Computationally efficient contextual processing for handwritten forms

In handwriting recognition, recognisers often deliver several alternative recognition results for each component with confidence values attached. The decision on the correct interpretation is then deferred until after use of syntactic and/or contextual constraints, so that undue loss of information is avoided. This paper outlines a semantic neural network approach to a generic contextual postprocessing module which draws input from uncertain character recogniser results and produces an optimum form-level result. The root node corresponds to form-level recognition and the leaf nodes to the uncertain character recogniser output, while intermediate nodes implement intermediate levels of syntactic and/or contextual knowledge at arbitrary intra- or inter-field levels. At every node, there is a recursive occurrence of entity and component relationships and the technique of lazy evaluation is utilised at all levels. The postprocessing is activated at the root node by a request for the best possible form interpretation. This request is decomposed and passed recursively down the hierarchy. Lazy evaluation makes sure that, at all nodes, only the necessary evaluations are carried out to satisfy the request. Subsequent requests for further form interpretations will successively retrieve further results in global best-first order.