[3Xin4] Poster session 1

[3Xin4-45] Evaluation of Gender Bias in Japanese Word Embedding Model

〇Shintaro Sakai¹, Yasuhiro Suzuki¹ (1.Nagoya University)

Keywords:word embedding, AI bias, gender bias

With substantial progress in Natural Language Processing(NLP), word embedding made it possible to represent semantic relationships in a low-dimensional vector space. Although word embeddings are widely used in various applications, they have been reported to inherit and amplify social biases such as gender and racial biases. While a growing number of studies have been conducted to analyze biases in word embeddings, research examining bias in Japanese word embedding model. is scarce. The present research aims to fill this gap by empirically evaluating gender bias in occupation with Japanese word embedding trained on Wikipedia articles. Our experiments show some of the occupations such as nurse and dental hygienist show strong women biases. We also found these strength of biases are highly correlated with percentages of women in each occupation in Japan(ρ = 0.78 with p < .001). Our research shows Japanese word embedding model also inherits gender bias in occupations and they accurately reflect real-world gender disparities in occupations.