Evaluating Bias Objectives in Multi-Objective Optimization for Gender-Fair Word Embeddings

Word embeddings are widely used in natural language processing systems, but they may encode and propagate gender stereotypes present in training corpora. This paper investigates how different bias metrics affect multi-objective optimization for gender-bias mitigation in word embeddings. We compare three bias objectives within the same NSGA-II framework: Mean Cosine Bias (MCB), SC-WEAT Effect Size (SCW-ES), and RIPA Projection. Semantic preservation is modeled through Spearman correlation on a word similarity benchmark, while downstream utility is evaluated on hate speech and sexism detection tasks. The results show that the choice of bias metric substantially changes the optimization behavior: improvements under one metric do not necessarily transfer to the others. These findings highlight the importance of cross-metric validation when evaluating bias mitigation methods and contribute to a more rigorous assessment of fairness-aware optimization in natural language processing.