On the Automatic Multilingual Detection of Persuasion Techniques in the News: A Natural Language Processing Approach

This thesis presents a Natural Language Processing (NLP) approach to automatically detect persuasion techniques in multilingual news content. As the dissemination of propaganda becomes increasingly prevalent in digital media, identifying these techniques is critical to promoting information literacy and countering propaganda. Through a comprehensive literature review, this thesis first explores the landscape of propaganda, introducing the concept and its history, and then dives deeper into computational propaganda and its dissemination and detection, highlighting gaps and challenges in current methodologies, both from text and network analysis perspectives. The collective research efforts known as shared tasks are also discussed in detail. The core of this work focuses on two experiments, involving the participation to the shared task CheckThat! Lab 2024 at CLEF-2024, that challenges participants to develop models to classify persuasion techniques in news across multiple languages, and an attempt conducted months later to enhance the performance of the first system developed. The dataset provided for training includes news articles in multiple languages on several topics annotated with various persuasion techniques, and multiple models are evaluated to determine their effectiveness in a competition setting with a public leaderboard. The methodology of our team UniBO encompasses data preprocessing, data augmentation, and the development of a sophisticated system for detection of persuasion techniques using state-of-the-art NLP tools. Results show that, while our proposed model achieves competitive performance in multilingual settings, challenges such as data scarcity, explainability, and model generalization persist. Ethical considerations, including biases in detection algorithms, are also addressed. The thesis concludes by discussing limitations and potential avenues for future research.