Research
Research focus
- We work on timely problems: building state-of-the-art models for analyzing and tracking global threats:
- Analysis of propaganda and disinformation
- We organize:
- Tracking outbreaks of epidemics in global media:
- Collaboration with MediSys: European Commission’s Joint Research Centre
- PULS information extraction from text media.
- We work on fundamental problems in LLMs — model explainability, confidence, and sustainability — please see Publications
Detecting Propaganda and Persuasion Techniques in Media
Rapid advances in technology make possible the easy establishment of direct communication channels between producers and consumers of information. This raises the potential to expose the latter to deceptive content and mass manipulation. Huge audiences are affected online, and major crisis events are subjected to the spread of harmful disinformation and propaganda.
We develop methods for supporting users in analyzing the media ecosystem, and characterizing the emerging manipulation attempts. The goal is to learn about propaganda and disinformation methods, and how they produce harmful effects.
Knowledge-Aware and Retrieval-Augmented Language Models
We develop methodologies for multilingual global media monitoring that produce reliable real-time alerts on an open set of event types that indicate global threats — including:
- pandemic outbreaks,
- terrorist activity,
- socio-political unrest,
- human trafficking, arms smuggling,
- natural and man-made disasters.
LLMs offer unparalleled quality of analysis, but are trained on “legacy” data, which — by design — is not up-to-date. We explore how the LLM can draw on external knowledge bases and live up-to-date media streams. To achieve these goals, we leverage High Performance Computing (HPC) to build and refine LLM methods, and a critical mass of international collaborators who bring complementary expertise and resources.
AI-Enhanced Knowledge Management for Business Applications
We work on state-of-the-art LLMs for processing large collections of documents, with practical focus and clear societal or business impacts. The aim is to combine structured data with unstructured data, i.e., internal company documents and knowledge bases that to make tailored applications for specific business needs.