Evolving in a world of Neural Machine Translation: "Automatic for the People"

Neural MT has pushed anyone working in the translation and localization industry to change their game plan. So, how do we adapt to the age of the enhanced multilingual machine?

Text by Dominique Puls


Image: © everything possible/123rf.com

It is not always easy to find the right words for significant discoveries in human history. How to describe new-found models and techniques? I recently spoke about my work and the topic of Machine Translation to a friend, who described the technology as "automatic for the people" – an album title by the band REM. This seminal piece of art, released on October 5th, 1992, sold 18 million copies and was a milestone in the band’s history – a completely new direction of their sound!

In the localization and translation industry, we have reached a significant milestone in past years and have broken down the language barrier. With the introduction of Neural Machine Translation, we have made translation "automatic for the people": a powerful productivity tool that simply cannot be ignored. Progressive LSPs are integrating the technology into their workflows to stay competitive, to speed up project turnaround times and to offer services that would have been impossible in earlier days.

To truly exploit the benefits of automatic translation, it is necessary to embrace new approaches in our business. We have to learn to adapt and embrace new working skills, and to upgrade our thinking.

"Try Not to Breathe"

But let’s start with the basics: Where are we at in MT and which solutions are actually available today?

We can differentiate between two main types of MT products: On the one hand, there are "off-the-shelf" solutions, such as Google Translate or DeepL. They can be used in stand-alone environments or be integrated into existing localization workflows. However, the available language combinations and subject areas are limited, and often they cannot be adapted to special needs or requirements.

On the other hand, there are Machine Translation companies that offer customizable solutions such as Systran, KantanMT, and Iconic Machine Translation, to name just a few. These solutions can be specifically trained and adapted to meet individual requirements such as field specifics, language combinations, customer-specific data (i.e. from post-editing processes) etc. Also they offer a different level of support to users.

"Everybody Hurts"

Of course, technical advances are important, but in many cases, we are led to believe that solutions can be integrated seamlessly. Moreover, what about the human factor? The "holy grail of Machine Translation" is something everybody wants to be a part of, but most of us have no clue as to how this could be effectively introduced into our processes and used by our teams. Essentially, what is needed is a cookbook for human behavior in order to be able to evolve, progress, and adapt.


So how do we go about this? Here I would like to share with you our experience, from the perspective of a "progressive LSP".

Over the past two to three years, Neural MT has become a game changer. But, as the drive to deploy this technology has accelerated, new people skills are required to operate it. Most of the time, this doesn’t just concern one single person, but a team or the whole company. Or, as we like to think of it in our own case, a band. We are a successful company with a history of over 40 years, and we are always on the lookout for a new sound.

When we set out on our search for this new sound, our goal was to implement a WORKING MT Solution WITH the existing team (band). Our requirements included compatibility with our current Translation Management Systems and integration into our existing workflows. We also wanted something that was "easy to use", customizable, and had high data security options.

Additional key questions were: Do we have enough data to train our customized systems? How will the MT solution be integrated? Do we want statistical MT or Neural MT? How can we get our team on board and make everyone contribute? How extensively will we need to train them? How can post-editing be supported? How can we get the best results? And, what about high fuzzies in the translations from the Translation Memory – where will we draw the line towards pre-translation from MT?

In the end, we found a solution, our "machine", which helped us become automatic. This is where our journey really started!


The introduction of Neural MT meant that existing roles and functions needed to evolve. For us, this meant that we had to find the right approaches within our company to guide our band towards these new challenges. As an organization, our core is built around the belief that our team can function in self-learning and mostly self-organized patterns. Adapting to completely new processes meant that we had to push ourselves even further into a more agile and emancipated way of working together. Our aim was for the teams to find the best, most fitting approaches themselves – a process that in essence is still underway. However, what we’ve learned so far has been breathtaking at times – but, most importantly: it worked!

Lead Singer – a/k/a Translator

Translators are our guardians of quality, language, and domain expertise. Having evolved from being "solely" translators to being experts in language, we have labeled this team "language specialists". They are central contributors to training, deployment, evaluation and development of our MT solutions, processes and post-editing approaches. How did the roles of translators change?

  • They needed to embrace post-editing and review as a concept.
  • Together, we redefined productivity vs. quality.
  • Language specialists have various tasks within their everyday jobs going beyond translating and review.
  • Embracing constant learning and acquiring new technological skills with regard to MT adaptation and optimization.

This team underwent the most radical changes of all. Many approaches within our organization made adaptation easier, and it felt organic. Nonetheless, it was in parts – and still is – quite a feat. Our internal team quickly adapted a mindset for the new tasks, such as post-editing. This was triggered by our expectation management. We developed our own training and workshops in post-editing, quality requirements, and the use of MT solutions.

