Utopia or dystopia?

A longer version of this post, with comments, is available on Kirti Vashee’s blog.

CharonTechnology is commonly defined as the practical application of knowledge[1]. And yet, a misconception still prevails that technology is limited to physical devices. Indeed, according to Encyclopedia Britannica, hard technology is concerned with physical devices and soft technology is concerned with human and social factors. Technology is also divided in basic and high.

Language is to all effects a technology, a soft, basic technology, although highly sophisticated.

For over half a century, we have been experiencing an exponential evolution of hard technology that we can hardly master.

This exponential technological evolution is the daughter of the Apollo program, universally acknowledged as the greatest achievement in human history, that stimulated practically every area of technology. Even the so-called translation industry is, in some ways, a spinoff of it.

Indeed, the birth of the translation profession as we know it today can be traced back to the years between the two world wars of the last century, with the development of world trade, while the birth of the translation industry can be set around the late 1980s with the spread of personal computing and office automation.

The products in these categories were aimed at new customers, SMEs and SOHO, rather than the usual customers, the big companies that had the resources and the staff to handle bulky and complex systems. These products could be sold to a larger public, even overseas, but for worldwide sales to be successful, they had to speak the languages of the target countries. Translation then received a massive boost, and the computer software industry was soon confronted with the problem of adapting its increasingly multifaceted products to local markets.

The translation industry as we know it today is then the abrupt evolution of a century-old single practice into shops. As a matter of fact, intermediaries (the translation agencies) existed even before tech companies helped translation become a global business, but their scope and ambition were strictly local. They were mostly multiservice centers and their marketing policy was essentially to renew an ad on the local yellow pages every year.

With the large software companies, translation memories (TMs) also burst onto the scene. The software industry saw a way to cut translation costs in the typical TM feature of finding repetitions.

So far, TMs have been the greatest and possibly the one disruptive innovation in translation. As SDL’s Paul Filkin recently recalled, they came from the research of Alan Melby and his team at Brigham Young University in the early ‘80s.

Unable to sustain the overhead that the large volumes from big-budget clients was producing, translation vendors devised a universal way to damp profit loss by asking for discounts to their vendors, regardless of the nature of jobs.

In the late 1990s, TMSs began to spread; they were the only other innovation, way less important and impacting than TMs.

The surge in demand for global content of the last three decades has made the need for translation grow way more than talents, while free online machine translation (MT) engines finally releasing “good-enough” outputs at the end of 2000s spurred the spread of MT to the point that today machines meets 99 percent of the global translation demand.

We are now on the verge of full automation of translation services. Three main components of the typical workflow might, indeed, be already almost fully automated: production with MT, management and delivery with TMSs.

Although not immune to waves and hypes, the translation industry is historically very conservative, little reactive and a late adopter of technology and innovation at large.

A typical example is the infatuation with the agile methodology, and the consequent excitement affecting some of the most prominent industry fellows. As a matter of fact, agile is rather a brand-name, with the associated marketing hype, and as such, more than as a management fad, it has a lifespan. Localization can hardly be suitable for agile methodology, for its typical approach and process, and for agile to be viable, a century-old teaching and practicing attitude should be profoundly reformed. Also, although agile has become the most popular software project management paradigm, it is known for not having even really improved software quality, that is generally considered low. In contrast, the translation industry has always been claiming to be strongly focused on and committed to quality.

Increasing the stake is the primary motive for the adoption of a new working methodology and translate it to more, faster and cheaper, although not necessarily better. Indeed, higher speed, greater agility, and lower cost of processes are supposed to make reworks and retrofitting expedient.

Also, the explosion of content has been posing serious translation issues to (would-be) global companies. The relentless relocation of businesses on the Web made DevOps and continuous delivery the new paradigms, pushing the demand for translation automation even further.

Many in the translation community speak and act as if they were and will be living in imaginary and indefinitely remote place that possesses highly desirable or nearly perfect qualities for its inhabitants. Others see the future, whatever it is depicted, as an imaginary place where people lead dehumanized and often fearful lives.

In the present and real, a survey presented a few weeks ago in an article in the MIT Technology Review reports a 50% chance of AI outperforming humans in all tasks in 45 years and of automating all human jobs in 120 years. Specifically, researchers predict AI will outperform humans in translating languages by 2024.

After all, innovation and translation have always been strange bedfellows. Innovations come from answering new questions, while the translation community has been struggling with the same old issues for centuries. Not surprisingly, any innovation in the translation industry is and will most certainly be sustaining innovation, perpetuating the current dimensions of performance.

Nevertheless, despite the fear that robots will destroy jobs and leave people unemployed, the market for translation technologies is increasing. Possibly translation technologies will not endanger translation jobs any time soon more than the lack of skillful professionals, but the translation industry resembles a still life painting, with every part of it seemingly immutable. A typical part of this painting is quality assessment, still following the costly and inefficient inspection-based error-counting approach and the related red-pen syndrome.

In this condition of increasing automation and disintermediation, a tradeoff on quality seems the most easily predictable scenario. As for the software industry, increasing speed and agility, while controlling costs could make reworks and retrofits acceptable. MT will be more and more spreading and post-editing will be the ferry between the banks of global communication, allowing direct transit between ends at a capital cost much lower than bridges or tunnels.

