Amalgamating Knowledge from Two Teachers for Task-oriented Dialogue System with Adversarial Training

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 3498–3507, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 3498 Amalgamating Knowledge from Two Teachers for Task-oriented Dialogue System with Adversarial Training Wanwei He1,2, Min Yang2∗, Rui Yan3, Chengming Li2, Ying Shen4, Ruifeng Xu5 1University of Chinese Academy of Sciences, China 2Shenzhen Key Laboratory for High Performance Data Mining, Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, China 3Wangxuan Institute of Computer Technology, Peking University, China 4 School of Intelligent Systems Engineering, Sun Yat-Sen University, China 5Harbin Institute of Technology (Shenzhen), China {ww.he, min.yang, cm.li}@siat.ac.cn, ruiyan@pku.edu.cn sheny76@mail.sysu.edu.cn, xuruifeng@hit.edu.cn Abstract The challenge of both achieving task comple- tion by querying the knowledge base and gen- erating human-like responses for task-oriented dialogue systems is attracting increasing re- search attention. In this paper, we propose a “Two-Teacher One-Student” learning frame- work (TTOS) for task-oriented dialogue, with the goal of retrieving accurate KB entities and generating human-like responses simul- taneously. TTOS amalgamates knowledge from two teacher networks that together pro- vide comprehensive guidance to build a high- quality task-oriented dialogue system (student network). Each teacher network is trained via reinforcement learning with a goal-specific re- ward, which can be viewed as an expert to- wards the goal and transfers the professional characteristic to the student network. Instead of adopting the classic student-teacher learn- ing of forcing the output of a student network to exactly mimic the soft targets produced by the teacher networks, we introduce two dis- criminators as in generative adversarial net- work (GAN) to transfer knowledge from two teachers to the student. The usage of discrim- inators relaxes the rigid coupling between the student and teachers. Extensive experiments on two benchmark datasets (i.e., CamRest and In-Car Assistant) demonstrate that TTOS sig- nificantly outperforms baseline methods. For reproducibility, we release the code and data at https://github.com/siat-nlp/TTOS. 1 Introduction Task-oriented dialogue systems (TDSs), which help users to complete specific tasks with natural lan- guage, have attracted increasing attention recently due to the broad applications such as event schedul- ing and flight booking. Conventional TDSs have ∗ Min Yang is corresponding author a complex pipeline (Williams and Young, 2007), which consists of modularly connected components for natural language understanding (NLU), dia- logue state tracking (DST), and dialogue policy (DP). A limitation of such pipelined design is that errors made in upper stream modules may propa- gate to downstream components, making it hard to identify and track the source of errors. In addi- tion, these methods usually require a large number of handcrafted features and labels, which may re- strict the expressive power and learnability of the models. To ameliorate the limitations with the conven- tional pipeline TDSs, great efforts have been made in designing deep neural network-based end-to-end solutions (Bordes et al., 2017; Eric et al., 2017; Madotto et al., 2018). Recent advances are over- whelmingly contributed by sequence-to-sequence (seq2seq) models (Bordes et al., 2017; Eric and Manning, 2017; Eric et al., 2017), which have taken the state-of-the-art of TDSs to …

Content truncated to 500 words