Stylized Knowledge-Grounded Dialogue Generation via Disentangled Template Rewriting

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 3304 - 3318 July 10-15, 2022 ©2022 Association for Computational Linguistics Stylized Knowledge-Grounded Dialogue Generation via Disentangled Template Rewriting Qingfeng Sun Can Xu Huang Hu Yujing Wang Jian Miao Xiubo Geng Yining Chen Fei Xu Daxin Jiang∗ Microsoft, Beijing, China {qins,caxu,huahu,yujwang,jianm, xigeng,yinichen,fexu,djiang}@microsoft.com Abstract Current Knowledge-Grounded Dialogue Gen- eration (KDG) models specialize in producing rational and factual responses. However, to es- tablish long-term relationships with users, the KDG model needs the capability to generate responses in a desired style or sentiment. Thus, we study a new problem: Stylized Knowledge- Grounded Dialogue Generation (SKDG). It presents two challenges: (1) How to train a SKDG model where no <context, knowledge, stylized response> triples are available. (2) How to cohere with context and preserve the knowledge when generating a stylized response. In this paper, we propose a novel disentan- gled template rewriting (DTR) method which generates responses via combing disentangled style templates (from monolingual stylized cor- pus) and content templates (from KDG corpus). The entire framework is end-to-end differen- tiable and learned without supervision. Exten- sive experiments on two benchmarks indicate that DTR achieves a significant improvement on all evaluation metrics compared with pre- vious state-of-the-art stylized dialogue genera- tion methods. Besides, DTR achieves compara- ble performance with the state-of-the-art KDG methods in standard KDG evaluation setting. 1 Introduction Every good conversational agent needs the ability to generate good responses, which are not only knowledgeable and coherent with contexts but also have abundant and desirable styles and sentiments (Rashkin et al., 2018; Smith et al., 2020; Zhou et al., 2020). Such an agent can deliver depth dialogues on various topics and yield more engaging and vivacious conversations to attract more users. In other words, rational and perceptual thought are all necessary for a perfect dialogue agent. Neverthe- less, most existing Knowledge-Grounded Dialogue Generation (KDG) methods (Dinan et al., 2019; ∗ Corresponding author. Harry potter is a series of fantasy novels written by J. K. Rowling . A: Harry potter is a really solid set of novels . B: Certainly. I have seen the movies yesterday. A: You are a real loyal fan ! Do you know who wrote the books ? Knowledge: Context: Reponses: KDG Model: Polite: Positive: Negative: I know the author of Harry potter is J. K. Rowling. Thank you ! The author is J. K. Rowling. Without any doubt ! J. K. Rowling is my favorite. No, I am not sure whether the she is J. K. Rowling. Human: She is my favorite author - J. K. Rowling. Figure 1: The KDG models only produce a pedantic response, which lacks emotion and attraction compared with the responses with polite style, positive and nega- tive sentiments. Kim et al., 2020; Zhao et al., 2020b) pay more attention to the former and ignore the latter. Specif- ically, let’s claim our motivation: The previous KDG works mainly focus on selecting knowledge and expressing knowledge …

Content truncated to 500 words