The future belongs to AI. You hear it all the time. It might even be true.
Some people in the translation industry are predicting that AI will soon be bringing instantaneous, free and automatic translations to the mass consumer market. That prediction is probably true. Even now, pictures of text we take with our phones can be translated immediately, and there are earbuds on the market that interpret spoken language in an instant. The results are ‘good enough’ for casual use, and these services answer a real need in society.
The underlying technology of machine translation, that is, translation performed by computers, is hardly new of course. Google Translate was launched as long ago as April 2006, for example. We are all aware it offers some handy features, but we also know from experience that the translations it produces are not exactly the thing of dreams, and certainly not what you would like to serve to your customers. Recognisable? Yes. Precise and eloquent? Not really!
But even if machine translation doesn’t give you a professional result, many language services providers are claiming that it can attain the same level of quality as a human translator, but at much lower cost. This can be achieved, they claim, by combining the machine translation with human intervention in the form of a post-edit. The advent of ‘Neural Machine Translation’, so they say, has greatly improved what machines can do, and the machine translation plus human edit workflow is the way to go.
In my opinion, this statement is little more than a marketing ruse.
This ‘magical’ cost-saving, quality-enhancing workflow is called ‘Post Editing Machine Translation’, or PEMT. In this scenario, your source text is first translated by a machine, with no human intervention. The translated result is then sent to a human translator, who checks the machine translation against the source text and corrects this as required. The idea is that this workflow is cheaper because there is less human intervention.
In theory, all well and good. Salespeople working for large translation agencies will vehemently insist that it works. And judging by the number of requests I receive to perform such work, I am often left wondering whether I’m a fool for not embracing the technology.
On the other hand, many projects get sent to me where NOT using machine translation is an explicit condition. Some clients even ask me to put that promise in writing.
So who is right? If you want to have a business document translated, should you, on the one hand, go for a company that promises you speed and savings through the use of post-edited machine translation? Or, on the other hand, should you opt for a company that promises a more traditional human process? Which selling proposition should you follow?
Let’s find out.
Option A: let the humans handle it
In a traditional translation workflow, your text will be translated by a human translator, who will take into account any context you have added to it. You – as the customer – should provide the translator with information on your intended audience, the required tone of voice, and perhaps any preferences you have in terms of specific terminology. Even if you don’t, a good human translator will almost certainly consult your company website to understand how you present yourself to the outside world, and will try to match that style.
A human translator will also find it easier to deal with any typos that may have crept into your source text and will know how to make sense of clunky or disjointed sentences. If they don’t, they will know how to ask for clarification.
This translation will then be checked by a second expert pair of eyes, and any remaining issues or misunderstandings will be ironed out by the team of humans who handle your text.
If this is done right, you will receive a reliable translation that you can put to use straight away and with confidence.
The downside to this process is that it is slow and relatively expensive (as I argued before, if it is cheap and fast, you might not get what you think you are paying for).
Option B: trust the machines
Because ‘slow and expensive’ is difficult to sell, no matter how good the end result is, ‘Post Editing Machine Translation’ is often presented as the better solution.
This means that your text will initially be processed by a computer. Without getting too technical (for a more comprehensive and historical overview, there is this excellent article by Vasily Zubarev), there are basically two ways in which this can be achieved.
One way is to use statistical machine translation, which will compare your text against patterns in millions of documents to help decide between the best translation from different groups of words. These translated chunks of text are then pieced together.
The much newer method, neural machine translation, “uses broader context to help it figure out the most relevant translation, which it then rearranges and adjusts to be more like a human speaking with proper grammar” (to quote Google). This is also called ‘deep learning’.
Both methods can bring about results we all love to ridicule: expressions get butchered, nuance gets lost, and literal translations put things on menus we rather would not eat.
The more recent ‘neural’ machine translation tends to get more things right however, and makes for better reading. Taking into account what Google asserted above, this would seem to make sense. Neural machine translation tries to predict what a human would like to read, and shapes it in such a way as to make it more appealing.
In other words, the results from neural machine translation are still very often objectively wrong, but also equally often quite plausible.
And that is a problem when it comes to PEMT.
After your text has been processed by the system, the machine outcome will be reviewed by a human translator. Of course, to maintain the advertised benefits of speed and low cost, this phase of the process needs to be quick, and can’t be entrusted to translators who ask for higher rates. Indeed, any request I have ever received for PEMT work was offered at a much lower rate than I consider viable, and requiring a much higher turnaround than what I consider feasible, even when reviewing the work of a highly qualified and extremely competent colleague.
In other words, a text you have spent so much time on to get right in your own language first gets translated by an unemotional machine with a limited sense of context, and subsequently reviewed by translators who do not really have sufficient time to think about the text and to make sure it reflects what you want to express.
At the same time, while neural machine translation may produce its flawed results in an ever more plausible and confident way, spotting the mistakes becomes even trickier and a task for which a specialist translator would be required. Indeed, it will be much easier to pick out a glaring error than to notice the machine happily omitted a “not” or a “no” – which may result in some unintended consequences for a contract clause.
In other words, what is essentially a difficult task is expected to be undertaken quickly, and by people who get paid so little that they are simply unable to scrutinise every sentence to ensure it conveys the correct meaning whilst also taking into account the context of your business.
At this point, two additional points need to be made, because this is information translation buyers are simply not aware of.
The first is that hardly any translators I talk to believe that machine translation is a viable alternative, by any stretch of the imagination. The vast majority dislike the task of post-editing, and therefore refuse to work on PEMT projects, almost universally described as mind-numbing, financially unviable, and annoying. As such, you are not getting the best people to review your text.
Second, PEMT is only ever advocated by large companies and intermediaries who don’t actually translate themselves. Indeed, very few small agencies, where the owners have a background in translation, would ever embrace this system. If they do, I believe they do so only because they are subject to the same price squeezes experienced by independent translators. As it stands, the human workflow is still preferred by most agency owners who have actual experience and expertise as translators.
Bearing all this in mind, it suddenly no longer becomes a matter of trusting the machines, but of trusting those people who claim that using machine translation gives you the best possible deal.
Working in the back-office of the translation workflow, I believe machine translation only benefits those who advocate it. It offers them a great sales pitch (cheaper and faster), lower outsourcing cost (reduced rates to mostly junior translators), and the possibility of handling and selling larger volumes.
Machine translation combined with human post-edit can potentially save you some money in the overall translation process, but at a great risk. In almost all cases, the result will be a sub-par translation at a marginally lower price. Sub-par quality, because the process is not conducive to anything else, and with marginal savings, because the real profits are made by the people selling this process at the expense of freelance translators.
Machine translation holds great promise. When used effectively, it can bring information to people in ways that were before unimaginable, in places and circumstances where they need it most.
However, machine translation is not a tool that can yield beautiful translations of a source that has been carefully crafted. For that, you need humans throughout the entire process.
Not convinced? Let’s put it to the test! Send me a short excerpt of your own text (no more than 150 words) and I will gladly provide you with a sample translation. To make the comparison useful, you should send me a passage of actual text from your documentation. You should also be ready to discuss the results with me, and to send me feedback on how I fared. This offer is subject to availability.