BEIJING, May 13 (Xinhua) -- A Tibetan large language model (LLM) emerged as a highlight at the recently concluded 28th China Beijing International High-tech Expo, spotlighting recent progress in the digitization of the Tibetan language.
The expo's organizing committee said a dedicated regional innovation exhibition area was established to showcase tech products tailored to plateau scenarios, including the Tibetan-language AI model and an intelligent patrol system for uninhabited areas, aiming to introduce such innovations to broader markets.
At the expo, attendees experienced the model, named DeepZang, by simply speaking into a smart mouse. The model then recognized the speech and responded in fluent Tibetan. It can handle a variety of everyday tasks, such as drafting a yak trading contract or composing a poem to praise one's parents.
"General-purpose LLMs are mostly trained on languages like Chinese and English, so their ability to understand and generate Tibetan is limited," said Tsering Dondrup, a doctoral student at Xizang University who is researching Tibetan LLMs. He added that the LLMs, trained on Tibetan speech data and designed to think in Tibetan, can bridge this gap.
According to Dorje Migyur, head of DeepZang's product department, the model has accumulated nearly 70 million accurately matched Tibetan-Mandarin parallel corpus entries and over 30,500 hours of Tibetan speech data, covering three major Tibetan dialects.
He said that while the written Tibetan language is mutually intelligible across different regions, spoken dialects vary significantly. Notably, by training on speech data from all major Tibetan dialect areas, the AI enables cross-dialect communication in Tibetan.
The developer of DeepZang is Xizang Jueluo Digital Industry Management Co., Ltd., which is based in Lhasa, capital of southwest China's Xizang Autonomous Region. According to the website of the China National Intellectual Property Administration, a patented technology filed by the company can effectively solve communication difficulties caused by dialect differences by combining voiceprint recognition with dialect classification.
Losang Donyu, who works in Tibetan-Mandarin translation in the city of Shannan in Xizang, believes that strong cross-dialect speech recognition capabilities can effectively lower the barrier to using Tibetan LLMs. "You don't need a high level of proficiency in written Tibetan. You can use AI simply by speaking, which can help more people," he said.
Losang Donyu and his colleagues now routinely integrate Tibetan LLMs into their daily work. "Previously, translating a document would take two or three people about 40 minutes. Now, with AI assistance, one person can get it done in just over 20 minutes," he said.
According to the company, DeepZang currently has more than 300,000 users, with young people aged 18 to 40 making up over 70 percent of its user base. "Our users are primarily in Xizang and the provinces of Qinghai, Sichuan and Gansu, and a significant number of them live in quite remote areas," Dorje Migyur said. "Many languages are moving toward digitization, and we don't want to be left behind."
This digital leap is fundamentally underpinned by China's robust infrastructure development. Data shows that all county-level areas in Xizang are now connected to the main power grid, and 5G networks have reached every township and 70 percent of administrative villages. With wider electricity and internet access, Tibetan-language AI can now help more Tibetan speakers.
Despite the successes, developers acknowledge the hurdles ahead. "There's still a long way to go from being usable to being truly useful," Dorje Migyur said, noting that some users reported that after just a few rounds of exchanges with DeepZang, the app warned them of an "insufficient token balance."
That's one reason the company debuted at the Beijing tech expo this year -- to find partners and solve problems like high computing costs, financial strain, and the fact that a sustainable business model hasn't yet taken shape.
Tenzin Norbu, the company's board chairman, said that participating in the event could help the company better integrate into the national scientific and technological innovation ecosystem. Regional collaboration has also given a significant boost to the company's development. For instance, the city of Beijing has a long history of providing Lhasa with targeted scientific and technological aid and support.
"Tsangyang Gyatso's poetry has fans all over the world, and it's these poems that bring them closer to Xizang," Tsering Dondrup said. In his view, AI raised in a Tibetan-language environment will one day introduce the world to many more literary treasures, inspiring people to visit Xizang and fall in love with the region. Enditem
京公网安备 11010802027341号