Decoding Ancient Oracle Bone Script via Generative Dictionary Retrieval

Yin Wu; Gangjian Zhang; Jiayu Chen; Chang Xu; Yuyu Luo; Nan Tang; Hui Xiong

cs.IRApr 1, 2026

Decoding Ancient Oracle Bone Script via Generative Dictionary Retrieval

Authors: Yin Wu, Gangjian Zhang, Jiayu Chen, Chang Xu, Yuyu Luo, Nan Tang, Hui Xiong

Organizations: Information Hub, The Hong Kong University of Science and Technology (Guangzhou), Guangzhou, China.

Abstract

Understanding humanity's earliest writing systems is crucial for reconstructing civilization's origins, yet many ancient scripts remain undeciphered. Oracle Bone Script (OBS) from China's Shang dynasty exemplifies this challenge: only approximately 1,500 of roughly 4,600 characters have been decoded, and a substantial portion of these 3,000-year-old inscriptions remains only partially understood. Limited by extreme data scarcity, existing computational methods achieve under 3% accuracy on unseen characters -- the core palaeographic challenge. We overcome this by reframing decipherment from classification to dictionary-based retrieval. Using deep learning guided by character evolution principles, we generate a comprehensive synthetic dictionary of plausible OBS variants for modern Chinese characters. Scholars query unknown inscriptions to retrieve visually similar candidates with transparent evidence, replacing algorithmic black boxes with interpretable hypotheses. Our approach achieves 54.3% Top-10 and 86.6% Top-50 accuracy for unseen characters. This scalable, transparent framework accelerates decipherment of a pivotal undeciphered script and establishes a generalizable methodology for AI-assisted archaeological discovery.

Figures & tables

Figure 1: Overview of the CipherOBS system for synthesizing and retrieving OBS variants. a, Two-stage synthesis process: FAD generates initial OBS drafts from modern Chinese character images using a denoising U-Net conditioned on font-augmented inputs; SR then refines these drafts with IDS, derived from the modern character to produce high-fidelity variants, accounting for historical erosion and stylistic differences. b, Dictionary-based retrieval: A encoder embeds the input OBS image and synthesized dictionary entries into a feature space; cosine similarity identifies Top- $k$ matches, with voting on modern character labels for robustness against degradation. This pipeline creates a visual dictionary of OBS-modern pairs for scalable decipherment. English translations are also provided for each OBS’s Chinese explanations.

Figure 2: Validation of the visual fidelity of the synthetic OBS dictionary. a, 2D t‑SNE of retrieval‑encoder features for 15 characters. Authentic glyphs (circles) and synthetic counterparts (squares) co‑localize within character‑specific clusters; insets enlarge two regions. b, Quantitative comparison between real OBS images and three candidate sets: modern character renderings, entries from our OBS dictionary, and dictionary entries after 20 refinement iterations. Metrics: FID, LPIPS, L1 loss (lower is better); SSIM (higher is better); Values are means across 20 runs; standard deviations are in parentheses. c, Results of a Turing‑style discrimination task (n = 12 participants; 1,200 judgments; balanced 5 real/5 synthetic per set).

Figure 3: Zero-shot generalization to newly deciphered script characters. a, Sankey diagram summarizing retrieval outcomes for 14 glyphs grouped by graph group—Qi (‘wife’; 妻), Qiang (‘Qiang people’; 羌), Lü (‘shoe’; 履), Chu (‘chick’; 雛). Flows indicate whether the structurally matched modern character appears at Top-1, within the Top-10, or outside the Top-10 shortlist. b , Top- $N$ structural matching accuracy as a function of $N$ ; accuracy reaches 100% within $N=100$ , consistent with expert-in-the-loop inspection of ranked candidates. c, Quality metrics for the corresponding synthesized dictionary entries, min-max normalized across the 14 cases. For FID, LPIPS, and L1 loss, lower values indicate better alignment; for SSIM, higher is better. Perceptual similarity remains comparatively strong across categories, while pixel-aligned agreement is weaker for Chu, reflecting erosion and stylistic variability. d, Qualitative links between recently studied OBS glyphs (top), synthesized OBS-style variants (middle), and modern Chinese characters (bottom) for Chu, Qiang, Lü, and Qi. We note that structural similarity to a modern character is a hypothesis-generating step; full decipherment requires broader philological analysis including phonological and contextual evidence.

