Analyzing Discourses in Portuguese Word Embeddings: A Case of Gender Bias Outside the English-Speaking World

In this paper we meticulously examined a Word Embedding model in Portuguese, endeavoring to identify gender biases through diverse analytical perspectives, employing SC-WEAT and RIPA metrics that is widely used in the English realm. Our inquiry focused on three primary dimensions: (1) the frequency-based association of words with feminine and masculine terms; (2) the identification of disparities between grammatical classes pertaining to gender sets; and (3) the categorisation and grouping of feminine and masculine words, including their distinctive attributes. In regard to frequency groups, our investigation revealed a pervasive negative association of words with feminine terms in most subsets, indicative of a pronounced inclination of the model’s vocabulary towards the masculine references. Notably, among the 100 most frequent words, 89 exhibited a stronger association with masculine terms. In the scrutiny of grammatical classes, our analysis demonstrated a predominant association of adjectives with feminine references, underscoring the imperative for supplementary description when referring to women. Furthermore, a conspicuous prevalence of participle verbs associated with feminine terms was observed, a phenomenon distinct from their male counterparts and one that requires further expert attention to be properly explained. The categorisation process underscored the existence of gender bias, as exemplified by the association of words with masculine terms within the domains of sport, finance, and science, while words related to feelings, home furniture, and entertainment were associated with feminine terms. These findings assume significance in fostering a discourse on gender analysis within non-English models, such as Portuguese models, thereby encouraging the Brazilian community to actively investigate biases in NLP models.