How AstraZeneca Brought Home Over 1 Billion Words in Scattered Translation Memories

AstraZeneca Translation Workflow

Christos Makropoulos, Global IT Training and Language Services Lead at AstraZeneca, closed a day of interesting presentations during SlatorCon November 2023 with his introduction of the company’s AZ Translate language service, which was launched in 2019.

Makropoulos’ presentation, titled “Translation for the Enterprise — a Story in Three Acts” began with his description of the company as led by science and innovation. He explained that a lot of what AstraZeneca does revolves around accelerating access to life-saving medicines for patients around the world, and that translation plays a central role in all the activities conducive to that objective.

Remarking on the dramatic improvements made in translation technology in the past 15 years, Makropoulos talked about how resources such as translation memories and AI, machine learning, and style guides and glossaries are essential in the translation process but are many times in the hands of translation agencies, not the enterprise.

The executive mentioned that the company embarked on a transformation involving technologies and automations to be deployed internally, which would then be integrated into collaborative workflows with vendors. 

In the process, they discovered a high degree of fragmentation in language operations, and Makropoulos recounted how as many as 11 main translation use cases were identified across multiple areas, including clinical studies, training for staff, patients, and providers, communications, labeling, patient safety, and IT, to name a few.

At the time, the company was using global and local language service providers (LSPs) through multiple third-party portals or email, or via public translation engines and APIs, with serious safety and privacy implications, added Makropoulos. All these factors made the case for the centralization of linguistic technology and operations that would become AZ Translate.

Act 1: Benefits of Centralization

Centralizing TMs across a limited number of translation management systems (TMS) was a key initial step to take advantage of TM leverage, explained Makropoulos.

The plan was to create a model that could be applied across the board to as many use cases and as many business areas as possible, using the whole of the TM corpus and feeding it to AI in a sort of “unified translation memory plus AI feedback loop,” he added.

In AZ Translate, translations were classified as low, medium, and high impact, guided by principles of data security, privacy, and confidentiality. Makropoulos explained that AI alone was considered adequate for low impact while TM leverage plus the option of human input was suitable for medium impact. In the case of high-impact translations (such as legal contracts, medical reports, and financial statements), TM leverage plus AI and full human post-editing was required.

Act 2: Challenges of Regulation

Makropoulos described some of the challenges that apply to regulated content translation, including highly technical language and phraseology, strict guidelines, regulations and regulatory domains, strict regulatory limits for timelines, and fully auditable workflows. 

The solutions AstraZeneca implemented included globally-shared terminology databases and glossaries, AI trained with internal and public domain data to ensure pre-translation is “right first time,” heavy TM leverage, and QA automation. As a result, “linguistic data is no longer fragmented across the same language pairs and content types around the world,” remarked Makropoulos.

The executive also told attendees about additional challenges that needed to be tackled during the implementation of the company’s centralized translation model, including convincing stakeholders and others of its benefits, describing the effort as “guiding the enterprise through change” and adding that “we needed to prove, in many different ways, that the financial benefits were much more impactful than the financial cost of introducing such technology in such an ecosystem to the organization.”

Act 3: Change

And prove they did, as Makropoulos illustrated with figures: 700 million words translated, 1.2 billion words in TMs, four global and over 20 local LSPs, and adoption by more than 40 business units. He added that gains were made in streamlined translation, costs, and efficiencies.

In closing, the AstraZeneca executive addressed the incorporation of generative AI in translation. Makropoulos said that generative AI is both in the present and future of translation in terms of its role in the workflow, including continuous learning and file regeneration, with AI trained on enterprise-owned data. The company has already experimented with the latter, with promising results.

For those who missed SlatorCon Remote November 2023 in real-time, recordings will be available via Slator’s Pro and Enterprise plans in due course.