Towards a spatial instance learning method for deep web pages

A large part of information available on the Web is hidden to conventional research engines because Web pages containing such information are dynamically generated as answers to query submitted by search form filled in by keywords. Such pages are referred as Deep Web pages and contain huge amount of relevant information for different application domain. For these reasons there is a constant high interest in efficiently extracting data from Deep Web data sources. In this paper we present a spatial instance learning method from Deep Web pages that exploits both the spatial arrangement and the visual features of data records and data items/fields produced by layout engines of web browsers. The proposed method is independent from the Deep Web pages encoding and from the presentation layout of data records. Furthermore, it allows for recognizing data records in Deep Web pages having multiple data regions. In the paper the effectiveness of the proposed method is proven by experiments carried out on a dataset of 100 Web pages randomly selected from most known Deep Web sites. Results obtained by using the proposed method show that the method has a very high precision and recall and that system works much better than MDR and ViNTS approaches applied to the same dataset. © 2011 Springer-Verlag.