Semeval 2019 Task 4 - Hyperpartisan News Detection

Second trial dataset for the SemEval 2019 Task 4: Hyperpartisan News Detection.

The dataset contains ~1 million articles. It is split in training and validation, where <strong>no</strong> publisher that occurs in the training set also occurs in the validation set. Due to imbalance in our raw data, the training dataset of this version contains more articles that are hyperpartisan (533334: 26667 left and 26667 right) than not (26667). The validation set is balanced as the test set will be: 50% hyperpartisan (33333 left and 33333 right) and 50% not (66666). All articles are labeled by the overall bias of the publisher as provided by BuzzFeed journalists or MediaBiasFactCheck.com.

The trial data is not fully cleaned. Due to some encoding error, some characters are replaced by question marks. However, all files are already fully compatible with the XML schema files. Unlike the first trial version of this dataset, the &lt;q&gt; tag is used instead of &lt;quote&gt; (to be compatible with HTML).