Knowledge-Grounded Dialogue Generation with Pre-trained Language Models

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 3377–3390, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 3377 Knowledge-Grounded Dialogue Generation with Pre-trained Language Models Xueliang Zhao1,2, Wei Wu3, Can Xu4, Chongyang Tao4, Dongyan Zhao1,2, Rui Yan1,2,5∗ 1Wangxuan Institute of Computer Technology, Peking University, Beijing, China 2Center for Data Science, AAIS, Peking University, Beijing, China 3Meituan, Beijing, China 4Microsoft Corporation, Beijing, China 5Beijing Academy of Artificial Intelligence (BAAI), Beijing, China {xl.zhao,zhaody,ruiyan}@pku.edu.cn {wuwei19850318,chongyangtao}@gmail.com Abstract We study knowledge-grounded dialogue gen- eration with pre-trained language models. To leverage the redundant external knowledge un- der capacity constraint, we propose equip- ping response generation defined by a pre- trained language model with a knowledge se- lection module, and an unsupervised approach to jointly optimizing knowledge selection and response generation with unlabeled dialogues. Empirical results on two benchmarks indi- cate that our model can significantly outper- form state-of-the-art methods in both auto- matic evaluation and human judgment. 1 Introduction With advances in neural machine learning (Sutskever et al., 2014; Gehring et al., 2017; Vaswani et al., 2017) and availability of the huge amount of human conversations on social media (Adiwardana et al., 2020), building an open do- main dialogue system with data-driven approaches has attracted increasing attention from the commu- nity of artificial intelligence and natural language processing. In this work, we are interested in gen- erative approaches. Generative models for open domain dialogues are notorious for replying with generic and bland responses, resulting in mean- ingless and boring conversations (Li et al., 2015). Such deficiency is particularly severe when human participants attempt to dive into specific topics in conversation (Dinan et al., 2019). As a result, there is still a big gap between conversation with existing systems and conversation with humans. Very recently, there emerge two lines of research that seem promising to bridge the gap. One is to apply large-scale pre-trained language models, such as GPT-2 (Radford et al., 2019), to the task of open domain dialogue generation. Prototypes ∗Corresponding author: Rui Yan (ruiyan@pku.edu.cn). Context A I just discovered star trek and I really like watching star trek . B Gene Roddenberry created it based upon ...science fiction and it is American media. A If I remember Captain Kirk was not the original captain . B The Star Trek Canon of the series an ani- mated had 5 spin offs. A I watched a little of the next generation but could not get into it like i did with the original show . Response Human These adventures went on but were short lived and six feature films. DialoGPT I think it’s worth it. Table 1: An example from the test set (Test Seen) of Wizard of Wikipedia (Dinan et al., 2019) . such as DialoGPT (Zhang et al., 2019c) have ex- hibited compelling performance on generating re- sponses that make sense under conversation con- texts and at the same time carry specific content for keeping the conversation going. While the gi- ant language models can memorize enough pat- terns …

Content truncated to 500 words