Due to essential maintenance, access to Trados cloud will be unavailable on Saturday 30 August from 00:00 to 12:00 UTC.

DeepL NMT producing poor output on segments split up by multiple tag pairs.

Hi there

We're currently testing DeepL outputs on files for a project and have noticed that the output translations provided by the DeepL plugin seem to be poor, like it's having trouble recognizing the tag pairs correctly.

We're testing on Trados Studio 2024 Professional (on a trial) and using the newest version of the DeepL Translation Provider plugin (7.0.2.0). Our source file is an .xml, but uses non-standard tag pair identifiers.

Example 1:

Screenshot showing text with non-standard tag pairs and a repeated translation error in the output.
In this first screenshot you can see there's a repetition in the output text. The translation within the first tag pair is actually already correct for the entire segment.
Screenshot of a plain text translation without errors, contrasting with the previous image.
Here we can see the same segment in plain text as handled by the DeepL online translator.

Example 2:
Screenshot showing text with non-standard tag pairs and a translation output with separate sentences.
Here we can see the output is 3 separate sentences. While the translation isn't essentially wrong, the translation we got from the online translator using plain text is miles better:
Screenshot of a plain text translation without sentence separation, contrasting with the previous image.

Example 3:
Screenshot showing text with non-standard tag pairs and both repetition and sentence separation errors.
Here we have both issues from example 1 and 2 showing up in the same segment. And again the translation from the online translator with plain text doesn't have this issue:
Screenshot of a plain text translation without repetition or sentence separation errors.

We also noticed that not every segment had this issue.

Example 4:
Screenshot showing text with non-standard tag pairs with correct tag handling and acceptable translation output.

In this segment the tag handling is correct and it's presented as a single complete sentence. The plain text translation is a bit more concise but the output we got from the plugin in Trados Studio is acceptable.
Screenshot of a plain text translation that is more concise but still acceptable.

Additional Notes:

1. We prepped the Studio Project using a custom file type. While the source file is an xml file, it uses some non-standard tag pairs like <st1> <st2> that you can see in the images above.

2. We tried changing the tag handling setting for the DeepL plugin to XML, HTML and None but got the same output with all 3 settings.

3. We found that the DeepL API is supposed to be able to handle this with some adjustment: https://developers.deepl.com/docs/xml-and-html-handling/structured-content. We believe the instructions for restricted splitting would be the solution, but do not know whether this can be set up in the Studio environment using the plugin.

Are there any solutions anyone knows of to resolve this issue we're experiencing?

Thanks

Eddie



Generated Image Alt-Text
[edited by: RWS Community AI at 10:43 AM (GMT 1) on 14 Aug 2024]
emoji