Word embeddings are biased. But whose bias are they reflecting?

Abstract From Curriculum Vitae parsing to web search and recommendation systems, Word2Vec other word embedding techniques have an increasing presence in everyday interactions human society. Biases, such as gender bias, been thoroughly researched evidenced be present embeddings. Most of the research focuses on discovering mitigating bias within frames vector space itself. Nevertheless, whose is reflected embeddings has not yet investigated. Besides it also important examine whether a feminine or masculine-centric view represented biases This way, we will only gain more insight into origins before mentioned biases, but novel approach investigating Natural Language Processing systems. Based previous social sciences studies, hypothesize that masculine-centric, otherwise known androcentric, are dominant To test this hypothesis used largest English association data set publicly available. We compare distance responses male female participants cue words space. found biased towards viewpoint, predominantly reflecting worldviews set. Therefore, by conducting research, aimed unravel another layer considered when examining fairness algorithms.