This research is aimed at supporting interaction with large databases that include handwritten documents. The objective is to advance understanding of methodologies for interactively querying large document databases. Examples of such documents include medical and insurance information systems into which volumes of paper records have been scanned, government security systems, and electronic historical document collections, such as the collections created by the Model Editions Partnership. An existing, high-performance handwritten recognition algorithm will be adapted to the research problem. The questions to be addressed include: Which global, word-level training objective functions result in better performance? Which character class confidence measures result in better system performance? And, which character-based, word level confidence estimators results in better system performance? Experiments will focus on comparison of different methodologies for training; of different objective functions, optimization criteria, and optimization methodologies; and of different methods of estimating character confidence.
