ARElight

The escalating volume of unverified user-generated content across social media has intensified the spread of mis and disinformation, undermining public trust in online information sources. Particularly on platforms like X/Twitter, hostile actors can quickly propagate manipulated narratives, fueling confusion during critical events such as elections or public health crises. Traditional topic-modelling approaches (e.g., LDA) have been widely used to map and summarise these discussions, yet they often fail to capture the nuanced relationships and sentiments between key entities — critical signals for identifying coordinated disinformation campaigns. 

In this work, we propose a novel graph-based topic-modelling framework to address these limitations by representing not only the most salient terms and named entities but also the directed and sentiment-rich relationships between them. As such, it becomes feasible to identify unusual patterns, such as a sudden surge in negative sentiment linking specific actors or issues — a common hallmark of disinformation. Furthermore, the system supports intuitive graph operations (e.g., union, intersection, difference) that allow investigators to compare different timelines or user communities, revealing shifts or overlaps in suspect narratives. A publicly available prototype (ARElight) demonstrates the method’s ability to handle social media text at scale, offering interactive graph visualisations to facilitate human-driven investigations into potential disinformation activities.