An ontology-based approach for web information extraction

Extracting structured information from pages published on the worldwide web is a problem with many facets that has gained growing interest in recent years. We propose a novel ontology-based approach that can achieve both the extraction and the semantic description of data contained in a web page. Existing methods addressing these issues range from pure manual methods based on rules to systems that can achieve wrapper induction automatically. Automatic systems require web pages with a list of records or a set of similar web pages to deduce the template used to generate them. In most cases, they cannot assign labels to extracted data. Ontology-based systems can automatically extract semantics, but they require an intensive and expertise consuming work to build anthologies that contain syntactic descriptions of attributes to be extracted. Our approach is fully automatic and is based on a "seeded" ontology that contains the basic information about the defined domain of interest. It uses an instance-based classifier to characterize the attributes of the ontology. In opposition to existing methods, our approach does not make any prior assumption on the design and the format of web pages, it is totally independent and it is able to achieve semantic extraction from a single web page with a single instance. Our method for localizing the data frame combines Information Retrieval techniques, instance-based classifying and a similarity-measure to the ontology. This combination permits achieving data-region localization, data extraction and label assignment, all in one single step. Experimental results obtained from different web pages of different web sites show that our approach is effective.