Leveraging Large Language Models for multilingual disinformation detection: the 2022 Kenyan General Elections case study

Abstract

The rise of disinformation on social media, particularly during elections, poses challenges to democratic processes. This study examined the effectiveness of Large Language Models (LLMs) in detecting disinformation during the 2022 Kenyan General Elections. It analyzed the structure of disinformation networks, key actors’ influence mechanisms, and the synergies between these methodologies. Using a dataset from X (formerly known as Twitter), the study applied Hugging Face’s mBERT mBERT model for its multilingual capabilities. The data used to train the model had 20,000 rows and 37 features. Textual data was tokenised and classified as 'Hate Speech' or 'Not Hate Speech,' with key performance metrics, including accuracy, precision, recall, and F1-score indicating strong performance but potential overfitting due to dataset similarities. Further analysis revealed that disinformation networks relied on repetitive messaging and high-retweet amplification, with key influencers shaping discourse. SNA identified coordinated clusters, while SA detected strategic use of neutral and coded language strategically used to reframe discussions. LLMs enhanced detection by recognising subtle framing techniques, including word substitutions and sentiment shifts. Findings showed that disinformation was amplified through coordinated behaviour involving high-profile figures and media accounts. The integration of SA, SNA, and LLMs demonstrated the potential of AI-driven tools in misinformation mitigation. The study highlights the need for a multi-pronged detection approach, combining network analysis, sentiment evaluation, and language modeling, with implications for policymakers, journalists, and researchers.

Keywords: Large Language Models (LLMs), Disinformation, Hate Speech