‘All the right notes – just not necessarily in the right order’: how machine translation could damage your business

                                            ? Cienpies Design & Communication | Dreamstime.com

Towards the end of last week I found myself editing a machine translation of a website. The text was 500 words long and I re-wrote the whole thing from start to finish. Not a single sentence of the machine’s version was left intact.

I’ve been post-editing machine translation (known in the language industry as MT) for three years now, and it is fair to say the machines have made significant progress in that time. However, I still believe customers should think very carefully before entrusting promotional websites and marketing texts to a computer.

Why? Because with marketing copy, the key is not so much what you say as how you say it. Marketing texts have to do more than just convey information; they have to provoke a response and convince the reader to take action. Your text can be packed with relevant details, but if it doesn’t engage your audience and persuade them to buy your product, it’s not fit for purpose.

Establishing this connection with your reader is where MT falls down. The machines are getting better all the time and they can be very useful, especially if your text is highly repetitive or if you only need the gist of the original. However, they still tend to translate text word-for-word, aping the structure, word order and phrasing of the source language. This approach to translation frequently results in text that makes no sense whatsoever. Even if the translated text is grammatically correct, it often reads very awkwardly. In the immortal words of Morecambe and Wise’s famous 'Grieg Piano Concerto' sketch, machine translation often ‘plays all the right notes – but not necessarily in the right order'.

The problem is that while you can teach a computer to translate mechanically by substituting one word for another, it’s much harder to teach it good writing style: most translation engines stick like limpets to the letter of the source text because they don’t know a better way of phrasing it. Even more importantly, machines have no sense of context, so even if they can put grammatical sentences together, they don’t know what those sentences actually mean.

A little knowledge is a dangerous thing

To see just how dangerous this lack of appreciation for context can be, click on this article from MIT Technology Review and take a look at what a state-of-the-art AI algorithm ‘thinks’ of Black Lives Matter, the Holocaust and gender equality. The deeply racist, sexist and anti-Semitic ‘opinions’ expressed by the machine (which, for the record, I absolutely do not endorse in any way) are a result of the way it approaches language. It doesn’t put sentences together on the basis of what the words mean, because it can’t understand meaning. Instead, it tries to work out which words go well together based on how often they appear next to each other in its database. Unfortunately in this particular case, that database contains rather too much of what British comedian Dave Gorman used to call ‘the bottom half of the Internet.’

Now, I’ve never seen anything that offensive in a machine translation, because MT systems are guided by the source text rather than creating copy from scratch, but these examples are a stark illustration of the worst mistakes the machine might produce. Would you want to see these kinds of errors in your marketing copy? I thought not. So next time you need a marketing translation, at least consider choosing the highly-trained professional linguist over the machine. The computer will be cheaper and it may be just about good enough to get the job done, but using it for customer-facing documents could be a serious risk to your reputation.

Daniel Williams (AITI, MCIL) is a freelance translator, transcriber, editor and proofreader, based in the UK and specialising in sports, legal, business and marketing content. He works from French, German and Russian into his native language of English.