Busting the Myths of NMT

By John Paul Barraza

Today, machine translation (MT) is evolving at the fastest rate since its inception 74 years ago thanks to the application of machine learning to NLP and the resulting AI models built for its current iteration, neural machine translation (NMT). MT has improved more in the past eight years than it had in the previous 50+ years. The move to neural network structures has been a game changer for the technology, and perhaps the irony is that machine learning itself was first posited in 1949, the same year the first proposals for MT were published.

Previous MT technologies relied on linguistic rules (rule-based machine translation, or RBMT) and/or statistical modeling (statistical machine translation, or SMT). Both demonstrated limitations in terms of tone, fluency, context, and accuracy. As these drawbacks were well known for decades, it is not surprising that the language services industry took a cautious, risk-averse approach to the adoption of NMT.

Myth: NMT yields poor or unreliable quality

Suddenly positioning your translation business to heavily rely on a technology that has been relatively unreliable in creating “translation perfection” for so long is a difficult step to take for any LSP. But the perception that MT simply does not produce the quality necessary for client work is strongly felt, and quite common, particularly among those with no experience using newer versions of NMT.

Busted (maybe)

The evolution of MT over the past eight years has been particularly pronounced when it comes to quality. With the development of NMT, output quality has improved at a greater rate than ever. Running NMT without sufficient language model training around industry, style and previous translations is never a good idea, and one that can doom its effectiveness from the outset.

Myth: NMT lacks specialization

It’s easy to perceive MT as inflexible and difficult to use to develop data sets for translation models that could provide the necessary quality suited for highly specialized niche domains and markets.


Model training is an absolute game changer when it comes to quality results, particularly in uncommon niches or those that feature heavy reliance on specific industry terminology. The more historical and industry-specific data an LSP feeds into a customer’s specific language model training process, the more effective NMT becomes.

According to a recent survey by CSA Research, 51% of language service providers (LSP) use some form of MT, while 80% have tried some form of NMT. The same study notes that 93% of the most mature LSPs are MT capable, meaning the companies do use it in many of their projects.

But for newer, smaller, or more domain-specialized LSPs, NMT should still have a place, and for larger “mature” LSPs, it can provide a far larger contribution to operations than it currently does. It is critical to understand LSPs’ perceptions on MT so that we can gain a sense of why the latter’s usage remains somewhat sparse after nearly a decade from when NMT was introduced.

After decades of slow improvement, the near exponential advances of the past eight years have been quite easy for LSPs to miss. But even those agencies that have seen the potential of NMT do face strong objections from a variety of sources, some of them more unexpected than others. The reasons for this hesitancy when it comes to NMT adoption are based on a few key misperceptions.

Myth: NMT is incompatible with translator workflows

The idea that NMT is incompatible with translation management systems and general workflow is a concern that may have been based on older iterations of the software.


NMT does require a level of adaptability for translators. In effect, strong NMT demands that translators think differently — instead of approaching translations on a word-by-word basis, translators must become editors, checking and reshaping NMT output to create the same “perfect translation quality.” It’s a big change for some agencies, but one that bears out its worth in productivity increase.

Myth: NMT is too expensive.

The perception that NMT is too expensive for regular use by all but the largest of LSPs also may have been born with previous generation systems. Also based on older MT generations is the idea that MT usage is expensive on a word-by-word basis.


Large LSPs developing their own flavors of NMT understand that the investment required for effective results with self-developed systems is quite large. But for smaller LSPs the range of systems available from NMT developers today are distinctively affordable and have no specialized hardware requirements. Savvy NMT providers understand that adoption must be both easy and affordable.

Oh, the humanity!

Language services are all about providing human quality service, so it’s understandable that many might hesitate to leverage MT. However, the world’s largest LSPs have already spoken, and have adopted NMT for very clear reasons. First, because they’ve learned how to harness the efficiency and power of placing the human (expert) in the driver seat behind a trainable NMT engine, and second, they understand that with NMT, they can have ultimate control over reviewing and training the engine to ultimately produce translation results that are of better human quality.

Generic results come from generic MT

For NMT to achieve its highest level of accuracy, there are a number of factors which must line up. First off, Translation memories should always be used. This is not a new concept, but one that does make a big difference for quality when compared to using a generic MT model alone.

Secondly, high-quality NMT systems must have the ability to learn from previous translations and industry specific language domains. Translation quality scores can easily double (or more)with effective language model training, and MT system developers that make it easy for LSPs to quickly train models with deep specificity can offer a product that wildly outperforms generic MT solutions and creates an intuitive and efficient post editing process.

Pricing models must evolve

The language services industry is extremely competitive. Pricing models have remained mostly constant over the past 75+ years, with a tight relationship between human productivity, quality, and complexity as well as the number of words that can be translated within a period of time. As more LSPs begin to adopt the support and assistance of NMT, productivity numbers can increase significantly in a short period of time. The rates these LSPs charge can remain the same, or be reduced to benefit the customer, which has a significant improvement on the sustainability and competitiveness of the LSP.

And yet, according to the same CSA survey referenced above, only 2% of LSPs use NMT systematically, meaning they run almost all their translation work through the technology. This is a shockingly low number, but one that fairly reflects the surprising strength of the industry-wide perceptions mentioned above.

Ready or not, here MT comes

Like it or not, NMT has already made a massive impact on larger, mature LSPs. Quick turnaround or emergency-priority translation contracts are already being won daily by firms that employ NMT technology. When it comes to high-speed translation processing, nothing beats NMT.

For these LSPs, using NMT effectively always breaks down to the efficiency of their postediting process, where the human translator comes in to clean up the translations to ensure quality. The more generic the language model NMT works with, the more time and translator-intensive the postediting process. But by using domain specific language models and client-specific historical translations to train a custom model before use (a process that takes a few hours to a day in most cases), postediting procedures can be dramatically efficient, enabling agencies to profit handily in these speed-priority situations.

Smaller LSPs, on the other hand, are beginning to understand that winning and then delivering work for large volume contracts may be in reach despite their small firm size, particularly with properly trained NMT. By “promoting” their translators to translator-editors, these LSPs can use NMT to complete the key foundational work on large projects at high speed, and then edit their way to successful results.

The future of LSPs and NMT

Like any business, LSPs require healthy growth to survive. Because of the people-first organizational structure of LSPs, however, finding that growth is often more top-line focused (increased sales) than it is bottom-line focused (exposing efficiencies and cutting costs). NMT is already proving itself a worthy efficiency generator in larger LSPs, and one that enables these agencies to increase their margins while staying on top of large, time-sensitive translation projects.

The key here is adaptability — LSPs that stick to rigid translation workflows will miss out on the benefits that come with NMT adoption and the inherent growth potential it creates. Meanwhile, those that empower their linguists to drive and train their NMT to new quality levels experience the best of both worlds — the ability to take on large, lucrative short-term projects while performing the work quickly and doing so with the least amount of billable translator hours.

John Paul (JP) Barraza is CIO of SYSTRAN Group globally, CEO of SYSTRAN Americas, and leads the company’s global technical operations.



Subscribe to stay updated between magazine issues.

MultiLingual Media LLC