共情对话论文集合

目录：[TOC]
Conference Papers
Survey
论文学习笔记

还有一个更全的论文网址 https://github.com/Sahandfer/EMPaper/blob/master/README.md

Conference Papers

• EMNLP22 Improving Multi-turn Emotional Support Dialogue Generation with Lookahead Strategy Planning, 2022. [paper] [code]
• NAACL22 Mitigating Toxic Degeneration with Empathetic Data: Exploring the Relationship Between Toxicity and Empathy, 2022. [paper]
• NAACL22 EmpHi: Generating Empathetic Responses with Human-like Intents, 2022. [paper] [code]
• Empathetic Response Generation with State Management, 2022. [paper] [code]
• IJCAI22 Control Globally, Understand Locally: A Global-to-Local Hierarchical Graph Network for Emotional Support Conversation, 2022. [paper]
• ACL22 MISC: A MIxed Strategy-Aware Model Integrating COMET for Emotional Support Conversation, 2022. [paper] [code]
• AAAI22 CEM: Commonsense-aware Empathetic Response Generation, 2021. [paper] [code]
• AAAI22 Knowledge Bridging for Empathetic Dialogue Generation, 2021. [paper] [code]
• EMNLP21 Improving Empathetic Response Generation by Recognizing Emotion Cause in Conversations, 2021. [paper] [code]
• EMNLP21 Perspective-taking and Pragmatics for Generating Empathetic Responses Focused on Emotion Causes, 2021. [paper] [code]
• EMNLP21 Constructing Emotion Consensus and Utilizing Unpaired Data for Empathetic Dialogue Generation, 2021. [paper]
• Exemplars-guided Empathetic Response Generation Controlled by the Elements of Human Communication, 2021. [paper] [code]
• WWW21 Towards facilitating empathic conversations in Online Mental Health Support: A Reinforcement Learning Approach, 2021. [paper] [code]
• EMNLP20 MIME: MIMicking Emotions for Empathetic Response Generation, 2020. [paper] [code]
• COLING20 EmpDG: Multi-resolution Interactive Empathetic Dialogue Generation, 2020 [paper] [code]
• EMNLP19 MOEL: Mixture of empathetic listeners, 2019. [paper] [code]
• ACL21 Towards Emotional Support Dialog Systems, 2021. [paper] [code] - 最开始提出情感支持的论文，相当于开辟了新赛道，数据量比共情对话少了一些，但是轮次多了很多，是清华和心理所合作的新数据集。
• ACL21 PsyQA: A Chinese Dataset for Generating Long Counseling Text for Mental Health Support, 2021. [paper] [code]
• COLING20 A Taxonomy of Empathetic Response Intents in Human Social Conversations, 2020. [paper] [code]
• ACL19 Towards Empathetic Open-domain Conversation Models: a New Benchmark and Dataset, 2019. [paper] [code] - 最开始提出Empathetic的论文
• EMNLP20 A Computational Approach to Understanding Empathy Expressed in Text-Based Mental Health Support, 2020. [paper] [code] - 提出了几种共情对话的评价方式，但是目前影响力还不大。

Survey

• Empathetic Conversational Systems: A Review of Current Advances, Gaps, and Opportunities, 2022. [paper]
• ACL22 The AI Doctor Is In: A Survey of Task-Oriented Dialogue Systems for Healthcare Applications, 2022. [paper]
• The Effectiveness of Artificial Intelligence Conversational Agents in Health Care: Systematic Review, 2020. [paper]

论文学习笔记

EMNLP2022-前瞻性启发式-动态建模-Improving Multi-turn Emotional Support Dialogue Generation with Lookahead Strategy Planning

2022-10-25
title: Improving Multi-turn Emotional Support Dialogue Generation with Lookahead Strategy Planning,EMNLP 2022[paper] [code]

Authers: Yi Cheng1∗ , Wenge Liu2∗ , Wenjie Li1† , Jiashuo Wang1, Ruihui Zhao3, Bang Liu4,Xiaodan Liang5, Yefeng Zheng3

institute: 1-Hong Kong Polytechnic University；2-Baidu Inc., Beijing, China； 3-Tencent Jarvis Lab；4-RALI & Mila, Université de Montréal； 5-Sun Yat-sen University.

Topic: Emotional Support Conversation

Motivation: 提供情感支持（ES）以安抚情绪困扰的人是社交互动中的一项基本能力。现有的关于构建ES会话系统的研究大多只考虑与用户的单回合交互，这过于简化了。相比之下，多回合情感支持对话系统可以更有效地提供情感支持，但面临着一些新的技术挑战，包括：i）如何进行支持策略规划，以获得最佳支持效果；ii）如何动态地建模用户的状态。本文提出了一个名为MultiESC的新系统来解决这些问题。

Details: 应用背景/面向对象是多轮次的情感支持对话(ESC, Emotional Support Conversation)，作者想在其中构建一种前瞻性启发式并能动态建模的算法。对于策略规划，作者从A*搜索算法中汲取灵感，提出了前瞻性启发式算法，以估计使用特定策略后的未来用户反馈，这有助于选择能够产生最佳长期效果的策略。对于用户状态建模，MultiESC专注于捕捉用户的微妙情绪表达，并了解其情绪原因。

Experiments：

1. 数据集 Experimental dataset: ESConv，每个对话平均29.8个语句，共8个策略（例如，提问、反映情感和自我表露）。

2. NRC VSD词典大小20000英文单词，三个维度（pleased-displeased, excited-calm, dominant-submissive).

3. 策略=策略得分函数中的最大值。策略得分函数=基于历史的参数g(st) +启发式预估未来用户反馈的前瞻得分h(st)×权重超参。个人理解：基于历史=概率，未来预估=数学期望。

   历史：g(st) = − log Pr(st|Ht, Ut). 已知Ut是第t轮之前所有用户状态嵌入的级联，st是策略，Ht为对话历史嵌入。Pr猜测是probability , 这个函数可能需要理解A*搜索。

   未来：h(st) = E[f(st, s>t, Ut)|st, Ht, Ut] 函数f表示连续应用st和s>t以安慰先前状态为Ut的用户之后的用户反馈得分。

4. 话语解码器的目标是生成下一个对话xt，架构和策略序列生成器一样，只是输入序列不一样（在话语序列前加入一个策略嵌入）

5. 实验结果：MultiESC在所有指标（流畅、共情、识别、建议）上都由于MOEL。它在整体支持效果上也优于BlenderBotJoint，尽管在流畅性方面相对较差，作者认为原因是BlenderBot关节的主干在大规模对话语料库上经过了广泛的预训练。与同样运用了启发式的“w/o lookahead”相比可以说是不相上下，所以启发式起到了很大作用。

Comments: 本文的novelty是面向多轮对话而非单轮交互，考虑的主要是（在多轮对话中）长期减少用户的情绪困扰，并提出了一种在战略规划和对话生成方面均显著优于baseline的算法MultiESC。但是文章也提到这种算法的限制limitation是无法提供个性化对话。个人思考：本文算法MultiESC对于对象的情绪分析有很好的效果，那么是否可以通过知识图谱和用户数据库进行个性化对话，然后通过语音指纹或者密码登陆来识别用户（这个涉及隐私和安全问题），但是这种个性化方法对资源要求很高。

by Wanyue Zhang