From NLP to Large Language Models

For most of the last two decades, machine translation in the region meant statistical systems that guessed the next word from patterns. Large language models changed the question entirely: the machine now carries a model of language itself.

That shift is real, and it is useful. Drafts arrive faster. Large volumes of repetitive institutional content, reports, forms, standard clauses, can be moved through at a speed that was not possible before. For a firm that has handled regional language work since 2004, this is a tool worth using well.

Where the models still fall short

General models are trained on the languages that dominate the internet. Albanian, Serbian and Macedonian institutional language, with its specific legal register, its terminology, and the context a certified document must satisfy, is thinly represented. The model will produce fluent text; fluent is not the same as correct.

A model can write a confident sentence in Albanian. Whether a court accepts it is a human question.

This is why Trankos treats AI as an accelerator, not a decision-maker. The model speeds the first pass. Qualified linguists own accuracy, register and the final, certification-grade judgement about what ships.

What this means for institutional clients

The practical answer is a workflow that keeps humans in the loop: AI for speed and scale, people for judgement and responsibility. It is the same discipline Trankos applies to translation memory and Trados, use the tool, but never outsource the decision.