利用深度去噪自编码器深度学习的指令意图理解方法  

Deep Learning of Instruction Intention Understanding Using Stacked Denoising Autoencoder

作　　者：李瀚清 房宁[3] 赵群飞[1,2] 夏泽洋[4] 

机构地区：[1]上海市北斗导航与位置服务重点实验室,上海200240 [2]上海交通大学自动化系,广东深圳518055 [3]上海交通大学人文学院,上海200240 [4]中国科学院深圳先进技术研究院,广东深圳518055

出　　处：《上海交通大学学报》2016年第7期1102-1107,共6页Journal of Shanghai Jiaotong University

基　　金：国家自然科学基金资助项目(51305436)

摘　　要：提出了一种利用深度去噪自编码器(SDAE)的自然语言指令意图理解方法.根据家庭服务机器人的使用环境和应用场景构建了一个自然语言文本指令语料库,并对语料库中各类指令进行意图标注,从而把文本指令理解问题转化为文本分类问题;在传统的文本向量空间模型的基础上,融合了文本指令的词性信息,定义了一种文本表示模型——词性向量空间模型;将SDAE应用于文本指令意图理解,提取指令的高阶特征;用高斯核支持向量机进行训练和预测,进而实现了自然语言指令的意图理解.在所建语料库上进行多折交叉验证,结果表明指令意图理解平均准确率达到96%以上.This paper proposed a natural language based text-instruction intention understanding method by using the stacked denoising autoencoder （SDAE）. A corpus of text-instructions in natural language was created according to the working environment and application scenarios for the home service robot. Then each text instruction was labelled with a corresponding intention so as to transform the problem of inten- tion understanding in natural language into a simple classification one. Based on the traditional text vector space model, the part-of-speech vector space model which contains the information of part-of-speech was defined. SDAE was applied to text-instruction intention understanding for extracting the higher level features of the text-instructions. Finally, the support vector machine was used for training and prediction in order to achieve text-instruction intention understanding in natural language. The result of 10-fold cross validation on the corpus shows that the average accuracy of text-instruction intention understanding reaches more than 96 %.

关 键 词：意图理解 向量空间模型 支持向量机 深度去噪自编码器 

分 类 号：TP24[自动化与计算机技术—检测技术与自动化装置]