Problem

TMs are used to re-use previous translations and thus lower translation costs. However, terminology work in a manufacturing or other industry company is an ongoing process. When terms are changed, all segments affected are no longer up to date.  In case terminology errors were found, 100% match pretranslations should not be locked and high fuzzies need to be adjusted by linguists who are familiar with CAT tool work.

But in many scenarios TM and MT usage are mixed and not linguists but in-country colleagues do the textual review of MT and TM pretranslations. They are not familiar with adjusting terminology in outdated segments. 

Solution

After pretranslation, the AutoQA should check all pretranslated segments for terminology (most important, also other qualities like spelling or tag errors are implemented if only little more effort). All segments with error(s) should receive a penalty in match rate by a configurable value. Example: 100% match with terminology error (in target language) should be a 79% match and a segment pre-translated with a 90% fuzzy match would be a 69% fuzzy match, if penalty 21 was configured. This will be called "AutoQA pre-translation penalty" from now on.

When the new match rate is below the MT threshold (e.g. 80%), then

• the pre-translation would optionally be removed AND MT should pretranslate this segment again (using terminology functionality as with DeepL or LLM).
• Thus the review will become more feasible for non linguists.

Details:

• Create one new system config "AutoQA pre-translation penalty" for each language resource type. If config is opened for editing it opens a grid with one row for each AutoQA quality. And the the following colums
  • Match penalty, if error in pre-translated segment
  • Empty the segment, if error found
  • If emptied, pre-translate segment again. Tooltip (developer hint: Logic to be implemented is in tooltip text!):
    • For new pre-translation other language resources than the one used so far will be used
    • AutoQA is done for new translation.
    • Redo in loop until no AutoQA error with penalty is found anymore or no language resource is left for pre-translation.
      (implementation hint: Store best match from every language resource in temp-Table (call all language resources already in first pre-translation) in RAM on initial pre-translation. Then just ignore the one match, that was deleted from the pre-translation and look for the next-best match in temp-Table. I would not run AutoQA over all translations from all language resources always (so not over content of complete temp-table), as this increases time for AutoQA to much). Redo AutoQA only for segments, that have been filled with new translation to save processing time.
      For this to work we also need to implement https://jira.translate5.net/browse/T5TMS-302 for t5memory (Implement switch for fuzzy request to always also look for fuzzy matches). And then in fuzzy queries to t5memory set this flag to "true" if an AutoQA pre-translation penalty is applied for an import.
• Config is overwriteable on client and import level
• Penalties from AutoQA Qualities are added, if there are multiple errors from different or the same qualities in same segment
• The match rate in the match rate column of the editor includes the AutoQA pre-translation penalty
• if a possible AutoQA pre-translation penalty is active for an AutoQA-Quality, the check in question is also run for segments, that should be locked because of match rate (so above 100%)
• After AutoQA and penalty only those are locked (if locking of matches above 100% is configured) that are still >=100% after the penalty is applied 
  • ensure that the locking of 100% matches is done after AutoQA and not as part of the pretranslation - or always send locked segments to the AutoQA and open the lock if match rate is reduced
• If AutoQA pre-translation Penalties are active for a task, in the fuzzy match panel of the editor the following text is displayed:
  • DE: In der Vorübersetzung wurden "AutoQA Vorübersetzungsabzüge" angewandt. Daher kann es sein, dass der Match-Wert eines Treffers im Fuzzy-Panel höher ist, als der beim Segment angezeigte Vorübersetzungs-Match-Wert.
  • EN: “AutoQA pre-translation penalties” were applied in the pre-translation. It is therefore possible that the match rate of a match in the fuzzy panel is higher than the pre-translation match rate displayed at the segment.

Note: Penalty is NOT applied in fuzzy match panel by purpose, because this increases complexity very much and users of fuzzy panel should be experts that respect AutoQA anyway. Therefore called "AutoQA pre-translation penalty".