Meneame Media Bias Dataset: Interaction Features and Bias Labels
收藏资源简介:
A processed dataset of news articles submitted to Meneame (Spanish social news aggregator) with automatic media bias labels and rich interaction features derived from user comments. Contents articles_with_features.parquet: 14,995 articles with 38 columns including bias labels (from DistilBERT trained on MBBMD), interaction features (karma statistics, comment engagement metrics), and metadata (outlet, tags, timestamp). articles_labeled.parquet: Articles with bias probability scores. karma_features.parquet: Advanced karma distribution features per article (entropy, Gini, bimodality, skewness) for 183K+ articles. comments_with_sentiment.parquet: 20K comment sample with sentiment (POS/NEG/NEU) and emotion (joy, anger, sadness, fear) scores from pysentimiento/robertuito. user_profiles.parquet: User-level bias exposure metrics. user_outlet_interactions.parquet: Bipartite graph data (user-outlet comment counts). Pipeline Data was collected from meneame.net (2005-2021), processed through a 5-step pipeline: ingestion, filtering, automatic bias labeling (franfj/fdtd_media_bias_E), interaction feature extraction, and statistical analysis. See the GitHub repository for full reproducibility. Key Statistics 14,995 articles from 2,868 media outlets 13.2M comments from 96K unique users 61.5% articles labeled as biased (automatic labeling) Timespan: 2005-2021
本数据集为提交至西班牙社会新闻聚合平台Meneame的新闻稿件经处理后的数据集,附带自动生成的媒体偏见标签,以及源自用户评论的丰富交互特征。 内容: 1. `articles_with_features.parquet`:包含14995篇新闻稿件,共38个字段,涵盖偏见标签(由在MBBMD数据集上训练的DistilBERT生成)、交互特征(用户声望(karma)统计数据、评论参与度指标)以及元数据(媒体来源、标签、时间戳)。 2. `articles_labeled.parquet`:附带偏见概率得分的新闻稿件数据集。 3. `karma_features.parquet`:针对18.3万余篇新闻稿件的高级用户声望(karma)分布特征数据集,包含熵、基尼系数、双峰性、偏度等指标。 4. `comments_with_sentiment.parquet`:包含2万条评论样本,附带由pysentimiento/robertuito生成的情感(积极/消极/中性)与情绪(喜悦、愤怒、悲伤、恐惧)得分。 5. `user_profiles.parquet`:包含用户层面偏见接触指标的数据集。 6. `user_outlet_interactions.parquet`:二部图数据(用户-媒体来源的评论数量)。 数据处理流程: 数据采集自meneame.net(2005年-2021年),并通过包含数据摄入、过滤、自动偏见标注(采用franfj/fdtd_media_bias_E模型)、交互特征提取与统计分析在内的五步流程完成处理。完整可复现的处理流程请参阅对应GitHub仓库。 关键统计数据: - 来自2868家媒体机构的14995篇新闻稿件; - 来自9.6万独立用户的1320万条评论; - 经自动标注后,61.5%的稿件被判定为存在偏见; - 时间跨度:2005年-2021年。



