Smoothing probabilistic automata: an error-correcting approach

In this paper we address the issue of smoothing the probability distribution defined by a probabilistic automaton. As inferring a probabilistic automaton is a statistical estimation problem, the usual data sparseness problem arises. We propose here the use of an error correcting technique for smoothing automata. This technique is based on a symbol dependent error model which guarantees that any possible string can be predicted with a non-zero probability. We detail how to define a consistent distribution after extending the original probabilistic automaton with error transitions. We show how to estimate the error model's free parameters from independent data. Experiments on the ATIS travel information task show a 48% test set perplexity reduction on new data with respect to a simply smoothed version of the original automaton.