CATCH-ME
收藏资源简介:
CATCH-ME是由布鲁诺·凯斯勒基金会与天主教圣心大学联合构建的首个大规模、专家策划的多语言对话数据集,专注于应对仇恨言论与虚假信息的交叉领域。该数据集包含2015个虚构多轮对话,总计12,298轮交互,其中6,149轮为基于外部知识的反驳言论,覆盖英语、意大利语、马耳他语、波兰语和西班牙语五种语言,并针对穆斯林、犹太人、有色人种等七个边缘化群体。数据创建过程采用了四种人机协作策略,包括预编译、交互式、手动和自动翻译方法,并依托23名专家在18个月内完成,确保对话自然且基于事实核查文章与非政府组织报告进行外部知识锚定。该数据集旨在支持检索增强生成系统研究,为训练和评估更具说服力、事实依据的反驳言论模型提供资源,以解决在线仇恨与虚假信息交织传播的社会治理难题。
CATCH-ME is the first large-scale, expert-curated multilingual dialogue dataset jointly developed by the Bruno Kessler Foundation and the Catholic University of the Sacred Heart, focusing on the intersection of hate speech and misinformation. It encompasses 2015 fictional multi-turn dialogues, totaling 12,298 interaction turns, among which 6,149 are external knowledge-backed counterarguments. Covering five languages—English, Italian, Maltese, Polish and Spanish—the dataset targets seven marginalized groups including Muslims, Jews, and people of color. The dataset was constructed via four human-machine collaboration strategies: pre-compilation, interactive, manual and automatic translation methods, and was completed by 23 experts over an 18-month timeline. To guarantee natural dialogues and external knowledge grounding, all content is anchored to fact-checking articles and non-governmental organization (NGO) reports. This dataset aims to support research on retrieval-augmented generation (RAG) systems, offering resources for training and evaluating more persuasive, fact-based counterargument models to address the social governance challenge of the intertwined propagation of online hate speech and misinformation.
数据集概述
该数据集与论文《Assisted Counterspeech Writing at the Crossroads of Hate Speech and Misinformation》相关,专注于针对仇恨言论和错误信息的自动反言论生成与专家修订。
数据条目结构
每条数据包含 13 个字段,具体如下:
pair_id:言论-反言论对的唯一标识符。claim:包含错误信息的仇恨性言论。CS_gen:针对该言论自动生成的反言论。CS_ed:专家修订后的反言论版本。generation_strategy:生成反言论所使用的策略,可选值为Fact-Checkers、NGO或Mixed。target:该言论所针对的边缘化群体。antistereotype_id:作为外部知识使用的反刻板印象的唯一ID列表。fc_article_id:作为外部知识使用的事实核查文章的唯一ID。annotator_type:修订反言论的专家类型,为事实核查员(FC)或非政府组织运营者(NGO)。fc_article_url:事实核查文章的来源URL。antistereotype_url:所使用的反刻板印象的来源URL列表。antistereotype_location:反刻板印象在对应来源URL中的具体位置(例如,数值"2"表示该页中的 Myth #2)。HTER:使用pyter3库计算的CS_gen与CS_ed之间的 HTER 分数。HTER 值 ≥ 0.39 的配对被选中用于针对研究问题1(RQ1)的人工评估。





