deb2viz: Debiasing gender in word embedding data using subspace visualization
Enoch Opanin Gyamfi, Yunbo Rao, Miao Gou, Yanhua Shao · 2020
Word embedding have been used in numerous Natural Language Processing and Machine Learning tasks. However, it is a high-dimensional vector field that propagate stereotypes to software applications. Its current debiasing frameworks do not completely capture its embedded patterns. In this paper, we propose deb2viz, a visual debiasing approach that explores and manipulates the high-dimensional patterns of word embedding field. First, we partition this vector field into interrelated low-dimensional subspaces to equalize and neutralize distances between gender-definitional and gender-neutral words. To further reduce gender bias, we update the distances of appropriate nearest neighbors for gender-neutral words to be arbitrarily close. Experimental results on several benchmark standards show the competitiveness of our proposed method in mitigating bias within pre-trained word2vec embedding model.