Effect of Debiasing on Information Retrieval

Word embeddings provide a common basis for modern natural language processing tasks, however, they have also been a source of discussion regarding their possible biases. This has led to a number of publications regarding algorithms for removing this bias from word embeddings. Debiasing should make the embeddings fairer in their use, avoiding potential negative effects downstream. For example: word embeddings with a gender bias that are used in a classification task in a hiring process. In this research, we compare regular and debiased word embeddings in an Information Retrieval task. We show that the two methods produce different results, however, this difference is not substantial.