A field study of an endangered language

NaxiAI 纳西语 · naq xi

Preserving the Naxi language through AI — a Tibeto-Burman language spoken by roughly 30,000 fluent speakers (out of ~350,000 ethnic Naxi) in the Lijiang valleys of Yunnan, and home to Dongba, the world's last living hieroglyphic script. 用人工智能保护纳西语。纳西语属藏缅语族,约30万人使用,主要分布于云南省丽江。纳西族拥有世界上唯一仍在使用的象形文字——东巴文。

Plate I · Dongba pictographs 东巴文
sun
mountain
water
to see
person
词典词条 Dictionary Entries
平行语料 Parallel Corpus
已评估模型 Models Evaluated
试试翻译 Try the translator ↓ The fine-tuned model on HuggingFace →

Photo · Naxi drummers, Jade Dragon Snow Mountain — CEphoto, Uwe Aranas · CC BY-SA 3.0

§ I Translate

纳西语翻译 Translate with Naxi

使用微调翻译模型进行对话。支持纳西拼音、中文和英文。
Chat with our fine-tuned translation model. Supports Naxi Pinyin, Chinese, and English.

纳西语翻译 Naxi Translation

试试纳西语、中文和英文之间的翻译
Try translating between Naxi, Chinese, and English

纳西拼音声调标记:l=高调, q=低调, f=升调, 无标记=中调
Naxi Pinyin tone markers: l=high, q=low, f=rising, (none)=mid

从 From
到 To
§ II Where the Naxi comes from

引擎的来源 What Feeds the Engine

翻译模型背后的两个项目——印刷典籍、地名实体。
Two projects stand behind the translator: the printed record it learns from, and the named places it must never invent.

Digital Library纳西文化数字图书馆

Naxi cultural and linguistic books scanned, OCR'd in Chinese and translated into English page by page — with full-text concordance search across every one.

 

Naxi Gazetteer纳西语地名

Naxi place names with their etymologies, Pinyin, and coordinates — the named entities a translation model must get right rather than hallucinate.

 

Also from the archive: 研究论文与文章 Research Hub →

§ III Evaluation

NTQS 模型排行榜 Model Leaderboard

纳西语翻译质量评分——AI模型翻译纳西语的能力如何?
Naxi Translation Quality Score — how well can AI models translate Naxi?

#
模型 Model
NTQS
声调 Tone
语义 Semantic
完整 Complete
Plate II The Naxi world

东巴之乡 A Homeland of Living Script

The language this project serves belongs to a place — the Lijiang basin beneath Jade Dragon Snow Mountain, and the people who still read its pictographs. 纳西语所属之地——玉龙雪山下的丽江坝子,以及至今仍能读写东巴文的纳西人。

Jade Dragon Snow Mountain rising above pine forest near Lijiang
Fig. 1 Jade Dragon Snow Mountain 玉龙雪山 CEphoto, Uwe Aranas · CC BY-SA 3.0
The Moon-Embracing Pavilion and stone bridge at Black Dragon Pool, Lijiang
Fig. 2 Black Dragon Pool 黑龙潭 Colin W · CC BY-SA 3.0
A Naxi elder in traditional dress in Lijiang Old Town
Fig. 3 A Naxi elder, Lijiang chensiyuan · CC BY-SA 4.0
Back of a Naxi woman's dress showing the seven embroidered star discs of the sheepskin cape
Fig. 4 The seven-star cape 披星戴月 CEphoto, Uwe Aranas · CC BY-SA 3.0
Elderly Naxi musicians performing Naxi Ancient Music
Fig. 5 Naxi Ancient Music 纳西古乐 Peter Morgan · CC BY 2.0
A Dongba-script antithetical couplet manuscript
Fig. 6 A Dongba manuscript 东巴经 Daderot · Public domain
Dongba pictographs painted on a stone wall on Qiyi Street, Lijiang
Fig. 7 Living pictographs, Qiyi St. BrokenSphere · CC BY 3.0
Performers carrying baskets before Jade Dragon Snow Mountain
Fig. 8 Beneath the mountain CEphoto, Uwe Aranas · CC BY-SA 3.0
§ IV The project

关于项目 About the Project

为濒危语言构建AI工具 Building AI tools for an endangered language

数字词典 Digital Dictionary

3,800+纳西语词条,含拼音、IPA转写、声调标记及中英文释义,数据来源于Webonary、Pangloss和田野录音。
3,800+ Naxi entries with Pinyin, IPA transcription, tone markers, and parallel Chinese/English definitions sourced from Webonary, Pangloss, and field recordings.

微调模型 Fine-Tuned Models

基于60,000+纳西语样本训练的LoRA微调语言模型。最佳模型达到82.3 NTQS——比零样本前沿模型提升41分。
LoRA-adapted language models trained on 60,000+ Naxi examples. Our best model achieves 82.3 NTQS — a 41-point improvement over zero-shot frontier models.

文化保护 Cultural Preservation

纳西族的东巴文是世界上唯一仍在使用的象形文字。七星背包象征着知识的旅程——七颗星照亮前路。
The Naxi people's Dongba script is the world's last living hieroglyphic writing system. Their seven-star backpack symbolizes the journey of knowledge — seven stars to light the way.