Machine translation of health text degrades readability before it degrades meaning, with sentence complexity as the moderator

Narrative synthesis #

When machine translation of health text fails, readability degrades before meaning does: Khanna found a large fluency gap (3.4 vs 4.7) with no accompanying gap in adequacy, meaning, or serious clinical errors. Turner's error taxonomy supplies the mechanism — word-sense (40%) and word-order (22%) errors dominate English-to-Chinese MT — and both Khanna and Chen independently identify sentence complexity as the moderator. The Turner evidence is a single research program, not independent replication.

This claim is drafted apart from the companion effect-size claim (Machine translation of health materials approaches professional quality in high-resource languages…): that one is about which languages, this one about what breaks first, and they would be graded on different evidence despite sharing four EVDs — collapse into one conditional-quality claim if preferred. Merge candidates (proposal only — do not touch): C-0199ClaimC-0199Initial AI draftUnsupervised machine translation produces less grammatically fluent text than professional human translationUnsupervised machine translation of written medical information produces text that is grammatically less fluent (poorer grammar and readability) than professional human translation., C-0200ClaimC-0200Initial AI draftMachine translation of medical text introduces more errors overall than professional translationUnsupervised machine translation of written medical information introduces more errors overall (errors of any severity) than professional human translation., C-0216ClaimC-0216Initial AI draftMachine translation of English into Chinese is dominated by word-sense and word-order errors that are the most cognitively demanding to correctMachine translation from English into Chinese is dominated by word-sense (mistranslated meaning) and word-order errors, the error types that require the most cognitive effort to correct., C-0202ClaimC-0202Initial AI draftPreference for professional over machine translation depends on sentence complexityThe preference for professional over unsupervised machine translation depends on sentence complexity: professional translation is preferred for complex, multi-clause sentences, but the two are judged comparable for simpl…, C-0209ClaimC-0209Initial AI draftA voice-enabled machine translation app matches professional human translators on simple patient-education sentences but makes more errors as sentence complexity increasesA voice-enabled mobile machine-translation app produces translation quality comparable to professional human translators for simple, short patient-education sentences, but makes more errors as sentence complexity increas….