Abstract: A considerable amount of literature exists on bias detection and mitigation in natural language processing models, especially word2vec embeddings, which represent words as vectors of real numbers. The most common methods used compare cosine similarity between vectors corresponding to words from protected groups and (vectors corresponding to) attributes that are considered to be stereotypical or harmful in some way. Such methods will be in our focus. We will argue that the existing methods are for various reasons deficient and propose a better approach.
Speaker Bio: This event’s speakers are Alicja Dobrzeniecka and Rafal Urbaniak. Alicja Dobrzeniecka is an MA student in AI at Vrij Universiteit Amsterdam. Urbaniak is a scholar interested in applications of formal (esp. probabilistic) methods to philosophical and social problems, such as those related to legal applications of probabilistic methods, algorithmic fairness and formal epistemology.
In the Ethics Institute Speaker Series, hosted by the Department of Philosophy and Religion, visiting scholars will share their research with faculty and students alike. Intended to help build a stronger community of scholars, this series will de-mystify the research process by sharing works-in-progress.
