2026-07-28T23:36:31Z https://drops.dagstuhl.de/oai

oai:drops-oai.dagstuhl.de:26705 2026-07-28T09:32:41Z ddc:004 open_access

Polarizations in Static Word Embeddings: Investigating Bias in Marginalized Dialects of Portuguese Ferro Junior, Raimundo Juracy Campos Ribeiro, Eugénio Rocha, Leonardo Sampaio word embeddings bias detection Portuguese LGBTQIAP+ pajubá NLP lexical association static embeddings This study investigates bias in static word embeddings applied to Pajubá, a dialect spoken within the Brazilian LGBTQIAP+ community. Four models from the NILC repository (Word2Vec, GloVe, FastText, and Wang2Vec) were evaluated across two dimensions: representational capacity and polarity analysis. The vocabulary coverage test revealed that approximately 74% of dialectal terms are present in the models' vector spaces. The RND and SC-WEAT tests consistently point toward negative associations for identity terms across all models, with the SC-WEAT further suggesting that reappropriated dialect terms carry their standard Portuguese positive valence into the embedding space, while explicit identitarian markers remain negatively encoded. An adjectivation test confirms stereotypical associations, including the linkage of travesti with criminality and the pathologisation of gay. These findings suggest that static word embeddings fail to capture the sociolinguistic complexity of Pajubá and systematically reflect the dominant and often discriminatory discourses present in the training corpora. Schloss Dagstuhl – Leibniz-Zentrum für Informatik Raimundo Juracy Campos Ferro Junior and Eugénio Ribeiro and Leonardo Sampaio Rocha 2026 Is Part Of OASIcs, Volume 144, 15th Symposium on Languages, Applications and Technologies (SLATE 2026) InProceedings Text doc-type:ResearchArticle publishedVersion application/pdf doi:10.4230/OASIcs.SLATE.2026.7 urn:nbn:de:0030-drops-267055 https://drops.dagstuhl.de/entities/document/10.4230/OASIcs.SLATE.2026.7 eng https://creativecommons.org/licenses/by/4.0/legalcode