大语言模型和知识图谱协同的跨域异质数据查询框架

Summary

The paper addresses querying data that lives in several related but separate domains. Such data comes in different formats and quality levels and often cannot be pooled centrally for privacy reasons, which makes it hard for large language models to use. The authors propose a three-stage pipeline that pairs an LLM with a knowledge graph (KG).

1. Integration. Format-specific OSLC adapters parse structured, semi-structured and unstructured sources, assign unified identifiers, store content as JSON-LD linked data, and build a KG from the result.
2. Restructuring. The HKGE algorithm groups same-named, same-shaped records from different domains into "homologous" subgraphs. It then converts the graph into a knowledge line graph, where each node stands for a triple, to speed up retrieval.
3. Querying. The LLM extracts query nodes from a user request. TrustHKGM then scores source confidence, detects conflicts and discards low-quality nodes. MKLGP feeds the remaining trusted subgraph back to the LLM as a prompt to produce the answer.

Experiments on four public datasets cover conflict-confidence estimation and multi-domain querying. The authors report better effectiveness and efficiency than state-of-the-art baselines. The work matters for trustworthy, cost-aware LLM question answering over siloed enterprise data.

Key points

• Publication and funding. Published in 计算机研究与发展 (Journal of Computer Research and Development), vol. 62(3), pp. 605–619, 2025 (DOI 10.7544/issn1000-1239.202440634). The work is funded by civil space technology pre-research project D020101.
• Three named algorithms:
  • HKGE (homogeneous knowledge graph extraction) builds homologous subgraphs in O(n log n) time.
  • TrustHKGM performs trusted candidate subgraph matching.
  • MKLGP performs multi-domain knowledge line graph prompting.
• Knowledge line graph. Each triple of the original KG becomes a node. Two nodes are linked when their triples share an entity.
• Cross-domain homologous data. Records count as homologous when they come from different domains but share a name and a metadata shape. Isolated nodes with no homologous counterpart are still kept in the line graph.
• Results. The evaluation uses four public datasets. The extracted text omits the experimental section, so specific scores and dataset names are not available here.