基于信息熵的半监督领域实体关系抽取研究  

A semi-supervised learning method based on information entropy to extract the domain entity relation

作　　者：郭剑毅[1,2] 雷春雅[1] 余正涛[1,2] 苏磊[1,2] 赵君[1] 田维[1] 

机构地区：[1]昆明理工大学信息工程与自动化学院,云南昆明650051 [2]昆明理工大学智能信息处理重点实验室,云南昆明650051

出　　处：《山东大学学报（工学版）》2011年第4期7-12,共6页Journal of Shandong University（Engineering Science）

基　　金：国家自然科学基金项目(60863011);云南省自然科学基金重点项目资助项目(2008CC023);云南省中青年学术技术带头人后备人才项目资助项目(2007PY01-11)

摘　　要：针对监督机器学习方法抽取实体关系受限于标注语料的规模问题,提出采用信息熵方法来不断扩展小规模训练数据的半监督领域实体关系抽取。结合领域词汇选取小规模训练数据,构建了一定准确率的初始最大熵分类器,用来从未标记数据中预测出候选新实例。采用信息熵方法,通过设定不同熵值,多次循环以选取可信度较高的新实例来扩展训练数据。使用扩展后的训练数据重新迭代训练分类器,分类器性能趋于稳定迭代终止,实现了半监督学习的领域实体关系抽取。实验表明,和已有方法相比,本文提出的半监督领域实体关系抽取通过结合信息熵方法,在小规模标注样本环境中取得了较好的学习效果。To solve the limitation by the scale of labeled corpus of the supervised learning method,a semi-supervised method based on information entropy was proposed to extract entity relation using small-scale training data.First,combined with field vocabulary to select small-scale training data,an initial maximum entropy classifier of certain accuracy was constructed to predict some new candidate instances from unlabeled data.Second,the method of information entropy was applied by setting different entropy value and cycling many times,and some new instances of the higher credibility from candidate instances were selected to expand the training data.Finally,the training classifier was re-iteratived with the expanded training data until classifier performance tended to a stable iteration termination,which achieved field entity relation extraction.Experimental results showed that the semi-supervised learning method based on information entropy achieved better learning results compared to other methods.

关 键 词：信息熵 半监督 最大熵分类器 未标记 可信度 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]