moses-for-mere-mortals

This site offers a set of Bash scripts and Windows executables add-ins that, together, create a basic translation chain prototype able of processing very large corpora. It uses Moses, a widely known statistical machine translation system.

The idea is to help build a translation chain for the real world, but it should also enable a quick evaluation of Moses for actual translation work and guide users in their first steps of using Moses. The scripts cover the installation, the creation of representative test files, the training, the translation, the scoring and the transfer of trainings between persons or between several Moses installations.

A Help/Short Tutorial (http://moses-for-mere-mortals.googlecode.com/files/Help.odt) and a demonstration corpus (too small for doing justice to the qualitative results that can be achieved with Moses, but able of giving a realistic view of the relative duration of the steps involved) are available.

Two Windows add-ins allow the creation of Moses input files from *.TMX translation memories (Extract_TMX_Corpus.exe), as well as the creation of *.TMX files from Moses output files (Moses2TMX.exe). A synergy between machine translation and translation memories is therefore created.

The scripts were tested in Ubuntu 9.04 (64-bit version). Documents used for corpora training should be perfectly aligned and saved in UTF-8 character encoding. Documents to be translated should also be in UTF-8 format. One would expect the users of these scripts, perhaps after having tried the provided demonstration corpus, to immediately use and get results with the real corpora they are interested in.

Though already tested and used in actual work, this should be considered a work in progress. So as to protect the users not yet completely acquainted with Moses, these scripts try to avoid mistakes that would cost them dearly in terms of time and/or results, but do not completely insulate them (especially from the consequences of malformed corpora files).