DiffusEmp: A Diffusion Model-Based Framework with Multi-Grained Control for Empathetic Response Generation

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 2812–2831 July 9-14, 2023 ©2023 Association for Computational Linguistics DIFFUSEMP: A Diffusion Model-Based Framework with Multi-Grained Control for Empathetic Response Generation Guanqun Bi1,2, Lei Shen3, Yanan Cao1,2∗, Meng Chen3∗, Yuqiang Xie1,2, Zheng Lin1,2, Xiaodong He3 1Institute of Information Engineering, Chinese Academy of Sciences, Beijing, China 2School of Cyber Security, University of Chinese Academy of Sciences, Beijing, China 3JD AI Research, Beijing, China {biguanqun,caoyanan,xieyuqiang,linzheng}@iie.ac.cn {shenlei20,chenmeng20,xiaodong.he}@jd.com Abstract Empathy is a crucial factor in open-domain con- versations, which naturally shows one’s caring and understanding to others. Though several methods have been proposed to generate em- pathetic responses, existing works often lead to monotonous empathy that refers to generic and safe expressions. In this paper, we propose to use explicit control to guide the empathy ex- pression and design a framework DIFFUSEMP based on conditional diffusion language model to unify the utilization of dialogue context and attribute-oriented control signals. Specifically, communication mechanism, intent, and seman- tic frame are imported as multi-grained sig- nals that control the empathy realization from coarse to fine levels. We then design a spe- cific masking strategy to reflect the relationship between multi-grained signals and response to- kens, and integrate it into the diffusion model to influence the generative process. Experi- mental results on a benchmark dataset EMPA- THETICDIALOGUE show that our framework outperforms competitive baselines in terms of controllability, informativeness, and diversity without the loss of context-relatedness. 1 Introduction Empathetic response generation, as a conditional text generation task, aims to endow agents with the ability to understand interlocutors and accurately express empathy in their communication (Rashkin et al., 2019; Lin et al., 2019; Li et al., 2020; Shen et al., 2021). However, the generated responses tend to be generic and monotonous (Chen et al., 2022), i.e., showing shallow empathy and few con- nections to the context. As shown in the upper part of Figure 1, “I’m sorry to hear that.” is used as a reaction to different contexts with negative feelings. To alleviate the problem, existing works mainly in- corporate emotion or knowledge modules into the encoder-decoder framework and train their models ∗Corresponding authors. We all get nervous on our first day of work. Try to look at new things as adventures. You will do fine! | | Acknowledging | IT | Suggesting IT DiffusEMP | Interpretation | I was really nervous when starting a new job. Context 3 I once lost my job and got mad. Context 1 I did not get a single gift during Christmas. Context 2 angry I’m sorry to hear that. sad nervous I was really nervous when starting a new job. Context CM Intentionally_act Desirability Pronoun Transition_to_state Experiencer_focus Work Calendric_unit Spatial_contact Age Perception_active Encouraging IT Pronoun SF Figure 1: A monotonous empathetic response (upper) and an informative empathetic response (lower). “CM”, “IT”, and “SF” are abbreviations for “Communication Mechanism”, “Intent”, and “Semantic Frame”, which represent control signals at the utterance, sentence, and token level respectively. with …

Content truncated to 500 words