NLP helps Eli Lilly work at a global scale

With more than 30,000 employees spread across more than 60 affiliate locations and 14 manufacturing sites around the world, pharmaceutical company Eli Lilly operates at a truly global scale. Operating at that scale comes with issues, not the least of which is sharing accurate and timely information internally and externally.

“From internal training materials to formal, technical communications to regulatory agencies, Lilly is translating information often,” says Timothy F. Coleman, vice president and information officer for information and digital solutions at Eli Lilly and Co.

For years, Lilly relied on third-party human translation providers for the bulk of its translation needs. Although public web translation services are available, confidentiality requirements meant those services did not meet Lilly’s standards for information security. Even with a footprint of more than 400 translation vendors, the process was slow.

“Lilly engages with costly third-party human-translation providers across the organization to provide verified and reliable translations,” Coleman says. “Depending on the requirements, the planning, translation, and verification of these engagements can take weeks to complete.”

Coleman adds that many bilingual Lilly employees were also being tapped to provide translations in addition to their current scope of work.

To address these challenges, the pharmaceutical firm developed Lilly Translate, a home-grown IT solution that uses natural language processing (NLP) and deep learning to generate content translation via a validated API layer, Coleman says.

“This innovative application of natural language technology enables Lilly to achieve greater efficiency gains, significant cost reduction, higher quality content, and lead the way for future tech innovations using natural language technology to achieve enterprise value at scale,” he says.

Passion project pays off

Coleman says Lilly Translate started as a passion project by a curious software engineer who had an idea for addressing a pain point of the Lilly Regulatory Affairs system portfolio: Business partners continually experienced delays and friction in translation services.

“That married up well with an opportunity to explore and learn emerging technologies,” he says. “It became a great opportunity that a Lilly software engineer picked up and ran with, initially as a great learning opportunity.”

After sharing the idea and technical vision with peers and managers, the project immediately garnered support from leadership at Eli Lilly Global Regulatory Affairs International, who advocated for investment in the tool. Coleman’s team worked closely with Regulatory Affairs to identify requirements around document types, languages, and so on.

The Lilly Translate API and UI are delivered via a serverless tech stack built on Node.js, Python, .NET, and Docker. It can be accessed via mobile devices, web browsers, and programmatically through the secure API.

The service, which earned Eli Lilly a CIO 100 Award in IT Excellence, provides real-time translation of Word, Excel, PowerPoint, and text for users and systems, keeping document format in place. Deep learning language models trained with life sciences and Lilly content help improve translation accuracy, and Lilly is creating refined language models that recognize Lilly-specific terminology and industry-specific technical language, while maintaining the formatting requirements of regulated documentation.

“The product was developed via a DevSecOps agile framework,” Coleman says. “Initially, we did not have a dedicated Scrum master and product owner, but later we were able to adjust that. The increased focus helped us accelerate our delivery efforts.”

The project took about a year to get to MVP, with several iterations and pilots needed to achieve the level of translation quality needed to meet business expectations.

“The level of quality of the translation output was an initial challenge we faced where the team had to work through various services to figure out how to improve the overall general level of quality,” Coleman says. “Once improved, we had to work diligently to ramp up our [organizational change management] efforts to gain confidence in the tool.”

With the tool fully deployed, a process that required the creation of work orders and days or weeks to complete now takes just a few seconds or minutes. The automation has also led to a big cost savings.

“In surveys distributed across the company there is consistent feedback that Lilly Translate is saving time across multiple business processes as well as getting answers to questions faster,” Coleman says. “Lilly Translate touches every area of the company from HR to Corporate Audit Services, to Ethics and Compliance Hotlines, Finance, Sales and Marketing, Regulatory Affairs, and many others. The time savings is extensive. Translations are now taking seconds instead of weeks, providing key resources time to focus on other business-critical activities.” 

CIO 100, Natural Language Processing