Designing syntactic pattern classifiers using vector quantizationand parametric string editing

摘要

We consider a fundamental inference problem in syntactic patternnrecognition (PR). We assume that the system has a dictionary which is ancollection of all the ideal representations of the objects in question.nTo recognize a noisy sample, the system compares it with every elementnin the dictionary based on a nearest-neighbor philosophy, using threenstandard edit operations: substitution, insertion, and deletion, and thenassociated primitive elementary edit distances d(.,.). In this paper, wenconsider the assignment of the inter-symbol distances using thenparametric distances. We show how the classifier can be trained to getnthe optimal parametric distance using vector quantization in thenmeta-space. In all our experiments, the training was typically achievednin a very few iterations. The subsequent classification accuracy wenobtained using this single-parameter scheme was 96.13%. The power of thenscheme is evident if we compare it to 96.67%, which is the accuracy ofnthe scheme which uses the complete array of inter-symbol distancesnderived from a knowledge of all the confusion probabilities