TM-Augmented Translation - TTN Translation Network

In TM-augmented translation, the text is not generated by the AI but by the translation memory. Full matches are adopted unchanged, partial matches are passed to the AI provider as a specification, and the termbase entries relevant to the segment are inserted into the translation prompt. The AI supplies only what is missing from the existing material. With highly repetitive texts – bulletins, reports, standards, catalogs – that is often no more than a few percent of the text. TTN supplies the framework, the AI the filler words.

How the pre-translation is produced

Before the pre-translation, TTN TMS compares every segment with the translation memories attached to the project. Matches of 99 % and 100 % are adopted unchanged in the target text; they are never sent to an AI provider at all. For matches between 60 % and 98 %, the TM entry found is attached to the prompt as a specification, together with the instruction to translate the segment on that basis. In parallel, the system looks up the termbase entries relevant to each segment and inserts them into the prompt as well – the term injection. For the short avalanche bulletin published daily in winter by the SLF in Davos, that amounts to up to 80 terms.

All newly generated translations flow into a second translation memory, which is attached as a secondary resource. The material maintained by human translators thus remains the primary source, while the result of the machine pre-translation is kept separate and nevertheless reused.

With this way of working, the percentage shown in the Trados editor has a different meaning than in the classic process without pre-translation. An orange 74 % match does not mean that 26 % of the segment differs from the source text, but that 26 % of the segment was translated or adapted by the AI, while 74 % was taken unchanged from the translation memory.

The effect is immediately measurable: term injection and TM hints improve the quality of the pre-translation considerably. It is therefore worth adding as many terms as possible to the termbase – every entry saves work on the next order. There is also an effect that is gaining weight over time: AI models are increasingly trained on their own output, which is referred to as AI cannibalization or the Habsburg phenomenon and tends to reduce the quality of purely machine translation. Term injection and TM hints break this cycle, because the point of reference remains the material maintained by humans.

Further arguments in favor of TM-augmented translation

1. Repeatability. An AI translates the same bulletin differently every time. Execution depends, among other things, on the computing time available, which results from the total number of pending tasks divided by the available computing capacity. After one or two seconds a timeout takes effect and the evaluation of the candidates is aborted. Anyone translating during the football World Cup gets a good translation; in rush hour, a poorer one. As far back as the 17th century, René Descartes established that a phenomenon is scientific only if it is repeatable. At TTN, the text is generated primarily from the TM maintained by human translators; all newly generated translations flow into a second TM that is attached as a secondary resource. The same input produces the same output at TTN, 100 % of the time. With ChatGPT, something different comes out every time.

2. Traceability. After a major avalanche accident, for example, a public prosecutor’s investigation usually follows; the warnings, the safety measures and their procedures and processes are examined closely for weaknesses. Numerous Federal Supreme Court rulings attach central importance to traceability. A model such as Kimi K3 with 2.8 trillion parameters is a black box – the provenance of a translation cannot be evidenced. With the TTN model, the log shows how many 99 % and 100 % matches, how many TM hints at what percentage and how many injected terms went into the result. TTN supplies the framework, the AI the filler words – which, in a highly repetitive bulletin, account for only a few percent.

3. Correctability. With a model containing trillions of parameters, nobody knows exactly where to intervene when a text is unsatisfactory and should be generated differently in future. In TM-augmented translation, all data resides in a Microsoft SQL database. A single command changes every occurrence of a text token, so that the new term or the new fuzzy match is used from then on.

4. Technology updates. As a high-technology translation agency, TTN processes texts in series and, in an extremely fast-moving technological environment, is better placed to stay at the technological forefront.