China’s DeepZang AI Aims to Improve Tibetan Translation

Choknor, a Chinese technology company founded in 2016, announced upgrades to DeepZang, described as China's first Tibetan-language large language model, at a Monday seminar in Hohhot, Inner Mongolia, with new features scheduled between October and December including offline language packages for Tibetan, Chinese, and English and a December pilot of 50 simultaneous interpretation devices among at least 300 participants in Lhasa. DeepZang, launched in March in Lhasa, has attracted nearly 400,000 users across Xizang, Qinghai, Gansu, and Sichuan, and was trained on nearly 70 million pairs of standardized Tibetan and Chinese texts plus more than 30,500 hours of Tibetan speech covering the Ü-Tsang, Kham, and Amdo dialect groups. Choknor said the model has cut the time for some Tibetan-Chinese translation tasks involving two or three people from 40 minutes to just over 10 minutes for one person, and founder Tenzin Norbu said the company aims to bring DeepZang into "pastoral tents, tourist service counters, and grassroots hospitals.

NEWS https://www.sixthtone.com/features/26166/news?source=normal article China’s DeepZang AI Aims to Improve Tibetan Translation Jiang Xinyi https://www.sixthtone.com/users/1010512/jiang-xinyi?source=normal article DeepZang, China’s first Tibetan-language large language model, is being upgraded to better adapt to Tibetan dialects, enable real-time translation between Tibetan and Mandarin Chinese, and support offline use. At a Monday seminar in Hohhot, capital of northern China’s Inner Mongolia Autonomous Region, the model’s developer unveiled a series of upgrades scheduled between October and December. The seminar featured representatives from companies, research institutions, and universities. DeepZang, developed by local technology company Choknor, was launched in March in Lhasa, capital of the southwestern Xizang Autonomous Region. Its app supports AI-powered conversations, translation, and speech transcription in Tibetan, Mandarin Chinese, and English. At the model’s launch, organizations including the Xizang branches of China Mobile and the People’s Insurance Company of China, as well as Sichuan Minzu Publishing House, signed agreements with Choknor to promote its adoption. According to Choknor, DeepZang has attracted nearly 400,000 users across Xizang, the northwestern Qinghai and Gansu provinces, and southwestern Sichuan province, where most of China’s Tibetan speakers reside. The company has said the model has cut the time needed for some Tibetan-Chinese translation tasks involving two or three people from 40 minutes to just over 10 minutes for one person. It also reported that the model has been used in such varied settings as helping non-Tibetan-speaking doctors communicate with local patients, generating Tibetan-language contracts for yak sales, and assisting grassroots officials in drafting Tibetan work reports. The model has been trained on nearly 70 million pairs of standardized Tibetan and Chinese texts, along with more than 30,500 hours of Tibetan speech covering its three major dialect groups — Ü-Tsang, Kham, and Amdo. The team collected the data from libraries, medical institutions, and monasteries, among other settings. At Monday’s seminar, Choknor unveiled plans to develop offline language packages for Tibetan, Chinese, and English, allowing the app to function with limited or no internet access. The company also plans to pilot 50 simultaneous interpretation devices among at least 300 participants in Lhasa this December. Tenzin Norbu, Choknor’s founder, said the company aims to bring DeepZang into “pastoral tents, tourist service counters, and grassroots hospitals,” enabling Tibetan speakers to “enjoy the benefits of artificial intelligence on an equal footing.” Choknor said on Sunday that beyond fostering exchanges and integration among ethnic groups, DeepZang could help ensure the accuracy of Tibetan-language information and contribute to social stability and long-term security. Founded in 2016, Choknor specializes in Tibetan-Chinese bilingual speech technology, scanning and translation tools, and optical character recognition. Its products include smart computer mice, scanning and translation pens, and an AI-powered Tibetan subtitling system. Editor: Marianne Gunnarsson. Header image: VCG