摘要
在问答系统中用户的查询是以自然语言问句的形式出现的,查询分类对生成合适的答案有着重要的指导性作用。现有文献大多基于SVM统计学习模型实现查询分类。文章详细分析了中文查询分类的典型特征及其编码过程,并给出了LibSVM分类器的参数优化及核函数选取方法。比较了词袋特征(bag-of-word)和词性与词袋绑定特征(bag-of-word/pos)在LibSVM(RBF)、LibSVM(Linear)和Liblinear三个分类器上的分类精度。实验结果表明,在问题训练集规模较大、特征维数较高的情况下,Liblinear分类器具有更好的性能。同时,得出一个结论:bagof-word/pos特征对英文查询分类有一定的贡献;对于中文查询分类,虽然理论上增加特征有利于提高SVM分类器的精度,但由于绑定词性特征后可能会引入噪声,进而降低查询分类的精度。
In question answering system,user query is in the form of natural language sentence,query classification (QC) plays an important role in generating appropriate answers.In existing literatures,many implement QC based on machine learning method such as SVM.In this paper we analyze the typical features and their coding of QC.The method of parameter optimization and kernel function selecting in SVM is also proposed.With the bag-of-word(/pos) features,we compare the accuracies of Chinese query classification on LibSVM(RBF),LibSVM(Linear) and Liblinear respectively. Experimental results show that the Liblinear Classifier has a better performance in the case of larger scale training set and higher dimension features.In the meantime,we also draw a conclusion that the bag-of-word/pos feature has certain contribution to English query classification.For Chinese query classification,the increase of features will help to improve the accuracy of SVM classifier theoretically,but in our experiment,binding the part of speech feature introduces the noise which reduces the query classification accuracy.
出处
《情报学报》
CSSCI
北大核心
2011年第9期946-950,共5页
Journal of the China Society for Scientific and Technical Information
基金
计算机软件新技术国家重点实验室开放课题基金项目(KFKT2010B02)
安徽省高校省级自然科学研究项目(KJ2007B245)
安徽省高校省级自然科学研究重点项目(KJ2011A48)
关键词
问答系统
查询分类
SVM
核函数
question answering system
query classification
SVM
kernel function