Researchers have been critical of word embeddings, which are used to represent words as vectors, for capturing undesirable associations such as gender stereotypes. However, methods for measuring and removing these biases are not well understood. A new study has found that a common debiasing method is equivalent to training on an unbiased corpus under certain conditions. The researchers also discovered that the most commonly used association test for word embeddings systematically overestimates bias. They propose a new measure of association called the relational inner product association, which reveals that some word embedding models can amplify existing gender associations. --- Why it matters: This study matters to AI engineers and researchers because it sheds light on how word embeddings can perpetuate biases in language data. Understanding these issues is crucial for developing more inclusive and fair AI systems. Source: https://arxiv.org/abs/1908.06361