Problem

When a task has a pivot language and an AI language resource (OpenAI, Azure OpenAI, translate5 AI) is used as pivot resource, the pivot pre-translation prompt asks the model to translate from the source language into the pivot language. This part is correct. The terminology that is added to the same prompt is collected for the task target language instead of the pivot language. The model then receives the instruction "translate into English" together with term pairs from German to Slovak, and it tends to follow both. In such cases the pivot column can contain words in the target language.

Example (task de-DE to sk-SK, pivot en-GB, AI resource as pivot resource, term collection with de, sk and en terms): the source "Halterung, für Schiene" can get the pivot text "Bracket, for koľajnica". A segment that consists of a single term, for example "Netzkabel", can get the pivot text "sieťový kábel". Segments without a term hit and segments that come from a translation memory are not affected.

The situation occurs only when an AI resource is used as pivot resource and term collections are attached to the task. Pivot content from translation memories is correct.

Solution

Collect the terminology for the language pair of the connector, not for the language pair of the task.

• The terminology provider gets an optional target language. When it is given, the target terms are looked up in that language and the same-language check compares the source language with that language. Without the parameter the task target language is used, so all other callers (for example TQE) keep their behaviour.
• The term retrieval of the TermTagger processor gets the same optional target language and passes it to the term lookup, which already supports a target language. The term tagging of the source text itself does not change.
• The OpenAI connector passes its own target language, which is the pivot language during pivot pre-translation. Both call sites change: the single segment query and the batch request.
• Terms without an entry in the pivot language are skipped, as today for missing target terms.
• Verification: a de-DE to sk-SK task with pivot en-GB, an AI pivot resource and a term collection with de, sk and en terms. After pivot pre-translation the pivot column must contain English terms only.