Dialogue Distillation: Open-Domain Dialogue Augmentation Using Unpaired Data

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 3449–3460, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 3449 Dialogue Distillation: Open-Domain Dialogue Augmentation Using Unpaired Data Rongsheng Zhang1∗, Yinhe Zheng2,3∗, Jianzhi Shao4†, Xiaoxi Mao1, Yadong Xi1, Minlie Huang2‡ 1 Fuxi AI Lab, NetEase Inc., Hangzhou, China 2 Department of Computer Science and Technology, Institute for Artifical Intelligence, State Key Lab of Intelligent Technology and Systems, Beijing National Research Center for Information Science and Technology, Tsinghua University, Beijing, China. 3 Samsung Research China - Beijing (SRC-B), Beijing, China 4 Alibaba Group, Hangzhou, China zhangrongsheng@corp.netease.com, yh.zheng@samsung.com Abstract Recent advances in open-domain dialogue sys- tems rely on the success of neural models that are trained on large-scale data. However, collecting large-scale dialogue data is usually time-consuming and labor-intensive. To ad- dress this data dilemma, we propose a novel data augmentation method for training open- domain dialogue models by utilizing unpaired data. Specifically, a data-level distillation pro- cess is first proposed to construct augmented dialogues where both post and response are retrieved from the unpaired data. A ranking module is employed to filter out low-quality dialogues. Further, a model-level distillation process is employed to distill a teacher model trained on high-quality paired data to aug- mented dialogue pairs, thereby preventing dia- logue models from being affected by the noise in the augmented data. Automatic and man- ual evaluation indicates that our method can produce high-quality dialogue pairs with di- verse contents, and the proposed data-level and model-level dialogue distillation can improve the performance of competitive baselines. 1 Introduction Open-domain dialogue systems have attracted much research attention (Shum et al., 2018; Huang et al., 2020), thanks to the success of neural gener- ation models trained with large-scale data. Exist- ing research has been endeavored to address var- ious aspects in dialogue systems, such as model- ing persona (Qian et al., 2018; Zheng et al., 2019; Zhang et al., 2018), expressing emotion (Zhou et al., 2018a), or generating knowledge-grounded dialogues (Ghazvininejad et al., 2018; Zhou et al., 2018b, 2020). ∗ Equal contribution. Order determined by swapping the one in Zheng et al. (2020b) † Work performed while at Fuxi AI Lab, NetEase Inc. ‡ Corresponding Author: aihuang@tsinghua.edu.cn Today’s flight seems to be on time. You are lucky, mine is delayed. Post Response Today’s flight is not delayed. Lucky guy, but my flight is delayed. Sounds nice, but mine is delayed. Lucky for you, mine is delayed. … Unpaired data Today’s flight is not delayed. Lucky for you, mine is delayed. Anchor pair Augmented post-response pairs … Pair 1 Pair K Sounds nice, but mine is delayed. Today’s flight is not delayed. 1 2 3 4 Figure 1: Process of constructing augmented post- response pairs. The sentence in blue rectangle is used to match the anchor pair and the corresponding re- sponse is then used to retrieve similar sentences in un- paired data. Each augmented pair contains two sen- tences both from unpaired data. In general, training neural open-domain dialogue models requires …

Content truncated to 500 words