From terminological inconsistencies to major mistranslations

This article presents a qualitative analysis of frequent NMT errors in public health communication. Its primary aim is to sensitise users and decision-makers to common issues and raise public awareness for potential pitfalls and limits of current state-of-the-art NMT systems. In this context, the article also addresses the usability of raw NMT output in emergency situations. The investigation itself focuses on pandemic-related WHO texts and consists of a fine-grained manual error analysis encompassing three languages (English, French, and Spanish) and two NMT systems (DeepL and Google Translate). The five most frequent error types observed in this investigation included mistranslations, inconsistent use of terminology, unidiomatic or awkward style, untranslated text, and other internal inconsistencies. These error types are illustrated with examples and analysed in terms of their severity, their underlying causes, and their potential consequences. The findings show that raw NMT output is useful only to a very limited extent and that the risks and benefits associated with its use should be assessed extremely carefully.