EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

STAR-RAG: Structure-Aware Graph Retrieval for Retrieval-Augmented Generation on Knowledge Graphs

Zhenshu Zhong, Kan Chen, Yechen Yin, Song Ming, Qian Sun

Electronics · 13.11.2026 · 15 (22) · 5289

DOI: 10.3390/electronics15225289 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Electronics, online erschienen am 13.11.2026. Verfasst von Zhenshu Zhong und 4 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: National Natural Science Foundation of China.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Knowledge graph retrieval-augmented generation (KG-RAG) provides structured evidence to support large language models (LLMs) in complex question answering. However, multi-hop graph expansion introduces weakly relevant relations, making it challenging to identify answer candidates supported by both question semantics and graph connectivity. Although existing graph retrievers exploit relational and structural information, effectively coordinating semantic message propagation, multi-order topological information, and candidate-level scoring remains challenging. To address these issues, we propose STAR-RAG, a structure-aware graph retrieval framework for multi-hop question answering. Specifically, Semantic noise routing (SNR) employs question-conditioned gates to regulate semantic messages before aggregation. Decoupled topological propagation (DTP) explicitly extracts directed degree and multi-order connectivity features and propagates structural information through a separate stream. Semantic-structural adaptive fusion (SSAF) then adaptively combines semantic and structural scores to rank answer candidates. The ranked candidates are connected to topic entities through evidence paths, which are provided to an LLM for answer generation. Experiments on WebQSP and CWQ show that STAR-RAG achieves retrieval Hit@1/F1 scores of 78.0/73.4 and 55.4/52.4, respectively. With Llama2-7B, generation Hit/F1 reaches 85.8/71.9 on WebQSP and 70.5/62.3 on CWQ. Ablation studies and diagnostic analyses further demonstrate the contributions of semantic routing, multi-order structural modeling, and adaptive fusion to retrieval performance.

Förderung

  • National Natural Science Foundation of China

Zitieren

APA

Zhong, Z., Chen, K., Yin, Y., Ming, S., & Sun, Q. (2026). STAR-RAG: Structure-Aware Graph Retrieval for Retrieval-Augmented Generation on Knowledge Graphs. Electronics, 15(22), 5289. https://doi.org/10.3390/electronics15225289

Vancouver

Zhong Z, Chen K, Yin Y, Ming S, Sun Q. STAR-RAG: Structure-Aware Graph Retrieval for Retrieval-Augmented Generation on Knowledge Graphs. Electronics. 2026;15(22):5289. doi:10.3390/electronics15225289

BibTeX
@article{Zhong2026,
  author = {Zhong, Zhenshu and Chen, Kan and Yin, Yechen and Ming, Song and Sun, Qian},
  title = {STAR-RAG: Structure-Aware Graph Retrieval for Retrieval-Augmented Generation on Knowledge Graphs},
  journal = {Electronics},
  year = {2026},
  volume = {15},
  number = {22},
  pages = {5289},
  doi = {10.3390/electronics15225289}
}
Kostenlos starten →