面向功能信息的相似专利动态聚类混合模型  

SIMILAR PATENT DYNAMIC CLUSTERING HYBRID MODEL FOR FUNCTIONAL INFORMATION

作　　者：马建红 张少光 曹文斌 王晨曦 Ma Jianhong;Zhang Shaoguang;Cao Wenbin;Wang Chenxi(School of Artificial Intelligence,Hebei University of Technology,Tianjin 300401,China)

机构地区：[1]河北工业大学人工智能与数据科学学院,天津300401

出　　处：《计算机应用与软件》2021年第5期201-207,共7页Computer Applications and Software

基　　金：河北省科技厅创新软件设计及公共应用服务平台项目(15240118D)。

摘　　要：专利的功能信息对专利规避及创新具有重要意义。为了完成对专利基于功能的聚类,提出一种面向功能信息的相似专利动态聚类混合模型。该模型筛选功能信息句,运用双向长短期记忆网络联合注意力机制、卷积神经网络、改进的权值主题模型抽取专利文本的序列特征、文本嵌入特征、主题特征,同时加入功能词语关注部分,采用并行融合的方式提取特征。将所有特征表示作为聚类的输入,聚类结果作为反馈信息实现动态聚类。选取无机化学类方面的专利进行实验,实验结果表明,该模型能够提高聚类的F-measure值,同时也能提高时间效率。The functional information of patents is significant for patent evasion and innovation.This paper studies the function-oriented patent clustering method,and proposes a similar patent dynamic clustering hybrid model for functional information.The model filtered the functional information sentences,and used bidirectional long short-term memory joint attention mechanism,convolutional neural network,improved weight topic model to extract the sequence features of patent texts,text embedding features,and topic features.Meanwhile,we added functional words attention,and the features were extracted by parallel fusion.All feature representations were used as input of clustering,and clustering results were used as feedback information to achieve dynamic clustering.This paper selected the patents of inorganic chemistry for experiments.The experiments show that the proposed model can improve the F-measure value of clustering and improve the time efficiency.

关 键 词：深度学习 神经网络 注意力机制 主题模型 动态聚类 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]