EvaHan 2026 Ancient Books Multimodal OCR and Layout Analysis System Technical Report

Abstract

This paper introduces our system proposal and experimental results for the 5th International Evaluation of Ancient Chinese Information Processing (EvaHan 2026). This evaluation focuses on ancient books OCR tasks using multimodal large language models, including three subtasks: Printed Text Recognition (Task A), Layout Element Analysis (Task B), and Handwritten Text Recognition (Task C). To address core challenges such as numerous variant characters, complex handwritten ligatures, dense layout elements, and annotation noise, we propose a Supervised Fine-tuning (SFT) scheme based on data synthesis augmentation and multi-stage curriculum learning. We also optimized the data preprocessing workflow, resolving key issues like repetition mark recognition and annotation quality improvement. We completed a 9:1 train-validation split on the official dataset and verified the effectiveness of our methods through 6 groups of comparative experiments. Finally, we selected the model with the best comprehensive performance for submission. The code and synthetic dataset are available at https://github.com/zhengningch/EvaHan2026-data.

Resources

Details