It’s no surprise that Neural Machine Translation (NMT) is one of the hottest topics in localization. Based on neural network technology for deep and representational learning, the latest NMT offers 30% improved translation quality compared to the previous Statistical Machine Translation (SMT) approach. Today, NMT fully embraces artificial intelligence (AI) and machine learning, and is one of the top-rated approaches to help solve very complex language translation issues. For example, NMT addresses the complexity of distant language pairs, such as Chinese to English and Japanese to English.
It can be daunting to find the right MT solution, and one that’s right for your business, but 61% of respondents to a recent SDL survey agree that machine translation is essential to coping with increasing translation demands. I spoke with Mihai Vlad, VP of Machine Learning at SDL, about the latest breakthroughs in MT, and what makes SDL’s approach to secure NMT unique.
NMT is based on neural network models that produce more natural word order, compared to previous “phrase based” statistical methods. Can you further explain the differences between SMT and NMT?
SMT and NMT are based on very different algorithms and architectures. Let’s take the design of a self-driving car as an analogy. With the “statistical” approach, you would get a car that is able to drive really well on a specific track based on training data that is collected from numerous drivers for that specific road. With “neural”, you get a car that is able to drive well on any track based on training data that is collected from different tracks, and different drivers. A self-driving car uses a “neural” paradigm and is not tied to a specific track, as the algorithm itself is able to generalize better than previous algorithms.
SDL was the first to bring SMT to market in 2005. Building from this work and over 40+ MT patents, please explain why SDL offers a superior NMT engine compared to other solutions in the market.
The development of MT actually began in the 1970s using rules to get the desired MT output. You can encode a set of rules, but then you realize that there are so many exceptions to support numerous language pairs that models became incredibly complex. In 1993, machine learning was introduced for MT whereby the algorithms learned from data (bilingual text called “parallel data”) rather than being pre-programed.
In 2005, SDL was the first company to build a SMT product based on machine learning, and for the subsequent 15 years, this approach was constantly enhanced through the creation of sub-components (like language models or syntax models), and language models that contributed to SDL’s 40+ MT patents.
While NMT is a revolution for MT, the fundamentals are the same — it is still Machine Learning. Mastering machine learning and all the sub-components of SMT are transferrable to NMT. Additionally, the past 15 years of data and linguistic expertise help fine tune the NMT engine and improve it.
Artificial Intelligence is a latest buzzword and everyone seems to be claiming to offer AI. Interestingly enough, MT is actually an application of AI that SDL has been pioneering for years. How does MT expertise offer SDL an advantage for developing additional AI applications moving forward?
AI is our collective ambition to get computers to replicate what humans do. What makes us human? Seeing, hearing, moving, planning are examples of typical human behavior. But one of the most complex human tasks is communicating.
And maybe the hardest task is being able to translate – and to convey the same meaning from one language to another. Here’s some interesting data: approximately 40% of the global population can speak only one language, 43% can speak two languages, 13% can speak three languages, 3% can speak four languages, and only 1% can speak more than four languages (according to ilanguages research).
We can all move, run, eventually coordinate our movements to drive cars, but we cannot speak all the languages on the planet. Asking machines to do this is a very tall order. And that’s why AI researchers see solving the MT problem as “AI complete”.
What we are doing now at SDL is unbundling the MT technology stack and components to address more natural language processing challenges. For example, we are building technology for content creators to help them with summarizing text, extracting themes from content and even generating language. We want to bring the same level of technology available today for translators to content creators and marketing teams around the world. Exciting AI applications are underway from SDL.
Open source has contributed to a set of online MT that are really geared towards consumers. Why should organizations consider using enterprise-grade MT over open source?
NMT code is more compact compared to SMT. And there are a few open source projects that developers can use. You can actually create a system that is able to learn from data, and then be able to translate with a few hundred lines of code. That’s actually incredible. It’s almost like a Raspberry Pi for Machine Translation enthusiasts.
However, the needs of an enterprise solution span beyond a few hundred lines of code. Scalability, integration, and customization for higher quality are just a few examples that need to be built into the NMT code. For example, generic NMT struggles with document formatting, which is a key requirement for any organization that wants to maintain the integrity of corporate documents. Also, the generic open source NMT output sometimes repeats words over and over again (i.e. Neural Babble). Other areas such as using dictionaries must also be addressed in NMT. One of the key issues with generic NMT however is the cost of training and the cost of translation. It is order of magnitudes higher than SMT. And enterprises wanting to deploy such a system don’t want to pay the hardware penalty for using NMT.
In a nutshell, with an open source NMT system, you are able to get a very fluent, yet very expensive translation system that makes quite a lot of mistakes. This is where the SDL proprietary NMT code helps provide a real enterprise-grade solution.
Originally developed and deployed for use in government, law enforcement, and intelligence agencies for over 15 years, SDL ETS has now been upgraded to a fully enterprise grade commercial product. What are the strengths that the government has validated that the commercial market can now enjoy?
Building government solutions must be reliable and stable. Calling customer support is not an option when debugging the translation of classified documents. Secondly, it has to be scalable, and able to deal with the most demanding workloads. We’ve brought these elements to the Enterprise and added features to make it easy to use, manage and integrate. For example, our Microsoft Office integration makes it very simple for employees to translate documents through a single click button.
Why are enterprises turning to SDL?
With such amazing results, SDL NMT clients — across all sectors — are taking advantage of the boost in translation quality to identify more MT automation workflows to increase customer satisfaction and engagement, gain newer insights, lower costs and more. SDL NMT is available with SDL Enterprise Translation Server (ETS), an enterprise MT solution that is deployed on-premise, or in a private cloud. As organizations transition from free online MT to enterprise MT, SDL ETS fulfills enterprise requirements for security and privacy, reliability and performance, quality and customization, and interoperability with the rest of the enterprise application ecosystem.
Hear more from Mihai Vlad on “Demystifying AI and NMT”.