Technical Summary: Towards Computational Chinese Paleography

1. Problem Statement

Chinese paleography, the study of ancient Chinese writing systems (ranging from Oracle Bone Inscriptions to Han Dynasty clerical scripts), faces a "golden age" paradox: an explosion in the quantity of unearthed materials has outpaced the capacity of traditional manual scholarship. The field is characterized by three primary bottlenecks:

• Scale and Fragmentation: Unearthed artifacts are often physically fragmented, incomplete, and noisy. The sheer volume of materials makes comprehensive collation, rejoining of fragments, and transcription via manual means increasingly impractical.
• Decipherment Complexity: The core intellectual challenge lies in deciphering undeciphered characters. Unlike phonographic scripts where restoration can be framed as sequence modeling, ancient Chinese is a logographic (morphemo-syllabic) system. Decipherment requires a holistic synthesis of graphic form, phonological properties, semantic meaning, and contextual usage.
• Data Scarcity and Imbalance: Existing datasets are heavily skewed toward Oracle Bone Inscriptions (OBIs), leaving bronze and bamboo/silk scripts under-resourced. Furthermore, data distribution follows a long-tail pattern where many characters appear only once or twice, and high-quality, standardized, multimodal datasets (linking images, transcriptions, and semantic analysis) are scarce.

The paper argues that the field is undergoing a "computational turn," evolving from isolated task automation (e.g., simple character recognition) toward the creation of integrated digital ecosystems capable of supporting complex scholarly inquiry.

2. Methodology and Technical Pipeline

The paper structures the computational landscape into a methodological pipeline progressing from foundational visual analysis to advanced reasoning:

A. Foundational Visual Analysis

• Image Processing: Techniques range from traditional filters (K-SVD, anisotropic diffusion) to deep learning models like Structure-Texture Separation Networks (STSN) and OBIFormer (using channel-wise self-attention) to denoise rubbings and restore damaged strokes. Generative models (GANs, Diffusion models like Diff-Oracle) are employed for data augmentation to address scarcity.
• Character Detection & Recognition: The field has shifted from handcrafted features (SIFT, HOG, topological graphs) to deep learning. Current state-of-the-art utilizes Convolutional Neural Networks (CNNs) enhanced with attention mechanisms, and Transformers (e.g., QBSC Transformer for Qin slips). A distinct trend involves radical-based recognition, where models decompose characters into constituent components to support zero-shot learning and handle structural variants.

B. Contextual Analysis

• Artifact Rejoining: Computational methods automate the matching of fragmented artifacts (e.g., oracle bones, bamboo slips) by analyzing physical fracture contours and character sequences. Approaches include Siamese networks, the SUM algorithm (contour slope), and physics-driven deep learning models (WisePanda).
• Deduplication: AI models identify duplicate rubbings of the same artifact by treating inscriptions as visual patterns rather than text, enabling the discovery of duplicates even when textual content varies slightly.
• Periodization and Classification: Deep learning models (e.g., recurrent graph neural networks) are used for font classification and automated dating of artifacts. Some frameworks (e.g., BIRD) leverage textual content and glyph-biased sampling to date bronze inscriptions, achieving high accuracy in fine-grained classification tasks.
• Segmentation and Language Modeling: Efforts include component-level segmentation for oracle bones and the development of multi-modal tokenizers for bamboo slips. Language models are being adapted for ancient texts, often utilizing transfer learning from received texts or constructing specialized corpora (e.g., OracleCS) with expert annotations.

