[1L4-OS-4a] OS-4

[1L4-OS-4a-03] Visualizing the Effects of Gender Bias and Debias in Word Embedding

〇Arisa Sugino¹, Takayuki Itoh¹ (1. Ochanomizu University)

Keywords:Word Embedding, Gender Bias, Visualization

In recent years, services utilizing AI such as chatbots and image analysis for translation have been increasing daily, and AI is playing an active role in various aspects of our lives. Word embedding is a technique in natural language processing that converts words into numerical representations for computational manipulation. Here, biases inherent in the data set used to pre-train the language model may affect the resulting model is one of the drawbacks of word embedding. There are a limited number of prior studies on biases and debiasing in Japanese word embedding, with most studies focusing on English word embedding. Therefore, this study aimed to detect gender bias from a pretrained Word2vec dataset trained on the Japanese version of Wikipedia, visualize the biased results, and classify the biases. The experimental results demonstrate differences between categories with high bias and those susceptible to the effects of debiasing.