4 years ago
August 14, 2017
How SDL’s Secure Neural Machine Translation Helps Enterprises Reduce Risk
“It seems we can’t go a day without a data breach taking place somewhere in the world. And we have definitely seen enterprises realizing that the reputational damage outweighs financial penalties or regulations,” said Thomas Labarthe, EVP of Corporate Business Development at language service and technology provider SDL, when asked about how views on cybersecurity within the enterprise have evolved over the past couple of years.
Labarthe observes that “enterprises are taking much more proactive actions to have better control around privacy and security and that now definitely includes translation. The conversation has dramatically changed over the past couple of years.”
In multinational corporations, employees’ use of public, cloud-based machine translation systems is being recognized as one area where sensitive company data and information could be compromised.
Meanwhile, progress in language technology is moving fast. New machine translation systems based on neural networks have resulted in dramatically improved output. But many of these systems run on the cloud.
The challenge for the enterprise requiring translation of large amounts of sensitive information, therefore, is how to leverage new technology without sacrificing data security.
“Enterprises are taking much more proactive actions to have better control around privacy and security and that now definitely includes translation” — Thomas Labarthe, EVP of Corporate Business Development, SDL
Over the past two years, SDL has begun to successfully roll-out its SDL Enterprise Translation Server (SDL ETS) solution among a growing list of enterprise users in life sciences and financial services, as well as among global law firms and e-discovery service providers. SDL ETS is available as an on-premise or private-cloud 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.
“It is important for us to have proven SDL ETS in government,” Labarthe said, adding “that is one of the most demanding environments.” Its roots also mean SDL ETS has been developed to require close to zero support after installation.
“We actually have customers where we cannot go and provide support so we have to be absolutely hands off because there is no way for us to go there,” Labarthe explained. “That has allowed us to develop a really a hands-off solution. It’s a mature and reliable MT solution with very little administrative overhead to maintain high services.”
Researchers Collaborate with Practitioners
SDL ETS’ origins in government and law enforcement ensures the product is secure and reliable. To deliver superior translation quality, SDL leverages its position as both a leading translation technology provider and a global language service provider, which employs over 1,000 internal linguists.
“We often involve our expert linguists quite early in the development and deployment process of a SDL ETS project,” Labarthe said. “That helps us fix any algorithm errors like high-frequency error patterns early in the development process so it immediately raises the standards of the output going beyond what other solutions can do.”
“We often involve our expert linguists quite early in the development and deployment process”
“To give you an example,” Labarthe explained, “we recently used our linguists to work with our engineering team to improve the English to German translation quality by addressing the problem of translating verbs.” SDL’s linguists worked hand-in-hand with the company’s machine translation researchers on developing and testing mathematical methods to improve word order in this language combination. Furthermore, SDL leverages this collaboration between researchers and practitioners to customize machine translation engines for specific clients.
Neural Machine Translation. On Premise
What about neural machine translation? How does it impact current workflows and how are SDL enterprise users benefiting from advances in neural machine translation (NMT)? Labarthe explains: “NMT is something we’ve added behind the scenes. It doesn’t change anything from the user experience end points.”
According to Labarthe, SDL can already do NMT on premise currently, with the same deployment, installation wizard, and user interface.
“The SDL ETS system can basically work with both statistical machine translation and NMT technology at the same time and it is completely transparent to the end user,” Labarthe concluded.
Easy to Deploy, Scale, and Use
The admin and easy-to-use features of SDL ETS are designed in a way that makes it extremely easy to deploy within the enterprise.
“Step one would be requirement discovery,” Labarthe said, discussing the steps they take towards an installation. Requirement discovery involves looking at the language pairs required, quality needed, as well as volume requirement and usage estimates.
“The SDL ETS is very easy to scale. Changes to the system in order to accommodate higher loads at a later time are not technically challenging,” Labarthe explained, “but upfront estimation of volume helps facilitate the initial license negotiation, the investment calculation, and also, specification of hardware.”
Another key component of requirement discovery is to analyze specific integrations that will be required in the enterprise systems, where SDL ETS will be deployed in.
“Step two goes into deployment specification and stakeholder alignments,” Labarthe continued, “We would look at the acquisition of hardware if required for the deployment of NMT.”
Running NMT at scale requires so called graphical processing units (GPU), i.e. high-performance chips with enough computing power to crunch the mountains of data required to run neural networks.
Asked if SDL supports this hardware setup Labarthe says: “Yes, absolutely. It is actually very simple. We have a tool to configure GPU hardware depending on the requirement from the customer.”
SDL will also consider aggregating locations that can share a private cloud. “It is quite common to have a central or regional data center or private cloud that is existing and that can be used to deploy infrastructure,” Labarthe says.
“We have a tool to configure GPU hardware depending on the requirement from the customer”
Step three is straightforward installation considerations. Finally, step four is deployment to users within the enterprise either via SDL’s intuitive browser-based UI or via API for programmatic interactions.
Deployed in Finance and Retail
A large financial organization partnered with SDL to deploy SDL ETS. They were concerned with the IP risks associated with employees using free online translation tools and wanted to optimize the quality of output from the machine translations.
Particularly appealing to the financial organization was that SDL ETS can be deployed on-premise and keeps confidential data safe. They also highlighted SDL ETS’ out-of-the-box large selection of generic language pairs, and the direct access to the SDL development team for feature enhancements.
A large retailer voiced a slightly different set of concerns around machine translation before rolling out SDL ETS. Like the financial organization, the retailer was concerned about employees using Internet MT to translate sensitive information.
In addition, the retailer faced escalating costs, legal risks, and slower time to market due to silos of translation processes across the company requiring reconciliation of inconsistent translation results.
They chose SDL ETS because it is an on-premise solution guaranteeing data security, and because expert MT engine customization results in higher quality output.
Finally, the retailer pointed out that SDL’s integrated products and services provide a complete and streamlined solution.
For both the financial organization and the retailer, SDL ETS provides a translation hub to enable global communication and collaboration among globally distributed departments.
For more information on SDL ETS 8.0 click here.