Tool-aided translation has evolved considerably over the past few decades. The idea of machine-assisted translation has always been a kind of holy grail for game localization, with its promise of faster resolutions and the ability to take on greater quantities of work.
However, perfecting the process has been a bumpy road, as with any initiative involving human language. Idioms, localized slang, and the peculiarities of the way people communicate present challenges for non-humans when it comes to proper understanding. The goal is clear, but it’s been a trial-and-error procedure over many years to reach the current level of proficiency... and it still requires a lot of human involvement. Here’s a look at the history of machine translation (MT), our path to adoption, and the benefits and challenges it presents for the future.
Historically, human translation, editing, and proofreading was a time-consuming, manual process. Teams had a ratio of 2:1 between translators and proofreaders.
The earliest instances of MT were mainly government-funded. Over the last 70 years, huge investments have been made in evolving MT systems, with recent developments incorporating deep learning and artificial neural networks improving accuracy and reliability.
Neural machine translation engines use artificial neural networks that mimic the brain’s linguistic rules by translating whole sentences at a time, rather than in parts. This results in higher-quality translations which are more accurate, human-sounding, and easier to read.
The benefits of MT use are clear: more content can be translated faster. More time is gained for complex and creative translations (like in video game content). There is also a significant productivity increase for professional translators who can post-edit higher volumes and faster, compared to traditional human translation methods. As well as saving time, the cost saving can be significant with post-editing and sample review workflows that substitute the traditional translation and revision workflow.
Our continuing process as a company has always been to stay on top of industry trends and technology developments. We had many discussions and conducted analysis around the rapidly changing MT phenomenon. With the proliferation of neural machine translation engines and the benefits that could be offered to our partners, we decided to incorporate it into our own processes.
In the beginning, we tested MT as a tool to boost productivity of our internal teams, translating internal communications, corporate documents, and bug reports from Japanese into English and vice versa.
We initially invested our time and effort in one specific MT provider, but after several tests, we noticed that some MT engines performed better on specific language pairs and topics.
We trialed several engines, testing each one across six languages with internal comms between our development and Player Support teams.
We gathered, analyzed, and interpreted the data based on accuracy, fluency, number of errors, percentage of segments that didn’t require post-editing, and most common error categories. The results provided us with a list of the top four engines for the different language combinations we were interested in. A lot of work and time was also spent experimenting with plug and play engines that are more accessible and can be customized.
While it might have been easier and less complex to go all-in on just one solution, we didn’t want to compromise on the quality of our translation offering in any one language. So, we decided to create bespoke workflows using the best engines for the type of text and the specific language combinations required.
The move to being engine-agnostic provides us with freedom and flexibility to choose the right engine for the right language and text style. This means we’re able to support different languages and different types of game projects in a way that we wouldn’t have been able to before.
Our technology teams have been heavily involved, helping with data curation, engine training, and deployment, but also connecting the engines to existing tools and even creating new tools where necessary. We have also made substantial investments of time and money in quality assessment and evaluation, both for European and Asian languages. This involves collating and curating data for staff training, training engines, and human intervention for post-editing and evaluation.
Our process involves defining and implementing suitable MT solutions that fit our partners’ localization workflow, from engine customization and quality testing to matching each language to the right engine for a particular game or franchise.
The selected engines are then trained and seamlessly connected to the relevant CAT tool or TMS. The workflow always ends with human –post-editing by our game specialist linguists to ensure quality is maintained.
The developments in MT technology and the significant quality improvements brought by Neural MT were some of the key factors that motivated us to invest in MT tools. That investment has paid off with increased production and decreased execution times. However, there are still many areas for improvement and many opportunities for further use.
We are in the process of launching an in-house tool that enables the quick and effective translation of player support tickets, leveraging MT solutions. This will allow for our Player Support agents to answer tickets that are not in their native language and optimize our support teams for spikes in tickets of a certain language.
As the market for video game translation widens, we’ll also be using MT for less common or low resource languages like Finnish and Turkish, among others traditionally not considered viable for game localization.
Finally, we are exploring high-quality output by integrating MT for translation between Asian languages (i.e., Japanese into Simplified and Traditional Chinese; Simplified Chinese into Korean, etc.). Testing will begin with internal documents, but the focus will move to game content as our tools improve.
We may never reach a place with machine translation where the process simply involves inputting a script in one language into the algorithm and having a perfectly-translated text appear on the other end. Human language constantly evolves on its own at a rapid pace, to the extent that such an outcome might be impossible. Yet the benefits are evident, even at the stage in which MT exists now, and the research is additive, with each improvement building on the previous one and pointing the way to the next.