基于图神经网络的专利关键词提取算法研究  

Research on Patent Keyword Extraction Algorithm Based on Graph Neural Network

作　　者：王玉叶 王玙[1] Wang Yuye

机构地区：[1]西安电子科技大学经济与管理学院,陕西西安710126

出　　处：《情报理论与实践》2023年第5期202-208,共7页Information Studies:Theory & Application

摘　　要：[目的/意义]专利关键词提取是专利挖掘任务中非常重要的前置子任务,基于图模型的关键词提取是目前最有效的算法。传统图模型只考虑了单词的局部上下文信息,为了捕获单词的全局信息,提出一种基于图神经网络的专利关键词提取算法,结合词向量与图模型实现专利关键词的提取。[方法/过程]首先,用专利数据集构建异构网络,以专利分类号为标签,训练图神经网络模型,使得同一主题下的单词具有相似的向量表示,获取包含主题信息的词向量;然后,根据专利摘要在滑动窗口内的单词共现关系和词向量相似度,构建融合了单词主题信息的文本图,利用词向量中的主题信息捕获单词的全局联系;最后,在文本图上使用PageRank算法,获取关键节点,构成专利的关键词。[结果/结论]与基线方法相比,该算法在提取专利关键词时,能够检测到新颖性与准确性更高的关键词。[Purpose/significance]Patent keywords extraction is a very important precursor subtask in patent mining tasks,and extract keywords on graph is currently the most effective method.Traditional text-graph models only consider the local contextual information of words,in case to capture global information of words,a keyword extraction algorithm based on graph neural network is proposed in this paper,which extracts keywords for patent text by combining word embedding and graph model.[Method/process]Firstly,a heterogeneous network is constructed with patent data set,and the graph neural network model is trained with International Patent Classification as label to make the words under the same topic have similar vector representation,which means the word embedding contains topic information.Then,according to the word co-occurrence in the sliding window of the patent abstract,and embedding similarity between words,a text graph which constructed with the global relation of words is captured by the topic information in the word embedding’s.Finally,PageRank is used in the text graph to find the key nodes which forms the keywords.[Result/conclusion]The algorithm in this paper can detect keywords for patent with higher novelty and accuracy than the baseline method.

关 键 词：关键词提取 图神经网络 专利 词向量 PAGERANK 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术] TP183[自动化与计算机技术—计算机科学与技术] G255.53[文化科学—图书馆学]