Neural Computing and Applications Программное обеспечение
LLM-based hierarchical knowledge enhancement with task-differentiated adaptation for few-shot knowledge graph completion
Аннотация
Few-shot knowledge graph completion (FKGC) aims to infer missing facts in a limited number of reference entity pairs, effectively alleviates the impact of long-tail relations. Large language models (LLMs), with their extensive knowledge bases and reasoning capabilities, can make promising potential for mitigating the limitations of previous FKGC methods. However, current research generally lacks domain knowledge related to specific query tasks and uniformly processes heterogeneous query tasks, resulting in the failure of differentiated adaptation. In this work, we introduce the HKEDA framework, which integrates traditional FKGC method and LLM with hierarchical knowledge enhancement and task-differentiated adaptation. Specifically, HKEDA constructs hierarchical knowledge context expression systems for query and candidate entities respectively. Additionally, to achieve differentiated adaptation for query tasks, we design a query difficulty evaluation module based on the dimensions of entity ambiguity and relation polysemy. Finally, the specific task prompts constructed based on the above method are integrated to enhance the performance of LLM-based re-ranking inference for FKGC tasks. Experiments conducted on the NELL-One and Wiki-One datasets highlight the advantages of the proposed HKEDA, it shows 1.1%/4.7% increase in Hits@1 and 1.3%/3.6% increase in MRR compared to the best benchmark model.