Variable length vector pattern recognition

Variable length vector pattern recognition deals with the problem of recognizing words of various symbol lengths (characters, phonemes, etc.) embedded in unsegmented strings of words (sentences, messages, etc.). A mathematical model is developed to describe the generation of such messages, assuming that successive words are chosen independently. A loss function is defined which leads to the development of an iterative decision procedure for minimizing the risk. The computational complexity of this procedure is proportional to message length, whereas the message length would appear exponentially in the expression for computational complexity if the iterative procedure did not exist. Alternative decision models and alternative loss functions can be employed in the development of similar iterative procedures. The optimum iterative procedure was programmed in ASA FORTRAN IV. An illustrative experiment was conducted using the program to compare performance of the procedure with and without segmentation information.