基于直推式学习的半监督属性抽取  

Semi-supervised method for attribute extraction based on transductive learning

作　　者：苏丰龙 谢庆华[2] 黄清泉[1] 邱继远[1] 岳振军[1] 

机构地区：[1]解放军理工大学通信工程学院,江苏南京210007 [2]解放军理工大学国防工程学院,江苏南京210007

出　　处：《山东大学学报（理学版）》2016年第3期111-115,共5页Journal of Shandong University(Natural Science)

摘　　要：针对在文本信息抽取研究中传统的监督学习方法存在标注工作量大和时间代价高等缺点,提出一种改进的半监督学习模型。该模型利用支持向量机的分类优势以及直推式学习在未标注样本上的泛化特点,先用少量标注语料进行学习,同时测试新语料,然后再加入到模型当中一起训练,调整预测规律。在领域实体属性抽取试验中,与传统的支持向量机学习方法相比,该模型能够在小语料条件下取得较好的抽取效果,泛化学习能力较强,可以节省大量的人力成本。In the study of text information extraction,aiming at the default of traditional supervised learning method that manual text annotation is high time-consuming and large workload,an improved semi-supervised learning model was proposed. The model combined support vector machines advantage of classification with transductive learnings advantage of generalization,which used a small amount of tagged corpus to study and test and then added them to the training model to adjust the prediction gradually. In attributes extraction experiments,this model performed better compared with traditional support vector machine methods. The generalization ability of the model was strong,which saved a great deal of manual cost.

关 键 词：信息抽取 半监督学习 直推式支持向量机 属性抽取 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]