A statistical-estimation method for stochastic finite-state transducers based on entropy measures

Abstract

The stochastic extension of formal translations constitutes a suitable framework for dealing with many problems in Syntactic Pattern Recognition. Some estimation criteria have already been proposed and developed for the parameter estimation of Regular Syntax-Directed Translation Schemata. Here, a new criterium is proposed for dealing with situations when training data is sparse. This criterium is based on entropy measurements, somehow inspired in the Maximum Mutual Information criterium, and it takes into account the possibility of ambiguity in translations (i.e., the translation model may yield different output strings for a single input string.) The goal in the stochastic framework is to find the most probable translation of a given input string. Experiments were performed on a translation task which has a high degree of ambiguity. © Springer-Verlag Berlin Heidelberg 2000.