Meticulously Unsupervised Entity Alignment With Large Language Models

Zhihuan Yan

doi:10.1111/exsy.70406

Return

Meticulously Unsupervised Entity Alignment With Large Language Models

2026-09-30

PRE

DOI：10.1111/exsy.70406

Original

Original request for help

Save

Abstract

En 中文

Knowledge graph entity alignment refers to the process of identifying and linking entities that refer to the same real-world object from different knowledge graphs. Structural heterogeneity and scarcity of training data have always been two major challenges that impede entity alignment task. The advent of Large Language Models presents new avenues for entity knowledge completion and unsupervised EA, inspired by their extensive background knowledge and comprehensive capability to process semantic information. However, it is nontrivial to directly apply Large Language Models in addressing the aforementioned two challenges due to the following reasons: (1) it will blindly enrich entity knowledge in the absence of appropriate constraints; (2) as a probability based model, it could generate noisy labels that may mislead the alignment. To this end, this paper presents a novel unsupervised entity alignment framework, named LLM-Align, to effectively leverage Large Language Models to precisely annotate entity pairs. First, Constrained Entity Information Enrichment (CIE) technique is devised, which employs attributes and relationships existing in the KGs to constrain the completion process, which alleviates structural heterogeneity between aligned entities. Subsequently, a Code-formatted Prompt Template (CPT) module was designed to assist the Large Language Models in labelling entity pairs from candidates obtaining via both semantic and structural similarities. Ultimately, Combinatorial Optimization method based entity pair Refinement (COR) technique was conceived to further enhance the quality of the labelled entity pairs, which are used to train a base EA model. Extensive experiments on various benchmark datasets demonstrate the advantages of LLM-Align.

Keywords:

entity alignment

entity information enrichment

entity pair refinement

entity-code encoding

LLM-based entity pair labelling