasks yet benefiting the regularization between them, we further design hierarchical inductive transfer that en- ables new tasks to use general knowledge in the base adapter without being misled by diverse knowledge in task-specific adapters. Empirical evaluation and analysis indicate that our frame- work obtains comparable performance under deployment-friendly model capacity. 1 Introduction Neural dialogue models (Shang et al., 2015; Ser- ban et al., 2016; Li et al., 2016) have drawn in- creasing attention due to their high commercial value. Previous work usually makes efforts to im- prove the diversity and coherence of the responses (Serban et al., 2017; Zhang et al., 2018a,c; Feng et al., 2020; Sun et al., 2021). However, the ap- plication of neural dialogue models also requires advanced conversation skills, and recently, a lot of work tries to enable models to express empa- thy (Zhou et al., 2018; Rashkin et al., 2019), be knowledgeable (Ghazvininejad et al., 2018; Dinan et al., 2019), and demonstrate consistent personali- ties (Qian et al., 2018; Zhang et al., 2018b, 2019). Specifically, the dialogue model is trained on a task- specific dataset to learn the corresponding conver- sation skill. However, with the increasing number of online chit-chat scenarios, the dialogue system is further expected to continually specialize in new tasks without sacrificing the performance on old tasks. Meanwhile, the dialogue system must keep its capacity as small as possible for the deployment on the computation resource–limited embedded de- vices. Pre-trained models (Radford et al., 2018; Devlin et al., 2019) have successfully facilitated the learn- ing of the downstream tasks in various fields. To address the challenge of continual dialogue learn- ing, directly fine-tuning pre-trained models on each of the new dialogue tasks is a straightforward way to equip the dialogue system with new conversa- tion skills continually. However, it explodes the capacity of …

Content truncated to 500 words