Translators and machine translation – the next chapter

This year marked the launch of our flagship post-editing certification program powered by Language Weaver.

 

It also marks 15 years since I started working in machine translation. Talking to translators about post-editing in the early days of MT was not an easy task and required a lot of perseverance and convincing. Translators were understandably wary of a technology that seemed to threaten their livelihoods – we were asking them to accept lower rates to adjust output that was a far cry from what we see nowadays with neural machine translation. In these early days, there was undeniable tension between Language Service Providers who were keen to introduce MT and the translator community who felt that MT was possibly a risk too far.

First and foremost, translators were worried that they would lose their jobs. There was a lot of debate as to whether the arrival of MT would signify the end of the translation profession. Would MT replace the human touch?

 

Another main concern – not only among translators – was the quality of the MT output. Early MT output was often very literal and not sensitive to context at all, which regularly resulted in unintended hilarious translations.

 

Related to the quality concerns was the very real fear that wrong MT output would lead to serious consequences, including actual injury or bodily harm. Early MT systems often translated the exact opposite of what was said in the source text. A good example would be the instruction to turn a nut on a car wheel clockwise instead of counterclockwise. If this mistake in the MT output would have gone unnoticed, the wheel of the car would have come off while driving.

 

But despite these concerns, people like myself who were working in MT already suspected what we know now: that MT is the future.

Despite early predictions, there are more translators now than ever before. Although the number varies across languages and domains, it is believed there are 640,000 translators working today as opposed to estimates of 300,000 10 to 15 years ago. A 2020 report by the US Bureau of Labor reported that employment of interpreters and translators is projected to grow 24 percent from 2020 to 2030, much faster than the average for all other occupations.

 

Why is that? In today’s world, an estimated 2.5 quintillion bytes of data are created every day. This sheer volume of content is astonishing and there are simply not enough human translators to handle it all. MT allows enterprises to translate larger volumes and enables the translation of content that would otherwise remain untranslated, helping them to reach much larger audiences. Some of this content will be post-edited by humans while other content requires a less traditional approach.

 

At RWS, we have defined different post-editing service levels to help our clients decide how much human touch is needed for a particular content. A client can require anything from a full post-edit followed by a separate review to a light post-edit for content where understanding and usability are more important than grammatical correctness and a polished style.

 

The quality concerns that kept people up at night in the early stages of MT almost seem a little bit quaint now. This is not to say that MT always produces perfect output, far from it.

 

However, over the past 15 years, we have seen MT evolve from a rules-based system to first statistical MT and now neural MT. Every new iteration of MT came with enhanced quality, to the point where neural MT output often sounds like a human translation and we advise translators that very fluent output can be deceptive and may not be a correct reflection of the source!

 

What about the fears that bad MT output could jeopardize safety?

 

MT developments have allowed the reach of MT to expand far beyond what was thought possible in the beginning. At the time, very few people believed that MT would move beyond a narrow set of content from the automotive and perhaps the travel industry field.

 

Nowadays many more content types, domains, industry verticals and languages are suitable for post-editing.

 

These developments have only accelerated during the global pandemic. A good example is the life sciences sector, where the development and successful roll-out of vaccines has hugely increased the need to translate content quickly and make it available across language barriers. Another example is Translators without Borders who recently published an introduction to machine translation and post-editing in response to increased demand.

 

As these examples show, MT is no longer regarded as harmful but as a pathway to safer outcomes.

MT and NMT are part of the much larger framework of artificial intelligence. MT itself is a subset of natural language processing, relying on high quality data. This data, which is being created or labelled by professional translators, will be at the heart of new AI capabilities.

 

The natural language processing disciplines that have produced NMT now enable functionalities that go much further than pure text-in-text-out automated translation. They elevate MT into the AI technology space with linguistic AI features that provide additional informative context and actionable intelligence such as quality estimations, greater adaptation capabilities and content insights. These new linguistic AI capabilities build on the need for a holistic approach to address today’s content challenges, helping us to deliver greater value as shown in the graph below: