I am not a great fan of Formula 1 racing. For me it is a lot of going around in circles interspersed with some moments of confused drama. But the one part of it I do enjoy is the moment when all of those powerful machines are lined up on the starting grid. The highly tuned turbo-charged engines are roaring ready to spring forward at tremendous speeds when unleashed. It is wonderful to behold such a superb collection of state-of-the-art technology, brilliantly designed and programmed to achieve the pinnacle of success.

Well, according to the http://www.slator.com 2019 NMT industry report the domain of turbo-driven engines is not only to be found in Formula 1 racing. Apparently, developers of neural machine translation solutions are also “turbo-charging” their engines as they attempt to capture huge swathes of business in an ever-expanding market.

Turbo-Diamond-Blades (1).png

The report, published, earlier this year, titled “Neural Machine Translation Report: Deploying NMT in Operations”, says that there are now a dozen global tech companies “aggressively” pursuing enhanced solutions in machine translation and natural language processing. The list of global companies pouring millions into high-tech research reads like a Who’s Who of the IT world; Microsoft, Amazon, Salesforce, eBay and Facebook are just some of those who have poured millions into developing the technology.

For some – such as Microsoft and Amazon – the rationale behind their initial investment was self-serving, driven by internal needs to handle billions of words of their own translation projects. However, these global companies now see the advantages of taking this same self-built technology and monetizing it as a product they can now sell to smaller companies and also make available to individual users around the globe.

In parallel to this corporate push is a rapidly expanding level of research into the technology across many colleges in the USA, China and Europe. A chart in the report illustrates clearly this phenomenal growth, revealing that there was a notable rise in papers on NMT published in 2018. That year saw the publication of 391 papers, almost double that of 2017 and a six-fold increase of the 2016 numbers. Asia is to the forefront in the development of the technology.

Indeed, some Chinese academics controversially argue that the whole NMT industry was started by them and is now being propelled by their research. Undoubtedly, the exponential growth of the technology has been remarkable, growing from only a handful of champions just five years ago to a plethora of them today. The report talks of the current development spurt as being the “Third Wave” of MT technology. It suggests that with the huge investment by so many global companies – and growing levels of academic research – that this is a technology with a promising future.

As with the growth of any technology, there is the beginning of diversification processes as companies seek unique ways to monetize their product in an effort to recoup some of their investment. The industry is being driven by the global goliaths whose researchers were initially tasked with creating a solution for internal needs. These solutions are now seen as an opportunity for these corporations to develop another line of revenue. To harness this growing potential many of the tools’ companies, the companies behind such products as Catalyst, MemQ, Déjà Vu, Across, Wordfast etc have created APIs so that their product has the ability to interface smoothly with these NMT solutions.

Development efforts have also gone in many directions depending on the perceived needs of different markets. Some developers have concentrated energies in creating more and more language pairs; Google, for example, supports 50 pairs bidirectionally. Microsoft supports 41 pairs but uniquely allows its users to upload data if the do not have bilingual data. Amazon takes the lead by supplying up to 127 language pairs but is quite restrictive in how its users employ their engines.

One of the things all of these technology suppliers have it common is a menu of charges, although each menu differs as to when the charges kick in, and as to how is much is free. The report briefly highlights how many LSPs are struggling with how they should charge their customers for NMT services. Most have opted to go with the traditional per word method. However, the report suggests that this is something that might evolve into another method of charging as the service matures.

One thing that does jump out of the report is the number and scale of the Asian companies who are serious players in this field. In fact, Chief Scientist at Chinese search giant Baidu, Andrew Ng, claims that “Neural machine translation was a technology first pioneered and developed and shipped in China” and that US companies only came in “well after Baidu.” A comment that has stung many of the developers in the USA. The report seems to suggest there is a bit of a technology race between East and West development giants.

And as in the case of the huge sharks of the ocean, there are “pilot fish” companies swimming around the huge NMT companies seeking to feed off their efforts. This sub-industry is made up of companies who have seen an opportunity for supplying data-hungry developers with the billions of words they need from different market-types. Other enterprising companies have developed training courses aimed at addressing the growing demand for qualified post-editors. There has also been an increase in the number of “boutique” suppliers. These are smaller companies who offer the NMT technology to businesses who have neither the capital nor time to invest in developing their own NMT solution.

The report quotes several CEOs who argue that the technology will only become truly efficient when a comprehensive, qualitative form of integrated testing technology is available. The report interviews CEOs across a range of user companies and gives interesting feedback on their experiences in using NMT. Many of the reports are upbeat and give optimism that the technology is going in the right direction. A few of the CEOs have commented that after some initial reluctance many translators are now seeing that NMT is something that can help expand their earning power and are beginning to feel comfortable about where they now fit in the new L10n workflow.

The report is an easy read. It is not heavy on jargon, and it gives an interesting insight into the industry. Although the authors themselves don’t declare that machine translation technology is the future, taken in the round, it clear that NMT has become established, is making huge strides and is expanding from translating text to also translating voice. Clearly, this is an exciting time to be involved in the world of neural machine translation.

Aidan Collins, Marketing Manager at KantanMT.