Host: The Japanese Society for Artificial Intelligence
Name : The 26th Annual Conference of the Japanese Society for Artificial Intelligence, 2012
Number : 26
Location : [in Japanese]
Date : June 12, 2012 - June 15, 2012
A huge amount of consolidated information on the World Wide Web are embedded in HTML pages as they are generated dynamically from databases through some search form. This paper proposes a page-level web data extraction system FiVaTech2 that extracts schema and templates from these template-based web pages automatically. The proposed system, FiVaTech2, is an extension to our previously page-level web data extraction system FiVaTech. FiVaTech2 uses a machine learning (ML) based method which compares HTML tag pairs to estimate how likely they present in the web pages. We use one of the ML techniques called J48 decision tree classifier and also use image comparison to assist templates detection. Each HTML tag in the web page has several features that can be divided into the three types: visual information, DOM tree information, and HTML tag contents. Our experiments show an encouraging result for the test pages when combinations of the three types of tag features are used. Also, our experiments show that FiVaTech2 performs better and has higher efficiency than FiVaTech.
