The thorn in the side of machine translation

October 10, 2018
The thorn in the side of machine translation

Machine translation has come on by leaps and bounds since the vast power of dual learning and deep neural networks were applied to it. Indeed, earlier this year, Microsoft revealed that it had managed to achieve parity of quality between machine and human translators on a test project of 2,000 sentences. However, despite this progress, machine translation has not yet been perfected in a way that has been possible to roll out commercially. The current offerings still result in suspect grammar and plentiful inaccuracies. And that’s before anyone has even begun to think about the thorny issue of localization.

Machine translation progress

According to Microsoft, it is the application of deliberation networks, joint training and agreement regularization that has contributed to machine translations finally sounding more natural. Deliberation networks are essentially a way of the machine checking over its own work, and then making improvements to it. Agreement regularization, meanwhile, reads sentences both forwards and backwards as a further means of creating more natural language patterns. 

It’s no longer a question of trying to make a machine that imitates the way humans translate, but of finding a way to imitate the results. Humans don’t read each sentence backwards, but if that’s how machines can improve their translation abilities, then doing so can help them to imitate the results that human translators deliver. 

The localization issue

Professional linguists and technology experts have been trying to perfect machine translation since the 1950s. Progress has been unexpectedly slow, to say the least, and that’s before the issue of localization is considered. Even when machines are finally able to translate with the same level of accuracy as humans (and we don’t expect that to be anytime soon, despite recent advances), they won’t be able to consider a text’s cultural fit in the same way that a human mind can. 

Localization takes into account cultural, political, religious and other sensitivities, as well as everything from pop culture references to assumed historical knowledge. It is a skilled service that requires extensive knowledge of a document’s intended readership and an instinctive feel for those words and phrases that won’t sit well with the planned audience. 

What does this mean for the translation industry?

In short, this means that the translation industry won’t be disappearing anytime soon. In particular, translators who are able to localize the copy they’re working with can look forward to a continuing stream of work. 

That said, the translation sector is already changing as a result of machine translation. Translation agencies and freelance translators alike are adapting and flexing their services in order to meet the demands of this brave new world. Those staying ahead of the curve are now offering, for example, services such as post-editing machine translation. 

The role of post-editing machine translation

At this stage in machine translation’s development, progress has been such that many companies are happy to give it a try – although most still baulk at the results. This has led to a surge in demand for post-editing machine translation. This is where translators take the translation that the machine has produced and turn it into natural language. It’s a hybrid approach that can potentially save companies money, assuming the quality of the machine translation is passable. 

Likewise, post-editing machine localization can help companies to keep their costs down a little, with a human localization expert reviewing the machine-generated text in order to ensure that it is appropriate to its intended audience. This hybrid style of service delivery looks set to become more and more sought after over the years ahead, as translators work with the current state of the machine translation industry to continue delivery the services that companies need. 

Final thoughts

Do you undertake post-editing machine translation and localization work? How does it compare with regular translation and localization work?

One more thing, while you’re here… Have you discovered the Tomedes Text Summarizer Tool yet? It’s a handy free resource for all those who work with content, whether online or offline. Why not check it out for yourself and see?