Research

Knowledge-Aware and Retrieval-Augmented Language Models

The project aims at developing large-scale algorithms for highly-multilingual media monitoring that will produce reliable real-time alerts on an open set of event types indicating global threats—including:

  • pandemic outbreaks,
  • terrorist attacks,
  • socio-political unrest,
  • human trafficking, smuggling of arms,
  • natural and man-made disasters,
  • etc.

Objectives

The objectives in this theme are:

  1. We pursue a highly timely and topical application: designing state-of-the-art large language models (LLMs) for detecting events in the context of global social threats, with clear societal impacts.
  2. We explore fundamental problems in LLMs—in the context of this application, namely: model explainability, confidence, and sustainability.

LLMs offer unparalleled quality of analysis, but are trained on “legacy” data, which — by construction — is not up-to-date. We explore how the LLM can draw on external knowledge bases and live media streams — which are up-to-date. To achieve these goals, we leverage High Performance Computing (HPC) to build and refine LLM methodologies, and a critical mass of international collaboration, with collaborators who bring complementary expertise and resources.

Detecting Propaganda and Persuasion Techniques in Media

Advancing technology opens vast possibilities to easily create communication channels between information producers and consumers. This creates 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, to analyze the media ecosystem and characterize manipulation attempts. The goal is to identify what methods the propaganda is using, and what kind of harmful effect it aims to achieve.

Seminar

…