Impact of Debiasing Word Embeddings on Information Retrieval

Word embeddings are a core technology in neural methods for information retrieval. However, previous work has suggested undesirable biases in word embeddings, in particular against gender. In this paper, we look at the extent of the bias in different cases. Presumably, not all biased analogies are ‘robust’ and can sometimes give unexpected results. We discuss some ways in which bias in word embeddings could affect systems in information retrieval, which is the topic of our future research.