C. Advanced Reasoning and Knowledge Systems

• Automated Decipherment: This is defined not merely as recognition but as the scholarly process of interpreting undeciphered characters (kao shi). Approaches include:
  • Generative Methods: Using diffusion models to generate plausible modern character forms or "semantic typography" from ancient glyphs.
  • Compositional Methods: Treating decipherment as a "jigsaw puzzle" (e.g., Puzzle Pieces Picker, CoLa) where characters are deconstructed into radicals or latent components and reconstructed.
  • Cross-Reference: Retrieving similar characters across different historical scripts (Cross-Font Image Retrieval) or using Large Vision-Language Models (LVLMs) to bridge visual form and semantic meaning.
• Human-AI Collaboration: The paper emphasizes a shift from full automation to human-AI synergy. Systems like Ithaca (for Greek) and emerging Chinese tools (e.g., RejoinX, GenOV) act as intelligent assistants, handling computationally intensive tasks (e.g., fracture matching) while leaving complex integrative reasoning to human experts.
• Knowledge Graphs: Multi-modal knowledge graphs are being constructed to unify heterogeneous data (form, radical, sememe, vocabulary) and model relationships such as diviner social networks or inscription layouts.

3. Key Contributions

The paper makes three primary contributions to the field:

1. A Problem-Oriented Pipeline Analysis: It maps the full spectrum of computational paleography, moving from foundational visual tasks (image restoration, recognition) to contextual analysis (rejoining, dating) and culminating in advanced reasoning (decipherment, knowledge graphs). It highlights the critical shift from isolated character datasets to full-document, multimodal resources.
2. Technological Trajectory Mapping: It traces the evolution from classical feature engineering to modern deep learning, specifically analyzing the adaptation of architectures (CNNs, Transformers, Diffusion models) to the unique logographic nature of Chinese script. It highlights the emergence of methods grounded in radical-based reasoning and compositional analysis.
3. Synthesis of Challenges and Future Agenda: It identifies core obstacles—specifically data scarcity, modality imbalance, and the disconnect between AI's focus on graphic form and the holistic nature of humanistic inquiry. It advocates for a future research agenda focused on multimodal, few-shot, and human-centric systems designed to augment, rather than replace, scholarly expertise.

4. Results and Current State

The paper synthesizes existing literature to report on the current state of the field:

• Dataset Landscape: A proliferation of datasets exists, primarily for Oracle Bone Inscriptions (e.g., HUST-OBS, Oracle-50K, OBC306). While datasets for bronze and bamboo scripts are emerging (e.g., CHUBS, BIRD), they remain less comprehensive. Multimodal datasets (e.g., OBIMD) that link images with transcriptions and semantic annotations are critical but still limited.
• Performance Benchmarks: Recent benchmarks (e.g., OBI-Bench) indicate that current large multimodal models barely reach the level of an untrained human in complex decipherment tasks. While models achieve high accuracy in specific sub-tasks (e.g., 89.8% accuracy in coarse bronze vessel dating), they struggle with the open-set problem of undeciphered characters and the "long-tail" distribution of rare glyphs.
• Collaborative Success: Pilot studies in human-AI collaboration (e.g., artifact rejoining) demonstrate that delegating "simple nature" tasks (fracture matching) to AI while retaining "complex nature" tasks (stylistic analysis) for humans yields successful new discoveries.

5. Significance and Claims

The paper positions itself as a comprehensive survey and a call to action for the field of Computational Chinese Paleography. Its significance lies in:

• Defining the Discipline: It articulates the transition from "paleographic digitalization" (creating infrastructure) to "computational paleography" (integrating computation into the core logical chain of inquiry).
• Highlighting the Logographic Challenge: It emphasizes that the unique nature of Chinese script (logographic, morphemo-syllabic) necessitates a multimodal approach that cannot rely solely on sequence modeling or text-only representations. Visual information is indispensable for tasks like rejoining and decipherment.
• Advocating for Human-Centric AI: The paper argues that the ultimate goal is not full automation but the creation of intelligent scholarly infrastructure. It posits that the future of the field depends on developing systems that act as intellectual partners, using few-shot learning and multimodal fusion to help scholars overcome cognitive bottlenecks and discover large-scale historical patterns invisible to manual inspection.

The paper concludes that while technical innovation is vital, the success of this field hinges on a deep, respectful integration of computational methods with the rich traditions of humanistic inquiry, requiring a spirit of interdisciplinary collaboration between computer scientists and paleographers.