Maximum mutual information and conditional maximum likelihood estimation of stochastic regular syntax-directed translation schemes

Formal translations have become of great interest for modeling some Pattern Recognition problems, but they require a stochastic extension in order to deal with noisy and distorted patterns. A Maximum Likelihood estimation has been recently developed for learning the statistical parameters of Stochastic Regular Syntax-Directed Translation Schemes. The goal of this paper is the study of estimation criteria in order to take into account the problem of sparse training data. In particular, these are the Maximum Mutual Information criterion and the Conditional Maximum Likelihood criterion. Some experimental results are reported to compare the three criteria.