Probabilistic qualitative spatial reasoning with applications to GeoQA

This paper explores the use of probabilistic and conventional qualitative spatial reasoning (QSR) in the context of geospatial question answering (GeoQA) systems. The paper presents a thorough empirical investigation of the performance of a probabilistic and a conventional qualitative spatial reasoner, across a range increasingly sophisticated scenarios with real data and synthetically generated questions. The results indicate the potential of probabilistic QSR to provide more detailed information about spatial configurations than conventional QSR; but at the cost of less frequent errors in estimating the relative likelihood of different reasoning conclusions. Errors in probabilistic reasoning also tend to be systematically associated with lower probability conclusions. The results have implications for reliable and flexible automated spatial reasoning systems, especially where neither conventional geographic information retrieval (GIR) techniques nor large language models (LLMs) are able to provide a satisfactory solution to GeoQA problems.