Quantifying and Debiasing Gender Bias in Japanese Gender-specific Words with Word Embedding

Machine Learning is playing a significant role in modern life. However, the problem that Machine Learning has biases and stereotypes has also drawn the researcher's attention. Word2Vec, a popular framework in the NLP field to encode the word's meaning as a real-valued vector, has been used in many machine learning and natural language processing tasks. Still, it also has been proved that it contains severe biases toward women. In this paper, we used Word2Vec to analyze the relationship between gender-specific words and personality adjectives in Japanese to Figure out the latent gender bias in those gender-specific words. We first found that the Word2Vec model trained by Japanese Wikipedia data shows that some occupation gender-specific words strongly connect with negative personality adjectives. The experiment results reflect that people commonly use these gender-specific words to criticize women in these specific occupations. Then we eliminated the projection of word vectors of personality adjectives on the gender subspace and reduced the relationship between negative personality adjectives and gender-specific words by word vector calculation.