Abstract
Entity Alignment (EA) is critical for merging disparate Knowledge Graphs (KGs). Traditional embedding-based methods often suffer from limited structural context and a lack of global reasoning. This paper presents IMEA (Informed Multi-context Entity Alignment), a framework that leverages Transformers to capture diverse structural features—including triples, paths, and neighborhoods. Crucially, IMEA introduces a "holistic reasoning" feedback loop where high-confidence alignment evidence is used to iteratively refine training labels. Experimental results on the OpenEA benchmark demonstrate that IMEA achieves state-of-the-art performance across various cross-lingual and monolingual datasets.
Introduction
Knowledge Graphs (KGs) such as DBPedia and Wikidata provide structured representations of real-world knowledge but are often incomplete and siloed. Entity Alignment (EA) aims to identify equivalent entities across different KGs. Current embedding-based EA methods face two primary challenges: (1) they often focus on a single structural context, failing to capture the rich multi-faceted nature of KGs; and (2) they typically treat alignment as a local similarity search, ignoring global logical constraints and failing to integrate reasoning into the representation learning process.
Methodology
IMEA addresses these gaps through a multi-context Transformer architecture and an informed reasoning loop.
The framework consists of two main components:
1. Multi-context Transformer
Instead of standard Graph Neural Networks, IMEA uses parameter-sharing Transformers to encode neighborhood and path information. This allows the model to dynamically attend to relevant structural features while filtering noise. To incorporate triple semantics, a TransE-style regularization ($h + r \approx t$) is applied at the input layer. Furthermore, "equivalent paths" are generated using seed alignments to bridge the gap between different KGs, forcing aligned entity embeddings to converge.
2. Holistic Reasoning & Soft Label Editing
Unlike post-processing methods, IMEA integrates reasoning directly into training. It calculates functionality weights for relations to assess alignment evidence (e.g., highly functional relations like birthplace provide stronger signals). High-confidence alignments derived from this reasoning are used to "edit" the soft labels of the training data, creating a feedback loop that informs the embedding learning process with global evidence.
Experimental Evaluation
The model was evaluated on the OpenEA benchmark (V1), covering cross-lingual (EN-FR, EN-DE) and monolingual (D-W, D-Y) datasets at 15K and 100K scales. Performance was measured primarily using Hits@1 and Mean Reciprocal Rank (MRR).
Results and Discussion
IMEA consistently outperforms state-of-the-art baselines like KE-GCN and AliNet. On 15K datasets, it achieved Hits@1 improvements ranging from 1% to 7.9%. The model's performance on 100K datasets highlights its scalability and robustness in large search spaces.
Ablation studies confirmed the importance of the informed feedback loop; removing holistic inference led to a performance drop of 1% to 2.5%. Analysis of path lengths (as seen in the figure above) shows that while longer paths improve accuracy by providing more context, they also increase computational costs, presenting a clear trade-off for practical applications.
Conclusion
IMEA demonstrates that combining multi-context Transformer representations with iterative reasoning significantly improves entity alignment. By treating reasoning as a training signal rather than a post-process, the model effectively bridges the gap between local embeddings and global logic. Future work includes incorporating non-structural "side information" such as entity attributes and improving the explainability of the alignment decisions.
Summary updated