Marcos Paulo Silva Gôlo, Paulo Viviurka do Carmo, Edgard Marx, Ricardo Marcondes Marcacini · · 2026
DOI: 10.5753/kdmile.2026.30058
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
Large language models have made notable progress in translating natural-language questions into SPARQL queries for knowledge graphs. Recent studies from the first Text2SPARQL challenge have explored zero-shot and few-shot scenarios, as well as in-context learning techniques. With state-of-the-art (SoTA) results, the challenge exposes significant gaps, including limited use of open-source models and a scarcity of open-source fine-tuning, especially for reasoning. In this sense, we propose a new method for the Text2SPARQL task that combines pre-fine-tuning with explicit reasoning and a subsequent reward-based fine-tuning stage. First, we perform a fine-tuning to generate both a reasoning chain and the corresponding SPARQL query. Next, we apply reinforcement learning guided by reward functions. Our experiments show that the proposed model outperforms all baselines and six SoTA methods.
No comments yet — start the discussion below.