5 Things that Machine Translation Can’t Do

Machine translation (MT) has made tremendous progress in the last few years. And for certain kinds of texts, it may even produce nearly perfect results. But there are still some things that it still can’t do, while human translators can.

MT will not find mistakes in the source text

Nothing is 100% perfect. This certainly holds true for technical manuals, medical device IFUs, employee handbooks and other lengthy publications. MT software will not stop and ask, hey what does this mean? It will apply the translation exactly as it is written in the source text and repeat or even amplify the error. On the other hand, a professional translator will send a query to her/his project manager who will in turn refer the question to the owner of the content. This will solve two issues: (a) it will prompt the owner to correct the mistakes in the source and (b) will render an accurate translation.

MT does not translate (that well) in obscure languages

Machine translation thrives on training. Which means that software engineers load huge volumes of aligned, bilingual text into the system (referred to as corpora). These huge corpora are readily available for languages which are widely spoken worldwide like English, French, German, Spanish and Italian. So, the quality of these languages is translated very well by MT. But for languages that are spoken by fewer people like Icelandic, Maltese and Finnish? Not so much.

Facebook came out with a solution that they say does not require large training corpora for obscure languages. But these systems have not been widely tested and are furthermore, not available to the public as a standalone translation application like Google or Microsoft.

MT does not have a sense of humor

MT is better for translation of controlled text. Like technical instructions that have short sentences and bullet points. But once the text becomes more idiomatic and conversation-like in structure, MT does not do a good job. Only a human translator will be able to translate content such as blog posts, poems and jokes in a sensible manner.

MT does not account for regional differences in language

French is spoken in both France and Canada. But the differences between the two language variants are significant. The same is true for Portuguese, which is spoken in both Brazil and in Portugal. And what about Spanish? Not only are there major differences between Latin American and Castilian (European) Spanish; but there are dozens of other Spanish-speaking countries, each with their own regional nuances. The main online MT applications, like Google and Microsoft, don’t offer regional language options based on country. Only a professional translator will capture the precise language variant that you need.

MT won’t sign an NDA

Some customers, especially law firms and government agencies, rely on strict confidentiality and information privacy. Using a public online MT platform is not suitable for them, even though the possibility of data compromise is infinitesimal. If your vendor has a private MT server, then that is OK as long as they sign a non-disclosure agreement which commits them to maintain the confidentiality of your data. You also need to ascertain that they will not align your bilingual data and use it to train their translation memories and MT servers and use it for other clients.

In Summary

MT systems and technology are becoming harder to ignore and are excellent choices for some translation tasks. In all cases, NEVER rely on MT unless you are a professional linguist or at least have excellent command of both the source and target languages. If you want to use MT for mission-critical tasks, turn to a professional translation company for help.