Ramstad: AI is English-centric, but it’s picking up Hmong quickly

A couple of weeks ago at the public library in Maplewood, tech consultants May Yang-Her and her husband Dao Her were teaching digital marketing to Hmong business owners when May asked how many had used AI to translate something.

Nearly everyone had, and the class came alive with the business owners describing moments when ChatGPT and other AI applications accurately handled the Hmong language.

“It has come a long way,” May said. “The word choices they use are fascinating.”

“Sometimes I think ChatGPT knows more Hmong than I do,” Dao said.

The absorption and manipulation of language is a key element of artificial intelligence. Until recently, the large language models, or LLMs, that are the vanguard of AI systems focused on English and pretty much ignored less-used languages.

That’s particularly challenging for languages like Hmong, which was purely an oral language until the 1960s, and which many native speakers still only know as a spoken form of communication.

“We want these systems to be able to know our language, for our elders to use, and for future generations to still be able to learn our language,” Mai Lee Chang, an AI and robotics researcher, said during a panel on AI at the Hmong National Development conference in Minneapolis in April.

The imbalance of language and cultural representation is a well-known problem in technology. English still dominates on more than half of all websites, even though only 5% of the world’s nearly 8 billion people speak it as their first language. While the imbalance is more pronounced in AI at the moment, it is likely to shrink.

AI is now in the stage of its development where public discussion and media portrayals swing between doom and hype. We talked about the internet the same way 30 years ago.

Like the internet, AI is becoming a utility. Its weaknesses, including the risk of cultural imperialism, are being identified and addressed.

For instance, in the past year, computer scientists announced huge progress on small language models, or SLMs. In contrast to LLMs, they use smaller datasets that can be directed to singular tasks such as translating a language.

May Yang-Her and Dao Her are watching the effects of all this in their work and at home. They own DaoTech Solutions, a provider of tech integration, networking and other services for small businesses and charter schools.

Both promote AI to their clients and in the Hmong community. They’ve used it to help school clients with lesson plans and to diagnose plumbing problems at their cabin in northern Minnesota. They’ve watched their teenage son use AI to create podcasts to learn philosophy and their daughter use it to distill her notes from nursing classes.

May also participated in the AI panel at the Hmong conference, where she predicted that people in the future will be judged less by their degrees and credentials than by their problem-solving abilities, which AI will aid.

There are about 12 million Hmong speakers in the world, most of them in China. A large diaspora fled Laos after helping the Americans in the Vietnam War, winding up in Thailand, France and the U.S.

In America, one of the largest concentrations of people who are Hmong, nearly 100,000, is in Minnesota. The state this year has been celebrating the 50th anniversary of the arrival of Hmong people.

“With the 50th anniversary this year, there’s been a lot of looking back and talk about where we’ve been,” May said when I visited the couple in their St. Paul office. “The next 50 years is going to be so different.”

On a trip to Thailand last year, May and Dao noticed the way the Thai language had influenced new words for Hmong speakers there.

“They were using Hmong words we didn’t know. We were asking, what do you mean by that?” Dao said.

That’s when the Hers began to think AI models could have an enormous role sorting out differences in the Hmong language and steering it in the future.

May said ChatGPT doesn’t “know the nuances and the culture and the different dialects” of Hmong. “But they’ve gotten a lot better with interpretations,” she said.

One of the students in the Hers’ class, Bon Xiong, who works in sales of industrial systems, told me he’s tracked the improvement because he frequently writes a document in English, then runs it through ChatGPT for a Hmong translation.

“I noticed it in the last couple months,” Xiong said. “I was working with a Hmong friend, and I spent a half an hour trying to translate myself and then I thought, ‘What am I doing? Let me just plug this into ChatGPT.’ And it came back so fast, and I was like, whoa. I was impressed. And she read it too and she said, ‘Did you write this?’”

AI models have not figured out how to pronounce Hmong words, however. And Hmong isn’t an easy language in that respect; there are seven or eight tones, depending on the dialect.

“Right now, we’re just looking at text,” May said. “I would like to see it do text-to-talk for our elders who can’t read English or can’t read Hmong.”