Why Thai is hard
There are no spaces between words. The system has to decide where each word begins and ends before it can translate anything, and Thai script does not mark it. Segmentation errors do not produce slightly wrong translations; they produce translations of words that were never in the source. Everything downstream inherits the error.
Nothing is inflected. Verbs do not mark tense. Nouns do not mark number. Where English encodes time, quantity and definiteness in word forms, Thai encodes them in context, particles and optional markers — or leaves them implicit. Translating into Thai means deciding what to make explicit; translating out of Thai means recovering information the source never stated. Engines do both by pattern-matching, and they are frequently wrong in ways that read perfectly.
The register system is layered and socially consequential. Thai distinguishes multiple levels of formality, with different vocabulary for the same concept depending on who is speaking to whom, including registers reserved for monastic and royal contexts. Using the wrong level is not a stylistic slip; it can be offensive, and in some contexts it is a legal matter. An engine has no access to the social information that determines the choice.
Classifiers are obligatory. As in Chinese, counting requires a measure word selected according to the category of the thing counted. Errors mark text as machine-produced instantly.
Training data is comparatively thin. Thai is a low-resource language in the technical sense — not for lack of speakers, but for lack of the digitised parallel text that engines learn from. Fewer examples, more guessing.
Rendering and line breaking are their own problem. No spaces means no obvious break points. Thai text in interfaces, subtitles and fixed-width layouts breaks in the wrong places, and the failure appears after release rather than in review.
What we do
- We check segmentation-driven errors explicitly, because they are the category that produces confidently wrong output rather than awkward output.
- We fix the register decision at the start — audience, setting, relationship — and score register errors as their own category with their own severity scale.
- We verify what the source left implicit: tense, number, definiteness, and whether the Thai target commits to something the English never said.
- We review rendering as well as text, including line breaking and layout behaviour in interfaces and constrained spaces.
- We use reviewers matched to the domain, which in Thai is a real supply constraint and part of why this pair costs what it costs.
Where Thai work is worth most
Product and software content, where the interaction of segmentation, rendering and constrained layouts produces failures no engine can see. Manufacturing and industrial documentation for Thailand's substantial production base. Maritime and transport documentation, given Thai crewing. Medical and pharmaceutical content, where the register decision and the implicit-information problem are both consequential.