Enhancing Multilingual Fine-Tuning for the Detection of Persuasion Techniques

A new paper authored by the Kempelen Institute of Intelligent Technologies (KInIT), one of the members of the VIGILANT consortium, delves into the topic of multilingual fine-tuning for the automatic detection of persuasion techniques. The research is anchored in the success of KInIT's participation in the esteemed SemEval2023 Data Challenge, where they proposed one of the winning solutions.

KInIT's work addresses the critical issue of malign online content, where persuasion techniques are often misused to sway public opinion on sensitive topics like vaccination or migration. Recognising the pivotal role of AI in addressing such content, the paper underscores how the automatic detection of persuasion techniques can help with the finding and uncovering of disinformation and other forms of online manipulation.

However, the task comes with its own set of challenges, ranging from the complexity of identifying multiple persuasion techniques within a given text to the inherent imbalance in data representation.

KInIT's solutions not only address these challenges but also contribute significantly to the VIGILANT project's mission of supporting law enforcement officers with timely and effective detection of malign online content.

As part of their approach, KInIT experimented with different methods, ultimately revealing the superiority of a fully multilingual model. Leveraging pretrained large language models like RoBERTa and XLM-RoBERTa, fine-tuned specifically for persuasion technique detection, KInIT's research showcases the effectiveness of their approach, providing a valuable contribution to the ongoing efforts against misleading and harmful online content.

For a comprehensive summary of KInIT's innovative solution, explore the full research paper here.

The data has been published at GitHUB according to Open Access principles.