cardiffnlp/super_tweeteval
收藏SuperTweetEval 数据集概述
数据集基本信息
- 名称: SuperTweetEval
- 语言: 英语 (en)
- 许可证: 未知
- 多语言性: 单语
- 大小: 小于50K
- 任务类别: 文本分类, 令牌分类, 问答, 其他
- 任务ID: 主题分类, 命名实体识别, 抽象问答
- 标签: super_tweet_eval, tweet_eval, 自然语言理解
数据集结构
数据文件配置
- config_name: tempo_wic
- train: data/tempo_wic/train.jsonl
- test: data/tempo_wic/test.jsonl
- validation: data/tempo_wic/validation.jsonl
- config_name: tweet_emoji
- train: data/tweet_emoji/train.jsonl
- test: data/tweet_emoji/test.jsonl
- validation: data/tweet_emoji/validation.jsonl
- config_name: tweet_emotion
- train: data/tweet_emotion/train.jsonl
- test: data/tweet_emotion/test.jsonl
- validation: data/tweet_emotion/validation.jsonl
- config_name: tweet_hate
- train: data/tweet_hate/train.jsonl
- test: data/tweet_hate/test.jsonl
- validation: data/tweet_hate/validation.jsonl
- config_name: tweet_intimacy
- train: data/tweet_intimacy/train.jsonl
- test: data/tweet_intimacy/test.jsonl
- validation: data/tweet_intimacy/validation.jsonl
- config_name: tweet_ner7
- train: data/tweet_ner7/train.jsonl
- test: data/tweet_ner7/test.jsonl
- validation: data/tweet_ner7/validation.jsonl
- config_name: tweet_nerd
- train: data/tweet_nerd/train.jsonl
- test: data/tweet_nerd/test.jsonl
- validation: data/tweet_nerd/validation.jsonl
- config_name: tweet_qa
- train: data/tweet_qa/train.jsonl
- test: data/tweet_qa/test.jsonl
- validation: data/tweet_qa/validation.jsonl
- config_name: tweet_qg
- train: data/tweet_qg/train.jsonl
- test: data/tweet_qg/test.jsonl
- validation: data/tweet_qg/validation.jsonl
- config_name: tweet_sentiment
- train: data/tweet_sentiment/train.jsonl
- test: data/tweet_sentiment/test.jsonl
- validation: data/tweet_sentiment/validation.jsonl
- config_name: tweet_similarity
- train: data/tweet_similarity/train.jsonl
- test: data/tweet_similarity/test.jsonl
- validation: data/tweet_similarity/validation.jsonl
- config_name: tweet_topic
- train: data/tweet_topic/train.jsonl
- test: data/tweet_topic/test.jsonl
- validation: data/tweet_topic/validation.jsonl
数据集任务详情
任务与数据集对应关系
| 任务 | 数据集 | 描述 | 实例数量 |
|---|---|---|---|
| 主题分类 | TweetTopic | 多标签分类 | 4,585 / 573 / 1,679 |
| 命名实体识别 | TweetNER7 | 序列标注 | 4,616 / 576 / 2,807 |
| 问答 | TweettQA | 生成 | 9,489 / 1,086 / 1,203 |
| 问题生成 | TweetQG | 生成 | 9,489 / 1,086 / 1,203 |
| 亲密分析 | TweetIntimacy | 单文本回归 | 1,191 / 396 / 396 |
| 推文相似度 | TweetSIM | 双文本回归 | 450 / 100 / 450 |
| 意义转移检测 | TempoWIC | 双文本二分类 | 1,427 / 395 / 1,472 |
| 仇恨言论检测 | TweetHate | 多类别分类 | 5,019 / 716 / 1,433 |
| 表情符号分类 | TweetEmoji100 | 多类别分类 | 50,000 / 5,000 / 50,000 |
| 情感分类 | TweetSentiment | ABSA五点尺度分类 | 26,632 / 4,000 / 12,379 |
| 命名实体消歧 | TweetNERD | 二分类 | 20,164 / 4,100 / 20,075 |
| 情感分类 | TweetEmotion | 多标签分类 | 6,838 / 886 / 3,259 |
数据集字段
数据字段统一描述
- tweet_topic
text: 字符串gold_label_list: 字符串列表date: 字符串
- tweet_ner7
text: 字符串text_tokenized: 字符串列表gold_label_sequence: 字符串列表date: 字符串entities: 字典列表,包含{"entity": "string", "type": "string"}
- tweet_qa
text: 字符串gold_label_str: 字符串context: 字符串
