Identifying biases in politically biased wikis through word embeddings

Abstract

With the increase of biased information available online, the importance of analysis and detection of such content has also significantly risen. In this paper, we aim to quantify different kinds of social biases using word embeddings. Towards this goal we train such embeddings on two politically biased MediaWiki instances, namely RationalWiki and Conservapedia. Additionally we included Wikipedia as an online encyclopedia, which is accepted by the general public. Utilizing and combining state-of-the-art word embedding models with WEAT and WEFAT, we display to what extent biases exist in the above-mentioned corpora. By comparing embeddings we observe interesting differences between different kinds of wikis.