I, Robot? – The Future of Translation
Robot doctors and automated cleaning staff. Legal automats and number-crunching accountancy droids. The subject is hardly a new one, but in recent times, the spectre of automation as a means of replacing direct human output has become an increasingly conceivable and tangible prospect. Though it never really left the limelight, AI, as a major buzzword in our rapidly advancing technological age, and the fruits of progress in this very area, seemingly loom large over the translation profession in particular. So what does the near future hold for the hermetic translator of yore? Will we see more and more translators simply ousted by software programs that can perform their task harder, better, faster and stronger? Will translators en masse be made redundant by lexically-enhanced and newly insight-imbued machine tools? For the moment at least, things don’t seem to be so cut and dry.
The altogether anachronistic image of the scholastic wordsmith wielding his quill as he leans crooked over a highly-perched workstation, introspectively searching for the most fitting word within a given body of text, is a curiously outdated image that nevertheless springs to the minds of many when trying to imagine a translator’s day-to-day. What makes this all the more amusing is that, if anything, translation is a profession that has had to modernise at incredible pace in recent years, constantly at the behest of rapidly changing technology.
Indeed, as an industry, it has been moulded and shaped by recent technical progress almost more than any other. With the steady rise and increasing sophistication of automatic translation tools, and the more recent development of deep learning and neural network technology to aid and abet these advances, many – often at their peril – make the assumption that in some ways, human translation is already functionally dead… this is of course an erroneous assumption, and it generally results in translations of a quality approximating that which you’ll see below. Aside from being the archetypal false economy, this results in time and energy lost, and it’s a corner certainly not worth cutting if you don’t wish to lose credibility.
But there may be some truth to the notion of partial human replacement even in the here and now. Or at least machine assistance. Heaven knows the marketplace for a translator has been made all the more competitive by technology and globalisation. The emergence of the world wide web at the turn of the last century has meant that any one translator can be competing with counterparts who provide the same language pairs from all four corners of the globe. Something not limited to the translation profession of course, but a pertinent example of a wider phenomenon. Unfortunately, being reduced to a simple profile in an online database swamped with competitors all vying for the same prospective client’s attention has had a negative impact on pay and conditions for many of those active in the profession. And this kind of disruption looks only set to continue as machine translation becomes ever more advanced and gains greater contextual understanding. So, as with any other sector affected by technological change, translators have had to adapt. And adapt they continue to do.
As it stands, CAT (computer-assisted translation) tools have in fact been immensely helpful and empowering for most translators. Tools such as Trados, which has already been around for many years, allow translators to work on vast swathes of text while reducing duplication of effort through their storage and memory capabilities. But what of the latest garde in machine translation? What do advances made in this area mean for translators further down the line? Since its initial launch in 2006, 500 million people have become daily users of Google Translate, using the automatic tool to convert roughly 100 billion words into another language each and every day. Quick-fire translations of indecipherable web content or short text in an unknown language can thus be input rapidly into the tool’s text box, and a plausibly understandable translation then appears.
Of course, the limitation of the tool’s utility has always been that it doesn’t understand context, and it certainly can’t capture anything resembling tone or seek to convey similar cultural specificities or notions. All vital when it comes to localisation. Instead, the metric in place has focused on finding equivalents for singular words or phrases, before automatically translating them and putting them back into their existing sequence. But such technology is becoming more sophisticated. Where the innovation seemingly lies today is in the area of neural networks, which, not just in terms of translation, merit everyone’s attention, as increased automation becomes the trending topic du jour in the global focus on the impending technological revolution of the labour market.
Google, for its part, has for the last few years been working on its most recent innovation in the field, the Google Neural Machine Translation (GNMT) system. As opposed to using singular words or phrases as a unit by which to translate, as before, GNMT considers the entire input entered into the automatic translation portal as a unitary measurement. The metric is said to be human-rated and side-by-side compared, meaning that human translations are used as a yardstick. It therefore purports to result in more accurate translations overall. The system’s attention mechanism and its recurrent neural networks are said to rely on previously entered word sequences to derive a better and more precise sense of how to treat certain phrases. So what of the resultant improvements to the tool in real world terms? Well, while it has undoubtedly made the machine translation mechanism in place more accurate for short and simple phrases, it could hardly be celebrated as one giant leap for translation kind.
Indeed, in a recent article published by Chinese newspaper Global Times, Chen Boxing, who is Research Fellow at the Institute for Information Technology under the National Research Council of Canada, suggests that Google’s GNMT makes an incremental improvement to Google Translate, but it is by no means a breakthrough. Context is still not accounted for, so trying to carry out meaningful translation on any significant scale will still result in fairly problematic output, to say the very least. Particularly if you wish to capture nuance or convey anything of significant cultural value. Boxing mentions that in the development of AI there are three development stages: developing the intelligence to compute (basic arithmetic), developing the intelligence to perceive (getting AI to listen and comprehend) and developing the intelligence to cognize (where AI can learn, remember and associate). The biggest hurdle is evidently this final stage, and it seems we may still be quite a bit off. After all, when one considers image and speech technology, machines have been able to reach as high as 99% accuracy levels in terms of listening and understanding, but they still can’t associate. With this, machine translation, as it stands, still seems a long way off the possibility of transcreation. That is to say, culturally and dialectically adapting messaging to any real human standard.
And this is really where it all gets a bit tricky, as translation is not merely about the connotative sense of words, but also their denotative sense. That is to say, it’s not a simple question of capturing a word or phrase’s abstract meaning, but rather, it’s about capturing this meaning in context and applying the right words to the right situations in order to elicit the same sentiment in the reader as would have been the case in the original text. This is undoubtedly more complicated, and until machine tools can develop their ability to cognize to a significantly higher level, human translators are far from being replaced. In some areas, this prospect still seems particularly remote. Ben Screen, a PhD researcher in linguistics at Cardiff University, recently posited in a widely circulated article that rather than being “replaced”, so to speak, by machines in the immediate term, translators can expect the phenomenon of machine encroachment and assistance to intensify in the years to come. Screen likens this increased machine intervention to similar changes in other industries over the years, but as with everything technological nowadays, likely at a more advanced pace. He explicitly references how machines help lawyers, doctors and teachers on a daily basis, but how those active in these professions have not (yet) been entirely supplanted by omniscient robot overlords with laser eye beams.
According to his own research, as well as research that he explicitly cites, translators who use machine mechanisms to translate and who then subsequently correct this output are indeed faster and more productive, without the quality of the translation itself being adversely affected. While on paper this looks almost certain to result in a potentially less interesting job for translators, who may eventually find themselves condemned to simply tweaking what a machine has had to say in the first instance, we mustn’t lose sight of the diversity of content that exists in the world of translation. The difference between a washing machine manual (with its repetitive terminology and purely instructional language) and a best-selling novel is about as great as the difference in scale and ingenuity between the Stari Most bridge and the Millau Viaduct. It’s quite a stretch. According to the figures cited in a recent piece from The Economist’s Johnson column, literary translation is among the areas that remains under virtually “no immediate threat” whatsoever, with sales of translated fiction shooting up over 600% between the years 2001 and 2015 in the UK alone. Difficult to imagine how C3PO might convey the subtleties of Joyce in Norwegian or Cantonese, for example. So, if machine tools continue along the same lines of progress, we can expect their involvement in a translator’s day-to-day to increase further, and maybe much faster than we think… but as for humans being definitively usurped, it seems that the lexical Terminator still has some way to go before annihilating us completely. So hang on to your quill, trusty wordsmiths.