- tweet_qg
text: 字符串gold_label_str: 字符串context: 字符串
- tweet_intimacy
text: 字符串gold_score: 浮点数
- tweet_similarity
text_1: 字符串text_2: 字符串gold_score: 浮点数
- tempo_wic
gold_label_binary: 整数target: 字符串text_1: 字符串text_tokenized_1: 字符串列表token_idx_1: 整数date_1: 字符串text_2: 字符串text_tokenized_2: 字符串列表token_idx_2: 整数date_2: 字符串
- tweet_hate
gold_label: 整数text: 字符串
- tweet_emoji
gold_label: 整数text: 字符串date: 字符串
- tweet_sentiment
gold_label: 整数text: 字符串target: 字符串
- tweet_nerd
gold_label_binary: 整数target: 字符串text: 字符串definition: 字符串text_start: 整数text_end: 整数date: 字符串
- tweet_emotion
text: 字符串gold_label_list: 字符串列表
评估指标与模型
评估指标
| 数据集 | 评估指标 | 黄金标签 |
|---|---|---|
| TweetTopic | macro-F1 | arts_&_culture, business_&_entrepreneurs, celebrity_&_pop_culture, <br />diaries_&_daily_life, family, fashion_&_style, <br />film_tv_&_video, fitness_&_health, food_&_dining, <br />gaming, learning_&_educational, music, <br />news_&_social_concern, other_hobbies, relationships, <br />science_&_technology, sports, travel_&_adventure, <br />youth_&_student_life |
| TweetNER7 | macro-F1 | B-corporation, B-creative_work, B-event, <br />B-group, B-location, B-person, <br />B-product, I-corporation, I-creative_work, <br />I-event, I-group, I-location, <br />I-person, I-product, O |
| TweettQA | answer-F1 | - |
| TweetQG | METEOR | - |
| TweetIntimacy | spearman correlation | [1 - 5] |
| TweetSIM | spearman correlation | [0 - 5] |
| TempoWIC | accuracy | no, yes |
| TweetHate | combined-F1<br /> (micro-F1 for hate/not-hate &<br /> macro-F1 for hate speech subclasses) | hate_gender, hate_race, hate_sexuality, hate_religion, hate_origin, <br />hate_disability, hate_age, not_hate |
| TweetEmoji100 | accuracy at top 5 | Full emoji list: ./data/tweet_emoji/map.txt |
| TweetSentiment | 1 - MAE^M<br /> (MAE^M : Macro Averaged Mean Absolute Error) | strongly negative , negative, negative or neutral, <br /> positive, strongly positive |
| TweetNERD | accuracy | no, yes |
| TweetEmotion | macro-F1 | anger, anticipation, disgust, fear, joy, love, optimism, <br />pessimism, sadness, surprise, trust |
模型卡片
引用信息
主参考论文
bibtex @inproceedings{antypas2023supertweeteval, title={SuperTweetEval: A Challenging, Unified and Heterogeneous Benchmark for Social Media NLP Research}, author={Dimosthenis Antypas and Asahi Ushio and Francesco Barbieri and Leonardo Neves and Kiamehr Rezaee and Luis Espinosa-Anke and Jiaxin Pei and Jose Camacho-Collados}, booktitle={Findings of the Association for Computational Linguistics: EMNLP 2023}, year={2023} }
个别数据集引用
- TweetTopic
@inproceedings{antypas-etal-2022-twitter, title = "{T}witter Topic Classification", author = "Antypas, Dimosthenis and Ushio, Asahi and Camacho-Collados, Jose and Silva, Vitor and Neves, Leonardo and Barbieri, Francesco", booktitle = "Proceedings of the 29th International Conference on Computational Linguistics", month = oct, year = "2022", address = "Gyeongju, Republic of Korea", publisher = "International Committee on Computational Linguistics",




