Incorporating global and local hierarchical information into text encoder: a novel approach for multi-label patent classification

Incorporating global and local hierarchical information into text encoder: a novel approach for multi-label patent classification

Journal article

1 volume.

Manual patent classification is labor-intensive and time-consuming. With the rapid growth of patent data and labels, traditional approaches struggle to handle large-scale classification. Existing automated methods partially address this issue, but most focus only on patent text semantics and ignore hierarchical label information. This limits models from fully exploiting label relationships, reducing both accuracy and generalization. In this study, a contrastive learning-based multilabel classification model, named CLMCM, is proposed, which incorporates the global and local hierarchical information into a patent text encoder to accomplish the classification task. Firstly, a hierarchical label-aware embedding module is deployed to exploit the global semantic information to generate label-wise patent text vectors. Secondly, a hierarchical adaptive label correlation learning module is designed to adaptively learn the semantic correlations around the label codes within the hierarchical taxonomy tree, and then aggregate the horizontal and vertical information to capture the local relationship of the labels. Finally, experimental evaluations on a real-world patent dataset demonstrate the superiority of the proposed method over existing multilabel classification methods, achieving 1.6% improvement in Mi-F1 score.

Crossref

World Patent Information ; 85, June, 2026

World Patent Information

Oxford [England] : Elsevier Ltd., 2026.

eng

https://www.sciencedirect.com/science/article/abs/pii/S0172219023000108

External Online Journals > World Patent Information