Elamur Research⌕
☰
home / papers / 2026 / The Structure-Content Trade-off in Knowledge Graph Retrieval: A Diagnostic Study of Question Decomposition

The Structure-Content Trade-off in Knowledge Graph Retrieval: A Diagnostic Study of Question Decomposition

Paper · 2026 · Valentin Six, Gaël de Chalendar and Evan Dufraisse

Summary

Question decomposition is a common way to handle multi-hop questions when LLMs answer over knowledge graphs. However, it is unclear how splitting a question changes the shape of the evidence that gets retrieved. The authors do not propose a new system. Instead, they run a diagnostic study.

They define a similarity score that blends two signals through a weight α in [0, 1]:

  • how close graph node and edge labels are to the original question, and
  • how close they are to the current subquestion, which is enriched with answers to earlier subquestions.

The best-scoring nodes and edges receive prizes. A Prize-Collecting Steiner Tree (PCST), borrowed from G-Retriever, then extracts one connected subgraph per subquestion. The merged result is deliberately left unrepaired.

On CWQ and WebQSP, a stable trade-off appears. Favouring subquestions makes the retrieved content more precise but breaks the graph into disconnected fragments. Favouring the full question keeps the graph coherent but lowers relevance. QA accuracy peaks at intermediate α values. The study argues that KG retrieval should be evaluated on structure as well as relevance.

Key points

  • Retrieval focus formula: sim_α = (1 − α)·cos(z, z_Q) + α·cos(z, z_q). α = 0 uses only the original question, α = 1 only the subquestion, and the default is α = 0.5.
  • Subgraph extraction: PCST is applied per subquestion. Top-k nodes and edges receive non-zero prizes, with defaults of 3 nodes and 5 edges and an edge cost of 0.5; CWQ preprocessing uses 5 nodes and 7 edges. Connectivity therefore emerges naturally rather than being forced post hoc.
  • Benchmarks: CWQ (first 1000 test examples) and WebQSP, both built on Freebase.
  • Models and resources:
    • Default setup: a graph_llm pipeline with a frozen LLaMA 2 7B backbone, a GNN encoder, and all-roberta-large-v1 embeddings.
    • LLaMA 2 13B and DeepSeek-R1-Distill-Qwen-32B also appear among the listed resources.
  • Code: released at github.com/cea-list-lasti/kg-retrieval-tradeoff.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓Georgia Institute of TechnologyAtlanta · United States133 works · 24,497 citations🎓Université Paris-SaclayGif-sur-Yvette · France45 works · 6,127 citations🏛️Commissariat à l'Énergie Atomique et aux Énergies AlternativesParis · France11 works · 416 citations🏛️Laboratoire d'Intégration des Systèmes et des TechnologiesPalaiseau · France8 works · 321 citations🔬CEA Paris-Saclay - Etablissement de SaclaySaclay · France4 works · 207 citations

 Fields, topics and concepts

📚 Shelves 5The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 worksInformation Systems · 1,437 works

🧭 Topics 2Research subjects and languages the work is about.Information retrieval & search · 473 worksKnowledge graphs & ontologies · 192 works

Cite

Cite as Valentin Six, Gaël de Chalendar and Evan Dufraisse. 2026. The Structure-Content Trade-off in Knowledge Graph Retrieval: A Diagnostic Study of Question Decomposition. Proceedings of the Knowledge Graphs and Large Language Models Workshop (KG-LLM) @ LREC26.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.63317/4aamore77jty

@inproceedings{six2026structure,
  title = {{The Structure-Content Trade-off in Knowledge Graph Retrieval: A Diagnostic Study of Question Decomposition}},
  author = {Valentin Six, Gaël {de Chalendar} and Evan Dufraisse},
  booktitle = {Proceedings of the Knowledge Graphs and Large Language Models Workshop (KG-LLM) @ LREC26},
  year = {2026},
  doi = {10.63317/4aamore77jty},
  url = {https://doi.org/10.63317/4aamore77jty}
}

