专利文本分类的基础问题研究

Fundamental Research Questions in Patent Text Categorization

对专利文本分类中的基础问题进行研究,包括术语作为专利文本分类特征的适用性,主权项字段分类研究和相近主题对分类结果的影响等。研究在两种朴素贝叶斯分类器、kNN、Racchio和支持向量机等5个分类器上进行,测试主要采用交叉验证的方法。研究结果显示,在同样的设定下,采用术语作为特征的分类结果优于使用一般特征词;使用摘要训练,对主权项进行分类有助于改善主权项的分类效果;相近主题会降低分准率,有必要设计层次的分类器进行分类试验。研究结果可以为专利文本分类研究和实践提供参考数据,并可作为信息分析等工作使用专利文本分类技术的参考。

The paper focuses on some fundamental problems in patent text categorization, including the feasibility of using terms for automatic categorization, the research on claim categorization, and the effect of classes with close - related topics on the categorization result. The research is executed on two Naive Bayesian classifiers, kNN, Racchio and SVM classifier, and cross validation is used for testing. The results of the paper are that terms are better than common features under the same settings, that training a classifier with s can improve the claim categorization results, and that clas- ses with close- related topics result in low precision and hierarchical design of classifier is necessary, correspondingly. The paper provides fundamental data for patent text categorization and can be referred by information analysis and other applications using patents.