Research on development of retrieval and classification system of documents on the Web which is based on analysis of teaching text books.

本研究的目的是比较现有的教育网络资源主题分类方法的效果。三种方法是利用余弦相似度的向量空间模型（VSM）、自组织映射上的向量距离（SOM）和三层感知器（MLP）。基于奇异值分解（SVD）方法，从教学导则中提取术语特征向量，估计文档向量。编制了403个教学方案作为学科区分的测试文件。比较了三种方法的查全率、查准率和分类性能指标。在大多数被试中，MLP的召回率最高。不同学科的方法准确率最高。结果表明，使用MLP方法的大多数受试者的F1测量作为总绩效指标最高。在此基础上，对基于多层感知器（MLP）的教育Web资源主题分类方法进行了测试，取得了较好的分类效果。编制了10种分类方法和403份试验文件，并用F1测度比较了不同条件下的分类性能。研究了奇异值分解（SVD）从词-文档矩阵中提取词特征作为国家教学导则描述特征的有效性。当以出现唯一主题的词来估计文档特征时，使用从SVD中提取的所有词特征进行训练的MLP方法表现出最高的分类性能。正确与错误分类文档的特征成分存在显著差异。

The purpose of this study is to compare performance among developed subject classification methods for educational Web re-sources.Three methods are vector space model (VSM) using degree of cosine similarity, vector distance on the self-organizing map (SOM) and three layer percptron (MLP). The document vectors were estimated by the term feature vectors which were extracted from the teaching guidelines based on the singular value decomposition method (SVD).The 403 teaching schemes were prepared as the test documents for the subject discrimination. The recall rate, precision rate and Fl measure as the classification performance measure, were compared across three methods. The recall rate using MLP is the highest for most subjects. The highest precision rate among the methods depend on subject. As a result, F1 measure as total performance index of most subjects are the highest for the method using MLP.According to the results, the subject classification methods for educational Web resources by using multi layer perceptron (MLP) is tested to achieye better performance. Ten classification methods and 403 test documents were prepared, then classification perfor-mance was compared among conditions by the F1 measure.As a result, it was examined effectiveness of singular value decomposition (SVD) to extract term feature from term-document matrix as description characteristics of the national teaching guideline. The highest performance of classification method was the use of MLP which trained with all term features extracting from SVD when document feature was estimated by terms appering unique subject.Also, it was suggested that there were significant differences of document feature components between correct and incorrect classified documents.