By default, Translation Hub uses the Google Neural Machine Translation (NMT) model to translate documents, which is suited for generic translation tasks. In cases where you have to translate a specific domain area and writing style, consider using a custom model (also known as an AutoML Translation model). Custom models can provide more tailored predictions.
When you train a custom model, AutoML Translation starts with the general Google NMT model and tunes the model to fit your training data, which includes pairs of sentences in a source and target language.
You create and manage custom models through AutoML Translation. Prepare a training dataset with your sentence pairs and then use AutoML Translation to create a custom model. Any costs that are associated with training custom models are charged separately by AutoML Translation. For more information, see the AutoML Translation documentation.
Translation Hub automatically makes new and existing AutoML Translation models available for administrators to assign to portals. After you add a custom model to a portal, portal users can choose to use the custom model for their translations.
If you have a translation memory, you can export that data and use it as training data, depending on the number of sentence pairs you have and their quality. AutoML Translation recommends around 6,000 sentence pairs. In general, higher quality data (like full sentences) results in higher quality models than more data.
Add models to portals
Administrators add models to portals by using the Google Cloud console. Portals users can choose to use these models when they request translations.
In the Translation Hub section of the Google Cloud console, go to the Resources page.
From the list of resources, select one or more models to add to one or more portals.
Click Assign to portals, which opens the Assign resource to portal pane.
From the portals field, select one or more portals to add the models to.
On the Resources page, you can confirm the addition by viewing the Portal names column for each resource.