The Gradual Rise of the (Translating) Machines

Anyone who had to quickly translate a text or an email has probably used Google Translate at least once. And, depending on the source material, they will either have got some good or some rather, um, funky results.  

Either way, texts translated by machine always require some human post editing; how much, exactly, depends mostly on the type of the text and the style.

Machine translation works well only with certain text types. It is not suitable for translating, for example, marketing or creative texts, where phrases, idioms, or style can change the meaning of the text.  

Machine translation doesn’t recognize fixed expressions, and cannot really distinguish between styles such as sarcasm, translating them word-by-word. It uses terminology quite well, though, meaning it is excellent for translating legal or technical texts, says Igor Varga, a freelance translator.  

“These two fields are the most important ones where it is supposed to be used. Nevertheless, companies are trying hard to force its use on other types of texts as well,” he notes. Even when using it for legal texts there is a risk that the translation could look good but mean something completely different, he adds.

Because it is unable to properly recognize style and register, machine translation always requires review and post editing, which can sometimes mean more work for an editor/translator than translating the text without it.  

Still, there are translation companies which, mainly with the aim of saving money, use machine translation and commission post-editing.  

OK vs Perfect

There used to be tutorials for post-editing which differentiated two quality levels – there were texts that were fine if you edited them just “OK” (resulting in a more raw text), and others required full editing (where the output text must be perfect). These, however, seem to have disappeared, probably because clients insisted on good quality and therefore only the second variant survived, Varga notes.

There are, however, some cases when texts can be translated entirely by a machine: usually for in-house use by, for example, a large car-making company, where a huge corpus of vocabulary has been accumulated throughout years.  

“In these cases, there is no need for translator; the machine translation will result in a quality that workers can use,” Varga says.

For now, though, machine translation is unlikely to replace translation by people, the expert says. It is simply not smart enough and, as already discussed, cannot handle register, types of style, etc. Clearly, care should be taken when using it, but it can be a nice complementary tool.  

It is illustrative looking at the mistakes Google Translate makes. This currently represents the highest level of artificial intelligence in the translation industry, and one can clearly see the problems currently unsolved. The quality also varies according to language; there is even more room for mistakes with those not part of the Indo-European language family.

Third-party Disclosure

Using machine translation has its limitations not only because of the capabilities of the system. If texts contain sensitive information, the client may require that data is not disclosed to any third parties.

For that reason alone, many customers ban the use of machines, explains freelance translator Barna Kanik. He says there have been cases where translators have been fined for using a machine to translate a text.

To what extent machine translation can be useful depends a great deal on the language pairs, or the texts that run through the system, Kanik says. In the case of a German-English language pair, it can give quite good results.  

For languages that are not a member of a large language group, such as Hungarian, it will give poorer results. At this point, machine translation is an addition, not a replacement for translation by people, he notes.

Internationally, therefore, it is common practice to use these tools, including Google Translate which has a much more sophisticated version behind a paywall.  

“It is expanding in Hungary as well, but in baby steps. Post-editing cannot be skipped though. It is a long way until that can be minimized,” Kanik says.

When it comes to the types of translation software, all major translation service companies have their preferred computer-assisted translation tools (CATs) such as SDL’s Trados. These are not to be confused with cloud-based machine translation, such as Google Translate, the best-known one, or Microsoft’s similar service. These cloud-based translation APIs can dynamically translate text between thousands of language pairs (but always via English). They let websites and programs integrate with the translation service programmatically. With certain plug-ins, CATs can be connected to these platforms. The programs with fully-fledged services are subscription based; Google Translate translation charges on a per-character basis, even if the character is multiple bytes.