National Weather Service uses artificial intelligence to reach Spanish, Chinese speakers

The National Weather Service is using artificial intelligence to reach Long Islanders in Spanish and simplified Chinese. 

The weather service’s New York Forecast Office is one of 14 offices participating in the program, which uses an AI algorithm to translate the agency’s weather forecasts and warnings from English, then posts the translations to a website.    

In the metro area that includes Long Island, “we have 20 million people with 800 languages,” said Nelson Vaz, warning coordination meteorologist for the Weather Service’s New York office. “If we want to build a weather-ready nation, we need to provide outreach and all these services, but if we’re not reaching some of our communities, it’s going to be really hard.” 

In Suffolk County, roughly 98,000 residents speak Spanish at home but have limited English proficiency, according to the Census; roughly 5,500 people speak Mandarin or Cantonese, the main Chinese dialects, with limited English proficiency. In Nassau County, the numbers are roughly 168,000 and 20,000, respectively. 

New York City has large Chinese-speaking communities in Queens and Manhattan, one reason the New York forecast office was chosen as the only one in the country to offer the simplified Chinese translations. 

Previously, the Weather Service had relied on human translators to do some of this work, with New York forecasters sometimes enlisting multilingual colleagues in other offices, Vaz said. But service-wide, human translation was “labor intensive and not sustainable,” Monica Bozeman, automated language translation lead with NWS, said in a news release.

Using AI, Spanish-speaking forecasters reduced the time needed to translate one batch of material from the National Hurricane Center from an hour to less than 10 minutes, she said. 

San Francisco-based Lilt, a machine learning company specializing in large language models, provided the technology for the project. In earlier pilot projects, weather service scientists trained the language model in weather terminology. 

In an interview, Spence Green, Lilt CEO and co-founder, said the company’s technology was similar to that used in products like Google Translate and ChatGPT.

After a pilot program last summer, where meteorologists provided training feedback, “we found the system could achieve a high level of accuracy for meteorological content,” Green said. 

The language of much of the weather service forecasting product released to the public is highly “conventionalized, routinized,” he said, making it “a really good application for AI.” 

Green said the push for AI came after 2021 when Hurricane Ida’s remnants killed more than 45 people. “People lost lives because they didn’t get weather alerts,” he said. 

Green said his company’s tech can “operate autonomously once it’s been trained.”

Maureen O’Leary, a spokeswoman for NOAA, the NWS parent agency, said there would be a human “in the loop on some of the material.” Forecasters train the model a few hours a week and will give feedback on its translations in the coming year, she said.

Lilt is working under a five-year, $5.5 million contract through 2028, O'Leary said. Lilt’s tech also works on the websites of companies like Intel and Asics. The company’s other government clients include the Department of Defense and intelligence agencies, and after the Russian invasion of Ukraine last year, it helped meet a need for translation capacity caused by a shortage of Ukraine analysts, he said. 

Minerva Perez, executive director of Ola of Eastern Long Island, a nonprofit Latino-focused advocacy organization, said the initiative could provide needed language access for Spanish-speaking Long Islanders.

"We saw in the pandemic, with alerts and mandates put out there, that information wasn’t consistently put out to community members in Spanish as well,” Perez said. 

But, she said, “no matter what we have for AI, there should always be a human native speaker” to check its translations, she said. 

Steven Skiena, Distinguished Teaching Professor of Computer Science at Stony Brook University, said that natural language processing technologies had gotten “wildly better in recent years,” and that the routinized language of much of the weather service’s forecasting product made it a good application for the technology.

“If you told me they were translating novels and expecting the public to read it and think it was as good as a translator, that’d be one thing.” But “especially on a restricted domain like the weather forecast, I imagine these translations can be made very accurate.”