The current article presents a novel language-independent approach to machine translation system, based on machine learning principles and inexpensive language resources, developed within the PRESEMT project. Unlike most modern MT systems, this methodology needs only a limited parallel corpus of a few hundred sentences and relies instead on models extracted from a large monolingual corpus using both statistical and machine-learning techniques, making it ideal for low-resourced languages (such as Greek) for which only limited linguistic resources are available. By design, PRESEMT emphasises the use of publicly available tools and resources, to simplify the development of the MT system. This allows it to be portable to new language pairs in a straightforward process.