Neural Machine Translation for Game Localization

Frees Developers From Language Restrictions


Language is one of the central elements that allows developers to create grand video game narratives. Ironically it’s also the element that keeps non-native speakers from enjoying the experience. This wasn’t as problematic when games were smaller and less dependent on dialogue and prose. 

This isn’t true today. Video games are frequently massive endeavors involving hundreds of hours of on-screen and spoken text, paired with packaging, manuals, and marketing and advertising material. Localizing all of this text, and making sure that it connects emotionally with players, regardless of their language, culture, and local traditions is a huge undertaking.

And that’s just one translation. If most game developers had their druthers, they would release in dozens of languages simultaneously. It’s a mind-boggling task that has always had to balance speed, accuracy, and cost.

Human translators are accurate, but they’re slow and expensive. Traditional machine translation techniques speed up the process dramatically, but the output often missed the mark, and it wasn’t reliable enough to stand on its own.

Enter NMT or neural machine translation. This burgeoning new technology promises to free game developers once and for all from the costly restrictions of language translation. NMT can now match traditional machine translation techniques on speed while also providing accurate, contextually-sensitive translations on par with human output.

If your previous experience with machine translation has you doubting this claim, we understand completely. But NMT will make you a believer.

What is Neural Machine Translation?

Pure Neural MT

First a bit about what it isn’t. NMT isn’t a rule-based translation scheme, and it isn’t based on statistical models. These are both older machine translation technologies, and while they both served their purpose in the past, there were costly pitfalls to each that left translation wanting more.

Pure Neural MT

Rule-based systems work reasonably well for small bits of text, but begin to fall apart when tasked with larger passages. Language is simply too complex, loaded with ambiguity and idiomatic expressions. It’s difficult to design a dependable rule-based translation system when the language itself is frequently bending and breaking its own rules.

Pure Neural MT

Statistical translation techniques are equally volatile. They work well within a rigidly-defined scope and style but lose accuracy quickly when the text strays. The method also requires a voluminous amount of source material in both languages and a training process that can take months.

So How is NMT Different?

Neural Machine Translation uses advanced deep learning algorithms that take advantage of the single greatest achievement in artificial intelligence research to date — the artificial neural network.

This technology replicates the way the human brain learns. Previous machine translation methods attempted to teach the computer how to translate between languages. With an artificial neural network (ANN), the computer learns on its own.

At their core, ANNs are composed of cascading layers of code-based neurons. The strength of the connections between them are dynamically weighted. These weights are called the parameters of the network.

These parameters are not predefined or set in stone. The network is capable of altering and correcting them throughout the training process. This begins by feeding the network a large collection of properly-translated passages in both target languages. 

The network analyzes the text for patterns and then tries to accurately translate an unrelated passage based on what it thinks it knows. Human operators give the machine feedback, rating its success. This corrective feedback is sent “backward” to adjust weights and tune the network’s connections. In plain English, the program uses human feedback to refine its understanding. Over time the machine gets better, and translations are faster. They begin to resonate your tone due to the ability to provide a human interface. 

This technology enables our PNMT™ (Pure Neural™ Machine Translation) engine to generate the rules of a language from a given translated text and produce a translation that bests the current state of the art and does better than a non-native speaker.

Why Is NMT Becoming Popular in the Video Game Industry?

Video games depend on the subtleties of dialogue and the written word to help build a believable world. The degree to which players can immerse themselves in the game depends on a clear, rich narrative flow.

This is easy to achieve in a game’s native language. But how do you maintain the proper tone, references, and narrative style when you have to shift all of the text to a different language? How do you guarantee that important story and character details will land with the proper emotional gravity when passed through the filter of a different culture? Games aren’t fun if the audience can’t relate to them — and certainly not if they can’t understand them.

Video games often take large teams of people months, and often years to create. To recoup the cost of production most design firms work to launch globally from day one. Given the stakes, they can’t risk losing the game’s appeal in the global market due to poor-quality translations.

In order to capitalize on marketing buzz, game producers need to produce localized versions of their games quickly. They often don’t have the resources that human translation requires. 

NMT is becoming popular with the video game industry because it provides the speed and the accuracy that is desperately needed. It allows for rapid translations that respect the subtleties of the language. It helps to guarantee that players around the world have an equally immersive, entertaining experience, unhindered by the limitations of language.

How Does NMT Adapt to Unique Game Terminology?

GIF NMT Job Description Looping Animation


Video games are often set in fantastical worlds, replete with novel creatures, places, weapons, races, and characters that don’t exist anywhere else. How does NMT deal with unique terms like these?

It learns. Our neural networks might start off as generalists, but we can train them to become experts in any subject we choose. That could be electrical engineering, gardening, or the various alien races scattered across a video game’s universe.

User Dictionaries

User dictionaries are extremely helpful in getting a neural network up to snuff on long lists of game-specific jargon. NMT systems will translate new text, based on what it has learned from previous text entries. If it encounters a large number of new words it will at first be confused, having no frame of reference to draw from.


Inline tags can be used to denote text elements that shouldn’t be translated, give context for ambiguous passages, and differentiate spoken text from onscreen text. 

These tags also help the computer understand the subtleties of performance, as we’ll see later.

What Hesitations Do Video Game Developers Have About Using NMT?

What Hesitations Do Video Game Developers Have About Using NMT?

Protecting IP

One of the most common concerns that developers raise involves data privacy. Proprietary text assets are often used to train neural networks. People are understandably worried about where this data goes. Will it be used to help train the network for competitors? Might it accidentally leak to the public?

