Existing metadata mainly describes datasets through thematic content and provenance, but provides limited insight into the analytical questions datasets can answer. He adapted a formal question grammar to represent datasets as structured spatio-temporal experiments consisting of measures, controls, and fixed conditions. Datasource questions were converted into parse trees and compared using Tree Edit Distance, and these results were compared with metadata-based similarity derived from sentence embeddings. The results show a moderate correlation between both approaches, indicating that grammar-based representations capture structural dataset relationships not fully reflected in metadata. A user study further showed that participants were able to match datasets to their intended analytical questions significantly above chance level. These findings suggest that grammar-based representations capture structural dataset relations not represented in conventional metadata — an approach that complements conventional metadata and provides a basis for future semantic dataset discovery and source selection.

She developed a workflow using transformer-based language models to extract and disambiguate the spatio-temporal and thematic references in each thesis, linking them to external knowledge bases such as GeoNames and the UNESCO Thesaurus. The resulting SGPL Thesis Knowledge Graph lets researchers ask historiographical questions — such as which topics were studied in which places, and in which decades — directly of the archive.

Geospatial data repositories provide access to large collections of map layers for urban planning, environmental analysis, navigation, and research. However, the usefulness of these layers depends not only on technical accessibility, but also on whether their meaning is explicit enough to support human interpretation and machine-actionability. Existing metadata often describes technical properties such as file format, geometry type, coordinate reference system, or publication information, but these properties do not always explain a layer’s semantic role in GIS analysis — for example, point layers may represent objects, measurement locations, or events, while polygon layers may represent administrative units, homogeneous zones, or contour intervals. Jesse studied how layer-level semantic metadata can be generated more scalably for heterogeneous geodata repositories, more specifically how Large Language Models can support the automatic annotation of geospatial layers with geometry- and entity-level semantic metadata. He evaluated an LLM-based annotation pipeline on 114 map layers from the Maps Amsterdam geodata repository by comparing model-generated annotations with expert-made reference annotations. The results show that LLMs reproduce expert annotations more reliably for geometry-level labels than for entity-level labels; entity classification remains more difficult because several classes require subtle semantic distinctions, especially between objects, lattices, patches, coverages, and contours. These findings suggest that LLM-based knowledge extraction can contribute to scalable semantic metadata enrichment, especially as a supportive method for accelerating annotation and identifying uncertain cases. In a broader perspective, improving layer-level semantic metadata can support data discovery, meaningful GIS analysis, and future automation of geospatial workflows.

Geospatial data is only reusable when its attributes are well described, but this semantic metadata is usually missing and cannot be produced by hand at scale. Quinten used a large language model (LLM) to enrich such metadata with semantic annotation, producing automated labels that can serve as a first pass for manual checking. The target is the attribute’s semantics, expressed as the MeasurementScale of the Core Concept Data (CCD) ontology, predicted from only a column’s name, declared datatype, and a few example values. He compared a naive zero-shot baseline against an error-driven, informed prompting method, with few-shot and chain-of-thought controls. On an expert-labelled reference set, the informed prompts raised accuracy from 79.7% to 98.0%, close to an estimated human-level agreement, with the largest gains coming from concrete confusion patterns rather than formal definitions. Inspecting the model’s disagreements also surfaced mislabelled attributes. In addition, Quinten adapted the pipeline to a larger national data infrastructure, where the method carries over.