情报杂志2025，Vol.44Issue(5)：199-206,封3,9.DOI:10.3969/j.issn.1002-1965.2025.05.024
基于多模态深度学习的专利自动分类方法研究
Research on Automatic Patent Classification Method Based on Multimodal Deep Learning
摘要
Abstract
[Research purpose]Developing a multimodal deep learning-based patent automatic classification method to enhance classifica-tion efficiency,expedite the patent examination process,and reduce resource and labor costs.[Research method]The proposed method leverages the BERT-for-Patents model to extract textual features and the VGG19 model to extract visual features from patent data.These multimodal features are integrated and subjected to joint training through a stepwise concatenation process.The citation network among nodes is optimized by constructing connections between low-degree nodes of the same type.Finally,the GraphSAGE model is utilized for automatic patent classification.[Research result/conclusion]The study demonstrates that the proposed method achieves a classification accuracy of 95.71％for all constructed categories,90.19％for major categories,and 84.42％for subcategories,significantly outperfor-ming baseline models across all metrics.This method effectively leverages both patent text and image information,addressing the limita-tions of unimodal data and enhancing the model's ability to comprehend complex patent content.Consequently,it improves the granularity and accuracy of automatic patent classification,while also providing a valuable supplement to the existing research framework for patent classification.关键词
多模态/深度学习/专利分类/文本特征/图像特征/特征融合/网络关系优化Key words
multimodal/deep learning/patent classification/text features/image features/feature fusion/network relationship optimiza-tion分类
计算机与自动化引用本文复制引用
谢小东,吴洁,盛永祥,唐健廷,于娱..基于多模态深度学习的专利自动分类方法研究[J].情报杂志,2025,44(5):199-206,封3,9.基金项目
国家自然科学基金面上项目"面向产业安全的产业创新生态系统韧性内涵、评价与优化策略研究"(编号:72171122) （编号:72171122）
江苏省研究生科研与实践创新计划项目"创新联合体潜在合作伙伴选择及合作方向研究"(编号:KYCX23_3817)研究成果. （编号:KYCX23_3817）
