Researchers have developed a framework called HKEDA that combines traditional few-shot knowledge graph completion techniques with large language models to improve factual inference for sparse relations. Few-shot knowledge graph completion is designed to infer missing facts from a limited number of reference entity pairs, helping alleviate the impact of long-tail relations across complex relational networks.
Existing methods often lack domain knowledge pertinent to specific query tasks and tend to process heterogeneous queries uniformly without task-differentiated adaptation. Large language models possess extensive knowledge bases and advanced reasoning capabilities, offering potential solutions to overcome these constraints when properly integrated into knowledge completion architectures.
To address these challenges, the HKEDA framework constructs hierarchical knowledge context expression systems independently for query entities and candidate entities. It also incorporates a query difficulty evaluation module structured around the dual dimensions of entity ambiguity and relation polysemy to achieve differentiated adaptation across varied query tasks.
The system integrates these tailored task prompts to enhance the performance of language-model-based re-ranking inference during completion tasks. In experimental evaluations conducted on the NELL-One and Wiki-One benchmark datasets, HKEDA demonstrated marked improvements over existing baselines, achieving a 1.1% increase on NELL-One and a 4.7% increase on Wiki-One in Hits@1, alongside corresponding improvements of 1.3% and 3.6% in Mean Reciprocal Rank compared to the best benchmark model.
Subject of Research: Few-shot knowledge graph completion using large language models
Article Title: LLM-based hierarchical knowledge enhancement with task-differentiated adaptation for few-shot knowledge graph completion
Article References: Chen, L., Hu, Z., Hou, W., & Kong, D. (2026). LLM-based hierarchical knowledge enhancement with task-differentiated adaptation for few-shot knowledge graph completion. Neural Computing and Applications, 38(22), Article 901. https://doi.org/10.1007/s00521-026-12625-8
Image Credits: AI Generated
DOI: 10.1007/s00521-026-12625-8
Keywords: knowledge graphs, few-shot learning, large language models, HKEDA framework, knowledge graph completion, hierarchical knowledge enhancement, task-differentiated adaptation, long-tail relations, entity ambiguity, relation polysemy, re-ranking inference, query difficulty evaluation, NELL-One, Wiki-One, Hits@1, Mean Reciprocal Rank
News Source: Denise Maddox. (November 21, 2026). HKEDA Framework Enhances Few-Shot Knowledge Graph Completion Using Language Models. Scienmag.



