MuTox
收藏资源简介:
MuTox是由Meta公司开发的第一个高度多语言的音频基础毒性数据集,包含21种语言的毒性标签。数据集主要由英语和西班牙语的20000条音频以及其余19种语言的4000条音频组成。创建过程中,使用了多种文本毒性分类器进行预筛选,并根据长度和认知负荷筛选音频文件。MuTox数据集旨在解决多语言音频毒性检测的问题,特别是在多模态和多语言翻译中添加或删除毒性的情况。
MuTox is the first highly multilingual audio-based toxicity dataset developed by Meta, which provides toxicity labeling for 21 languages. The dataset primarily consists of 20,000 audio clips in English and Spanish, plus 4,000 audio clips spanning the remaining 19 languages. During its development, multiple text toxicity classifiers were employed for pre-screening, and audio files were filtered based on their duration and cognitive load. The MuTox dataset aims to address the challenges of multilingual audio toxicity detection, particularly scenarios where toxicity is added or removed during multimodal and multilingual translation processes.




