Leveraging Data Shapes in Large Language Model Contexts for Question Answering on Public and Private Knowledge Graphs

Knowledge Graph Question Answering aims to make structured semantic data accessible through natural language interfaces. While recent Large Language Models can generate SPARQL queries from natural language questions, their effectiveness is limited by their reliance on prior exposure to vocabularies and datasets during pretraining. This work explores how Knowledge Graph specific schema information, so called shape constraints, can be used to provide Large Language Models with context about the dataset to be queried, enabling more accurate and generalizable query generation. We show that ShEx-augmented prompting achieves an F1-score of 0.28 on unseen knowledge graphs, compared to a baseline score of 0.00, demonstrating its ability to generalize beyond the training distribution. A pipeline is developed that integrates knowledge graph data shape extraction, prompt construction, and automatic SPARQL validation. Experiments were conducted on benchmarks with public and proprietary data sets using different state-of-the-art Large Language Models and demonstrate that shape-informed prompting improves the execution accuracy of generated queries. ... mehrThese results suggest that augmentation with data shapes can reduce the dependency on Large Language Model pre-training or fine-tuning on a specific dataset and offer a practical pathway toward domain-agnostic, robust Knowledge Graph Question Answering systems.