Bias in Knowledge Graph Embeddings

In this paper, we study bias in knowledge graph embeddings. We focus on gender bias in occupations, but our approach is applicable to other types of bias. We start by proposing measures for identifying bias in the dataset (i.e., in the KG) and then present two methods for testing whether any bias in the dataset is amplified by the embeddings. First, we look for gender-specific occupation analogies in the embeddings. Second, we test whether link prediction (i.e., occupation prediction in our case) aggregates gender bias by proposing gender-dominated occupations to people of the corresponding gender more often than expected. Then, we use a debiasing approach based on projections on the gender subspace. We present experimental results using the Wikidata dataset and pretrained TransE embeddings. Our results show that there exists gender bias in the dataset and that such bias is amplified by the embeddings. Our debiasing approach removes bias with a small penalty on accuracy.