CAROL: Document Structure Recognition: A Rule-Based and Neural Network ApproachApplied to Cataloging

摘要

The report presents two approaches to document structure recognition applied tothe special case of recognizing logical elements on title pages of grey literature for cataloging. Both approaches are promising. The latter has a lot of advantages such as its flexibility with noisy input and changing layout. Another advantage is the training ability, i.e. for new types of documents there is no need for writing a new set of rules, the system of neurons 'learns' their internal weights by giving some examples. To improve the performance and the results other ways of input representation for the neural net have to be found. One way could be to work with smaller nets which have learned special keywords that appear on title pages and then after having included the layout information, another net will produce the output. (Copyright (c) GMD 1992.)