Detecting and Tracking the Spread of Debunked Narratives Across Languages

Misinformation and disinformation during critical events, like the COVID-19 pandemic and geopolitical conflicts such as the Ukraine war, poses threats to public perception, social cohesion, and political stability. While fact-checkers strive to counter their spread, a multifaceted problem emerges: the enduring and widespread propagation of similar or nearly duplicate false narratives across multiple languages, modalities, and social media platforms, often persisting long after the initial debunking by a professional fact-checker.

First, this thesis utilises the CoronaVirusFacts Alliance database to identify and uncover repeatedly debunked false narratives related to COVID-19. The spatiotemporal analysis indicates the global prevalence of false narratives related to general medical advice, consistently shared by Facebook users despite the existence of fact-checks that have already debunked similar narratives across different languages. Additionally, the thesis analyses debunks related to the Ukraine conflict, revealing the wider spread of disinformation compared to its debunks and demonstrating the delayed but positive impact of debunks on reducing Ukraine-related disinformation. The thesis ultimately advocates for the implementation of a cross-lingual debunked narrative search tool in the fact-checking pipeline to efficiently identify previously debunked narratives in different languages.

Motivated by the challenges posed by the persistent spread of debunked narratives, this thesis delves into cross-lingual debunked narrative retrieval, aiming to enhance the performance and robustness of retrieval models across various languages. Firstly, it introduces the Multistage BiCross encoder for multilingual access to COVID-19 information, presenting experimental results and search query optimisation techniques. Subsequently, the thesis introduces novel benchmark datasets and computational methods to aid fact-checkers in detecting debunked narratives across multiple languages. It also emphasises the need for social media platforms to adopt similar technologies at scale to optimise fact-checker resources. Finally, the thesis proposes unsupervised methods for training debunked narrative retrieval models, offering effective real-time adaptation without relying on time-consuming and labour-intensive human annotations.

In summary, the research contributes to a comprehensive understanding of the spread of debunked narratives. It offers practical solutions and insights that can inform policy decisions and contribute to the ongoing global efforts against misinformation and disinformation.