Does a custom translator model work in both directions or do you have to train two models?

Question

For example, would an English to French model also work for French to English or do you have to train a separate model.

Answer 1

@Kenefick, Andrew Apologies for the confusion. The above snippet from our learn documentation was confusing. So, I have moved the previous response as a comment. I am also discussing this internally to see if the documentation can be improved.

.

As your rightly pointed out. You would need to train a separate model for the reverse translation.

When you create a custom translator model in Azure, it typically works in one direction. So, an English to French model would not automatically work for French to English.

.

This is because translation models are optimized for the specific nuances and context of the source and target languages in one direction. Training separate models ensures higher accuracy and better performance for each translation direction.

.

The documentation is highlighting that you can use the same set of parallel documents to train both directions of translation, but you still need to create two separate models: one for English to French and another for French to English.

Here’s a bit more detail:

• Parallel Documents: These are pairs of documents where one is the translation of the other. They can be used to train translation models in either direction.
• Training Models: When you train a model, you specify which language is the source and which is the target. For English to French, English is the source and French is the target. For French to English, French is the source and English is the target.

So, while the same parallel documents can be used for both directions, you need to train two distinct models to handle each translation direction effectively.

I hope this clears things up!

**

Please do not forget to "Accept the answer” and “up-vote” wherever the information provided helps you, this can be beneficial to other community members.