Figure 4: Robustness of CipherOBS to field-realistic image degradation. a , Representative digitized oracle-bone fragments exhibiting common surface conditions in archaeological contexts. b , Degradation types (top: field manifestations; bottom: synthetic operators applied to a test glyph): Blur (Gaussian defocus), Noise (additive Gaussian noise), Erode (stroke thinning/chipping), and Mask (random occlusion), each at three severities. c , Top-1 (upper) and Top-20 (lower) retrieval accuracy for FAD, FAD+SR without IDS and the full system (FAD+SR) across 12 conditions (4 types $\times$ 3 severities). Points indicate mean accuracy across test images; error bars denote 95% confidence intervals; background shading encodes severity (1 = light; 2 = medium; 3 = heavy).

Metric	Unaided Condition	AI-Assisted	Difference
	Mean $\pm$ s.d.	Mean $\pm$ s.d.
Overall Performance (n=12)
Accuracy (%)	16.5 $\pm$ 9.2	54.4 $\pm$ 11.9	+37.9
Response Time (s)	45.0 $\pm$ 18.7	37.8 $\pm$ 15.1	-7.2
Confidence (1–5 scale)	3.0 $\pm$ 0.7	4.0 $\pm$ 0.5	+1.0
By Participant Expertise

Table 2: Human evaluation of OBS interpretation with and without retrieval-based assistance. Values are mean

\pm

s.d. across participants (n = 12; 50 trials per participant). Response time is per trial in seconds; confidence is self-reported per trial on a 1–5 Likert scale (1 = very uncertain, 5 = very certain). Differences are absolute (pp for accuracy, seconds for time, points for confidence).

AlphaOracle: Oracle bone script decipherment via human-workflow-inspired deep learning
Jul 20, 2026Yuliang Liu, Haisu Guan, Pengjie Wang +11
School of Software Engineering, Huazhong University of Science and Technology, Wuhan 430074, China · Key Laboratory of Oracle Bone Inscriptions Information Processing, Anyang Normal University, Anyang 455002, China · School of Electronic Information Engineering, South China University of Technology, Guangzhou 510641, China
Specializing Large Models for Oracle Bone Script Interpretation via Component-Grounded Multimodal Knowledge Augmentation
Apr 8, 2026Jianing Zhang, Runan Li, Honglin Pang +5VLM Reasoning
College of Software, Jilin University · School of Artificial Intelligence, Jilin University · Graduate School of Information Science and Technology, The University of Tokyo +2
FROD: Feature Matching Residual Denoising Oracle Bone Decipher
Sep 15, 2026Yanbin Hou, Biao Xiong, Guojun Xu +4Image-to-Image Translation Diffusion Models
School of Artificial Intelligence, Wuhan University of Technology, Wuhan 430070, China

Jul 20, 2026J/K move · Enter open · S save

AlphaOracle: Oracle bone script decipherment via human-workflow-inspired deep learning

Yuliang Liu, Haisu Guan, Pengjie Wang, Xinyu Wang+10

School of Software Engineering, Huazhong University of Science and Technology, Wuhan 430074, China · Key Laboratory of Oracle Bone Inscriptions Information Processing, Anyang Normal University, Anyang 455002, China · School of Electronic Information Engineering, South China University of Technology, Guangzhou 510641, China

Open paper AskPDF arXiv

Approximately 3,000 of the 4,500 oracle bone script (OBS) characters remain undeciphered due to fragmentary inscriptions and sparse evidence. Current AI approaches fail to replicate expert workflows that integrate form analysis, contextual semantics, and philological reasoning. We introduce AlphaOracle, a human-workflow-inspired framework that systematizes OBS decipherment using the largest digitized corpus to date. Its multi-stage pipeline comprises: (i) rubbing parsing; (ii) radical-based morphological analysis with diachronic modeling; (iii) contextual retrieval with semantic alignment; and (iv) philological validation against classical sources. Each stage generates explicit, confidence-weighted evidence chains, culminating in interpretable reports for scholarly verification. Across multiple test characters, AlphaOracle's readings strongly agreed with expert interpretations. In a study of 86 domain specialists, it reduced analysis time by 64% and 79% of participants rated it highly useful. Notably, AlphaOracle resolves the character "Lao" as a toponymic or clan designation, offering concrete revisions to Shang administrative and social interpretations. These results suggest that computational methods aligned with philological practice can facilitate OBS research and provide a conceptual reference for studies of other undeciphered scripts.

Abstract

Explore similar work

AlphaOracle: Oracle bone script decipherment via human-workflow-inspired deep learning

Specializing Large Models for Oracle Bone Script Interpretation via Component-Grounded Multimodal Knowledge Augmentation

FROD: Feature Matching Residual Denoising Oracle Bone Decipher