基于少样本特征提取与互引网络的产品专利知识挖掘与专利推荐方法

To address the challenges of insufficient knowledge mining and low efficiency of knowledge reuse in the innovative design of complex industrial products, a complex product patent knowledge mining and recommendation method based on few-shot feature extraction and cross-citation network is proposed. Firstly, patent knowledge model is constructed by extracting three types of featured entities including domain(D), function(F), and technology(T), which is represented as a three-aspect model <D, F, T>. Meanwhile, the annotation rule of few-shot patent dataset is proposed for manual feature annotation. Secondly, a fine-tuned BERT model is constructed based on a few-shot annotated patent dataset to achieve automatic patent feature extraction. A graph neural network is employed to learn the representation of the patent co-citation network. Thirdly, the fine-tuned BERT model is subsequently used to extract feature entities in the input design problem or task description, which are vectorized and their cosine similarities computed to generate a recommendation list. Finally, a prototype based on Django framework is also developed, which realizes the two functions of feature extraction and knowledge push, and illustrates the use scenarios of this method based on a case of automobile windshield wiper strip design.