Exodus - Exploring SMT for EU Institutions

In this paper, we describe Exodus, a joint pilot project of the European Commission’s Directorate-General for Translation (DGT) and the European Parliament’s DirectorateGeneral for Translation (DG TRAD) which explores the potential of deploying new approaches to machine translation in European institutions. We have participated in the English-to-French track of this year’s WMT10 shared translation task using a system trained on data previously extracted from large inhouse translation memories. 1 Project Background 1.1 Translation at EU Institutions The European Union’s policy on multilingualism1 requires enormous amounts of documents to be translated into the 23 official languages (which yield 506 translation directions). To cope with this task, the EU has the biggest translation service in the world, employing almost 5000 internal staff as translators (out of which 1750 at the European Commission (EC) and 760 at the European Parliament (EP) alone), backed up by more than 2000 support staff. In 2009, the total output of the Commission’s Directorate-General for Translation (DGT) and the Parliament’s Directorate-General for Translation (DG TRAD) together was more than 3 million translated pages. Thus, it is not surprising that the cost of all translation and interpreting services of all the EU institutions amounts to 1% of the annual EU budget (2008 figures). According to our estimations, this is more than e 1 billion per year. 1.2 Machine Translation and Other Translation Technologies at EU Institutions In order to make the translators’ work more efficient so that they can translate more pages in the same time, a number of tools like terminology databases, bilingual concordancers, and, most importantly, translation memories are at their disposition, most of which are heavily used. http://ec.europa.eu/education/ languages/eu-language-policy/index en.htm In real translation production scenarios, Machine Translation is usually used to complement translation memory tools (TM tool). Translation memories are databases that contain text segments (usually sentences) that are stored together with their translations. Each such pair of source and target language segments is called a translation unit. Translation units also contain useful meta-data (creation date, document type, client, etc.) that allow us to filter the data both for translation and machine translation purposes. A TM tool tries to match the segments within a document that needs to be translated with segments in the translation memory and propose translations. If the memory contains an identical string then we have a socalled exact or 100% match which yields a very reliable translation. Approximate or partial matches are called fuzzy matches and usually, the minimum value of a fuzzy match is set to 65%–70%. Lower matches are not considered as usable since they demand more editing time than typing a translation from scratch. First experiments have shown that the quality of SMT output for certain language pairs is equal or similar to 70% fuzzy matches. Consequently, the cases where machine translation can play a helpful role in this context is when, for a segment to be translated, there is no exact match and the available fuzzy matches do not exceed a certain threshold. This threshold in our case is expected to be 85% or lower. To this end, there exists a system called ECMT (European Commission Machine Translation; also accessible to other European institutions) which is a rulebased system. However, only certain translation directions are covered by ECMT, and its maintenance is quite complicated and requires quite a lot of dedicated and specialized human resources. In the light of these facts and with the addition of the languages of (prospective) new member states, statistical approaches to machine translation seem to offer a viable alternative. First of all, SMT is data-driven, i.e. it exploits parallel corpora of which there are plenty at the EU institutions in the form of translation memories. Translation memories have two main advantages over other parallel corpora. First of all, they contain almost exclusively perfectly aligned segments, as each segment is stored together with its translation, and secondly,