Cargando...
Fecha
2026-02-01
Derechos de acceso
info:eu-repo/semantics/openAccess
Título de la revista
ISSN de la revista
Título del volumen
Editorial
Citas
Resumen
Desde la popularización de las redes sociales a comienzos del siglo XXI, la comunicación en línea se ha multiplicado, convirtiéndose en un canal esencial tanto para el intercambio de información como para la creación de comunidades. No obstante, este crecimiento ha venido acompañado de un aumento perceptible del lenguaje ofensivo y el discurso de odio, fenómenos que amenazan la salud del ecosistema digital. Aunque las principales plataformas cuentan con mecanismos de moderación, la mayoría resultan costosos, lentos o incapaces de adaptarse al ritmo al que evoluciona el lenguaje.
Este Trabajo de Fin de Máster analiza la aplicación de modelos generativos basados en arquitecturas Transformer para la detección automática de lenguaje ofensivo en redes sociales. En concreto, se estudia el rendimiento del modelo Mistral, al cual se le ha aplicado una fase de ajuste fino supervisado utilizando tres conjuntos de datos ampliamente utilizados en esta tarea: OffendES, MEX-A3T y OLID, que contienen ejemplos tanto en español como en inglés.
El objetivo principal del trabajo es evaluar la capacidad del modelo ajustado para generalizar sobre distintos contextos y tipos de mensajes ofensivos. Para ello, se han diseñado y ejecutado una serie de experimentos que permiten analizar la calidad de los modelos mediante métricas como la precisión, la cobertura y la medida F1, comparando los resultados con enfoques tradicionales.
Los resultados obtenidos muestran que el modelo ajustado sobre Mistral ofrece un rendimiento competitivo en la clasificación de contenido ofensivo, destacando su adaptabilidad a distintos dominios lingüísticos. Estos hallazgos respaldan el uso de modelos generativos como una alternativa prometedora y flexible para la moderación automática de contenidos, siempre que se acompañen de una adecuada estrategia de ajuste y validación.
Since the early 2000s, social media platforms have become one of the main channels for global communication and information sharing. However, this growth has also led to a significant increase in the spread of offensive language and hate speech. Although most platforms include moderation systems, many of them remain manual, inefficient, or difficult to scale, highlighting the need for more advanced automated solutions. This Master’s Thesis explores the use of generative Transformer-based models for the automatic detection of offensive language in social media. Specifically, it investigates the performance of the Mistral model, which has undergone supervised fine-tuning using three widely known datasets in this domain: OffendES, MEX-A3T, and OLID, covering both Spanish and English texts. The main goal of the study is to assess the model’s ability to generalize across different contexts and types of offensive content. To achieve this, a series of experiments were conducted to analize the quality of models by evaluation metrics such as precision, recall, and F1-score, and the results were compared with those obtained by traditional approaches. The findings show that the fine-tuned Mistral model achieves competitive performance in offensive content classification, demonstrating strong adaptability across multilingual domains. These results support the potential of generative models as a flexible and effective foundation for real-time content moderation, provided that an appropriate fine-tuning and validation process is carried out.
Since the early 2000s, social media platforms have become one of the main channels for global communication and information sharing. However, this growth has also led to a significant increase in the spread of offensive language and hate speech. Although most platforms include moderation systems, many of them remain manual, inefficient, or difficult to scale, highlighting the need for more advanced automated solutions. This Master’s Thesis explores the use of generative Transformer-based models for the automatic detection of offensive language in social media. Specifically, it investigates the performance of the Mistral model, which has undergone supervised fine-tuning using three widely known datasets in this domain: OffendES, MEX-A3T, and OLID, covering both Spanish and English texts. The main goal of the study is to assess the model’s ability to generalize across different contexts and types of offensive content. To achieve this, a series of experiments were conducted to analize the quality of models by evaluation metrics such as precision, recall, and F1-score, and the results were compared with those obtained by traditional approaches. The findings show that the fine-tuned Mistral model achieves competitive performance in offensive content classification, demonstrating strong adaptability across multilingual domains. These results support the potential of generative models as a flexible and effective foundation for real-time content moderation, provided that an appropriate fine-tuning and validation process is carried out.
Descripción
Categorías UNESCO
Palabras clave
lenguaje ofensivo, redes sociales, modelos generativos, Transformer, fine-tuning, procesamiento del lenguaje natural, offensive language, social media, generative models, Transformer, Mistral, fine-tuning, natural language processing
Citación
Cruz Casado, Adrián. Trabajo de Fin de Máster: Detección de lenguaje ofensivo en redes sociales mediante modelos generativos. Universidad Nacional de Educación a Distancia (UNED), 2026
Centro
Escuela Técnica Superior de Ingeniería Informática
Departamento
Inteligencia Artificial