POLOR: Leveraging Contrastive Learning to Detect Political Orientation of Opinion in News Media

Abstract

News articles are naturally influenced by the values, beliefs, and biases of the reporters preparing the stories and the policies of the publishing outlets. Numerous studies and datasets have been proposed to detect the political orientation of news articles. However, most of these studies ignore real textual clues and learn the textual signature of the source (commonly the publisher and rarely the writer) of the article instead. Moreover, a good volume of opinion pieces published by major news outlets do not reflect the political orientation of the publisher but rather reflect the political orientation of a non-professional writer. Existing methods are not built to correct this difference in the training data and, hence, perform poorly on human-annotated data. We propose, POLOR, a fine-tuned BERT model that employs contrastive learning to detect the political orientation of news articles even when the training data is labeled by the source (i.e. the publisher of the news article). Unlike previous work in the literature, the model learns features by employing different contrastive learning objectives where each sentence is contrasted with sentences from various sources simultaneously. POLOR achieves a 15\% increase on our dataset compared to previously proposed baselines. Finally, we release two datasets of opinion news: source-annotated and human-annotated datasets. The full paper including supplementary materials, code, and datasets can be found at https://www.cs.unm.edu/~ajararweh/.