UniQ-Gen: Unified Query Generation Across Multiple Knowledge Graphs

Generating SPARQL queries is crucial for extracting relevant information from diverse knowledge graphs. However, the structural and semantic differences among these graphs necessitate training or fine-tuning a tailored model for each one. In this paper, we propose UniQ-Gen, a unified query generation approach to generate SPARQL queries across various knowledge graphs. UniQ-Gen integrates entity recognition, disambiguation, and linking through a BERT-NER model and employs cross-encoder ranking to align questions with the Freebase ontology. We conducted several experiments on different benchmark datasets such as LC-QuAD 2.0, GrailQA, and QALD-10. The evaluation results demonstrate that our approach achieves performance equivalent to or better than models fine-tuned for individual knowledge graphs. This finding suggests that fine-tuning a unified model on a heterogeneous dataset of SPARQL queries across different knowledge graphs eliminates the need for separate models for each graph, thereby reducing resource requirements.