Mimic and Rephrase: Reflective Listening in Open-Ended Dialogue

Proceedings of the 23rd Conference on Computational Natural Language Learning, pages 393–403 Hong Kong, China, November 3-4, 2019. c⃝2019 Association for Computational Linguistics 393 Mimic and Rephrase: Reflective listening in open-ended dialogue Justin Dieter Square Inc. / Stanford University justindieter@cs.stanford.edu Tian Wang Square Inc. tian@squareup.com Gabor Angeli Square Inc. gabor@squareup.com Angel X. Chang∗ Simon Fraser University angelx@sfu.ca Arun Tejasvi Chaganty∗ Square Inc. arun@squareup.com Abstract Reflective listening—demonstrating that you have heard your conversational partner—is key to effective communication. Expert hu- man communicators often mimic and rephrase their conversational partner, e.g., when re- sponding to sentimental stories or to questions they don’t know the answer to. We introduce a new task and an associated dataset wherein di- alogue agents similarly mimic and rephrase a user’s request to communicate sympathy (I’m sorry to hear that) or lack of knowledge (I do not know that). We study what makes a rephrasal response good against a set of quali- tative metrics. We then evaluate three models for generating responses: a syntax-aware rule- based system, a seq2seq LSTM neural mod- els with attention (S2SA), and the same neu- ral model augmented with a copy mechanism (S2SA+C). In a human evaluation, we find that S2SA+C and the rule-based system are compa- rable and approach human-generated response quality. In addition, experiences with a live deployment of S2SA+C in a customer support setting suggest that this generation task is a practical contribution to real world conversa- tional agents. 1 Introduction Humans in conversation naturally engage in reflec- tive (or active) listening, where they indicate they have heard and understood their partner by repeat- ing or rephrasing what they have said. This strat- egy has its roots in Rogerian psychology (Rogers, 1951) as a counseling technique meant to build trust and empathy. Dialog agents benefit from the same strategy to keep conversations pleasant, especially when the agent cannot help. Reflective listening can be formalized into two aspects: (1) mimicking and rephrasing the conver- sational partner’s utterance, and (2) incorporating ∗These authors contributed equally. Prompt: Hmm...I’m curious as to whether the swimming pool is open after 7pm?  a. I do not know that  b. I don’t know when the swimming pool is open.  c. I don’t know as to whether the swimming pool is open after 7pm.  d. I don’t know if you are curious as to whether the swimming pool is open after 7pm.  e. I don’t know Hmm...I’m curious as to whether the swimming pool is open after 7pm? Table 1: Possible responses to indicate the dialog agent doesn’t know the answer to a question (the prompt). The last two responses (d,e) are incorrect while the first three (a,b,c) are all acceptable with varying levels of specificity. The best response (b) is the one that is nei- ther too vague (a) nor too verbose and repetitive (c). an expressive speech act (Searle, 1976) appropriate for the utterance. For example, in Table 1 (b) we incorporate the speech act I don’t know on top of …

Content truncated to 500 words