Three-Stage Extraction of Spatial Relationships Using Markers

We present SpaRel – a system for Spatial Relation Extraction (SRE), which is focused on intra-sentence Spatial Relation Extraction. It was evaluated on three datasets: English (CLEF2017 mSpRL, SpaceEval) and Polish (PST2.0). We propose a novel procedure for joint relation elements discovery and classification. It significantly reduces the number of possible triplets checked for representing relation instances. This results in significant improvement in the quality and speed of the relation extraction process. To represent instances of relation instances, we expanded approaches to Relation Extraction, but we tried to extract limited amount of the necessary intermediate information. We propose a special procedure for resolving issues of one spatial indicator involved in many relations. The best results were achieved with the DeBERTa v3 (large) language model. We also studied SRE from the point of evaluation reliability and reproducibility of the previous approaches. Several issues related to the SpaceEval dataset were identified and their potential impact on results was discussed. Problems with one of the SOTA systems were identified and discussed. In a broader context, we have identified and discussed all the factors that hamper comparison of results in Relation Extraction.