Multi-hop Selector Network for Multi-turn Response Selection in Retrieval-based Chatbots

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 111–120, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 111 Multi-hop Selector Network for Multi-turn Response Selection in Retrieval-based Chatbots Chunyuan Yuan1,2,†, Shangwen Lv1,2,†, Mingming Li1,2, Wei Zhou1,2,*, Fuqing Zhu2, Jizhong Han2 and Songlin Hu1,2 1School of Cyber Security, University of Chinese Academy of Sciences 2Institute of Information Engineering, Chinese Academy of Sciences {yuanchunyuan, lvshangwen, limingming, zhouwei, zhufuqing, hanjizhong, husonglin}@iie.ac.cn Abstract Multi-turn retrieval-based conversation is an important task for building intelligent dialogue systems. Existing works mainly focus on matching candidate responses with every con- text utterance on multiple levels of granularity, which ignore the side effect of using excessive context information. Context utterances pro- vide abundant information for extracting more matching features, but it also brings noise sig- nals and unnecessary information. In this paper, we will analyze the side effect of using too many context utterances and pro- pose a multi-hop selector network (MSN) to al- leviate the problem. Specifically, MSN firstly utilizes a multi-hop selector to select the rel- evant utterances as context. Then, the model matches the filtered context with the candidate response and obtains a matching score. Ex- perimental results show that MSN outperforms some state-of-the-art methods on three public multi-turn dialogue datasets. 1 Introduction Building a dialogue system that can naturally and consistently converse with humans has drawn in- creasing research interests in past years. Ex- isting works on building dialogue systems in- clude generation-based and retrieval-based meth- ods. Compared with generation-based methods, retrieval-based methods have advantages in provid- ing fluent and informative responses. Many indus- trial products have applied retrieval-based dialogue system, e.g., the E-commerce assistant AliMe As- sist from Alibaba Group (Li et al., 2017) and the XiaoIce (Shum et al., 2018) from Microsoft. Early studies (Tan et al., 2015; Yan et al., 2016; Wan et al., 2016) of retrieval-based dialogue system †Equally contributed. * Corresponding author. focus on response selection for single-turn conver- sation. Recently, researchers have begun to pay attention to the multi-turn conversation, aiming at selecting the most related response from a set of candidates given the context utterances of a con- versation. Some effective models, such as Sequen- tial Matching Network (SMN) (Wu et al., 2017), Deep Attention Matching network (DAM) (Zhou et al., 2018c), Multi-Representation Fusion Net- work (MFRN) (Tao et al., 2019), have been pro- posed to capture the matching features on multiple levels of granularity (words, phrases, sentences, etc.) and short-term and long-term dependencies among words. Previous works have shown that utilizing multi- turn utterances can further improve the matching performance than only using single-turn utterance (i.e., last utterance). But context utterance is a “double-edged sword”, it also provides a lot of noise while providing abundant information, which would influence the performance due to the sensi- tivity of these matching-based methods. Table 1: An error case of SMN (Wu et al., 2017), DAM (Zhou et al., 2018c) from E-commerce Corpus. …

Content truncated to 500 words