Scaling a language strategy
The Consumer Financial Protection Bureau (CFPB) serves consumers with limited English proficiency, who often face challenges engaging with the U.S. financial system. Although CFPB leadership identified multilingual content as a priority, translation and publishing workflows were ad hoc, highly manual, and prone to error.
The CFPB sought to define a scalable, modern process that maintains translation quality while dramatically increasing multilingual content.
About the project
Challenge
Research and propose a sustainable translation and publishing process for CFPB staff to offer thousands of pages of quality multilingual content on CFPB’s site to consumers with limited English proficiency
Project team
User researcher (me)
Product manager
Content strategist
My lead responsibilities
User research
I also contributed to…
Workshop facilitation
Articulating goals
To begin the project, we gathered background information. A CFPB project team had recently launched a suite of multilingual pages. We reviewed findings from their focus group research and usability testing, which affirmed that translations helped consumers with limited English proficiency understand important money topics. Participants reported that high-quality translations increased their trust in the CFPB and made them feel included.
We also connected with stakeholders to brainstorm the reasons for undertaking this project and who would benefit from the work. This helped us articulate project goals and ensure that the team and stakeholders were on the same page.
Workshop with stakeholders to articulate project goals
Journey mapping
To better understand the current translation and publishing experience, I interviewed the project team that had recently launched a suite of multilingual pages on CFPB’s site. The team’s content strategist described a complex, manual workflow with many handoffs and documents. She noted that reviewing pages prior to publication was particularly nerve-wrecking, as she didn’t speak the target language. I synthesized her workflow and experience into a journey map to share with my team and stakeholders.
Journey map I created to capture the current process and experience of publishing multilingual content on CFPB’s site
“There were a lot of moving parts and a lot of organization and version control was needed. Arabic key terms was missed in the process. I made the content docs, just missed sending them in the mega email.”
Researching approaches
To identify scalable approaches to multilingual content, we interviewed other government agencies, researched machine translation methods, and explored proxy translation.
Publishing
To explore options for publishing multilingual content in our content management system (CMS), we researched internationalization packages available for our CMS. We installed a localize package on a demo server to better understand the workflow it offered. We also brainstormed changes we could make to our site’s information architecture and URL paths to better support a large volume of multilingual pages.
We investigated non-CMS approaches. Some of the city governments we spoke to used a method called proxy translation, in which a third-party vendor crawled their site content and produced a translated mirror of it. The translated mirror was hosted on a separate server and presented to visitors in the browser. In this workflow, they used machine translation, which provided near-instant translation and publishing when they made a change to English content.
Translation
We learned that several city governments used machine translation, such as Google Translate API, without human review. We explored machine translation options and connected DeepL to a demo server to try it out.
We talked to a federal agency that employed an in-house linguistic staff of 13 employees to handle human translation. The agency developed a content prioritization system to inform vital and non-vital content to translate and developed glossaries to ensure consistent translations across content.
We talked to a proxy translation vendor who offered translation via human, machine, or a combination of the two. They were developing an “adaptive” translation model in which human translation would be supplemented by a machine-learning model trained on the existing human translations.
The result
Equipped with a wealth of research and considerations, we presented our findings and recommendations to stakeholders.
Our recommendation included two main parts:
Conduct a proxy translation pilot. Proxy translation provides the potential to offer a large amount of quality translated content to consumers with limited English proficiency and avoid the costs and workflow concerns involved in managing translated content in our CMS.
Conduct a content audit. Review site content to identify content types, owners, opportunities to make the content plain language, and content that may be a candidate for sunsetting. While auditing, create a prioritization model to help identify vital content that needs human translation and non-vital content that may be a good candidate for machine translation.
Final presentation and recommendation