The quality of MT outputs is quite impressive now, and post-editing is replacing translation memory leveraging as the primary production environment in industrial translation. In the immediate future, the urge for post-editing of MT might escalate and find translation industry players unprepared.

The key question regarding post-editing is how much longer it will be necessary or requested. Right now, most jobs are small, rush, basic, and unpredictable in frequency, and yet production models are still basically the same as fifty years ago, despite the growing popularity of TMSs and other automation tools. This means that the same effort is required for managing big and tiny tasks following the same old stiff paradigms.

The most outstanding forthcoming innovations in this area will be confidence scoring and data-driven resource allocation. They have been already implemented and will be further improved when enough data is going to be available. In fact, confidence scoring is almost useless if scores cannot be first compared with benchmarks and later with actual results. Benchmarks can only come from project history while results must be properly measured, and measures must be known to be read and then classified.

This is not yet in the skillset of most LSPs and is far to be taught in translation courses or in translator training programs.

However, this is where human judgement will remain valuable for a long time. Not quality assessment, which is still today not yet enough objective. Information asymmetry will remain a major issue, as there will always be a language pair totally outside the scope of any customer, who has no way of knowing the product would match the promises. Indeed, human assessment of translation quality, if based on the current universal approach, implies the use of a reference model, although implicit. In other words, everyone who is requested to evaluate a translation does it based on his/her own ideal.

With machines being soon better at almost everything humans do, translation companies will have to rethink their business. Following the exponential pace of evolution, MT will soon leave little room for translation business. This does not mean that human translations will not be necessary any longer. Simply that today’s 1 percent will shrink much further. Humans will most possibly be required where critical decisions must be made. This is precisely the kind of situation where information asymmetry plays a central role, in those cases where one party has no way of knowing if the product received from the other party would match the promises, for example when a translation should be handled as evidence in court.

With technology making it possible to match resources, find the most suitable MT engine for a particular content, predict quality, etc. human skills will have to change. Already today, no single set of criteria guarantees an excellent translation, and the quality of people alone has little to do with the services they render and the fees they charge. What most translators should be afraid of is that expectations on professional translators will be increasing.

This implies that vendor management will be an increasingly crucial function. Assessing vendors, of all kinds, will require skills and tools that have never been developed in translation courses. Today, vendor managers are mostly linguists who have developed their competences on their own, typically cannot dedicate all their time and efforts to vendor assessment and management, and are forced to do their best with spreadsheets without having the chance to attend HRM or negotiation courses. VMSs have been around for quite some time now, but they are still unknown to most LSPs and yet translation follows a typically outsourcing supply chain, down to freelancers.

If the industry has been growing more or less steadily even in time of general crisis, but the business still counts for a meagre 1 percent of the total, this means that the growth is going to remain linear even when translation buyers are deciding to waive the zero-translation option and have all or most contents translated.

Agile in translation is not the only mystification. Now it is the turn of “augmented translation” and “lights-out project management[2].” Borrowing terms (not concepts) from other fields is clearly meant to look cool and astonish the audience, but trying to look cool does not necessarily means being cool. Even trendy models are shaped with precise rules and roles: using them only as magic words may backfire.

Nonetheless, this habit still dominates industry events.

Localization World, for example, is supposed to be the world’s premiere conference when it comes to unveiling new translation technology and trends. Most of the over 400 participants gathered in Barcelona seemed to have spent their time in parties and social activities, while room topics strayed quite far away from the conference theme of continuous delivery and the associated technologies and trends, despite the demand for better automation and more advanced tools is growing steadily. Maybe it is true that social aspect in conferences are what conferences are for, but then why pick a theme and layout presentations and discussions?

Presentations revolve around the usual arguments, widely dealt with before and after the event, and are often slavish repetitions of commercial propositions. Questions and comments are usually not meant to be challenging or to generate debate, although stimulating and enriching it would be. Triviality rules because no one is willing to burn his/her stuff that is intended to be presented in other times to different audiences.

Change is coming fast and, once again, the translation industry might be found unprepared when the effects of the next innovation will mess it up. So, it is time for LSPs—and their customers—to rethink their translation business and awaken from the drowsiness in which they have always received innovations. Also, jobs are changing quickly and radically too, and the gap to bridge between education and business would be even wider than it is now, and it is large. It is making less and less sense to image for one’s own children a future in translation as a profession, and this is going to make harder and harder to find young talents who are willing and able to work with the abundance of technology, data and solutions available in the industry, however fantastic.

This said, it won’t be long before “skilled in machine learning” becomes the new “proficient in Excel”. And very few in the translation community are concretely doing something. Choosing an ML algorithm will soon be as simple as selecting a template in Microsoft Word, but so far, very few translation graduates and even professional translators seems that proficient. In Word, of course.

[1] As in the Merriam-Webster, the Oxford dictionary, the Cambridge
, the Collins dictionary and Wikipedia.

[2] Lights Out Management (LOM) is the ability for a system administrator to monitor and manage servers by remote control.