New Machine Learning Method Improves Patent Entity Matching Across Data Sources

Linking identical patents and inventions across separate databases is a fundamental challenge in intellectual property research and management. A new approach developed by researchers tackles this problem by fusing multiple types of information to identify when the same patent entity appears in different knowledge graphs.

The Challenge#

When patents are catalogued in different data sources—whether by competing organizations, different countries, or separate research initiatives—the same invention may be represented multiple times under slightly different descriptions, formats, or metadata. Identifying these equivalent entities across disparate systems has traditionally relied on either analyzing the structural relationships within databases or processing the text-based descriptions of patents individually, but not both simultaneously in a coordinated way.

The Solution#

The researchers propose a combined approach that leverages graph convolution networks (a type of neural network designed for graph-structured data) alongside the BERT language model to simultaneously process two classes of patent information:

• Structural patterns: The relationships and connections between patents within a knowledge graph
• Semantic attributes: The textual details including patent names, descriptions, and assigned categories

By feeding both streams of information through their integrated model, the system generates richer numerical representations (embeddings) of each patent entity that capture both how it connects to other patents and what it describes itself to be.

Performance and Testing#

The authors evaluated their method on three standard benchmark datasets, where it outperformed existing approaches on the Hits@K metric—a standard measure of ranking accuracy in information retrieval and entity matching tasks. This suggests the multi-channel approach more reliably identifies true matches between equivalent patent records.