Using learned conditional distributions as edit distance

Abstract

In order to achieve pattern recognition tasks, we aim at learning an unbiased stochastic edit distance, in the form of a finite-state transducer, from a corpus of (input,output) pairs of strings. Contrary to the state of the art methods, we learn a transducer independently on the marginal probability distribution of the input strings. Such an unbiased way to proceed requires to optimize the parameters of a conditional transducer instead of a joint one. This transducer can be very useful in pattern recognition particularly in the presence of noisy data. Two types of experiments are carried out in this article. The first one aims at showing that our algorithm is able to correctly assess simulated theoretical target distributions. The second one shows its practical interest in a handwritten character recognition task, in comparison with a standard edit distance using a priori fixed edit costs. © Springer-Verlag Berlin Heidelberg 2006.