Work

Translating by Meaning, Not by Words: A Two-Stage Method for Native-Quality Machine Translation

Year: 2026
Type: article
Abstract: Machine translation, including translation by strong LLMs, produces output that is grammatically correct and unmistakably foreign: calqued phrases no native speaker would write, invented compound words, English sentence structure wearing local vocabulary. Users in smaller-language markets read this as "the product was not made for me." We show that this is not a model-quality problem. It is caused by the instruction "translate", which makes the model preserve the source's structure. Change the task and the foreignness disappears. We give a two-stage method built on that insight: stage one reduces the English source to plain meaning; stage two hands that meaning to a native-speaker persona who authors the copy from scratch, never seeing or mirroring the original's structure. Two anti-rules remove the most reliable machine-translation tells (replace calques with what a native would actually write; never invent compound words for foreign concepts). We describe the placeholder and ICU-plural guards that make the method safe for software strings, a ship-blocking artefact scanner, and an LLM-as-native-reviewer verification loop that runs without native staff. We also report the negative result that keeps our language list honest: for one low-resource language we evaluated, no available model writes acceptably, and the correct engineering decision was to refuse to ship it. Self-published technical report by FRITS AI ApS (report FRITS-TR-2026-02). Not peer-reviewed. Canonical version: https://frits.ai/research/meaning-first-translation/
Language: English
Open Access status: green
OpenAlex ID: W7167786162
FWCI:
Cites:
Cited by:
Related to: 10