Machine translation isn’t perfect - that’s why we’ve got editors in the first place! Sometimes, though, the translation engines produce what we call “hallucinations” in the machine-translated text. These are glitches, and often mean that the meaning of the text changes completely unless they’re spotted and corrected.
The way that our machines are trained means that the more data you feed into them, the better they become at translating. Eventually, as the machine absorbs more sentences expressing different concepts and situations, it comes closer and closer to the quality of language produced by a human. However, sometimes if the machine is struggling, it can default back to language it might have learned before, completely ignoring the input and producing an error in the output. ThIs is a machine translation hallucination, and they can be difficult to spot and can change the meaning of sentences completely. This is why we really need human editors to do what only humans can do well - spot and neutralise these errors.
Here are some real-life examples that have been reported by clients:
We’re now going to provide you with some real-life examples. Although there can be all sorts of errors in the machine translation, hallucinations tend to fall into four main categories, which are the ones you need to look out for:
- Names, such as organisations, places (often countries), proper nouns, people;
- Numbers (pay particular attention to those involving money);
- Dates;
- Negatives (where the machine misses out a negative from a translation, which then inverts the meaning of the text).
|
English original |
German machine translation |
Description of issue |
|
We do not have any maps that cover Georgia or Armenia. |
Wir haben keine Karten, die Deutschland oder die Schweiz abdecken. |
Georgia and Armenia have been translated as “Germany” and “Switzerland” respectively. |
|
A refund of 194.73EUR was done on January 12th. |
Eine Rückerstattung von 19.73 EUR erfolgte am 12. Januar. |
There was a glitch and the translation engine changed the amount from 194.73 to 19.73, causing confusion for the customer. |
|
English original |
Spanish machine translation |
Description of issue |
|
We did not process a refund. |
Hemos procesado un reembolso. |
The negative was not translated and the meaning of the target has been inverted. |
|
Thanks in advance. I look forward to hearing from you again. Kind Regards, John |
Gracias de antemano. Espero tener noticias suyas nuevamente. Saludos cordiales, Juan |
The machine translation translated the name, when it should have been kept in the original English version. |
When put together in those tables, these mistakes seem really obvious. However, even editors who have been working with us for years still make these mistakes. It’s easy not to check numbers, dates, and names, assuming that the translation engine will have translated them correctly - but it’s imperative that you check every part of the text. Each of these could have been avoided with just a very, very brief read through. Unfortunately, in these situations, clients often complain and such major errors may needlessly result in poor evaluation scores if a task with an uncorrected MT hallucination is selected.