Rule-based word clustering for document metadata extraction

Text classification is still an important problem for unlabeled text; CiteSeer, a computer science document search engine, uses automatic text classification methods for document indexing. Text classification uses a document's original text words as the primary feature representation. However, such representation usually comes with high dimensionality and feature sparseness. Word clustering is an effective approach to reduce feature dimensionality and feature sparseness, and improve text classification performance. This paper introduces a domain Rule-based word clustering method for cluster feature representation. The clusters are formed from various domain databases and the word orthographic properties. Besides significant dimensionality reduction, such cluster feature representations show a 6.6% absolute improvement on average on classification performance of document header lines and a 8.4% absolute improvement on the overall accuracy of bibliographic fields extraction, in contrast to feature representation just based on the original text words. Our word clustering even outperforms the distributional word clustering in the context of document metadata extraction.