VERANTWORTUNGSVOLLEKI.DE

Goal of the project is to develop methods to learn and to use word embeddings (learned semantic representations of words) that are able to deal with a bias in the training data. First, bias in word embeddings has to be defined and methods have to be developed to detect various types of bias in word embeddings. Subsequently we aim to make biases visible, e.g. by transforming the latent dimensions of the word embeddings into interpretable dimensions, as was done by Rothe et al. (2016) and Hollis and Westbury (2016). Thus reasons for the classification of a word or for similarity between words can be made transparent. Finally, ways to debias word embeddings have to be found and it has to be investigated whether approaches like those of Bolukbasi et al. (2016) und Zhao et al. (2018) to remove gender bias carry over to other types of bias like those for age, skin color, but also biases for genre or text types. Applications, like the detection of offensive language, can benefit from the..