Introduction
In the era of rapid global innovation, the volume of science and technology patents has exploded, leading various organizations and nations to develop independent Knowledge Graphs (KGs). While these graphs are invaluable for tracking innovation, they are often siloed, redundant, and heterogeneous. To synthesize a comprehensive view of a technology, researchers must perform Entity Alignment—the task of identifying equivalent entities across disparate databases (e.g., matching a Chinese patent entity with its US counterpart).
Existing alignment methodologies often suffer from a narrow focus, typically prioritizing either graph structure (linkages and citations) or entity attributes (textual descriptions). Patent data, however, is uniquely complex, characterized by both intricate relational structures and specialized technical jargon. This paper introduces MIFEA (Multi-information Fusion Entity Alignment), a framework designed to bridge these gaps by fusing structural and semantic information.
The MIFEA Framework
The MIFEA framework integrates four primary modules to achieve robust entity alignment:
- Structural Embedding (GCN): A two-layer Graph Convolutional Network (GCN) captures the topological context of the patent KG. By aggregating information from an entity’s neighbors, the model creates a vector representation that reflects the entity's functional role and position within the graph.
- Semantic Representation (BERT): To process technical descriptions, the framework utilizes a pre-trained multi-lingual BERT model. This allows for the mapping of patent names and abstracts from different languages into a unified semantic space, effectively handling cross-lingual alignment without manual translation.
- Information Fusion: The core innovation of MIFEA is a weighted fusion mechanism. Rather than simple concatenation, it balances structural ($T^G$) and semantic ($T^B$) embeddings using a weighting factor $\tau$: $$T^C = \tau \cdot T^G \oplus (1-\tau) \cdot T^B$$ This ensures the model can adaptively prioritize "where the entity is" versus "what the entity says."
- Alignment & Candidate Pooling: For computational efficiency, the system generates a candidate pool of likely matches based on GCN embeddings before performing a final ranking using cosine similarity.
Experimental Results
The researchers evaluated MIFEA against several established baselines, including MTransE, JAPE, and AttrGNN, using both standard benchmarks (DBP15K) and a custom "Science and Technology Patent" dataset containing 160,000 entries.
As shown in Table 1, MIFEA consistently outperformed existing methods. On the DBP15K dataset, it achieved a Hits@1 accuracy of 82.94%, notably higher than AttrGNN's 79.60%. On the domain-specific patent dataset, MIFEA reached a Hits@1 of 80.74% and a Hits@10 of 92.59%, demonstrating its effectiveness in handling technical, cross-lingual data.
Efficiency and Limitations
While MIFEA is more complex than translation-based models, it maintains competitive efficiency. It was found to be approximately 6.4% faster than JAPE, proving that multi-information fusion is feasible without prohibitive overhead.
However, limitations remain. The model's reliance on BERT and GCN layers makes it more computationally intensive than simpler subgraph-matching algorithms like RNM. Furthermore, results in specific language pairs, such as Japanese-English, suggest that more specialized pre-training may be required to fully capture the nuances of certain technical vocabularies.
Conclusion
MIFEA represents a significant step forward in patent entity alignment by successfully fusing structural graph data with deep semantic text analysis. By leveraging the strengths of both GCNs and BERT, the framework provides a scalable and accurate solution for integrating global science and technology knowledge graphs.
Summary updated