GRISP: Guided Recurrent IRI Selection over SPARQL Skeletons
A small language model generates SPARQL skeletons and iteratively selects knowledge graph items...
Researchers Sebastian Walter and Hannah Bast have introduced GRISP (Guided Recurrent IRI Selection over SPARQL Skeletons), a novel method for answering natural-language questions over knowledge graphs using a small language model (SLM). The approach first generates a natural-language SPARQL query skeleton via the SLM, then iteratively replaces placeholders by re-ranking and selecting knowledge graph items based on constraints. The SLM is jointly trained on skeleton generation and list-wise re-ranking data derived from standard question-query pairs.
GRISP was evaluated on common Wikidata and Freebase benchmarks, achieving better results than other state-of-the-art methods in a comparable setting. This work demonstrates that small language models can effectively handle complex SPARQL query generation and knowledge graph navigation, offering a more efficient alternative to larger models for structured data retrieval tasks. The paper is available on arXiv under arXiv:2604.21133.
- GRISP uses a small language model (SLM) to generate SPARQL skeletons from natural-language questions.
- The method iteratively re-ranks and selects knowledge graph items to replace placeholders using constraints.
- Outperforms state-of-the-art methods on Wikidata and Freebase benchmarks in a comparable setting.
Why It Matters
GRISP shows SLMs can efficiently navigate knowledge graphs, enabling lighter, more accessible structured data QA.