The research presented in this PhD thesis is in the machine translation field. By studying the foundations of example-based machine translation, especially in the Aleph system, we bring to light the problem of example selection. The Aleph system uses exclusively the operation of analogy to produce new sentences and new translations. The problem is to select the adequate sentences from a large corpus of examples to allow for the production of new sentences by analogy. Our first contribution consists in the design of a method for the complete enumeration of all analogies contained in a text. This method allows us to complete a statistical study of the most frequent analogies between word trigrams and to bring to light the most frequent patterns of analogy. These results allow us to design a new smoothing technique for trigram language models based on a small amount of patterns of analogy. We report experiments which show that this new smoothing technique outperforms classical methods.
