Decoding Ancient Oracle Bone Script via Generative Dictionary Retrieval

A research team led by Yin Wu has published a groundbreaking paper titled 'Decoding Ancient Oracle Bone Script via Generative Dictionary Retrieval' on arXiv. The work addresses one of archaeology's most persistent challenges: deciphering Oracle Bone Script (OBS), the 3,000-year-old writing system from China's Shang dynasty. Of approximately 4,600 known OBS characters, only about 1,500 have been decoded, leaving a vast portion of humanity's earliest written records unreadable. Previous computational methods struggled with extreme data scarcity, achieving under 3% accuracy on unseen characters.

The researchers overcame this limitation by fundamentally reframing the problem. Instead of treating decipherment as a classification task, they approached it as a dictionary-based retrieval problem. Using deep learning models guided by principles of character evolution, they generated a comprehensive synthetic dictionary of plausible OBS variants for modern Chinese characters. When scholars encounter an unknown inscription, they can query this AI-generated dictionary to retrieve visually similar candidates, complete with transparent evidence trails. This replaces algorithmic black boxes with interpretable hypotheses that archaeologists can evaluate.

The results are transformative: their method achieves 54.3% Top-10 accuracy and 86.6% Top-50 accuracy for previously unseen characters. This represents an order-of-magnitude improvement over existing approaches. The framework is both scalable and transparent, establishing a generalizable methodology for AI-assisted archaeological discovery beyond just Oracle Bone Script. The paper is currently under review at Nature Machine Intelligence, signaling its potential impact across multiple disciplines.