Performance Degradation of NLP Models Caused by Bias Removal

Abstract

In this paper, we propose a new criterion for the study of bias mitigation (debiasing) in natural language processing (NLP). Recent studies of debiasing have focused only on improving the bias metrics defined within each paper, and have rarely considered the impact of debiasing on the ability to solve NLP tasks. We focused on this problem and found through numerical experiments that debiased word vectors degrade the ability to solve the NLP task. We also investigated the cause of the degradation by creating a Visual Analytics (VA) system that can compare the models before and after debiasing on a per-class, per-classification, and per-sample basis, and by using existing machine learning explanation methods. We found that the word vectors may be losing information that is important for solving classification and question answering tasks. This is the first study to investigate the degradation of models and word vectors due to debiasing in NLP.