StyleDGPT: Stylized Response Generation with Pre-trained Language Models

Findings of the Association for Computational Linguistics: EMNLP 2020, pages 1548–1559 November 16 - 20, 2020. c⃝2020 Association for Computational Linguistics 1548 STYLEDGPT: Stylized Response Generation with Pre-trained Language Models Ze Yang1, Wei Wu2, Can Xu3, Xinnian Liang1, Jiaqi Bai1, Liran Wang1, Wei Wang4, and Zhoujun Li1∗ 1State Key Lab of Software Development Environment, Beihang University, Beijing, China 2Meituan, Beijing, China 3Microsoft, Beijing, China 4China Resources Group {tobey,xnliang,bjq,wanglr,lizj}@buaa.edu.cn {wuwei19850318,ww.cs.tj}@gmail.com caxu@microsoft.com Abstract Generating responses following a desired style has great potentials to extend applications of open-domain dialogue systems, yet is re- frained by lacking of parallel data for train- ing. In this work, we explore the challeng- ing task with pre-trained language models that have brought breakthrough to various natural language tasks. To this end, we introduce a KL loss and a style classifier to the fine-tuning step in order to steer response generation to- wards the target style in both a word-level and a sentence-level. Comprehensive empirical studies with two public datasets indicate that our model can significantly outperform state- of-the-art methods in terms of both style con- sistency and contextual coherence. 1 Introduction With advances in neural machine learning (Sutskever et al., 2014; Gehring et al., 2017; Vaswani et al., 2017) and availability of huge amount of human conversations on social media, there has been significant progress on building open-domain dialogue systems with natural lan- guage generation techniques. Though neural gen- erative models are notorious for replying with bland responses (Li et al., 2015), some very recent work demonstrates that response generation models learned with pre-training techniques (Radford et al., 2019) can effectively overcome the deficiency suf- fered by previous models and are capable of having smooth conversations with humans through reason- able and specific replies (Wolf et al., 2019; Zhang et al., 2019b). The compelling performance exhibited by the pre-trained dialogue models encourages us to ex- plore more difficult yet important problems in con- versational AI. In this work, we study stylized re- sponse generation, that is responses provided by a ∗ Corresponding Author model should not only be coherent with the con- versation contexts, but also be consistent with a designated style. Such research could facilitate developers to customize their dialogue systems in terms of response styles, and thus broaden appli- cations of the systems, from a social companion (Shum et al., 2018) or a virtual assistant (Ram et al., 2018) to a variety of vertical scenarios such as customer service (requiring a polite style), virtual characters in games (requiring specific personas), assistants in specific domains (requiring domain knowledge), etc. Normally, a target style is speci- fied by a non-conversational corpus (e.g., novels, news, blogs, etc.) apart from the paired dialogue corpus (Luan et al., 2017; Niu and Bansal, 2018; Gao et al., 2019). Thus, the major challenge of the task lies in the scarcity of paired data for learning the correspondence between conversation contexts and proper responses in the desired style, which is a key factor in success of the neural dialogue mod- els developed so …

Content truncated to 500 words