Trois étapes pour un résultat professionnel

Pour avoir été à l’avant-garde de l’utilisation des mémoires de traduction, DAT se trouve tout naturellement préparée à l’intégration des technologies de traduction automatique basées sur les réseaux de neurones et l’intelligence artificielle dans son processus. Certains clients restent cependant focalisés sur les déboires de la traduction automatique à ses débuts. C’était la période des traductions anecdotiques de « Google Translate » où les systèmes de traduction automatique étaient basés sur les modèles statistiques. Mais depuis, les choses ont énormément évolué avec l’adoption des systèmes à apprentissage profond (le fameux « deep learning ») et les modèles à réseaux de neurones. 
Pour simplifier, nous dirons que les systèmes actuels ne calculent plus mais raisonnent et apprennent par conséquent de leurs erreurs. Et parmi les plus précieux matériaux servant à l’apprentissage de ces systèmes, on trouve tout naturellement les mémoires de traduction (TM pour « translation memory ») que DAT a accumulées depuis plus de 20 ans.

Etape 1 : Pré-traduction

1. From the interface of the translation environment (as SDL Studio), the user enters the text to be translated and chooses the language and domain settings. From this same interface, the user also selects the relevant memories and connects them.
2. The user also connects the automatic translation system API and initializes a pre-translation.
3. At this stage, the output is a partial translation that reflects the content of the translation memories. And it is from there that the machine translation system (MT) takes over to complete the missing translations.

Step 2: Machine translation

1. The machine translation system integrates the terminology and context of translations from translation memories into its process. It also stores them in the user's work instance. This is where some of the learning takes place.
2. For segments that are not in translation memories, the system will use its neural networks and previous learnings to translate them itself.
3. The output of this process is a raw translation where the segments translated by the machine are marked for the post-editing step.

Step 3: Human Post-editing

1. The reviewer receives a pre-translation package and opens it in the interface where the segments from the memories and those from the MT system are clearly identified.
2. He begins his post-editing work by being connected to the translation memories and the MT system. The two systems are enriched in parallel by the reviser corrections.
3. The output is a revised translation, with enhanced memories and an up-to-date MT system.
   Among the advantages of this process is the absence of contradictions and errors in the figures, which is already considerable, especially for technical translations.
   This process allows clients to translate more documents more quickly, at lower cost and with professional quality. For example, it is estimated that the cost savings are in the range of 40 per cent for some documents and that the time savings are in the range of 60 per cent.
   It is not, however, applicable to all documents. The areas of exclusion are for example commercial brochures, websites, advertising and any documents that require transcreation rather than translation.
   However, it is extremely effective for all  manuals, maintenance manuals and technical specifications that require greater fidelity to the text.