Machine translation is a sub-field of computational linguistics
that investigates the use of computer software to translate
text or speech from one natural language to another.
In the last decades, there has been a major boost for the use of
statistical techniques in the development of machine translation systems.
In order to be able to apply these methods on a given language pair,
a parallel corpus on those languages needs to be available.
These techniques are so attractive because new systems are developed
without the need for expert knowledge from linguistic experts.

Finite state models have been very successful in multiple areas from
natural language scientific research, machine translation included.
Finite state models have some advantages with respect to other
statistical models, such as their simple integration into speech
recognition environments, their application to computer assisted
translation systems, or their ability to process the information,
which is not required to be complete, by means of a pipelined
architecture that is based on the so popular assembly lines.

The research goal is the study and exploitation of machine
translation techniques which are based on finite state models.
The work that is presented in this thesis is a detailed analysis about
the GIATI methodology on the inference of stochastic finite-state transducers
for their effective and efficient application as translation models,
allowing their usage on translation tasks with a high volume of data.

On the one hand, a software toolkit that implements the GIATI methodology
in an efficient way has been developed, thus it allows the learning of
the structure of the models and the estimation of their probabilities.
It also implements different search methods for the evaluation of the models.
Moreover, several scalability techniques that allow the usage of
a voluminous parallel corpus have been included.

On the other hand, nowadays the state-of-the-art in statistical
machine translation is based on the so called phrase-based models.
Their inherent idea has been integrated into our framework,
allowing the generation of phrase-based transducers which
have been contrasted with the ones that are based on words.
Their application to GIATI has encouraged an efficient adaptation of the search
strategies that has also allowed the usage of more effective smoothing algorithms.
We have also incorporated the modern trends on log-linear modelling
into this finite-state technology. The approach does a rescoring
on transition probabilities in order to increase the system performance.

Finally, the infrastructure for a better exploitation
of the available language resources has been established.
As a consequence, a better estimation of the translation models
would be possible thanks to the use of morphological analizers
over the languages which are involved in the translation process.
The associated linguistic information allows the application of
clustering techniques, thus reducing the corpus variability, then
obtaining a statistically more robust model after the training procedure.
The experimental results under this approach are rather preliminar
but they establish somehow the bases for a future post-doc researching
line on this topic.