TOXIGEN
收藏资源简介:
TOXIGEN是由微软研究院创建的一个大规模机器生成的数据集,包含274,186条关于13个少数群体(如非洲裔美国人、女性、LGBTQ+等)的毒性和良性陈述。该数据集通过使用GPT-3语言模型,采用演示基础的提示和对抗性分类器循环解码技术生成,旨在覆盖更广泛的隐含毒性文本,并比以往的人类编写资源更全面地涉及更多少数群体。TOXIGEN不仅在规模上超越了以往的数据集,而且在平衡毒性和良性陈述方面也更为出色,为改善现有毒性检测分类器的性能提供了重要资源。此外,数据集的应用领域主要集中在提高对隐含毒性语言的检测能力,帮助解决在线环境中对少数群体的偏见和歧视问题。
TOXIGEN is a large-scale machine-generated dataset developed by Microsoft Research. It contains 274,186 toxic and benign statements targeting 13 minority groups, including African Americans, women, LGBTQ+ individuals, and others. The dataset was generated using the GPT-3 language model, leveraging demonstration-based prompting and adversarial classifier cyclic decoding techniques. Its core objectives are to cover a broader spectrum of implicitly toxic texts and address more minority groups more comprehensively than prior human-written resources. TOXIGEN not only outperforms existing datasets in terms of scale, but also achieves better balance between toxic and benign statements, making it a pivotal resource for enhancing the performance of current toxicity detection classifiers. Furthermore, the primary applications of this dataset focus on improving the detection of implicitly toxic language, and aiding in resolving bias and discrimination against minority groups in online spaces.




