Two Heads Are Better Than One: Integrating Knowledge from Knowledge Graphs and Large Language Models for Entity Alignment

TL;DR

This work tackles cross-KG entity alignment by integrating structural knowledge from knowledge graphs with semantic knowledge embedded in large language models. The proposed LLMEA framework uses a relation-aware graph attention network to learn robust structural embeddings, filters candidate alignments through both structure and name signals, and leverages LLMs to generate a virtual equivalent and to answer multi-round multiple-choice prompts for final alignment. Empirical results on three multilingual DBpedia datasets demonstrate state-of-the-art performance, with ablations confirming the additive value of each candidate set and the LLM-driven prediction stage. The study highlights the practical potential and challenges of combining KG reasoning with LLM inference for scalable, accurate knowledge integration.

Abstract

Entity alignment, which is a prerequisite for creating a more comprehensive Knowledge Graph (KG), involves pinpointing equivalent entities across disparate KGs. Contemporary methods for entity alignment have predominantly utilized knowledge embedding models to procure entity embeddings that encapsulate various similarities-structural, relational, and attributive. These embeddings are then integrated through attention-based information fusion mechanisms. Despite this progress, effectively harnessing multifaceted information remains challenging due to inherent heterogeneity. Moreover, while Large Language Models (LLMs) have exhibited exceptional performance across diverse downstream tasks by implicitly capturing entity semantics, this implicit knowledge has yet to be exploited for entity alignment. In this study, we propose a Large Language Model-enhanced Entity Alignment framework (LLMEA), integrating structural knowledge from KGs with semantic knowledge from LLMs to enhance entity alignment. Specifically, LLMEA identifies candidate alignments for a given entity by considering both embedding similarities between entities across KGs and edit distances to a virtual equivalent entity. It then engages an LLM iteratively, posing multiple multi-choice questions to draw upon the LLM's inference capability. The final prediction of the equivalent entity is derived from the LLM's output. Experiments conducted on three public datasets reveal that LLMEA surpasses leading baseline models. Additional ablation studies underscore the efficacy of our proposed framework.

Figures (10)

• Figure 1: An illustration of entity alignment between an English KG and a Chinese KG.
• Figure 2: An illustration about LLMs modeling semantic knowledge of entities.
• Figure 3: An overview of LLMEA. The framework comprises three distinct phases: knowledge embedding, candidates generation, and LLM-enhanced alignment prediction.
• Figure 4: Prompt for virtual equivalent entity generation from a LLM.
• Figure 5: Prompt for the multi-choice question task to make the LLM to predict equivalent entities.