Generating Activity Definitions with Large Language Models

Composite activity recognition systems reason over streams of low-level, symbolic events in order to detect instances of composite activities, based on their formal definitions. Composite activity definition construction is a highly involved task, as these definitions often involve numerous spatio-temporal constraints, while labels that may be used for learning them automatically are hard to obtain. To address this issue, we propose a method that generates composite activity definitions from natural language descriptions using pre-trained Large Language Models (LLMs). In order to assess the quality of LLM-generated definitions, we propose a novel similarity metric, which reflects the human effort required to correct them. We present a thorough experimental evaluation of our approach on the maritime domain, demonstrating its effectiveness.