Dialogue Response Selection with Hierarchical Curriculum Learning

Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 1740–1751 August 1–6, 2021. ©2021 Association for Computational Linguistics 1740 Dialogue Response Selection with Hierarchical Curriculum Learning Yixuan Su♠,∗ Deng Cai♥,∗ Qingyu Zhou♦ Zibo Lin♦ Simon Baker♠ Yunbo Cao♦ Shuming Shi♦ Nigel Collier♠ Yan Wang♦ ♠Language Technology Lab, University of Cambridge ♥The Chinese University of Hong Kong ♦Tencent Inc. {ys484,sb895,nhc30}@cam.ac.uk thisisjcykcd@gmail.com {qingyuzhou,jimblin,yunbocao,shumingshi,brandenwang}@tencent.com Abstract We study the learning of a matching model for dialogue response selection. Motivated by the recent finding that models trained with random negative samples are not ideal in real-world scenarios, we propose a hierarchical curricu- lum learning framework that trains the match- ing model in an “easy-to-difficult” scheme. Our learning framework consists of two com- plementary curricula: (1) corpus-level curricu- lum (CC); and (2) instance-level curriculum (IC). In CC, the model gradually increases its ability in finding the matching clues be- tween the dialogue context and a response can- didate. As for IC, it progressively strength- ens the model’s ability in identifying the mis- matching information between the dialogue context and a response candidate. Empirical studies on three benchmark datasets with three state-of-the-art matching models demonstrate that the proposed learning framework signifi- cantly improves the model performance across various evaluation metrics. 1 Introduction Building intelligent conversation systems is a long- standing goal of artificial intelligence and has at- tracted much attention in recent years (Shum et al., 2018; Kollar et al., 2018). An important challenge for building such conversation systems is the re- sponse selection problem, that is, selecting the best response to a given dialogue context from a set of candidate responses (Ritter et al., 2011). To tackle this problem, different matching mod- els are developed to measure the matching degree between a dialogue context and a response candi- date (Wu et al., 2017; Zhou et al., 2018; Lu et al., 2019; Gu et al., 2019). Despite their differences, ∗The main body of this work was done during internship at Tencent Inc. The first two authors contributed equally. Yan Wang is the corresponding author. Dialogue Context Between Two Speakers A and B A: Would you please recommend me a good TV series to watch during my spare time? B: Absolutely! Which kind of TV series are you most interested in? A: My favorite type is fantasy drama. B: I think both Game of Thrones and The Vampire Diaries are good choices. Positive Response P1: Awesome, I believe both of them are great TV series! I will first watch Game of Thrones. (Easy) P2: Cool! I think I find the perfect thing to kill my weekends. (Difficult) Negative Response N1: This restaurant is very expensive. (Easy) N2: Iain Glen played Ser Jorah Mormont in the HBO fantasy series Game of Thrones. (Difficult) Table 1: An example dialogue context between speak- ers A and B, where P1 and P2 are easy and difficult positives; N1 and N2 are easy and difficult negatives. most prior works …

Content truncated to 500 words