A new neuro-fuzzy system for logical labeling of documents

Logical labeling, i.e. matching of the physical and logical structure, is an essential part of any document processing system. Contrary to the classical rule-based approach, where inflexible, heuristic knowledge is involved, a new neuro-fuzzy system is proposed in this paper. The main module, called FasArt (fuzzy adaptive system ART-based), is based on the well-known supervised neural architecture fuzzy ARTMAP. This new proposed architecture maintains the positive characteristics of its predecessor, such as capacity for incremental learning, respect to the plasticity-stability dilemma and generation of a symbolic representation of the system performance. Additionally, FasArt has a consistent formulation as a fuzzy logic system, a proven ability for approximation of any continuous function and a parameter that permits us to regulate its fuzziness degree. This new architecture is combined with a module that implements the Viterbi algorithm with the statistical properties of a certain document class, in order to resolve ambiguities or inconsistencies of the classification results provided by FasArt. Experimental results are shown for a variety of business letters, within a system of automatic feeding of an institution mailing database FasArt shows better results than Fuzzy ARTMAP while the use of the Viterbi-based modules boosts the correct labeling results to a 95% from the initial 80%