Multi-view Response Selection for Human-Computer Conversation

Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 372–381, Austin, Texas, November 1-5, 2016. c⃝2016 Association for Computational Linguistics Multi-view Response Selection for Human-Computer Conversation Xiangyang Zhou1∗, Daxiang Dong1∗, Hua Wu1, Shiqi Zhao1, Dianhai Yu1,2, Hao Tian1,2, Xuan Liu1 and Rui Yan1 1Baidu Inc., Beijing, China 2School of Information Science and Technology, University of Science and Technology of China �zhouxiangyang, dongdaxiang, wu hua, zhaoshiqi, yudianhai, tianhao, liuxuan, yanrui � @baidu.com Abstract In this paper, we study the task of response selection for multi-turn human-computer con- versation. Previous approaches take word as a unit and view context and response as se- quences of words. This kind of approaches do not explicitly take each utterance as a unit, therefore it is difficult to catch utterance- level discourse information and dependencies. In this paper, we propose a multi-view re- sponse selection model that integrates infor- mation from two different views, i.e., word sequence view and utterance sequence view. We jointly model the two views via deep neu- ral networks. Experimental results on a public corpus for context-sensitive response selection demonstrate the effectiveness of the proposed multi-view model, which significantly outper- forms other single-view baselines. 1 Introduction Selecting a potential response from a set of can- didates is an important and challenging task for open-domain human-computer conversation, espe- cially for the retrieval-based human-computer con- versation. In general, a set of candidate responses from the indexed conversation corpus are retrieved, and then the best one is selected from the candidates as the system’s response (Ji et al., 2014). Previous Deep Neural Network (DNN) based ap- proaches to response selection represent context and response as two embeddings. The response is se- lected based on the similarity of these two embed- dings (Lowe et al., 2015; Kadlec et al., 2015). In ∗These two authors contributed equally these work, context and response are taken as two separate word sequences without considering the re- lationship among utterances in the context and re- sponse. The response selection in these models is largely influenced by word-level information. We called this kind of models as word sequence model in this paper. Besides word-level dependencies, utterance-level semantic and discourse information are also very important to catch the conversation top- ics to ensure coherence (Grosz and Sidner, 1986). For example an utterance can be an affirmation, negation or deduction to the previous utterances, or starts a new topic for discussion. This kind of utterance-level information is generally ignored in word sequence model, which may be helpful for se- lecting the next response. Therefore, it is necessary to take each utterance as a unit and model the context and response from the view of utterance sequence. This paper proposes a multi-view response selec- tion model, which integrates information from both word sequence view and utterance sequence view. Our assumption is that each view can represent rela- tionships between context and response from a par- ticular aspect, and features extracted from the word sequence and the utterance sequence provide com- plementary …

Content truncated to 500 words