WIR 2009 Workshop Information Retrieval 2009 Editors

Table extraction is the task of locating tables in a document and extracting their content along with its arrangement within the tables. The notion of tables applied in this work excludes any sort of meta data, e.g. only the entries of the tables are to be extracted. We follow a simple unsupervised approach by selecting the tables according to a score that measures the in-column consistency as pairwise similarities of entries where separator columns are also taken into account. Since the average similarity is less reliable for smaller tables this score demands a levelling in favor of greater tables for which we make different propositions that are covered by experiments on a test set of HTML documents. In order to reduce the number of candidate tables we use assumptions on the entry borders in terms of markup tags. They only hold for a part of the test set but allow us to evaluate any potential table without referring to the HTML syntax. The experiments indicate that the discriminative power of the incolumn similarities is limited but also considerable given the simplicity of the applied similarity functions.