Use this link to cite:
https://hdl.handle.net/2183/49631 Clasificación y ranking automático de titulares periodísticos políticos mediante modelos de lenguaje y técnicas de PLN
Loading...
Identifiers
Publication date
Authors
García López, Claudia
Advisors
Other responsabilities
Universidade da Coruña. Facultade de Informática
Journal Title
Bibliographic citation
Type of academic work
Academic degree
Abstract
[Resumen] Los titulares periodísticos actúan como el principal mecanismo de framing de una noticia, condicionando de forma directa la interpretación del lector. Este trabajo aborda el problema de clasificación y ordenación automática de titulares periodísticos políticos en español, planteado en la tarea compartida PoliticHeadlinES del Iberian Languages Avaluation Forum (IberLEF 2026). En esta tarea, dado el cuerpo de una noticia y diez titulares candidatos, un sistema debe identificar y priorizar el titular original frente a nueve distractores generados automáticamente, tanto en modalidad textual como multimodal.
Se desarrollan y comparan dos paradigmas de solución. En una primera fase, se emplean modelos de lenguaje de gran escala (LLM) en régimen zero-shot, mediante prompt engineering estructurado, tanto a través de API externa como en despliegue local con cuantización a 4 bits para la variante multimodal. En una segunda fase, se propone una arquitectura ligera de ensemble que combina indicadores léxicos, semánticos y de entidades nombradas.
Los resultados muestran que la escala del modelo correlaciona positivamente con el rendimiento en régimen zero-shot, que la cuantización a 4 bits introduce una pérdida de precisión que contrarresta la ganancia aportada por la información visual, y que el ensemble ligero propuesto supera consistentemente a sus componentes individuales, ofreciendo además mayor eficiencia computacional y explicabilidad.
[Abstract] News headlines act as the primary framing mechanism of a news article, directly shaping readers’ interpretation of its content. This work addresses the problem of automatic classification and ranking of Spanish political news headlines, as proposed in the PoliticHeadlinES at Iberian Languages Avaluation Forum (IberLEF 2026). In this task, given the body of a news article and ten candidate headlines, a system must identify and prioritise the original headline over nine automatically generated distractors, in both text-only and multimodal settings. Two solution paradigms are developed and compared. In a first phase, large language models (LLMs) are employed under a zero-shot regime through structured prompt engineering, both via external API and local deployment with 4-bit quantisation for the multimodal variant. In a second phase, developed as an independent contribution following the competition, a lightweight ensemble architecture is proposed, combining lexical, semantic, and named-entity signals. Results show that model scale correlates positively with zero-shot performance, that 4-bit quantisation introduces a precision loss that offsets the gain provided by visual information, and that the proposed lightweight ensemble consistently outperforms its individual components, while also offering greater computational efficiency and explainability.
[Abstract] News headlines act as the primary framing mechanism of a news article, directly shaping readers’ interpretation of its content. This work addresses the problem of automatic classification and ranking of Spanish political news headlines, as proposed in the PoliticHeadlinES at Iberian Languages Avaluation Forum (IberLEF 2026). In this task, given the body of a news article and ten candidate headlines, a system must identify and prioritise the original headline over nine automatically generated distractors, in both text-only and multimodal settings. Two solution paradigms are developed and compared. In a first phase, large language models (LLMs) are employed under a zero-shot regime through structured prompt engineering, both via external API and local deployment with 4-bit quantisation for the multimodal variant. In a second phase, developed as an independent contribution following the competition, a lightweight ensemble architecture is proposed, combining lexical, semantic, and named-entity signals. Results show that model scale correlates positively with zero-shot performance, that 4-bit quantisation introduces a precision loss that offsets the gain provided by visual information, and that the proposed lightweight ensemble consistently outperforms its individual components, while also offering greater computational efficiency and explainability.
Description
Editor version
Rights
Attribution-NonCommercial 4.0 International








