Author: Joseba Fernández De Landa Aguirre

Thesis: "Adimen Artifizialeko metodoak gizarte ikerkuntzarako: analisi demografikoa, jarreren detekzioa eta joera politikoen identifikazioa"

Director: Rodrigo Agerri Gascon

Day: 2024ko urriaren 16an
Time: 10:00h
Place: Ada Lovelace aretoa (Informatikako fakultatea)

Abstract:

"This thesis dissertation explores the intersection of social research and artificial intelligence (AI), investigating how AI technology can be leveraged to enhance the methodology and outcomes of social science studies. The research explores the capabilities of AI, particularly Machine Learning and Natural Language Processing (NLP), to analyze large datasets, identify patterns and infer features that would be challenging to obtain through traditional methods. To achieve this goal, we develop methodologies to automatically characterize social media users leveraging their text content and user interactions thereby enabling more accurate and generalizable predictions. The developed methodologies are then applied to three main applications including demographic characteristic identification, stance detection and political leaning inference. First, this thesis presents the first large scale computational approach for demographic analysis to characterize Basque social media users, including automatic age and community prediction.Second, we exploit the effectiveness of both textual and interaction data to perform stance detection on social media with state-of-the-art results. Specifically, we build the \textit{VaxxStance} dataset, the first crosslingual dataset for stance detection which includes interaction and text data. Furthermore, we present the \textit{Relational Embedding} (RE) interaction-based user representation method which enables to capture user-based features with optimal performance not just for stance detection but also for political leaning inference task. Third, REs outperform every other interaction-based method for multi-class political leaning inference across diverse contexts, allowing to distinguish with high-accuracy between users with different levels of political engagement. Finally, the ability of REs to be effectively combined with textual features demonstrates their robustness and adaptability to perform AI-based social research."