Capturing political and ideological biases with word embeddings

Newspapers provide us with information about the world and have significant influence on our opinions, as well as our behaviour. Despite the journalistic principle of objectivity, the extent at which certain topics are covered in the media, as well as the opinions expressed about political situations depend heavily on the newspaper and the country it is published in.

In this project we will investigate whether political biases in news articles can be captured with word embeddings. Training a neural network on a large corpus of text documents creates a vector space that is able to preserve semantic relations between words. These so-called word2vec models depend on the word context that is provided by the text documents they are trained on. Therefore, consistent political and ideological bias in newspapers should affect the relations in the word2vec model such that compared with a different model these biases can be identified.

Based on the assumption that media in the US and Russia express diverging opinions regarding certain political issues, two models trained on these databases will be analyzed and compared regarding their view on their own and the opposing state.

The full report can be viewed in pdf format here.

This project paper was written for a text mining course during my Master program at Radboud University.

It is the foundation of MediaUncovered, a project to uncover hidden bias in media with word embeddings.