deusexvelo/tweet_eval

TweetEval是一个多任务Twitter文本分类基准数据集，包含七个异构任务，均以多类别推文分类形式呈现。任务涵盖反讽检测、仇恨言论检测、冒犯性内容检测、立场分类、表情符号预测、情感分类和情感分析。所有任务被统一到同一基准中，每个数据集具有相同的格式和固定的训练、验证及测试分割。数据集基于英语Twitter文本，由Cardiff NLP团队整理，适用于文本分类模型的训练和评估。具体配置包括：emoji（20类表情符号分类）、emotion（4类情感分类：愤怒、喜悦、乐观、悲伤）、hate（2类仇恨言论检测）、irony（2类反讽检测）、offensive（2类冒犯性内容检测）、sentiment（3类情感分析：负面、中性、正面）以及多个立场分类任务（如堕胎、无神论、气候变化、女权主义、希拉里等，每类分为反对、支持或无立场）。数据规模从数百到数万条不等，适用于自然语言处理中的意图分类、多类别分类和情感分类等任务。TweetEval is a multi-task benchmark dataset for Twitter text classification, comprising seven heterogeneous tasks all formulated as multi-class tweet classification tasks. These tasks cover irony detection, hate speech detection, offensive content detection, stance classification, emoji prediction, emotion classification, and sentiment analysis. All tasks are unified under a single benchmark framework, with each dataset sharing identical formatting and fixed train, validation, and test splits. Built on English Twitter texts and curated by the Cardiff NLP team, the dataset is designed for training and evaluating text classification models. Its specific configurations include: emoji (20-class emoji classification), emotion (4-class emotion classification covering anger, joy, optimism, and sadness), hate (2-class hate speech detection), irony (2-class irony detection), offensive (2-class offensive content detection), sentiment (3-class sentiment analysis including negative, neutral, and positive), as well as multiple stance classification tasks (e.g., abortion, atheism, climate change, feminism, Hillary Clinton, etc.), where each instance is categorized into against, support, or none. The number of instances ranges from hundreds to tens of thousands, making the dataset applicable to tasks in natural language processing such as intent classification, multi-class classification, and sentiment classification.