ToxiREX: Toxic REasoning in ConteXt
收藏资源简介:
ToxiREX是由阿姆斯特丹自由大学构建的一个多语言、上下文感知的毒性推理数据集,旨在系统性地捕捉和分析社交媒体中的隐式毒性内容。该数据集包含约12.5万条Reddit评论,涵盖英语、阿拉伯语、土耳其语等六种语言,数据来源于与重大现实事件相关的帖子,并采用独特的子线程预处理方法以保留对话语境。其创建过程基于毒性推理模式,通过GPT-4o生成训练集标注,并由母语者完成测试集标注,实现了对毒性含义的细粒度结构化解析。该数据集主要应用于自然语言处理领域,特别是隐式毒性检测、多语言内容审核及大语言模型评估,致力于解决在线环境中因文化、语境依赖而难以识别的有害语言推理问题。
ToxiREX is a multilingual, context-aware toxicity reasoning dataset developed by Vrije Universiteit Amsterdam, which aims to systematically capture and analyze implicit toxic content on social media. Comprising approximately 125,000 Reddit comments across six languages including English, Arabic, Turkish and others, the dataset is sourced from posts related to major real-world events, and adopts a unique sub-thread preprocessing approach to preserve conversational context. Its construction is based on toxicity reasoning patterns: training set annotations are generated via GPT-4o, while test set annotations are completed by native speakers, enabling fine-grained structured parsing of toxic meanings. This dataset is primarily applied in the field of natural language processing (NLP), particularly for implicit toxicity detection, multilingual content moderation, and large language model (LLM) evaluation, and is dedicated to addressing the challenge of identifying harmful language reasoning in online environments that is difficult to recognize due to cultural and contextual dependencies.
数据集概述:ToxiREX(Toxic REasoning in conteXt)
基本信息
- 数据集名称:ToxiREX
- 用途:上下文中的有害性推理(Toxic Reasoning in Context)
- 来源地址:https://github.com/cltl/toxirex
数据统计(基于模式特征的基本映射后分类)
数据集包含 test(测试集) 和 train(训练集) 两个部分,覆盖以下语言:
- ar(阿拉伯语)
- de(德语)
- en(英语)
- es(西班牙语)
- nl(荷兰语)
- tr(土耳其语)
分类标签含义
| 标签 | 含义 |
|---|---|
| cat_maybe_ | 潜在有害 |
| cat_justIn | 仅不当 |
| cat_differ | 不同种类的有害性 |
| cat_intimi | 类似恐吓 |
| cat_threat | 类似威胁 |
| cat_infe_n | 低等本质 |
| cat_infe_b | 低等行为 |
| cat_supe_n | 高等本质 |
| cat_supe_b | 高等行为 |
| cat_toxic! | 有害 |
测试集(test)统计
| 分类 | ar | de | en | es | nl | tr |
|---|---|---|---|---|---|---|
| cat_maybe_ | 626 | 759 | 650 | 785 | 447 | 656 |
| cat_justIn | 145 | 126 | 92 | 135 | 70 | 378 |
| cat_differ | 209 | 384 | 165 | 162 | 100 | 86 |
| cat_intimi | 4 | 6 | 19 | 22 | 27 | 1 |
| cat_threat | 0 | 1 | 7 | 13 | 8 | 1 |
| cat_infe_n | 61 | 79 | 27 | 139 | 29 | 50 |
| cat_infe_b | 142 | 106 | 173 | 168 | 190 | 93 |
| cat_supe_n | 1 | 0 | 0 | 0 | 0 | 0 |
| cat_supe_b | 0 | 0 | 0 | 0 | 0 | 0 |
| cat_toxic! | 562 | 700 | 476 | 627 | 416 | 608 |
训练集(train)统计
| 分类 | ar | de | en | es | nl | tr |
|---|---|---|---|---|---|---|
| cat_maybe_ | 597 | 1944 | 30493 | 6858 | 1049 | 2005 |
| cat_justIn | 32 | 73 | 1110 | 351 | 42 | 421 |
| cat_differ | 0 | 0 | 0 | 0 | 0 | 0 |
| cat_intimi | 18 | 92 | 1432 | 337 | 37 | 78 |
| cat_threat | 16 | 80 | 1031 | 292 | 36 | 67 |
| cat_infe_n | 233 | 590 | 8790 | 2683 | 368 | 689 |
| cat_infe_b | 212 | 685 | 12100 | 2164 | 378 | 519 |
| cat_supe_n | 0 | 0 | 0 | 0 | 0 | 0 |
| cat_supe_b | 2 | 0 | 2 | 0 | 0 | 0 |
| cat_toxic! | 497 | 1445 | 23451 | 5550 | 827 | 1709 |
数据集特点
- 多语言:涵盖6种语言(阿拉伯语、德语、英语、西班牙语、荷兰语、土耳其语)
- 多维分类:包含9种有害性相关的细分类别
- 分类规模不均衡:例如"cat_supe_n"和"cat_supe_b"在大多数语言中样本数为0或极少

- 1ToxiREX: A Dataset on Toxic REasoning in ConteXt阿姆斯特丹自由大学 · 2026年



