Target-Guided Open-Domain Conversation

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 5624–5634 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 5624 Target-Guided Open-Domain Conversation Jianheng Tang1,2, Tiancheng Zhao2, Chenyan Xiong3, Xiaodan Liang1∗, Eric P. Xing2,4, Zhiting Hu2,4∗ 1Sun Yat-sen University, 2Carnegie Mellon University, 3Microsoft Research AI, 4Petuum Inc. {sqrt3tjh,xdliang328}@gmail.com, {tianchez,zhitingh}@cs.cmu.edu chenyan.xiong@microsoft.com, eric.xing@petuum.com Abstract Many real-world open-domain conversation applications have specific goals to achieve dur- ing open-ended chats, such as recommenda- tion, psychotherapy, education, etc. We study the problem of imposing conversational goals on open-domain chat agents. In particular, we want a conversational system to chat naturally with human and proactively guide the conver- sation to a designated target subject. The prob- lem is challenging as no public data is avail- able for learning such a target-guided strat- egy. We propose a structured approach that introduces coarse-grained keywords to con- trol the intended content of system responses. We then attain smooth conversation transition through turn-level supervised learning, and drive the conversation towards the target with discourse-level constraints. We further derive a keyword-augmented conversation dataset for the study. Quantitative and human evaluations show our system can produce meaningful and effective conversations, significantly improv- ing over other approaches1. 1 Introduction Creating intelligent agent that can carry out open- domain conversation with human is a long-lasting challenge. Impressive progress has been made, advancing from early rule-based systems, e.g., Eliza (Weizenbaum et al., 1966), to recent end-to- end neural conversation models that are trained on massive data (Shang et al., 2015; Li et al., 2015) and make use of background knowledge (Fang et al., 2018; Qin et al., 2019; Liu et al., 2018). However, current open-domain systems still struggle to conduct engaging conversations (Ram ∗corresponding authors 1Data and code are publicly available at https://github.com/squareRoot3/ Target-Guided-Conversation Not so good. I am really tired. Oh, I’m sorry to hear. why? I have too much work to do. What kind of work is it? I am writing a computer program. Interesting. I read about coding from a book. work book Target: e-books Hi there, how are you doing? Really? You are smart. e-books sorry I prefer e-books over paperback books. Figure 1: Target-Guided Open-Domain Conversation. The agent is given a target subject e-books which is unknown to the human. The goal is to guide the con- versation naturally to the target. Utterance keywords are highlighted in red (agent) and blue (human) and in italic. et al., 2018), and often generate inconsistent or un- controlled results. Further, many practical open- domain dialogue applications do have specific goals to achieve even though the conversations are open-ended, e.g., accomplishing nursing goals in therapeutic conversation, inspiring ideas in educa- tion, making recommendation and persuasion, and so forth. Thus, there is a strong demand to enable the integration of goals and strategy into open- domain dialogue systems, and it imposes chal- lenges to both: first, how to define the goal for an open-domain chat system; and second, how to encode dialogue strategy into the response produc- …

Content truncated to 500 words