AutoAlign: A Fully Automated and Efficient Knowledge Graph Alignment Method Driven by Large Language Models

Knowledge Graphs (KG) have been widely applied in fields such as question-answering systems, dialogue systems, and recommendation systems. However, different Knowledge Graphs often store the same real-world entities in various forms, leading to challenges in knowledge sharing and information integration. This is particularly problematic when merging these Knowledge Graphs, a core task in practical applications. This involves Entity Alignment—identifying entities across different Knowledge Graphs that represent the same real-world entity. Existing methods usually require manually crafted seed alignments, which are costly to obtain, lack portability, and introduce biases that can affect alignment accuracy.

To address these challenges, Rui Zhang and scholars from institutions such as Tsinghua University, University of Melbourne, Universitas Indonesia, Chinese University of Hong Kong, and Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences, have proposed a new method called AutoAlign. This method was formally published in the June 2024 issue of IEEE Transactions on Knowledge and Data Engineering. The paper presents the first fully automated Knowledge Graph alignment method that requires no manual seed alignments, leveraging Large Language Models (LLMs) to achieve efficient and accurate entity and predicate alignment.

Research Background

Knowledge in Knowledge Graphs is typically stored in the form of triples, consisting of a head entity (Head), a predicate (Predicate), and a tail entity (Tail). These triples include both relationship triples and attribute triples. The alignment task not only identifies the same entities across different Knowledge Graphs but also aligns their predicates. Most existing methods are based on representation learning but require manually crafted seed alignments, which pose challenges for large-scale applications. Hence, this research aims to develop an automated Knowledge Graph alignment method that does not require manual intervention.

Method Introduction

AutoAlign achieves fully automated Knowledge Graph alignment through two core components: the predicate alignment module and the entity alignment module.

Predicate Alignment Module

AutoAlign first achieves predicate alignment by constructing a Predicate-Proximity Graph. This graph captures the similarity between predicates in two Knowledge Graphs using a Large Language Model. The specific steps are as follows:

1. Construction of the Predicate-Proximity Graph: Each triple in the Knowledge Graph is transformed by replacing the head and tail entities with their corresponding types, creating a graph representing entity-type relationships. For example, the triple “〈dbp:kromsdorf, dbp:located_in, dbp:germany〉” is transformed into “〈village, dbp:located_in, country〉”.

2. Automatic Alignment of Entity Types: Large Language Models (like ChatGPT and Claude) are used to automatically align the entity types in the two Knowledge Graphs. For instance, relevant prompts are input to Claude to automatically obtain similar type pairs from both Knowledge Graphs.

3. Module Learning: A target function is defined and optimized to learn predicate embeddings, making similar predicates in different Knowledge Graphs have similar representations in the vector space. This process employs two aggregation methods: weighted sum and attention-mechanism-based functions. The attention-mechanism-based method performs better in experiments.

Entity Alignment Module

After predicate alignment, AutoAlign achieves entity alignment through the following steps:

1. Independent Computation of Entity Embeddings: The TransE algorithm is used to compute entity embeddings in each Knowledge Graph separately.

2. Joint Learning: By calculating the similarity based on entity attributes, the entity embeddings from both Knowledge Graphs are transferred to a unified vector space. Specifically, attribute embeddings are based on the textual content of attribute values, bringing similar attributes closer in the vector space.

3. Entity Alignment: Finally, the similarity between all entity pairs is calculated in the unified embedding space obtained through joint learning. A threshold is set to filter out non-similar entity pairs, achieving entity alignment.

Main Flow Overview

To achieve embedded Knowledge Graph alignment, AutoAlign first merges the two Knowledge Graphs in their original forms and generates predicate-proximity triples and attribute triples. AutoAlign then obtains unified predicate, structure, and attribute embeddings. After acquiring entity embeddings, the entity alignment module determines entity pairs whose similarity exceeds a threshold, ultimately achieving entity alignment.

Experimental Results

AutoAlign underwent comprehensive experimental validation on multiple real-world Knowledge Graph datasets, demonstrating its superior accuracy in entity alignment compared to state-of-the-art methods.

1. Entity Alignment Performance: AutoAlign outperformed existing methods like MultiKE and AttrE in terms of Hits@10 performance. For instance, AutoAlign exceeded the best baseline method by 10.65% in Hits@10 on the dw-nb dataset.

2. Impact of Embedding Modules: Ablation experiments assessed the independent contributions of the structure embedding and attribute embedding modules, showing that using attribute embedding significantly improved alignment accuracy.

3. Advantages of Using Large Language Models: AutoAlign harnesses Large Language Models for fully automated entity type and predicate alignment, achieving higher levels of automation and accuracy compared to manual intervention methods.

Conclusion and Future Work

AutoAlign demonstrates the potential of Large Language Models in enhancing Knowledge Graph alignment performance, reducing manual work, and leveraging the knowledge stored in large models for efficient Knowledge Graph alignment. Future research can explore the application of Large Language Model-driven Knowledge Graph alignment methods in more graph or hypergraph-based study areas, such as aligning feature graphs or regional graphs in recommendation systems, thus enriching their representation capabilities.

Through this research, AutoAlign provides a feasible pathway for fully automated and efficient Knowledge Graph alignment, offering new insights for academic research and advanced technical support for data integration and knowledge discovery in practical applications.