Institute of Formal and Applied Linguistics
Charles University, Czech Republic
Faculty of Mathematics and Physics
The vast amount of structured data available holds significant potential, as extracting meaningful insights remains a challenging task. Although Large Language Models (LLM) excel at many tasks, their ability to reason, plan, and generate faithful information is limited. In many cases, the generation task is often underspecified, causing models to make implicit presuppositions about the data, which can lead to errors. As a result, accuracy-critical domains like journalism, medicine, law, and decision support still rely on reliable but rigid and unscalable rule-based methods. Current benchmarks also focus mainly on surface-level information, and even the state-of-the-art models often overlook deeper insights.
Our project aims to develop a neuro-symbolic system that combines the strengths of neural and symbolic AI to generate deeper, more reliable, and explainable insights across diverse data domains. The recent advancements in LLMs enable us to leverage LLM to provide symbolic systems with general knowledge, context, and ontologies for the provided data, retrieve facts through code generation (e.g., SQL, SPARQL, Python), and generate insights that can be logically assessed for soundness (e.g., using Prolog). Our goals are to uncover useful insights, identify factors that make insights “interesting”, explore the role of presuppositions, and improve automatic evaluation metrics for deeper insights.