Exploring Romanian Misinformation with GraphRAG: Community-Level Narratives of COVID-19

Predefined thematic taxonomies can obscure how misinformation narratives emerge, overlap, and change over time. One way to address this limitation is by developing bottom-up data-driven approaches that identify narratives directly from data. We present a pipeline that extracts entities and relations from news articles using GraphRAG, identifies communities using the Leiden algorithm, and uses this information to explore and predict misinformation. We evaluate the pipeline on a set of Romanian COVID-19 pandemic news articles with expert-defined narrative annotations and adapt the extraction prompts to Romanian. Exploratory results show that in this corpus, the graph for fake news is more fragmented and less modular than the graph for real news; it also contains higher proportions of events, disinformation theory, and vaccine entities, whereas the real-news graph contains higher proportions of locations, hospitals, and public health measures. Communities in the combined graph align weakly with the predefined super-narrative labels, indicating that the two representations group articles differently. Furthermore, expert review of similar articles from communities identified within the combined real–fake graph suggests that an approach based on entity and community may offer a promising avenue for identifying subtle misinformation strategies. Regarding the prediction task, a three-component ensemble that uses graph information achieves 85.86% macro-F1 (95% CI [78.50, 91.27]) and 86.51% accuracy (95% CI [78.57, 91.27]) on classifying articles as fake or true. These point estimates suggest incremental gains over the two-component ensemble without the graph similarity signal, though these gains are not statistically confirmed at this sample size. Together, these exploratory results illustrate the use of GraphRAG as a tool in pipelines that support the exploration and prediction of misinformation.