Algorithmic Bias and the Biases of the Bias Catchers
share
Concerns about gender bias have captured most of the attention in the AI research literature on the topic of bias in word embeddings models. In this work, a systematic analysis of popular word embedding models shows that many of those concerns are probably exaggerated. Gender bias in these models is often mild and frequently reversed in polarity to what has been regularly reported. Interestingly, other types of so far unreported moderate biases in word embedding models have been identified. Specifically, biases against intellectual phenomena such as political orientation and religiosity. This mismatch in the literature could be due to another type of bias, the bias of an orthodox epistemic community with widely shared community blind spots that perhaps is mostly bent on exploring only zeitgeist-conforming regions of the research landscape.
READ FULL TEXT