It’s a very exciting time in the language services industry.
The application of Artificial Intelligence and Neural Networks to complicated natural language processing challenges like speech recognition and machine translation (MT) is leading to remarkably rapid advancements.
In October, for instance, Microsoft announced that their speech recognition system now recognizes conversational speech as well as humans do.
And Google’s Neural Machine Translation is all over the news. Gideon Lewis-Kraus’s recent New York Times piece, “The Great Awakening,” is an excellent account of Google’s efforts to develop Artificial Intelligence programs and highlights their application to automatic language translation.
Even when compared to their own most recent efforts in Statistical Machine Translation, Google’s Neural Machine Translation has made truly impressive progress in providing more fluent and grammatical translations.
Although they’re getting a lot of press, we shouldn’t forget that Google is not alone in the development of Neural MT: Microsoft, Baidu, Systran, and others have also seen the biggest jump in MT output quality since the changeover about 10 years ago from rules-based MT to Statistical MT. Lionbridge’s own internal testing of various Neural MT systems supports this claim as well.
Short-term effect of Neural MT: broader and more confident application of MT
As MT quality continues to climb, we can expect post-editing barriers to fall. This effect will be particularly pronounced for traditionally difficult language directions—or with content matter that has thus far been considered too complex or nuanced for efficient MT post-editing.
Lionbridge has seen explosive growth in the volume of words that we post-edit for our customers. We are very well positioned to use Neural MT to drive further efficiencies in the provisioning of very high-quality translations.
We’ll also see an increase in use cases for which unsupervised MT is an appropriate translation solution.
The accuracy and fidelity of an MT system’s quality is a function of (a) the quality and quantity of its training material, (b) the sophistication of its machine learning algorithms, and (c) the similarity of future texts to the training material. MT cannot be better than its training material, and it will remain far from perfect when translating new content that is not adequately reflected in the training material.
So even with increased base-line quality and the promise of being able train Neural MT systems on smaller amounts of data, the need for professional review and correction of MT output will remain, as will the need for linguistically sensitive curation of training material, expert supervision of the training processes, and—hugely important—expertise in enterprise-scale translation processes.
Lionbridge’s strategic approach to MT
Throughout its 20-year history, Lionbridge has aggressively pursued efficiencies in the provisioning of translations.
Our work with MT dates back to the late 1990s, and we’ve had a comprehensive post-editing offering since 2004. Despite—or perhaps because of—our deep experience with this technology, we recognized that the ongoing development of a truly world-class proprietary MT system requires a scale of data collection, computing power, and development resources that give very large tech players, such as Google and Microsoft, a huge advantage.
So, instead of building or acquiring an MT system and trying to compete with them, Lionbridge developed a unique MT strategy and focused our development efforts on a flexible API-driven infrastructure. This allows us to harness the best of MT technology and guide it with our linguistic expertise and our experience grappling with translation and localization challenges.
That’s a long way of saying that Lionbridge is thrilled by these recent advances in Neural Machine Translation. They will—without a doubt—allow us to improve the services that we offer our customers.
But MT isn’t the whole story …
The localization industry is experiencing a transition that we have witnessed in numerous other industries.
Over the past 5 – 7 years, file sizes have dramatically shrunk, given the changing nature of the content being published. File sizes have shrunk from an average weighted word count in the 1000s to an average in the 10s. The resulting reality is that the transaction costs often exceed the translation costs.
Accordingly, Lionbridge is focusing on reducing these transaction costs via the industry’s most efficient content localization lifecycle management solution. This system manages the entire process of getting content from the customer’s repository, translating it into multiple languages (whether MT, HT, or combination), QA’d, and returned to the customer.
This process is automated, with an intelligent infrastructure that can chunk content into pieces and get it to the right multilingual worker at the right time.
As the industry evolves toward multilingual communication assurance, the need for an intelligent infrastructure becomes even more paramount. Content has to be understood by the system so it can be routed through the right workflows, assembled into the right pieces, and, when human involvement is necessary, routed to the right worker.
As mentioned above, the key to successful MT is an industry-scale investment that only the largest tech companies can make. I’ve always liked the military truism that “amateurs talk tactics; professionals study logistics.”
At Lionbridge, we pay particular attention to the logistics: the science of producing translations with minimum overhead and transaction costs.
That’s why Lionbridge has invested in an entirely cloud-based technology platform that allows us to administer the localization lifecycle at a segment level, receiving content from any customer repository, processing it per different customer needs, chunking it based on configurable business rules, and getting it in front of the right worker in real time.
Automating each of these processes and reducing the labor cost of each of these touchpoints—that’s the power of Lionbridge’s cloud platform. And as machine translation quality increases, the value of this platform is stronger than ever.