Adaptive Bridge between Training and Inference for Dialogue Generation

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 2541–2550 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 2541 Adaptive Bridge between Training and Inference for Dialogue Generation Haoran Xu1, 2∗, Hainan Zhang3†, Yanyan Zou3, Hongshen Chen3, Zhuoye Ding3, Yanyan Lan4 1Institute of Computing Technology, Chinese Academy of Sciences, Beijing, China 2 University of Chinese Academy of Sciences, Beijing, China 3Data Science Lab, JD.com, Beijing, China 4Institute of AI Industry Research, Tsinghua University, Beijing, China xuhaoran18s@ict.ac.cn,zhanghainan1990@163.com,zouyanyan6@jd.com ac@chenhongshen.com,dingzhuoye@jd.com,lanyanyan@tsinghua.edu.cn Abstract Although exposure bias has been widely stud- ied in some NLP tasks, it faces its unique chal- lenges in dialogue response generation, the representative one-to-various generation sce- nario. In real human dialogue, there are many appropriate responses for the same context, not only with different expressions, but also with different topics. Therefore, due to the much bigger gap between various ground-truth re- sponses and the generated synthetic response, exposure bias is more challenging in dialogue generation task. What’s more, as MLE en- courages the model to only learn the common words among different ground-truth responses, but ignores the interesting and specific parts, exposure bias may further lead to the common response generation problem, such as “I don’t know” and “HaHa?” In this paper, we pro- pose a novel adaptive switching mechanism, which learns to automatically transit between ground-truth learning and generated learning regarding the word-level matching score, such as the cosine similarity. Experimental results on both Chinese STC dataset and English Red- dit dataset, show that our adaptive method achieves a significant improvement in terms of metric-based evaluation and human evaluation, as compared with the state-of-the-art exposure bias approaches. Further analysis on NMT task also shows that our model can achieve a significant improvement. 1 Introduction Auto-regressive models(ARM) are widely used for natural language generation(NLG) tasks, such as machine translation (Sutskever et al., 2014; Wu et al., 2018), dialogue response generation (Li et al., 2017), image captioning (Lin et al., 2014; Vinyals et al., 2015) and video description (Don- ahue et al., 2015). They utilize the encoder-decoder framework to predict the next token conditioned ∗Work done at Data Science Lab, JD.com. †Corresponding author. Dialogue 1 context 听说广州已成避暑胜地 ( I heard that Guangzhou has become a summer resort) response1 确实，这边很凉快。( Indeed, it’s cool here. ) response2 晚上睡觉都没开风扇了。 ( There is no need to turn on the fan at night. ) Dialogue 2 context 哈哈，看看可爱的小猫咪( Ha ha, look at this lovely kitten ) response1 这是什么品种的猫哇？好可爱，我也想要 ( What kind of cat is this? So cute, I want it too. ) response2 好想要一只这样的猫，可以陪我儿子玩 ( I really want a cat like this to play with my son. ) response3 哇，好可爱哇，我屋也有一只这样的小猫 ( Wow, It’s so cute, I have a kitten like this in my house. ) Table 1: The two Dialogues in STC dataset, and the red part of responses in Dialogue 2 are the common words. on the previous tokens, and minimize the cross- entropy between the generation and ground-truths as their objective function. Specifically, at training time, the ground-truth is utilized as …

Content truncated to 500 words