Identifying disinformation narratives: a systemic approach

Abstract

Disinformation narratives represent one of the most challenging threats to the contemporary information ecosystem, particularly in the Republic of Moldova, which has been the target of coordinated influence campaigns documented by DFRLab, Recorded Future, and EUvsDisinfo. Unlike isolated fake news, narratives operate through repetition and coordinated distribution across multiple channels simultaneously, making them difficult to detect through traditional methods. This paper describes the architecture of the RealInfo system, an automatic narrative detection platform built on a hybrid three-layer pipeline: local linguistic analysis based on rhetorical pattern matching, emotional classification through pre-trained models, and narrative analysis through RealInfo LLM - a 14-billion-parameter model running locally on a distributed GPU cluster. The system integrates a semantic similarity module for detecting coordinated narratives across multiple sources. The proposed architecture demonstrates that a multi-layer hybrid approach is superior to any isolated method, while providing complete explainability of the decision.