Transformer-Based Model for Detoxifying Spanish Text with Regional Expressions

Serruto-Castillo A.; Rivera-Flores V.A.; Ramos-Apaza F.D.; Berrios-Espezua M. · Publicaciones de la Facultad de Educacion y Humanidades del Campus de Melilla (2026) · DOI: 10.30827/publicaciones.v56i1.36862

Resumen

Introduction: The proliferation of toxic content on digital communication platforms has increased the need for methods that can transform offensive expressions without losing the central meaning of messages or overlooking the nuances of local Spanish varieties. Method: This study presents CondRoBERTuito, a conditional adaptation of RoBERTuito for Spanish text detoxification. The method combines a masked-token predictor and an embedding-similarity chooser to identify offensive segments and propose contextually coherent substitutions. Training integrates toxic and non-toxic comments, including examples with regional expressions such as “lonccos” and “ccalas.” Results: On the evaluated dataset, CondRoBERTuito obtained 0.99 accuracy, 0.97 semantic similarity, .52 fluency, and .51 on the joint metric J. These values were higher than those reported for CondBERT and ParaGeDi in semantic similarity and the joint metric. Conclusions: The descriptive results show the value of adapting NLP models to local linguistic contexts. However, generalization depends on the corpora, labels, and Spanish varieties represented; therefore, broader cultural and human evaluation is required.

Instituciones: Universidad Nacional de San Agustín de Arequipa