Visualization for interactively adjusting the de-bias effect of word embedding

TL;DR

This paper introduces an interactive visualization method to adjust gender bias in Japanese word embeddings, balancing bias removal with model performance preservation through user-guided parameter tuning.

Contribution

It proposes a novel visualization-based approach allowing users to control debiasing levels per category, addressing bias and performance trade-offs in word embeddings.

Findings

Debiasing effects vary across categories.

User-adjusted debiasing improves fairness.

Trade-offs between bias reduction and performance are manageable.

Abstract

Word embedding, which converts words into numerical values, is an important natural language processing technique and widely used. One of the serious problems of word embedding is that the bias will be learned and affect the model if the dataset used for pre-training contains bias. On the other hand, indiscriminate removal of bias from word embeddings may result in the loss of information, even if the bias is undesirable to us. As a result, a risk of model performance degradation due to bias removal will be another problem. As a solution to this problem, we focus on gender bias in Japanese and propose an interactive visualization method to adjust the degree of debias for each word category. Specifically, we visualize the accuracy in a category classification task after debiasing, and allow the user to adjust the parameters based on the visualization results, so that the debiasing can be adjusted according to the user's objectives. In addition, considering a trade-off between debiasing and preventing degradation of model performance, and that different people perceive gender bias differently, we developed a mechanism to present multiple choices of debiasing configurations applying an optimization scheme. This paper presents the results of an experiment in which we removed the gender bias for word embeddings learned from the Japanese version of Wikipedia. We classified words into five categories based on a news corpus, and observed that the degree of influence of debiasing differed greatly among the categories. We then adjusted the degree of debiasing for each category based on the visualization results.