The World is Not Binary: Learning to Rank with Grayscale Data for Dialogue Response Selection

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 9220–9229, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 9220 The World is Not Binary: Learning to Rank with Grayscale Data for Dialogue Response Selection Zibo Lin1,2,∗, Deng Cai3,∗, Yan Wang4,†, Xiaojiang Liu4, Hai-Tao Zheng1,2,†, Shuming Shi4 1Tsinghua Shenzhen International Graduate School, Tsinghua University 2Department of Computer Science and Technology, Tsinghua University 3The Chinese University of Hong Kong 4Tencent AI Lab {lzb18,zheng.haitao}@mails.tsinghua.edu.cn thisisjcykcd@gmail.com, xiaojiangliu84@hotmail.com {brandenwang,shumingshi}@tencent.com Abstract Response selection plays a vital role in build- ing retrieval-based conversation systems. De- spite that response selection is naturally a learning-to-rank problem, most prior works take a point-wise view and train binary clas- sifiers for this task: each response candidate is labeled either relevant (one) or irrelevant (zero). On the one hand, this formalization can be sub-optimal due to its ignorance of the di- versity of response quality. On the other hand, annotating grayscale data for learning-to-rank can be prohibitively expensive and challeng- ing. In this work, we show that grayscale data can be automatically constructed without hu- man effort. Our method employs off-the-shelf response retrieval models and response gener- ation models as automatic grayscale data gen- erators. With the constructed grayscale data, we propose multi-level ranking objectives for training, which can (1) teach a matching model to capture more fine-grained context-response relevance difference and (2) reduce the train- test discrepancy in terms of distractor strength. Our method is simple, effective, and universal. Experiments on three benchmark datasets and four state-of-the-art matching models show that the proposed approach brings significant and consistent performance improvements. 1 Introduction Building intelligent conversation systems (Shum et al., 2018; Kollar et al., 2018) is gaining more and more attention in recent years. A core module in such kind of conversation systems is response selection (Ritter et al., 2011; Hu et al., 2014; Wu et al., 2017; Tao et al., 2019): Identifying the best response from a set of possible candidates given a dialogue context, i.e., conversation history. For ∗Equal contribution. Work was done during internship at Tencent AI Lab. †Corresponding authors. Dialogue Context Between Speakers A and B Relevance A: Would you please share some useful experience for improving spoken English? B: Sure! Watching English movies helped a lot. A: Agreed. I watched Friends many times. B: Me too! I bought the DVDs and they went broken due to my frequent use. Ground Truth G: Hah! Then your English should be very good! + Distractor Response During Training R1: Why didn’t the British police come? −−− Distractor Responses in Real-world Scenario R2: It’s said that a DVD can be preserved for decades. − R3: Friends is an American television sitcom. − Table 1: Dialogue context (conversation history) be- tween Speakers A and B. R1 is a random sample used as a negative instance during training. R2 and R3 are real distractors during testing. the response selection problem, the trendy practice is to build neural matching models (Ji et al., 2014; Wang et al., 2015; Xu et …

Content truncated to 500 words