Illustration: Gabriella Turrisi/Axios
A startup that provides AI-powered translation is working with the National Weather Service to improve language translations of extreme weather alerts across the U.S.
The big picture: Gaps in language access to emergency alerts during extreme weather events have led to missed evacuations, injuries and loss of life for non-English speakers. Machine learning could mitigate that.
Driving the news: AI-translation service Lilt has recently started working on a pilot project with the weather service to help produce more comprehensive weather warnings.
How it works: Incorporating a mix of software and human translators, the service learns from linguists in real time using a neural network, or a computer system modeled loosely on the brain — which gets smarter with each use.
- The team behind Lilt, which earlier this year raised a $55 million Series C, markets its speed of translation at the rate of at least three times the speed of other translation services, while also picking up slang and regional dialects.
- The software gets used by human forecasters at the NWS forecasting office, with the AI engine suggesting translation for the translators to work with, while actively storing all of their input.
What they're saying: Phil Stiefel, solutions lead at Lilt, told Axios there is a longstanding need for better translation at the weather service.
- "If there's a translation error in a translated weather report, and somebody takes the wrong action based on that missed translation, then somebody could get hurt or even killed because of that," Stiefel told Axios.
- According to data by the Migration Policy Institute, in 2019, 22% of U.S. residents over the age of five spoke a language other than English at home.
The backstory: Past translation problems during floods, wildfires and tornadoes point to a legacy of language barriers within federal, state and local emergency weather alerting systems.
- In 2017, nearly all of the people in New York City to die from flash flooding from the remnants of Hurricane Ida were Asian and spoke limited English or Spanish, which the New York Times reported may have led to those residents not receiving warnings to evacuate.
- Seven members of a Guatemalan family in Oklahoma died from flash floods in 2013 after leaving their home to seek shelter in a storm drain upon hearing a tornado warning. NBC News reported "they hadn't heard or understood there had also been storm and flood warnings."
- And a 2020 study found that emergency warnings during California's 2017-18 Thomas Fire were initially only available in English. As a result, Latino residents living in the two most heavily impacted counties missed information about evacuation areas, road closures, unsafe air and boil water notices.
The intrigue: This isn't the only way the NWS is working to improve translation issues in the alert system.
- A 2021 research article published in the American Meteorological Society looked at issues in English-to-Spanish translation of weather alerts that didn't account for dialects, which led to "inconsistent risk messaging," in Spanish-language alerts.
- The NWS/NOAA has since updated the language it uses in hazardous weather communication, adopting "dialect-neutral" terminology suggested by the study authors, as reported by Noticias Telemundo.
The bottom line: Monica Bozeman, the automated language translation lead at the NWS Office of Central Processing, told Axios in an email that the agency has developed experimental pilot projects looking to introduce automation to translation — Lilt being one of them.
- "From these projects, we hope to learn the feasibility of applying automation to gain efficiency with translation turnaround times while reducing the burden of translation on our personnel, especially during critical hazardous weather events," Bozeman wrote.
No comments:
Post a Comment