Studies on Summarization of Multimedia Contents Based on Relational Structure between Text Tables and Images

基于文本表与图像关系结构的多媒体内容摘要研究

• 批准号: 13680452
• 负责人: ENDO Tsutomu
• 金额: $ 1.54万
• 依托单位: Kyushu Institute Of Technology
• 依托单位国家: 日本
• 项目类别: Grant-in-Aid for Scientific Research (C)
• 财政年份: 2001
• 资助国家: 日本
• 起止时间: 2001 至 2003
• 项目状态: 已结题
• 来源: https://kaken.nii.ac.jp/grant/KAKENHI-PROJECT-13680452/
• 关键词: multimedia document; WWW; summarization; information integration; information retrieval; information extraction; Kansei information; document generation

This research intends to develop a system which summarizes product (PCs) information retrieved from Web sites based on relational structure between text, tables and images, and presents products suitable for a user's request.1.Extraction of product specifications from HTML documents.We proposed a method for extracting specifications from HTML documents using TSVMs (Transductive Support Vector Machines). The elements of a feature vector are keywords with normalized TF-DF weighting. We achieved 95% recall with 99% precision.2.Characteristic-data extraction and support system for PC selection.The specifications written in HTML are converted into normal form called table structure. The quantitative attributes are extracted by comparing them with the mean or mode of all sample data, and the qualitative ones are extracted using knowledge provided manually. The recommended PCs are dynamically determined from the extracted data by a user's request and relevance feedback. Moreover, a radar chart and Japanese sentences are generated from specifications.3.Classification of images and feature extraction.We proposed a method for classifying the contents of images using weighted keywords extracted from their neighboring sentences. We achieved 79% accuracy by TF-IDF weighting. We also developed a system which eliminates the background from a PC image, and classify the color of PC using C4.5.

本研究旨在开发一个基于文本、表格和图像之间的关系结构的产品信息汇总系统，并根据用户的要求提供产品信息。1.从HTML文档中提取产品规格说明提出了一种基于TSVM（Transductive Support Vector Machines）的HTML文档规格说明提取方法。特征向量的元素是具有归一化TF-DF加权的关键字。2.特征数据提取与PC选型支持系统将HTML格式的规格说明转换为表格结构。定量属性通过与所有样本数据的均值或众数进行比较来提取，定性属性通过人工提供的知识来提取。通过用户的请求和相关性反馈从提取的数据动态地确定推荐的PC。3.图像的分类和特征提取提出了一种利用相邻句子中提取的加权关键词对图像内容进行分类的方法。通过TF-IDF加权，我们实现了79%的准确率。我们还开发了一个系统，从PC图像中消除背景，并使用C4.5分类PC的颜色。

项目成果

林晃司, 嶋田和孝, 遠藤勉: "WWWからの性能表抽出のためのキーワード獲得と重み付け"電子情報通信学会技術研究報告Tl2002-48〜53[思考と言語]. Vol.102 No.688. 13-18 (2003)

Koji Hayashi、Kazutaka Shimada、Tsutomu Endo：“从 WWW 中提取性能表的关键字获取和加权”IEICE 技术研究报告 Tl2002-48~53 [Thinking and Language] Vol.102 No.688 (2003)。

福本篤史, 遠藤勉, 嶋田和孝: "Information Extraction from Specifications on the World Wide Web"Proceedings of PACLING2001 (Pacific Association for Computational Linguistics 2001). 109-116 (2001)

Atsushi Fukumoto、Tsutomu Endo、Kazutaka Shimada：“从万维网规范中提取信息”PACLING2001 论文集（太平洋计算语言学协会 2001 年）109-116 (2001)。

嶋田和孝, 伊藤哲郎, 遠藤勉: "Multiform Summarization from Product Specifications"Proceedings of PACLING2003(Pacific Association for Computational Linguistics 2003). 83-92 (2003)

Kazutaka Shimada、Tetsuro Ito、Tsutomu Endo：“产品规格的多形式总结”PACLING2003 论文集（太平洋计算语言学协会 2003 年）83-92 (2003)。

嶋田和孝, 遠藤勉他: "Information Modeling and Knowledge Bases XV"IOS Press. 333 (2004)

Kazutaka Shimada、Tsutomu Endo 等人：“信息建模和知识库 XV”IOS Press 333 (2004)。

Atsushi Fukumoto, Tsutomu Endo, Kazutaka Shimada: "Information Extraction from Specifications on the World Wide Web"Proceedings of PACLING2001 (Pacific Association for Computational Linguistics 2001). 109-116 (2001)

Atsushi Fukumoto、Tsutomu Endo、Kazutaka Shimada：“从万维网规范中提取信息”PACLING2001 论文集（太平洋计算语言学协会 2001 年）。