How one state is expanding language access for UI applications
champpixs via Getty Images
New Jersey has released training materials that can help agencies improve Spanish translations of government resources and services using artificial intelligence.
To expand language access for government services, the Garden State is now offering open-source, no-code training materials for large language models that state and local agencies nationwide can leverage to improve translations of government resources.
The New Jersey Department of Labor and Workforce Development, in partnership with U.S. Digital Response and Google.org, has developed translation resources that agencies can use with artificial intelligence models to sharpen the accuracy of English-to-Spanish translations for unemployment insurance services and processes.
Artificial intelligence is playing a larger role in government services, from improving infrastructure to detecting wildfires, but “this technology is only as good as the information that it's being trained on,” said Marcie Chin, language access product delivery manager at U.S. Digital Response.
That’s why USDR worked with New Jersey policy experts and bilingual staff for more than eight months to develop and validate Spanish translations of UI content, leading to the creation of training materials that can be fed into large language models used for translating.
One training resource is a plain language Spanish glossary of UI terms that defines phrases like “laid off,” “worker misclassification,” “partial benefits” and “certify for benefits.” They also developed an LLM prompt library to translate UI terms from English to Spanish that are legally compliant with relevant laws and regulations, use conversational language at an 8th-grade reading level and reduce government jargon. Their work also led to the creation of a Spanish version of the state’s UI application.
The training resources can be used to create custom translations with off-the-shelf models like ChatGPT, Claude or Gemini, reducing the need for costly vendor solutions, Chin said. The Illinois Department of Employment Security, for instance, is leveraging the spanish glossary to translate their UI website, according to USDR.
The materials also help NJDOL staff better interact with Spanish-speaking residents by, for instance, answering their questions more effectively, said Gillian Gutierrez, senior advisor and director at the New Jersey Office of Unemployment Insurance Modernization. The resources can help staff feel more confident that they are providing residents with accurate information and build trust among residents in the state’s UI system.
The glossary and LLM prompts have increased the speed of producing translations by up to three times, compared to translating content from scratch, Chin said. Spanish-speaking residents also reported that they found “very little difference” between translations produced by AI and human experts, she said.
Preliminary data shows that the state’s translation improvement efforts are starting to pay off when it comes to residents’ ability to complete UI applications. For all users, the application time has reduced from more than three hours to 28 minutes, and authorities noted a 20% increase in the likelihood that applicants followed through with the next step in the application process.
“Those are the types of places where I’m trying to find opportunities to measure where offering something natively in Spanish could lessen the likelihood that someone gets stuck [when applying for UI],” Gutierrez said.
NEXT STORY: Alabama senator refiles bill to make body cameras, dashboard footage public records