基于Deep Belief Nets方法的中文名实体分类研究  

Study on Chinese Named Entity Categorization based on Deep Belief Nets

作　　者：陈宇[1] 郑德权[1] 赵铁军[1] 

机构地区：[1]哈尔滨工业大学计算机科学与技术学院,哈尔滨150001

出　　处：《智能计算机与应用》2014年第2期29-31,35,共4页Intelligent Computer and Applications

基　　金：国家自然科学基金(61073130)

摘　　要：DBN是一种快速全局最优的神经网络分类方法,包含数层无监督学习网络和一层有监督学习网络。本文验证了DBN方法很好地适用于中文名实体分类任务。首先,采用多层RBM方法无监督地从字特征向量提取结构信息,得到更具有表征能力的特征;然后,利用BP方法微调网络参数并对提取后的特征向量进行分类,以此构成分类器进行名实体分类。通过对ACE 04的中文名实体进行的分类测试,准确率达到91.45%,明显高于支持向量机和反向传播神经网络等传统分类算法。DBN is a classification of fast and global optimum neural network. It contains several layers of unsupervised networks and one layer of supervised network. The paper approves this novelty machine learning approach is suitable to the domain of named entity categorization. The paper applies RBM,an unsupervised learning method,to reconstruct more representative features from character-based features. Subsequently,the paper utilizes BP,a supervised learning method,to fine-tune parameters in whole network and accomplish the categorization task. In the end,the paper tests DBN on ACE 04 Chinese corpus and achieve 91. 45% precision,which is much better than Support Vector Machine and Back-propagation neural network.

关 键 词：名实体分类 神经网络 DBN 字特征 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]