Our team now has a number of dedicated experts working closely with our IT and language engineers. Anyone in the team interested in a topic can join the expert group and contribute to our overall approach to MT-related products and/or processes and products in general. This concept is what we call "topic ownership", and it is currently being rolled out as a company-wide strategy.

Lead Guitar – a/k/a Project Management

Project management is handled differently in various organizations. We see this role from a holistic perspective, giving the project management team the power to emerge in the whole process, from customer account management down to communicating with business partners. Other teams within the organization support the project management team, but in terms of the projects themselves, they have full responsibility.

Four key factors changed in their everyday work as a result of introducing MT solutions:

  • The team had to work with a completely new business model, which had an effect on time, cost, and quality.
  • New services were added to the portfolio.
  • Both customers and partners needed to be closely consulted on the new processes to ensure workflows were implemented correctly.
  • The team had to adapt to the concept of "translation of everything" (TOE).

Drummer – a/k/a Language Engineer

This team dealt with the most straightforward changes requiring a smaller degree of adaptability. Nonetheless, they found themselves the most challenged by expectations from management and all other internal team members:

  • They needed to manage new sophisticated technologies incorporating AI and Machine Translation solutions.
  • They had to support these systems throughout the organization and carry out rollouts.
  • They needed to embrace the concept of "good data is better than more data".
  • They faced the "multiple interface dilemma" – handling various interfaces at a time, which did not always harmonize well.
  • They began engaging in more interaction and collaboration with other divisions, and thus spent more time communicating.

The last point especially is related to one particular effect that we found most interesting: With the heightened need to communicate, the other internal teams learned a great deal about our technical infrastructure and how things are essentially configured. We currently see increasing engagement when planning new workflows or optimizing existing processes. Team members from different divisions work together, and thus optimize and positively influence their own work environments.

Keyboarder – a/k/a MT Expert

Along the way, new roles were created: the MT Experts. This new engineering role is critical to the successful deployment of any MT solution and to the creation of the new sound of our company (or band). We found it crucial to have someone within our organization who can operate as a direct contact for any questions related to Machine Translation topics.

  • The main tasks of this new band member include:
  • Quick integration of MT
  • Augmentation in productivity output
  • Setup of post-editing environments
  • Implementation and enabling of previously non-existent service categories (i.e. instant translations)
  • Thinking outside the box and mixing approaches

Altogether, all of our teams had a high degree of acceptance towards this strategic implementation of Machine Translation solutions within our company. As mentioned above, a number of developments are still underway but, due to our structure in general, we always adapt organically and with flexibility: A strategy that has proven to be worth the effort.

"It’s the End of the World as We Know It and I Feel Fine"

We felt no remorse whatsoever for our decision to try customizable MT. We have boosted productivity – in some cases, we even see that we only need to touch approximately two thirds of machine-translated segments (high PE requirements). For one of our larger MT installations, we currently process an average of over 600,000 words per month, and rising.

Now more than ever, clients approach us seeking guidance and consulting in the field of MT. In most corporate scenarios, the "off-the-shelf" solutions – as sophisticated as they might appear – do not always do the job. The human factor is not taken seriously enough and first attempts at using Machine Translation often fail and/or create enormous costs. Problems occur not only with regard to linguistic questions and post-editing, but also holistically. We have learned that how people interact with each other and how they form a working process are essential for a resilient and scalable workflow.

For our journey, a number of factors have contributed to our success. Our approach to implementing new solutions is based on the idea of purchasing the "best of breed" technology. Once this step has been taken, we adopt the following pattern multiple times: evaluate, integrate, measure, iterate. This approach gives us the opportunity to take the whole internal (and potentially even external) team on board. Through our concept of topic ownership, we push self-organization and new work practices in new areas. This gives us the opportunity to realize cross-team collaboration that not only drives the technology, but also processes and products, and ultimately the customer experience itself.

"Man on the Moon"

Where will things go from here? Do we really need to decide between the machine and the human? Or should we combine both into a hybrid superhuman?

Neural MT is good, even great in some instances. With post-editing, it achieves productivity boosts that demand consideration. Still, many look at either the one or the other as a nuisance – possibly because its full potential has not yet been harnessed.

It is time to rethink our approach to localization, using fully distributed workflows to accelerate the delivery of content to our clients. Is it possible to reinvent the post-editing paradigm, and work at a micro-task level? The best results may be found at the crossroads between what the machine achieves and what the human creates.

How automatic will our future be, and how will it affect our daily lives? Overall, I believe that we should not over-interpret and that time will tell. Machines are machines – they are neither evil, nor good – they are machines. MT is here to stay. We have found a way to make it truly automatic for our team and our customers – automatic for the people.