Spatially aware rapid retrieval system (SPARRS)

This paper describes an extended version of the syntactic neural network (SNN) rapid retrieval system - a system for the efficient application of dictionary type knowledge in pattern recognition systems. The original system, previously developed by the author (1995, 1996), has proven to be an excellent method of applying contextual knowledge to provide an optimal interpretation of the uncertain outputs of low-level pattern classifiers. The original model, however, assumed that the input data could be reliably segmented and was therefore unable to deal effectively with segmentation errors, or continuous inputs such as cursive script. The extension described in this paper is to add spatial or temporal constraints to the symbols processed by the network. The extended system can cope with insertion and deletion errors, whether they are due to poor segmentation of the input, or actual missing or added data.