Towards a Method for Unsupervised Web Information Extraction

The literature provides a variety of techniques to build the information extractors on which some data integration systems rely. Information extraction techniques are usually based on extraction rules that require maintenance and adaptation if web sources change. In this paper, we present our preliminary steps towards a completely unsupervised information extraction technique that searches for shared patterns between web documents and fragments them until finding the relevant information that should be extracted. Experimental results on 1230 realweb documents demonstrate that our system performs fast and achieves very high precision (96%) and recall (98%).