Stepan Konev

Have you installed the plugin and looked at the app description on the appstore?  I had a quick play with this when we first approved it for the appstore and it's very simple to do this without any additional information to that already provided.  Certainly a very interesting solution!

Hi Stepan,

Good to know that the English to Russian model is working. When you install OPUS-CAT plugin to Trados, it adds a batch task to the Trados batch tasks, and you can start the fine-tuning by running that batch task for your project. Running this batch task will initiate fine-tuning in the OPUS-CAT MT Engine (the standalone program which the plugin communicates with). The result of this fine-tuning will be a new fine-tuned model, which should perform better with the kind of texts it has been fine-tuned with.

There are detailed instructions on using the batch task here: https://helsinki-nlp.github.io/OPUS-CAT/tradospluginfinetune

You can also finetune with a TMX file directly in the OPUS-CAT MT Engine: https://helsinki-nlp.github.io/OPUS-CAT/enginefinetune

-Tommi

Yes, those steps were the first thing I did. I have installed both MT engine and the plugin, and even tested them with a translatable file. Those instructions are clear. And the plugin is great. I am just affraid that I have lack of knowledge in NMT training terminology and concept. Thank you for mentioning OPUS-CAT in another comment somewhere here that I came across with.

Oh, thank you so much Tommi!
Just one minor question before I go to read your links: according to my understanding, I should completely translate a file first so that all segments are translated and tidied up, and then I fine tune the MT with the said batch task. Is that correct?

Hi Stepan Konev

You might find this of interest: https://community.sdl.com/product-groups/translationproductivity/f/openexchange_applications/35718/opus-cat-mt-fine-tuning-graph

Daniel

Wow! Will carefully read this. Thank you!
But am I right in thinking that terms "fine-tuning" and "training" are used interchangeably?

The fine-tuning functionality is meant more as a preprocessing step before starting work a project. The workflow I had in mind when I designed was as follows:

1. You receive a translation job with existing translations or a useful TM (project-specific translations).
2. You initiate the fine-tuning of the base model in OPUS-CAT MT Engine with the project-specific translations using the batch task.
3. Wait an hour or so for the fine-tuning to complete.
4. When the fine-tuning is complete, select the fine-tuned model in the OPUS-CAT plugin (by selecting its model tag) and start on the project.

The fine-tuning is ideally suited for large-scale projects that have a lot of existing translations. The fine-tuning will usually pick up the terminology and sentence structures used in the existing translations, which makes the MT much more useful (depends on the project, though).

It might also be useful to train a fine-tuned model for a specific client, and then use this model for all jobs for this client. This would be easiest to do with the TMX fine-tuning in the engine.

Amazing! Thank you!

As for the difference between fine-tuning and training, fine-tuning means continuing the training of an existing model with a fairly small batch of highly-relevant data (thousans or tens of thousands of sentence pairs). Training usually means training a model from the scratch using a massive parallel corpus (tens of millions of sentence pairs). Then in-between those two terms there is customization, where you use a larger batch of domain-specific data to customize a model. But the boundaries between the terms are fuzzy.

Thank you all, gentlemen!