Create back-translation.md

[liashahnazaryan]

Description

Fixes # 81

Type of PR

• Creates the article [Back-translation]

Checklist:

• I have read the contributing guidelines.
• I have followed the style guide.

[cefoo]

Thanks so much for your hard work, @liashahnazaryan!

I hope my comments make sense. Let me know what you think. :)

[cefoo]

Perhaps this definition is not clear, should we say something like "Translation of the machine translation output back into its input language"?

[bittlingmayer]

Another option. would be something like "Creating parallel data by translating monolingual data from the target language to the source language"

[cefoo]

I'd simplify "the monolingual data in the target language" to make it less wordy.

What do you think of something like this?

Back-translation is the process of using machine translation to translate again a machine translation output to generate synthetic data.

Steps

1- First, the machine translation system translates a text from one language to another language.
2- Then, the system uses the machine translation output as input, and translates it back into the original language.
3- Finally, the system translates the resulting synthetic input text again into the output language.

[cefoo]

Do you think creating an image would make it easier to understand?

[bittlingmayer]

I disagree with this description, because it makes it sound like then it is translated from source to target. But the training data is only translated from target to source.

Translation from source to target only happens at inference time, but that's a given.

[cefoo]

This is a great place to introduce the "augment" term!
Perhaps we can say that synthetic data is necessary to augment parallel data so that there is more data to train machine translation systems and improve quality.

[bittlingmayer]

"necessary" is not always true, we can be less opinionated.

[cefoo]

I don't think it's clear how it differs with back-translation. Is the "existing" translation human-generated?

[bittlingmayer]

I don't want to make articles longer, but I see a few things missing:

• the motivation: there is a lot more monolingual data out there than parallel data
• the effect: if overdone, helps fluency more than accuracy, since it's kind of like a target-side language model
• adoption: Google, Microsoft, DeepL etc use this technique heavily
• references: can we link some papers or articles? there was one year at WMT where this really become known

[liashahnazaryan]

I'll make this a draft to work on the changes and add the suggested parts.