When we welcome new customers into the Unbabel family, we need to train our translation engines to make sure that they’re working exactly as they should be for that customer. Because everyone writes in a slightly different way, and each brand has their own corporate identity, we need to make sure that the work we produce fits with what each customer expects. We do this in two ways:
- Producing an extensive glossary to ensure that editors have access to help for terms that need to be consistent;
- Running domain adaptation so that each client has an AI engine that fits them.
Domain adaptation (DA) tasks help us build this AI engine - and like any good technology, it needs a skilled human working on it from the start. In these cases, the skilled human is you - the editor.
When you submit DA tasks, they go to our data team who use them as a source of data that they input into the machine, making sure that next time the machine translation is better than before. When we launch with the customer for real, this means that editing is quicker and it’s easier for you to produce high quality work.
What do I have to do?
Working with domain adaptation tasks really isn’t that different from what you’re used to - but what you’re aiming to do is adapt the generic machine translation so that it fits well with the source.
Here’s an example of one of these jobs:
There are strings of short sentences which you need to edit to make sure they’re native quality. However, they’re not necessarily going to make sense when you put all of them together. That’s okay - we just put lots of different sentences in one job so that you don’t have to submit lots of smaller jobs every time.
Once you’re sure that each segment is perfect, move onto the next task and you’ll get a new set of sentences.