Preserving Grammatical Gender when Debiasing Word Embeddings in Spanish

Word embeddings are widely used in Natural Language Processing but often encode gender biases, which can lead to discriminatory outcomes. Various debiasing techniques exist, especially focusing on English, thus failing to account for the complexities of languages with grammatical gender, such as Spanish. In this paper, we propose INLP-Gram, an algorithm designed to mitigate gender bias in Spanish word embeddings while preserving grammatical gender information. It is an adaptation of the Iterative Nullspace Projection (INLP). We evaluate INLP-Gram using the Word Embedding Association Test (WEAT) and a grammatical gender classification test. Our results demonstrate that INLP-Gram effectively reduces gender bias while maintaining grammatical gender distinctions. This work advances bias mitigation techniques for word embeddings in morphologically-rich languages.