External Machine Translation

[AgentOren]

Dear Community,

In one of our project, we currently use ChatGPT with a fine-tuned model to translate texts in our specific domain.
So far we have very good results, much better than other machine translation services we tried, due to training/fine-tuning based on human-translated content in our domain.

I'm currently pushing new translations to Tolgee using the "Create new key" API. Since new texts are added frequently and reviewed by technical writers and translators, I would like to have the same effect as auto-translation in Tolgee: the texts are marked as "Machine Translated" with a special icon, to distinguish between machine-translated and human-translated texts.

I thought about two paths to achieve this:

1. Develop a custom "Machine Translation" service for Tolgee: Unfortunately, I could not find documents about this.
2. Push translated texts like I do today, but mark them as "machine translated". Unfortunately, I could not find an API that does this.

Has any of you used an external MT with Tolgee?
If so, how?

[JanCizmar]

Hi Oren, we were thinking about adding support for custom translation services, also adding the endpoint for marking translation "machine translated" would be pretty simple. :)

Anyway, it would be great to have a call with you, we always planned to add fine-tuning as a native feature and it would be great to know how far you've got. Would you be open to? This is my Calendly 🚀

[AgentOren]

Hi Jan!

Actually the usage of our fine-tuning model attempts are currently on-hold. The reason is that fine-tuning is currently only possible with older GPT-3.5 models. The good results we've seen are actually coming from stock GPT-4-Turbo and GPT-4o models, using a custom prompt (sorry for misleading). We're waiting for fine-tuning to become available for GPT-4 or newer models to resume our fine-tuning attempts.

Our custom prompts works better than other MT services we've tried, as the prompts include special instructions for handling our texts, e.g.: some texts contain markdown format that must be preserved, including code-like blocks that must never be translated or modified.

MT services also don't allow us to provide per-language or per-context instructions, such as translation of internal terms and industry Jargon (e.g. there are Hebrew translations for terms like "Internet", "Logs", "Ports" but they're rarely used by IT personnel, who prefer a less-formal phonetic copy of these English terms).

I'll be happy to share some knowledge with you, let me know if you still want to talk/chat.