Customized Google Translate

[worldwide booking]

Hi Google Team,

We looking for Arabic translation  for my Web and App (IOS, Android)

And i want to understand that translation API advance and AutoML translation how will manage if get wrong translation?

Thanks,

Ishtiyaq Ahmad

[Victor Rodriguez Pastor]

Hello Ishtiyaq,

    I understand that you might be worried about the performance of your model. Because Arabic and English are quite different languages, it might be difficult to convey in both languages the exact same meaning and context even for a human expert. Also, please bear in mind that the way in which humans understand the language is different to how computers do so because of the way in which words are represented (as vectors of numbers).

    Nonetheless, before using your model in production, Google Cloud Platform provides you with the Bleu Score [1] so you can know beforehand how good your model is. There are two kinds of models you can use in GCP depending on whether you are going to use everyday language or if, instead, the scope of your app will focus on a certain domain.

    The Natural Language API [2] is a general model that is not specialised in any specific field of knowledge. Nevertheless, better results can be obtained by providing your own datasets and labels for a specific domain (for example IT, biology, jargon language, etc). If that was your case, I would suggest you also check the AutoML Natural Language product in GCP [3] if you want to provide your own domain-specific keywords or phrases.

    For troubleshooting further issues that might arise about this topic, please also consider using IssueTracker [4] or signing up for one of Google Support plans [5].

Have a nice day!

[1]: https://cloud.google.com/translate/automl/docs/evaluate#bleu

[2]: https://cloud.google.com/natural-language/docs

[3]: https://cloud.google.com/translate/automl/docs

[4]: https://developers.google.com/issue-tracker

[5]: https://cloud.google.com/support

[Adam Bittlingmayer]

Ishtiyaq and I are following up offline, but for the benefit of all, I'll give a brief overview of this topic.

(Full disclosure: I'm the CEO of ModelFront, and no longer affiliated with Google and Google Translate.)

The short answer is that no major translation API provider (Google, Microsoft, DeepL...) includes a "confidence" score, "quality estimate" or risk prediction with the translation.  That's understandable, because what's a good or bad translation depends on the context and use case.

So the only way to catch bad translations is with an independent system.  There are a few options: you can build and train your own with an open-source project like OpenKiwi, or use a ready-made solution like ModelFront.  It's possible that your TMS has its own QE (Memsource) or already integrates ModelFront risk prediction (translate5, GlobalDoc LangXpert, Homer Language Manager...)

The best solution will depend on your stack and your use case - hybrid translation (flagging vs. safely auto-approving raw machine translation), just estimating post-editing effort for the project, or comparing different APIs.

Adam