Supervisor: Ilaria Tiddi (i.tiddi@vu.nl)
Large-scale knowledge graphs (KGs) are powerful tools for representing structured information, but they often lack common-sense knowledge about the physical properties of everyday objects, such as size, color, material, and texture. These properties are critical for many applications, including robotics and AI systems that interact with the physical world. This project aims to investigate the extent to which physical object properties are represented in large-scale KGs, like Wikidata, DBpedia or ConceptNet, and explore methods to enhance their coverage. We will explore the creation of a distilled dataset that specifically includes common-sense knowledge about physical attributes.
The project will involve analyzing several prominent knowledge graphs (WikiData, ConceptNet, YAGO, DBpedia, WordNet) to assess their coverage of physical properties related to everyday objects. This will include identifying the relevant object classes and attributes, and extracting data related to these properties. You will then develop a methodology to refine and distill this information into a clean, structured dataset that contains essential physical attributes (e.g., size, color, material). The dataset will be evaluated for completeness and accuracy, and potentially augmented with external sources if necessary . (Finally, the project will investigate how this distilled knowledge can be integrated into existing KGs for downstream applications, which will be done by us).
RQ: To what extent do large-scale knowledge graphs capture common-sense physical properties of objects, and how can we create a distilled dataset that accurately represents these attributes?