RESEARCH ON KEY WORDS SYSTEMATIZATION FOR INFORMATION RETRIEVAL OF EDUCATIONAL RESOURCE ON THE INTERNET

本研究的目的是为学校教学提供有效的网络信息分类。本文在分析教学教材词频的基础上，将网站分类为学科单元，通过形态分析从单元文档中提取术语，并汇总术语的词频，生成术语文档矩阵。这个矩阵的每个元素是每个类别单元中每个术语的频率。建立了利用奇异值分解（SVD）从调整后的矩阵中提取词条和文档的特征向量，以及由词条向量合成任意文档的特征向量的方法。基于这些特征向量的相似性，对文档的类单元进行了估计，并对包含120多个网站的测试集进行了评估，每个网站都有高中教师的答案。程序中的参数和变量等，使用哪些word类，如何调整 ...更多信息 在此基础上，对特征向量的特征矩阵、特征向量的维数等进行了优化。结果以网站特征向量与单元向量之间相似度降序排列的单元标题列表的形式给出。为了说明网站的结构，使用了一种非监督神经网络学习的SOM（SelfOrganizingMap）。结果，教科书的概念空间是可视化的，和映射的每一个网页，构成一个任意的网站显示如何网站页面分布在这个空间。本文还讨论了该地图的有效性，并给出了阅读这些地图的指导原则。建立了从教科书描述中提取特征向量和合成任意文档特征向量的方法。通过比较特征向量之间的相似性来获取文档的匹配类单元。最后利用SOM对分类结果和网站结构进行了可视化处理，结果表明，利用这些方法对网站或文档进行分类和可视化处理是可行的。少

The aim of this research is to efficient categorization of Internet information for school teaching. This thesis conducts categorizing websites to subjects-units, based on the analysis of term-frequencies in teaching textbooks.A term-document matrix is generated by extracting terms from unit-documents with morphological analysis and summarizing terms' frequency. Each element of this matrix is the frequency of every term in every class unit. Methods to extract feature vectors of terms and documents from the adjusted matrix using SVD (Singular Value Decomposition), and to composite feature vectors of arbitrary documents from the terms' vectors are established. A suitable class-unit for the document is estimated based on the similarity of these feature vectors.A test-collection including over 120 websites, each of which has an answer made by high-school teachers, was used for the evaluation. Parameters and variables in the procedures, such as, which word classes to use, how to adjust the … More term-document matrix, the dimension size of feature vectors, etc. were optimized based on this evaluation. Results were given as a list of unit-titles in descending order of similarity between the site's feature vector and the units' vectors.To illustrate the structure of websites, SOM (Self Organizing Map), which is a kind of a non-supervised neural network learning, is used. As a result, the textbook concept space was visualized, and mapping every page that constructs an arbitrary site showed how the site-pages were distributed in this space. The validity of this map is also discussed and guidelines in reading these maps are shown.In conclusion, a document classification specialized to educational information was done. Methods to extract feature vectors from descriptions in textbooks and to composite vectors of arbitrary documents are established. Matching class-units of the document is acquired by comparing similarities between the feature vectors. Visualization of categorizing results and website-structures are also done using SOM.It is shown that using these methods, categorization and visualization of websites or documents is feasible. Less