Peer Matrix Alignment: A New Algorithm

Abstract

Web data extraction has been one of the keys for web content mining that tries to understand Web pages and discover valuable information from them. Most of the developed Web data extraction systems have used data (string/tree) alignment techniques. In this paper, we suggest a new algorithm for multiple string (peer matrix) alignment. Each row in the matrix represents one string of characters, where every character (symbol) corresponds to a subtree in the DOM tree of a web page. Two subtrees take the same symbol in the peer matrix if they are similar, where similarity can be measured using either structural, content, or visual information. Our algorithm is not a generalization of 2-strings alignment; it looks at multiple strings at the same time. Also, our algorithm considers the common problems in the field of Web data extraction: missing, multi-valued, multi-ordering, and disjunctive attributes. The experiments show a perfect alignment result with the matrices constructed from the nodes closed to the top (root) and an encourage result for the nodes closed to the leaves of the DOM trees of the test web pages.