ApuEmo: Emotion Classification in Spanish Through a Hybrid Model With Transformer and Recurrent Layer

• Autores: Mamani Coaquira, Yonatan, Ibarra Cabrera, Manuel J., Aquino Cruz, Mario, Mollocondo Flores, Wilson J.
• Formato: artículo
• Fecha de Publicación: 2026
• Institución: Universidad Nacional Micaela Bastidas de Apurímac
• Repositorio: UNAMBA-Institucional
• Lenguaje: inglés
• OAI Identifier: oai:null:20.500.14195/1493
• Enlace del recurso: https://doi.org/10.1109/ACCESS.2025.3623629; https://hdl.handle.net/20.500.14195/1493
• Nivel de acceso: acceso abierto
• Materia: Emotion classification; Spanish language; Transformers; Recurrent neural networks; https://purl.org/pe-repo/ocde/ford#2.02.04

Emotion classification in social networks is a crucial task, driven by the increasing need to analyze the opinions and feelings expressed across various platforms such as Facebook, YouTube, Instagram, and X. This work presents a novel hybrid approach for emotion classification in Spanish-language texts, integrating the pre-trained SaBERT embedding with recurrent neural networks and attention mechanisms. A rigorous evaluation using the TASS 2020 dataset from the Workshop on Semantic Analysis for Task 2: Emotion Detection, alongside a collection of Spanish comments sourced from Facebook related to the Apurimac region in Peru, was conducted. The results show that the proposed model outperforms representative state-of-the-art models, such as ELiRF-UPV and UMUTeam, achieving a maximum F1-Macro value of 0.49. Moreover, complementary lexical and emotional analyses allowed for validating the model’s behaviour in regional contexts, revealing an emotional distribution consistent with the cultural and linguistic content of the Apurimac region in Peru.