SYSTEM AND METHOD FOR AUTOMATED INTERPRETATION OF INPUT EXPRESSIONS USING NOVEL A POSTERIORI PROBABILITY MEASURES AND OPTIMALLY TRAINED INFORMATION PROCESSING NETWORKS

摘要

A method and system for forming an interpretation of aninput expression, where the input expression is expressed ina medium, the interpretation is a sequence of symbols, andeach symbol is a symbol in a known symbol set. In general,the system processes an acquired input data setrepresentative of the input expression, to form a set ofsegments, which are then used to specify a set ofconsegmentations. Each consegmentation and each possibleinterpretation for the input expression is represented in adata structure. The data structure is graphicallyrepresentable by a graph comprising a two-dimensional arrayof nodes arranged in rows and columns and selectivelyconnected by directed arcs. Each path, extending throughthe nodes and along the directed arcs, represents oneconsegmentation and one possible interpretation for theinput expression. All of the consegmentations and all ofthe possible interpretations for the input expression arerepresented by the set of paths extending through the graph.For each row of nodes in the graph, a set of scores isproduced for the known symbol set, using a complex ofoptimally trained neural information processing networks.Thereafter the system computes an a posteriori probabilityfor one or more symbol sequence interpretations. Byderiving each a posteriori probability solely throughanalysis of the acquired input data set, highly reliableprobabilities are produced for competing interpreta-tionsfor the input expression. The principles of the presentinvention can be practiced with cursively written character-- 66 -strings of arbitrary length and can be readily adapted foruse in speech recognition systems.