Automatically generated place descriptions for accurate location identification: a hybrid approach with rule-based methods and LLM

Abstract

This paper explores the potential of using machine-generated descriptions to characterize a place in a way that humans can identify it. It presents a hybrid approach for generating place descriptions by combining rule-based generation of spatial relation facts with a LLM that converts these facts into natural language descriptions. The study focuses on urban areas and street-level scale, using OpenStreetMap as the primary data source. The rule-based method is informed by an experimental study that analyzed human-made place descriptions to understand reference object types used, their quantities, distances, and spatial relations. An experiment is carried out to assess the quality of machine-generated descriptions compared to human-made descriptions in a place identification task. The evaluation involved 70 participants identifying locations based on both human and machine-generated descriptions across a 200-hectare urban area. The results show that the same average identification accuracy was not achieved. However, the proposed method reached lower variance and the difference in accuracy is not substantial enough to impede place identification in the anticipated use cases. The method shows promise for applications in navigation systems, virtual assistants, and location-based services, particularly in situations where visual media cannot be used.