融合表示学习与机器学习的专利科学引文标题自动抽取研究

Extracting Titles from Scientific References in Patents with Fusion of Representation Learning and Machine Learning

【目的】针对专利引文类别繁多的问题,研究自动识别其中专利科学引文这一特定类别的方法,进而准确抽取专利科学引文的标题特征项,支撑后续专利科学引文深度分析与挖掘。【方法】引入表示学习方法 Doc2Vec实现专利科学引文整体的语义向量表示,结合机器学习分类方法实现专利科学引文这一特定类别的识别;在此基础上,利用表示学习方法实现专利科学引文标题等内容元数据的语义向量表示,结合机器学习分类方法抽取专利科学引文标题。【结果】在基因领域专利的实验中,专利科学引文的识别精确率达到99.27%,专利科学引文标题抽取精确率达到92.59%,抽取精确率较单纯的机器学习方法提高5.96%。【局限】人工标注训练集较为耗时;对实验数据格式有一定要求。【结论】本文方法在专利科学引文识别和标题抽取上具有良好效果。

[Objective] This paper aims to automatically identify scientific references in patent(SRP), and then extract titles from SRP to support in-depth data mining.[Methods] Firstly, we used the Doc2Vec method to generate vectors for the patent citations. Then, we identified the SRPs with support vector machine(SVM). Third, we created vectors for the metadata(such as titles) of SRP, and extracted titles with SVM.[Results] We examined the proposed method with patent citations from the genetic field. The accuracy of SRP recognition and titles extraction reached 99.27% and 92.59% respectively. The latter was 5.96% higher than those of the traditional methods.[Limitations] Manually tagging the training set was very time consuming, and there are format requirements for the experimental data.[Conclusions] The proposed method could effectively identify and extract patent citations and titles.