Exploiting Structural Similarity for Automatic Information Extraction from Lists

Abstract

In this paper, we propose a novel technique to reduce dependency on knowledge base for ONDUX, the current state-of-art method for information extraction by text segmentation. While the existing approach mainly relies on high overlapping between pre-existing data and input lists to build an extraction model, our approach exploits structural similarity of text segments in the sequences of a list to align them into groups to achieve effectiveness with low dependency on pre-existing data. Firstly, a structural similarity measure between text segments is proposed and combined with content similarity to assess how likely two text segments in a list should be aligned in the same group. Then we devise a data shifting-alignment technique in which positional information and the similarity scores are employed to cluster text segments into groups before their labels are revised by an HMM-based graphical model. The experimental results on different datasets demonstrate the ability of our method to extract information from lists with high performance and less dependence on knowledge base than the current state-of-art method.