Prompting LLMs for the Run-Time Event Calculus (Short Paper)

Andreas Kouvaras, Periklis Mantenoglou, Alexander Artikis

Abstract: Composite activity recognition systems analyse streams of low-level, symbolic events to identify instances of complex activities based on their formal definitions. Crafting these definitions is a challenging task, as it often requires specifying intricate spatio-temporal constraints, and acquiring labeled data for automated learning is difficult. To address this challenge, we introduce a method that leverages pre-trained Large Language Models (LLMs) to generate composite activity definitions, in the language of the Run-Time Event Calculus, from natural language descriptions.