Context-Sensitive Generation of Open-Domain Conversational Responses

Proceedings of the 27th International Conference on Computational Linguistics, pages 2437–2447 Santa Fe, New Mexico, USA, August 20-26, 2018. 2437 Context-Sensitive Generation of Open-Domain Conversational Responses Wei-Nan Zhang∗, Yiming Cui†, Yifa Wang∗, Qingfu Zhu∗, Lingzhi Li∗, Lianqiang Zhou‡, Ting Liu∗ ∗Research Center for Social Computing and Information Retrieval, Harbin Institute of Technology, Harbin, China. †Joint Laboratory of HIT and iFLYTEK (HFL), iFLYTEK Research, Beijing, China ‡Joint Laboratory of HIT and Tencent Corporation, Shenzhen, China ∗{wnzhang, yfwang, qfzhu, lzli, tliu}@ir.hit.edu.cn †ymcui@iflytek.com ‡tomcatzhou@tencent.com Abstract Despite the success of existing works on single-turn conversation generation, taking the coher- ence in consideration, human conversing is actually a context-sensitive process. Inspired by the existing studies, this paper proposed the static and dynamic attention based approaches for context-sensitive generation of open-domain conversational responses. Experimental results on two public datasets show that the proposed static attention based approach outperforms all the baselines on automatic and human evaluation. 1 Introduction Until recently, training open-domain conversational systems that can imitate the way of human convers- ing is still not a well-solved problem and non-trivial task. Previous efforts focus on generating open- domain conversational responses as an unsupervised clustering process (Ritter et al., 2010), a phrase- based statistical machine translation task (Ritter et al., 2011) and a search problem based on the vector space model (Banchs and Li, 2012), etc. With the booming of deep learning, particularly the neural net- work based sequence-to-sequence models, generating open-domain conversational responses gradually turns into an end-to-end encoding and decoding process (Sutskever et al., 2014; Vinyals and Le, 2015; Shang et al., 2015; Serban et al., 2016b; Li et al., 2016a; Li et al., 2016b; Shao et al., 2017; Yao et al., 2017). Despite the success of the above research on single-turn conversational response generation, human conversations are usually coherent (Li et al., 2016c) and context-sensitive (Tian et al., 2017; X- ing et al., 2017). Table 1 illustrates how contextual information in conversations impact on the response generation. For instance, given a message1 “How should I tell my mom?”, as input, to a single-turn Conversation 1 Conversation 2 A: I got a high score on my exam. A: I failed to pass the exam. B: Oh! Great! B: That’s too bad. A: How should I tell my mom? A: How should I tell my mom? B: Go and give her a big surprise! B: Just tell her the truth and do well next time. Table 1: An example of the impact of contextual information on human conversations. “A” and “B” denote two speakers in the conversations. conversational response generation model, it should output a fixed response regardless of the content in previous utterances. However, as shown in Table 1, in the conversations2, the responses to be generated (the last utterance in Table 1) should not only dependent on the last one message (“How should I tell my mom?”), but also need to consider the longer historical utterances in the conversations. This work is licensed under a Creative Commons Attribution 4.0 International License. License details: …

Content truncated to 500 words