Cargando...
Fecha
2026-03-01
Derechos de acceso
info:eu-repo/semantics/openAccess
Título de la revista
ISSN de la revista
Título del volumen
Editorial
Citas
Resumen
Este trabajo aborda la detección de posturas similares en redes sociales, comparando modelos clásicos de aprendizaje automático con arquitecturas basadas en transformers, como BERT y RoBERTa, así como redes siamesas mediante SBERT. Los modelos se entrenan y evalúan sobre distintas temáticas para identificar posturas “similares”, “contrarias” y “no relacionadas”. Los resultados muestran que los transformers superan consistentemente a los enfoques clásicos, alcanzando valores de F1-weighted de hasta 0.61 en clasificación múltiple y 0.66 en binaria, mientras que los modelos clásicos rara vez los superan. La aplicación de backtranslation permitió aumentar el volumen y la diversidad léxica del conjunto de entrenamiento, reduciendo parcialmente el desbalance entre clases, aunque sin mejorar globalmente el rendimiento. SBERT mostró limitaciones al predecir con frecuencia la clase mayoritaria, evidenciando que la similitud basada en embeddings no captura de manera eficaz relaciones complejas de oposición semántica. Los modelos basados en transformers generalizan mejor hacia temáticas no vistas, especialmente cuando el entrenamiento se realiza en contextos con mayor diversidad de opiniones y la evaluación se centra en temas cercanos al corpus entrenado. La clasificación binaria mejora la homogeneidad del rendimiento y facilita la detección de posturas similares, donde la clase “contrarias” sigue siendo el principal desafío de detección. En conjunto, estos resultados confirman que los transformers son herramientas robustas para la detección de posturas similares en redes sociales y destacan la influencia tanto del esquema de clasificación como de la naturaleza de los temas en el desempeño de los modelos.
This work addresses the detection of similar stances on social media, comparing classical machine learning models with transformer-based architectures, such as BERT and RoBERTa, as well as siamese networks using SBERT. Models are trained and evaluated on different topics to identify “similar”, “opposing” and “unrelated” stances. The results show that transformers consistently outperform classical approaches, achieving F1-weighted values of up to 0.61 in multiple classification and 0.66 in binary classification, while classical models rarely exceed 0them. The application of backtranslation increased the volume and lexical diversity of the training set, partially reducing the imbalance between classes, but without improving overall performance. SBERT showed limitations in frequently predicting the majority class, demonstrating that similarity based on embeddings does not effectively capture complex semantic opposition relationships. Transformer-based models generalise better to unseen topics, especially when training is performed in contexts with greater diversity of opinions and evaluation focuses on topics close to the trained corpus. Binary classification improves performance homogeneity and facilitates the detection of similar stances, being “opposing” category the main challenge for detection. Taken together, these results confirm that models based on transformers are robust tools for detecting similar positions on social media and highlight the influence of both the classification scheme and the nature of the topics on model performance.
This work addresses the detection of similar stances on social media, comparing classical machine learning models with transformer-based architectures, such as BERT and RoBERTa, as well as siamese networks using SBERT. Models are trained and evaluated on different topics to identify “similar”, “opposing” and “unrelated” stances. The results show that transformers consistently outperform classical approaches, achieving F1-weighted values of up to 0.61 in multiple classification and 0.66 in binary classification, while classical models rarely exceed 0them. The application of backtranslation increased the volume and lexical diversity of the training set, partially reducing the imbalance between classes, but without improving overall performance. SBERT showed limitations in frequently predicting the majority class, demonstrating that similarity based on embeddings does not effectively capture complex semantic opposition relationships. Transformer-based models generalise better to unseen topics, especially when training is performed in contexts with greater diversity of opinions and evaluation focuses on topics close to the trained corpus. Binary classification improves performance homogeneity and facilitates the detection of similar stances, being “opposing” category the main challenge for detection. Taken together, these results confirm that models based on transformers are robust tools for detecting similar positions on social media and highlight the influence of both the classification scheme and the nature of the topics on model performance.
Descripción
Categorías UNESCO
Palabras clave
detección de posturas similares, redes sociales, transformers, same stance detection, social networks, transformers
Citación
Gil Díaz, Cristina. Trabajo de Fin de Máster: Detección de posturas similares en redes sociales: Evaluación de modelos de Aprendizaje Automático y Transformers. Universidad Nacional de Educación a Distancia (UNED), 2026
Centro
Escuela Técnica Superior de Ingeniería Informática
Departamento
Inteligencia Artificial