1. Collection – Collect language data from websites, glossaries, language databases, documents, TMS exports
2. Extraction – Extracting the language pairs from the collected data
3. Processing – Validating, cleaning and processing the language data for the trainings
4. AI-Training– Training the language models on the translation knowledge
5. Evaluation – Evaluation of the language models by the customer in the model blind test

1. Collection:

You would explain language rules to a new translator. You might show them a directory of product names. There may be a glossary in which you have defined technical terms and translations. In some cases, you can even include a document on corporate language and tone-of-voice. Or you don’t have a structure and instruct the translator to look at your publications and keep that style. We do this part of the collection fully automatically as soon as we receive the information.

2. Extraction:

From all the sources and data that are made available to us, we have to find pairs of sentences.
Sentence pairs are the currency of translation model training.

For example, an example of a training set at KWS Saat is:

Deutsch: “Und so widerstandsfähig wie Roggen ist kaum ein anderes Getreide.”
Englisch: “And there is hardly any other cereal that is so resistant as rye.”

This sentence contains both grammar and the linguistic tonality that KWS speaks of “cereal” while generic translation models would translate  “grain”.

3. Processing:

The data we receive is usually unsorted. Especially in published documents and websites, we find texts that have not been translated 1:1 in the same place and appropriately. That’s why we have developed our own technology to find and process these pairs of sentences. 
This makes training economically and time-wise.

4. AI Training:

We bring together the processed language data in a training corpus. Of course, with a control group and optimally prepared for the model to be trained. The training cycles are then run through until the results are good enough for evaluation.

5. Rating:

In the evaluation phase, we make the results available in a separate area. There, the translation managers can validate and check the results of the trained business models with those of generic models or previous translations.

During operation, we provide you with your trained translation models via API and web frontend. With each translation, you can add changes from the proofreading. In addition, we periodically use the training sources and check which changes have taken place in the vocabulary in order to train them back into the model.

And if we have ad-hoc changes and, for example, release a new product? In our training center, you can manage the changes directly and determine that translations have to be changed immediately. We also regularly include your adjustments in the training center in the model training.

We were able to complete the data exchange, training and evaluation phases in just a few working days. A common value is about 3-4 weeks. The critical path here is access to the data – especially in the case of third-party systems such as translation memory systems. And the availability of those responsible for translation for evaluation.

In der Evaluation werden etwa 200 Satzpaare geprüft, Sie sollten mit etwa 16 Stunden Aufwand rechnen.