Publikation
Advancing NL2SPARQL via Agentic Workflow Design and Dynamic URI Grounding
Franz Pawlus; Justin Weich; Markus Nilles
In: European Semantic Web Conference. European Semantic Web Conference (ESWC-2026), 23rd European Semantic Web Conference, May 12-14, Dubrovnik, Croatia, Pages 457-475, Springer, 2026.
Zusammenfassung
Large Language Models (LLMs) translating natural language to SPARQL (NL2SPARQL) frequently suffer from URI hallucinations. While Retrieval-Augmented Generation mitigates this, common approaches rely on pre-indexing entities, creating scalability bottlenecks for large, live Knowledge Graphs. To address this, we propose an agentic framework that utilizes on-the-fly URI grounding to query the live graph at runtime, eliminating the need for static indices. Unlike rigid execution pipelines, our agent employs an iterative reasoning workflow capable of self-correction. We evaluated the approach on a curated, stratified subset of the DBLP-QuAD benchmark against two baselines: a deterministic pipeline using the same grounding mechanism and a standalone LLM. Results show that while both grounding methods eliminate hallucinations, the agentic architecture significantly outperforms the deterministic pipeline (90.2% vs. 71.6% accuracy). This performance gap is most pronounced on complex queries, demonstrating that adaptive agentic reasoning is essential for providing a scalable, index-free solution for NL2SPARQL.