In fact, machine translation (MT) systems have advanced to the point where machine translation of general-language source texts works almost flawlessly. A fundamental requirement is that the AI is trained using large datasets. However, even the best MT system cannot guarantee that a machine-translated text is error-free. In many areas, the human factor still plays the most significant role. Translation errors are particularly likely when the operators of the MT system have little or no command of the target language.
Source of error: homographs
A major danger with machine-translated texts lies in an inconspicuous detail: homographs. These are words that have the same spelling but different meanings. Many rule-based and statistical MT systems cannot actually interpret the correct meaning of a homograph in such cases – whether a translation is correct is therefore purely a matter of chance. Here are some examples:
| Word | 1. Meaning | 2. Meaning |
| translate | translated into another language | go to the other side |
| modern | contemporary (adj.) | rot (verb) |
| seven | Number | filter, extract (verb) |
| Bug | Ship part | Programming error (translated into German) |
| Assembly | Day of the week (plural) | Assembly (Germanized) |
Neural machine translation systems represent an exception to the error-proneness associated with homographs. If these systems have been trained and fed large datasets beforehand, they can easily distinguish between the different meanings of a homograph. They infer the appropriate meaning from the context and then translate it into the correct equivalent in the target language.
Problems and opportunities of MT systems in specialized translations
In general, the risk of serious errors in specialized translations is increased simply due to the potential for incorrect application of the machine translation (MT) system. Furthermore, machine translations are of little use for texts in the fields of marketing or literature, regardless of the operator’s skill and experience. This is because linguistic nuances, idioms, and wordplay are often completely lost in the process. A literary masterpiece in Japanese can then degenerate into an emotionless novel in German.
Machine translation of legal, technical, or medical texts is possible. However, we advise against using machine translation without subsequent proofreading: You bear a high risk of resulting translation errors. Furthermore, many legal uncertainties still exist in the field of machine translation. For example, who assumes responsibility and liability for personal injury or property damage resulting from an error in a machine-translated text? If you prefer not to venture into this uncertain territory, you should probably rely on the services of a human translator.
Despite this, machine translation holds enormous potential in many of these specialized fields. Neural machine translation followed by post-editing by a subject-matter expert is already a viable and sensible alternative to human translation.
