- Archaeological News
-
Mogao Caves in Dunhuang, China, a major centre of Buddhist art and manuscript heritage along the ancient Silk Road.
Image Credit:
Zhangzhugang, CC BY-SA 4.0, via Wikimedia Commons
AI Helps Fill the Gaps in Ancient Tibetan Manuscripts from Dunhuang
A new study is testing whether artificial intelligence can help scholars reconstruct missing passages in damaged Tibetan manuscripts from Dunhuang, using a specially adapted version of the DeepSeek-R1 language model to suggest characters and syllables that have disappeared through centuries of deterioration.
Ancient Tibetan manuscripts preserve valuable information about history, religion, philosophy, medicine and other aspects of life, yet many survive only in damaged condition. Weathering, staining, pests, fading ink and physical tears can leave characters partly visible or completely missing, forcing specialists to reconstruct passages manually from the surviving context. The new research explores whether a language model trained to understand both modern and ancient Tibetan can make that process faster and more systematic.
The researchers developed a model called DeepSeek-R1-Tibet, based on the 32-billion-parameter DeepSeek-R1-Distill-Qwen model. Rather than training it from scratch, they adapted it to Tibetan through targeted fine-tuning, allowing the system to retain its broader language and reasoning abilities while learning the vocabulary, spelling patterns and contextual relationships needed for ancient Tibetan texts.
The ancient material came from Tibetan manuscripts preserved in volumes 1–7 of the Dunhuang documents held by the National Library of France. Their texts were first extracted through optical character recognition and then checked in two stages by Tibetan-language specialists against the original manuscript images, reducing the risk that OCR mistakes would become part of the training data. The processed ancient corpus contained 496,121 Tibetan syllables.
One obstacle was the limited amount of surviving ancient Tibetan material available for machine learning. To compensate, the team added a much larger 10.56-gigabyte corpus of modern Tibetan, while developing a syllable-based vocabulary capable of handling forms found in both modern and ancient texts. The resulting vocabulary covered about 98.6% of the Tibetan syllables in the corpus.
The system was tested by deliberately hiding one, two or three syllables from verified passages and asking different models to predict what belonged in the gaps. DeepSeek-R1-Tibet improved dramatically after fine-tuning: one evaluation score rose from 0.012 to 0.895 when a single syllable was missing, to 0.743 for two missing syllables and 0.667 for three.
The comparison revealed an important distinction. For simpler passages with only one or two missing syllables, the fine-tuned mBERT model achieved higher absolute scores on some measures. But as the gaps became larger and required more information from the surrounding sentence, DeepSeek-R1-Tibet gained an advantage. When three syllables were missing, the correct answer appeared among its top three suggestions in 66.7% of cases, compared with 45.6% for mBERT.
That doesn’t mean the model can independently “restore” a manuscript. Its most practical role may instead be to produce a short list of plausible candidates for specialists to examine. This is especially useful when several consecutive syllables are missing, because the model can consider broader relationships across the sentence rather than treating every gap separately.
The researchers also caution that important limitations remain. Even expert-checked OCR can contain errors when manuscripts are badly damaged, while adapting a very large model with a comparatively small ancient corpus raises the possibility that it may memorize patterns rather than truly generalize to unfamiliar texts. Current evaluation measures also test whether the correct candidate appears near the top of the list, but don’t fully judge whether a reconstructed sentence makes linguistic and historical sense. Human expertise remains essential. Even so, the approach may have uses beyond Tibetan. Similar methods could eventually assist work on other damaged manuscript traditions from the Silk Road, including Classical Mongolian, Tangut and Sogdian texts. The goal isn’t to replace the specialist reading an ancient manuscript, but to give that specialist a new way to approach what the surviving page can no longer show clearly.
Published on: 18-10-2026
Edited by: Abdulmnam Samakie
Source: npj Heritage Science
