Learning Finite-State Models for Machine Translation

Abstract

In formal language theory finite-state transducers are well-know models for “input-output” rational mappings between two languages. Even if more powerful, recursive models can be used to account for more complex mappings, it has been argued that the input-output relations underlying most usual natural language pairs are essentially rational. Moreover, the relative simplicity of these mappings has recently lead to the development of techniques for learning finite-state transducers from a training set of input-output sentence pairs of the languages considered. Following these arguments, in the last few years a number of machine translation systems have been developed based on stochastic finite-state transducers. Here we review the statistical statement of Machine Translation and how the corresponding modelling, learning and search problems can be solved by using stochastic finite-state transducers. We also review the results achieved by the systems developed under this paradigm. After presenting the traditional approach, where transducer learning is mainly solved under the grammatical inference framework, we propose a new approach where learning is explicitly considered as a statistical estimation problem and the whole stochastic finite-state transducer learning problem is solved by expectation maximisation.