PERSONACONFLICTS CORPUS
收藏资源简介:
PERSONACONFLICTS CORPUS是一个包含N = 5,772个模拟对话的数据集,这些对话涵盖了在朋友、家庭成员和恋人之间发生的各种冲突场景。该数据集由麻省理工学院、卡内基梅隆大学和艾伦人工智能研究所的研究团队创建,旨在研究关系背景对人类和模型感知对话冲突的影响。数据集中包含了自然模拟的对话,这些对话由LLMs生成,并针对对话中每个回合的沟通崩溃类型进行了细粒度的标注。该数据集为评估LLMs在检测有害沟通方面的能力提供了一个重要的框架,并为开发更个性化的AI系统以解决人际关系中的冲突提供了新的视角。
The PERSONACONFLICTS CORPUS is a dataset containing N=5,772 simulated dialogues covering a wide range of conflict scenarios between friends, family members, and romantic partners. Developed by research teams from the Massachusetts Institute of Technology, Carnegie Mellon University, and the Allen Institute for Artificial Intelligence, this dataset aims to investigate the impact of relational context on how humans and models perceive conversational conflicts. It consists of naturally simulated dialogues generated by LLMs, with fine-grained annotations for the type of communication breakdown in each dialogue turn. This dataset provides a critical framework for evaluating LLMs' ability to detect harmful communication, and offers new perspectives for developing more personalized AI systems to resolve conflicts in interpersonal relationships.

- 1Words Like Knives: Backstory-Personalized Modeling and Detection of Violent Communication麻省理工学院(美国马萨诸塞州剑桥市), 卡内基梅隆大学(美国宾夕法尼亚州匹兹堡市), 艾伦人工智能研究所(美国华盛顿州西雅图市) · 2025年



