Diversifying Dialog Generation via Adaptive Label Smoothing

Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 3507–3520 August 1–6, 2021. ©2021 Association for Computational Linguistics 3507 Diversifying Dialog Generation via Adaptive Label Smoothing Yida Wang1,2∗, Yinhe Zheng1,3∗, Yong Jiang2, Minlie Huang1 † 1 The CoAI group, DCST, Institute for Artificial Intelligence, State Key Lab of Intelligent Technology and Systems, Beijing National Research Center for Information Science and Technology, Tsinghua University, Beijing, China 2 Tsinghua-Berkeley Shenzhen Institute, Tsinghua Shenzhen International Graduate School, Tsinghua University, Shenzhen, China 3 Samsung Research China - Beijing (SRC-B) wangyd18@mails.tsinghua.edu.cn, yh.zheng@samsung.com, jiangy@sz.tsinghua.edu.cn, aihuang@tsinghua.edu.cn Abstract Neural dialogue generation models trained with the one-hot target distribution suffer from the over-confidence issue, which leads to poor generation diversity as widely reported in the literature. Although existing approaches such as label smoothing can alleviate this issue, they fail to adapt to diverse dialog contexts. In this paper, we propose an Adaptive Label Smooth- ing (AdaLabel) approach that can adaptively estimate a target label distribution at each time step for different contexts. The maximum probability in the predicted distribution is used to modify the soft target distribution produced by a novel light-weight bi-directional decoder module. The resulting target distribution is aware of both previous and future contexts and is adjusted to avoid over-training the dialogue model. Our model can be trained in an end- to-end manner. Extensive experiments on two benchmark datasets show that our approach outperforms various competitive baselines in producing diverse responses. 1 Introduction The success of neural models has greatly advanced the research of dialog generation (Huang et al., 2020; Wang et al., 2020; Zhang et al., 2020). How- ever, most of these models suffer from a low- diversity issue where models tend to generate bland and generic responses such as I don’t know or I’m OK (Li et al., 2016). Although various approaches have been proposed to tackle this issue (Li et al., 2016; Zhao et al., 2017; Du et al., 2018; Zhou et al., 2018; Welleck et al., 2020; Zheng et al., 2020b), there are still remarkable gaps between responses generated by neural models and those from hu- mans (Holtzman et al., 2020). Further, some ex- isting methods may even harm the fluency or co- herence when improving the diversity of generated ∗ Equal contribution † Corresponding Author: aihuang@tsinghua.edu.cn So, what exactly do you do around here ? I make the robots seem more ___ Post: Response: human 0.9 1.0 0.61 0.01 0.0 0.01 bank 0.01 0.0 0.01 fights 0.01 0.0 0.10 ugly 0.01 0.0 0.08 dull 0.01 0.0 0.11 fun … Hard Target (One hot) Label Smoothing AdaLabel (Ours) Figure 1: A dialogue sampled from the OpenSubtitles dataset. We demonstrate the hard target, label smooth- ing, and Adaptive Label Smoothing approach when learning to predict the next word (“human”). responses. (Ippolito et al., 2019; Massarelli et al., 2020; Zheng et al., 2020a). Recently, Jiang and de Rijke (2018); Jiang et al. (2019) show that there is a strong connection …

Content truncated to 500 words