Learning a Matching Model with Co-teaching for Multi-turn Response Selection in Retrieval-based Dialogue Systems

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3805–3815 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 3805 Learning a Matching Model with Co-teaching for Multi-turn Response Selection in Retrieval-based Dialogue Systems Jiazhan Feng1∗, Chongyang Tao1∗, Wei Wu2, Yansong Feng1, Dongyan Zhao1,3 and Rui Yan1,3† 1Institute of Computer Science and Technology, Peking University, Beijing, China 2Microsoft Corporation, Beijing, China 3Center for Data Science, Peking University, Beijing, China 1fengjiazhan@foxmail.com 2wuwei@microsoft.com 1,3{chongyangtao,fengyansong,zhaody,ruiyan}@pku.edu.cn Abstract We study learning of a matching model for response selection in retrieval-based dialogue systems. The problem is equally important with designing the architecture of a model, but is less explored in existing literature. To learn a robust matching model from noisy train- ing data, we propose a general co-teaching framework with three specific teaching strate- gies that cover both teaching with loss func- tions and teaching with data curriculum. Un- der the framework, we simultaneously learn two matching models with independent train- ing sets. In each iteration, one model transfers the knowledge learned from its training set to the other model, and at the same time receives the guide from the other model on how to over- come noise in training. Through being both a teacher and a student, the two models learn from each other and get improved together. Evaluation results on two public data sets in- dicate that the proposed learning approach can generally and significantly improve the perfor- mance of existing matching models. 1 Introduction Human-machine conversation is a long-standing goal of artificial intelligence. Recently, building a dialogue system for open domain human-machine conversation is attracting more and more atten- tion due to both availability of large-scale human conversation data and powerful models learned with neural networks. Existing methods are either retrieval-based or generation-based. Retrieval- based methods reply to a human input by select- ing a proper response from a pre-built index (Ji et al., 2014; Zhou et al., 2018b; Yan and Zhao, 2018), while generation-based methods synthe- size a response with a natural language model (Shang et al., 2015; Serban et al., 2017). In this ∗Equal Contribution. †Corresponding author: Rui Yan (ruiyan@pku.edu.cn). work, we study the problem of response selec- tion for retrieval-based dialogue systems, since retrieval-based systems are often superior to their generation-based counterparts on response fluency and diversity, are easy to evaluate, and have pow- ered some real products such as the social bot Xi- aoIce from Microsoft (Shum et al., 2018), and the E-commerce assistant AliMe Assist from Alibaba Group (Li et al., 2017). A key problem in response selection is how to measure the matching degree between a conver- sation context (a message with several turns of conversation history) and a response candidate. Existing studies have paid tremendous effort to build a matching model with neural architectures (Lowe et al., 2015; Zhou et al., 2016; Wu et al., 2017; Zhou et al., 2018b), and advanced mod- els such as the deep attention matching network (DAM) (Zhou et al., 2018b) have achieved …

Content truncated to 500 words