An AI that reads oracle bones the way scholars do: AlphaOracle traces form, context and classical texts to 3,000-year-old characters

Pinning down one unknown character on a Shang-dynasty oracle bone can take a scholar days of moving between rubbings, bronze inscriptions and classical commentaries. About 3,000 of the roughly 4,500 characters carved into bone and shell still have no agreed reading, and progress is slow because the evidence sits in different centuries, formats and archives. Writing in The Innovation, researchers led by Yuliang Liu at Huazhong University of Science and Technology, with South China University of Technology and Anyang Normal University, describe AlphaOracle, an AI system built to walk that same trail. In a study with 86 specialists it cut analysis time by 64%.

Earlier computational work mostly recognised characters that were already deciphered. Pipelines borrowed from optical character recognition infer a reading from shape alone and seldom ask whether it fits the sentence or survives comparison with transmitted texts. The results can look plausible and still be unusable, because an epigrapher's judgement rests on a chain of corroboration rather than a single resemblance.

AlphaOracle has four stages. A parsing module detects characters on a rubbing, separates settled forms from disputed ones, and restores reading order with a graph neural network. Morphological analysis then draws on three networks: a diffusion model that projects glyphs forward through bronze, seal, clerical and regular script; a decomposition model that maps components onto modern character structure; and a model that places period variants on one timeline. They return ranked candidates with confidence scores. Contextual alignment retrieves every occurrence of the target across 73,883 hand-traced rubbings and 45,364 archaeological scans, then uses a masked language model trained on divination text to judge which candidate fits the formula. Philological grounding checks survivors against 25 transmitted classics and 23,755 modern studies, weighting evidence by age and authority.

The stages narrow in sequence rather than vote. On 88 characters withheld from training, top-1 accuracy was 23.1%, against 13.1% for a reimplementation of Google's SR3 image model. Restoring masked characters from context reached 56.1%, versus 22.2% for the best general model; re-ranking candidates reached 95.2% against 65.4%. Explanations scored 0.491 BLEU, 0.586 ROUGE and 0.659 METEOR, against 0.171, 0.229 and 0.346. General models were tested zero-shot, which the authors say shows how much domain data matters. In blind comparisons, 86 specialists preferred AlphaOracle's contextual explanations 45.2% of the time, and 78.7% rated its usefulness at 4 or 5 out of 5.

The headline case is 勞, long split between "clan or place name" and "adverb marking completion". Morphology, ten parallel inscriptions and sixteen studies pointed to the toponym reading, and the analysis has entered an authoritative database; 22 other disputed variants were re-examined. The system still needs a body of transmitted literature to work with, does not adjudicate between candidates, and leaves the verdict to scholars. The authors present it as a step for scripts that are partially documented, not a general decoder.