We offer two levels of protection. First, if customers choose to use our cloud-based hosting options we guarantee that no one will ever see their IP, including us. We don’t peer into the source files, and we don’t use them for any other purposes besides training the customer’s specific neural net. No other customers will have access, nor will they benefit from them.

For those that are still leery about the cloud we also offer local, on-premise hosting. Customers can run the translation service on a local server place behind their firewall. We take IP privacy very seriously, and we labor to make certain that customers are comfortable at all points.

Those still worried that using machine-based translation services restrict your ability to control your IP, consider how little control you have when you release it to teams of human translators. Any one of those people could accidentally or intentionally cause a data breach, with terrible consequences.


One of the most common concerns that developers raise involves data privacy. Proprietary text assets are often used to train neural networks. People are understandably worried about where this data goes. Will it be used to help train the network for competitors? Might it accidentally leak to the public?

Except that they don’t.

Human translations can vary in cost depending on the volume and complexity of the text, and other special considerations. It’s also a slow process that frequently extends beyond allotted schedules. 

NMT providers like SYSTRAN offer a known cost that doesn’t fluctuate based on translation requirements.



NMT is not going to be perfect for every video game right out of the box. NMT, like a human, will need time to train and learn the nuances involved. With that said, NMT doesn’t take a lunch break or a vacation. It doesn’t have off days or make repetitive mistakes. If you were to find a human resource that only had to be told to correct a mistake once and was systematically incapable of making that same mistake again, that human would become a top performer very quickly. 

With NMT, you get that and more - consistent, efficient, and scalable performance that only get better with time. 

Want to learn more about how you can train your translation engine for gaming?

Check out our Roadmap to Training a Translation Engine eBook.

Download the eBook

Roapmap Gaming Ebook
How NMT, Rule-Based Translation, and Human Translation Fit Together

NMT, Rule-Based Translation, and Human Translation Fit Together

Up until a few years ago, rule-based translation engines were significantly faster than NMT systems. 

Neural networks are now at parity with rule-based systems in terms of speed, and handily outperform them when it comes to translation. Our NMT system can easily translate tens of thousands of characters per second. This bests other machine translation technologies and leaves human translators far in the distance.

Interestingly, NMT and human translators have the same capabilities and limitations. Both require training. Both learn on their own. The better trained they are, the more capable they are. A human translator that isn’t familiar with the subject matter at hand won’t do as good a job as a subject matter expert. The same holds for neural networks.

But NMT has a few advantages. Neural networks will provide consistent results every time. If you give the same text to five different people, you’ll get five different translations. NMT systems are also easier to train. In a matter of weeks, we can train a neural network to translate even dense technical texts accurately. Try getting a human up to speed that quickly.

It might take a human days just to learn the technical jargon needed to translate accurately. We can feed our systems a long list of terms for immediate use.

No system is perfect, but NMT currently offers the best mix of qualities for the video game industry and beyond by controlling quality and offering unparalleled repeatability.

How Does NMT Affect the Current Use of Human Translation?

With everything said about NMT and its alternatives, it’s extremely important to understand that it isn’t NMT vs humans. NMT should not be feared as a replacement for a human translator. Rather, NMT enables efficiency for human translators to do what they do best, "sound human." Allow NMT to do the heavy lifting, while learning, providing human translators the opportunity to provide quicker, more accurate, consistent tone with quicker delivery over time.

check (1)-1
Empower your human translators

Humans can proofread much faster than they can translate. And because of the technology’s general accuracy, the rate of mistakes and poor translations is much lower than for previous machine translation methods. This means human translators are freed to do more creative work.

check (1)-1
Increase throughput

Instead of brute force translation, humans can focus on nuance and performance. They’ll have the opportunity to take the computer’s translation and improve it. This process leverages the strengths of NMT and the human brain to create an efficient partnership that gets more done in less time.

check (1)-1
Quicker turnaround

Humans can be used to support machine translation in other ways, too. They’re needed to tag text with performance notes, fill user dictionaries with critical terminology, and work with translation memories.

How Are Game Companies Leveraging NMT?

Gamers competing

They’re using the technology both for in-game and out-of-game content. The former is fairly obvious. Developers are using NMT to provide rapid localization of the dialogue and written text found inside the game experience. This includes both the written dialogue and the synthesized audio that supports it. 

Textual tags embedded in the text give the game engine context, helping it to understand emphasis, usage, and characterizations. Our translation recognizes these tags and uses them to improve the translation, while simultaneously leaving them untouched so they’re present to guide the localized in-game experience.

Out-of-game content, like the fan pages, forums, marketing collateral, and customer support material also requires localization. These are generally needed just as rapidly as in-game content and can be accomplished quickly without rigorous specializations.

In reality, game companies are leveraging NMT across the whole of their business. And rightly so. The time savings afforded by the technology allows them to bring their games to a global market more rapidly than ever before. And these savings don’t come at the cost of quality. Developers will enjoy accurately localized products with all of the subtle emotional impacts that make the game worth playing kept intact. This makes for grander launches, higher sales, and droves of happy gamers around the world.

There’s a few basic, yet powerful reasons why SYSTRAN’s NMT is quickly flooding the game development market. The injection of this impactful, scalable resource will quickly create efficiencies and expedite projects. It will enable you to communicate consistently on a global level and allow your games to be experienced just as they were designed, no matter where in the world they’re being played.

We can't wait to speak with you!

Schedule time to discuss what SYSTRAN neural machine translation can do for your game development efforts.

Connect with us