References

32 references, 7 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Daniel Bienstock, Michel X Goemans, David Simchi-Levi, and David Williamson. 1993. A note on the prize collecting traveling salesman problem. Mathematical programming, 59(1):413– 420. find
  2. Kurt Bollacker et al. 2008. Freebase: a collaboratively created graph database for structuring human knowledge. In Proc. ACM SIGMOD, pages 1247–1250. find
  3. Chi-Min Chan et al. 2024. Rq-rag: Learning to refine queries for retrieval augmented generation. arXiv preprint arXiv:2404.00610. find
  4. Xinbang Dai et al. 2025. Large language models can better understand knowledge graphs than we thought. Knowledge-Based Systems, 312:113060. find
  5. Wentao Ding, Jinmao Li, Liangchuan Luo, and Yuzhong Qu. 2024. Enhancing complex question answering over knowledge graphs through evidence pattern retrieval. In Proceedings of the ACM Web Conference 2024, pages 2106–2115. find
  6. Na Dong, Natthawut Kertkeidkachorn, Xin Liu, and Kiyoaki Shirai. 2025. Refining noisy knowledge graph with large language models. In Proceedings of the Workshop on Generative AI and Knowledge Graphs (GenAIK), pages 78–86. find
  7. Ruiliu Fu et al. 2021. Decomposing complex questions makes multi-hop QA easier and more interpretable. In Findings of EMNLP, pages 169–180. find
  8. Xiaoxin He et al. 2024. G-retriever: Retrieval-augmented generation for textual graph understanding and question answering. Advances in Neural Information Processing Systems, 37:132876–132907.
  9. Lei Huang et al. 2025. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. ACM Transactions on Information Systems, 43(2):1–55.
  10. Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Ye Jin Bang, Andrea Madotto, and Pascale Fung. 2023. Survey of hallucination in natural language generation. ACM computing surveys, 55(12):1–38.
  11. Ehsan Kamalloo, Nouha Dziri, Charles Clarke, and Davood Rafiei. 2023. Evaluating open-domain question answering in the era of large language models. In Proceedings of the 61st annual meeting of the Association for Computational Linguistics (volume 1: long papers), pages 5591–5606.
  12. Petri Kotiranta, Marko Junkkari, and Jyrki Nummenmaa. 2022. Performance of graph and relational databases in complex queries. Applied sciences, 12(13):6490. find
  13. Mufei Li, Siqi Miao, and Pan Li. 2024a. Simple is effective: The roles of graphs and large language models in knowledge-graph-based retrieval-augmented generation. arXiv preprint arXiv:2410.20724. — [arXiv:2410.20724](https://arxiv.org/abs/2410.20724) find
  14. Yading Li et al. 2024b. A framework of knowledge graph-enhanced large language model based on question decomposition and atomic retrieval. In Findings of EMNLP, pages 11472–11485. find
  15. Yixin Liu, Kejian Shi, Katherine S He, Longtian Ye, Alexander R Fabbri, Pengfei Liu, Dragomir Radev, and Arman Cohan. 2024. On learning to summarize with large language models as references. arXiv preprint arXiv:2305.14239.
  16. Linhao Luo et al. 2023. Reasoning on graphs: Faithful and interpretable large language model reasoning. arXiv preprint arXiv:2310.01061. find
  17. Meta AI. 2023a. LLaMA 2 13B. Meta AI. Hugging Face. PID https://huggingface.co/meta-llama/Llama-2-13b-hf. Large language model (13B parameters), Hugging Face version. ↗
  18. Tobias Aanderaa Opsahl. 2024. Fact or fiction? improving fact verification with knowledge graphs through simplified subgraph retrievals. In Proceedings of the Seventh Fact Extraction and VERification Workshop (FEVER), pages 307– 316. find
  19. Meta AI. 2023b. LLaMA 2 7B. Meta AI. Hugging Face. PID https://huggingface.co/meta-llama/Llama-2-7b-hf. Large language model (7B parameters), Hugging Face version. ↗
  20. Sentence-Transformers. 2025. all-roberta-largev1. Sentence-Transformers project. Hugging Face. PID https://huggingface.co/sentence-transformers/all-roberta-large-v1. Neural sentence embedding model based on RoBERTa. ↗
  21. Ciyuan Peng, Feng Xia, Mehdi Naseriparsa, and Francesco Osborne. 2023. Knowledge graphs: Opportunities and challenges. Artificial intelligence review, 56(11):13071–13102. find
  22. Ethan Perez et al. 2020. Unsupervised question decomposition for question answering. arXiv preprint arXiv:2002.09758.
  23. Talmor et al. 2018. Complex Web Questions (CWQ). Tel-Aviv University. PID http://nlp.cs.tau.ac.il/compwebq. Question answering dataset over Freebase. ↗
  24. Sangjin Shin and Kyong-Ho Lee. 2020. Processing knowledge graph-based complex questions through question decomposition and recomposition. Information sciences, 523:234–244. find
  25. Yih et al. 2016. WebQSP. Microsoft Research. PID https://www.microsoft.com/en-us/research/publication/the-value-of-semantic-parse-labeling-for-knowledge-base-question-answering-2/. Question answering dataset over Freebase. ↗
  26. Wenqing Wu, Zhenfang Zhu, Jiangtao Qi, Wenling Wang, Guangyuan Zhang, and Peiyu Liu. 2023. A dynamic graph expansion network for multi-hop knowledge base question answering. Neurocomputing, 515:37–47. find
  27. Sohee Yang, Elena Gribovskaya, Nora Kassner, Mor Geva, and Sebastian Riedel. 2024. Do large language models latently perform multi-hop reasoning? URL https://arxiv.org/abs/2402.16837. ↗
  28. This appendix documents the default runtime configuration used in our codebase. Unless explicitly stated otherwise in the main text, experiments use these defaults. For all reported results, we select checkpoints by validation loss and evaluate the best checkpoint on the test set. find
  29. Michihiro Yasunaga et al. 2021. Qa-gnn: Reasoning with language models and knowledge graphs for question answering. In Proc. NAACL-HLT, pages 535–546. find
  30. Biao Zhang, Barry Haddow, and Alexandra Birch. 2023. Prompting large language model for machine translation: A case study. In International conference on machine learning, pages 41092– 41110. PMLR.
  31. Xiaohan Zou. 2020. A survey on application of knowledge graph. In Journal of Physics: Conference Series, volume 1487, page 012016. IOP Publishing. 10. find
  32. Language Resource References Default Hyperparameters and Experimental Setting A.1. Command-Line Defaults find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/six-2026-structure-content-trade-off-knowledge/