Introduction

This repository is for 'You Can't See Their Bias: Exploring Implicit Media Bias in Non-Political News Articles' paper. We publicly release 1) Implicit Media Bias Dataset (IMBD) and 2) source code of four classifiers for reproducibility.

1. IMBD

News articles in the section of information technology and science, which were published in two most-biased Korean news outlets.

• human_evaluated_news_articles_200.csv (760KB)

  • 200 labeled news articles (100 conservative and 100 progressive)
  • columns: id, date, topic, title, text, news_outlet_label, human_label, objectiveness, fairness, unbiasedness
    id - integer (1 to 200) identifier for news articles (same as the annotated label in K-means clustering; refer to the paper)
    news_outlet_label - integer {0: 'consevative', 1: 'progressive'}
    human_label - float (-1.00 'consevative' to 1.00 'progressive')
    objectiveness, fairness, unbiasedness - float (-3.00 to 3.00)

• trained_news_articles_24376.csv (74.2MB)

  • 24,376 labeled news articles (18,094 conservative and 6,282 progressive)
  • columns: date, topic, title, text, news_outlet_label
    news_outlet_label - integer {0: 'consevative', 1: 'progressive'}

2. Source Code of Four Classifiers

Classifiers which are trained/examined with large-scale non-political articles (trained_news_articles_24376.csv), to identify each article's political orientations (news_outlet_label of human_evaluated_news_articles_200.csv).

• requirements.txt

• models/naive_bayes.py

• models/cnn.py

• models/bilstm.py

• models/bert.py

For more details of news outlets and experiment settings, please refer to our paper.