A systematic pipeline for diagnosing and reducing gender-stereotype bias in Japanese PLMs for sentiment analysis.

Pre-trained language models (PLMs) are widely used in sentiment analysis, but they may inherit gender-stereotypical bias from large-scale text corpora and transfer such bias to downstream sentiment predictions. Despite growing attention to gender-stereotypical bias in PLMs, existing studies predominantly focus on English corpora and static word embeddings, limiting understanding of how such bias affects sentiment analysis models and the effectiveness of mitigation strategies. In this study, we present a three-stage task-oriented pipeline for diagnosing, mitigating, and evaluating gender-stereotypical bias in Japanese PLMs for sentiment analysis tasks. Specifically, the proposed framework diagnoses bias based on pro-stereotypical (PS), anti-stereotypical (AS), and neutral test sets, which are constructed to compare stereotype-aligned, stereotype-violating, and gender-neutral contexts under the same sentiment analysis setting. We further introduce two complementary evaluation measures, the Stereotype Bias Index (SBI) and Gender Sentiment Bias (GSB), to quantify stereotype-level bias between PS and AS samples, as well as sentiment prediction differences among male, female, and neutral groups. To mitigate bias, the framework performs debiasing fine-tuning using gender-swapped training data and then quantitatively evaluates bias reduction while monitoring sentiment classification performance. Experimental results on three Japanese BERT-based sentiment analysis models demonstrate that the proposed pipeline substantially reduces gender-stereotypical bias. For the SBI metric, the bias magnitude is reduced by 97.4%, 70.0%, and 76.9% for Tohoku BERT