Data-Centric Strategies for Ancient Chinese Text Recognition: Augmentation, Annotation Refinement, and Style Transfer in EvaHan 2026

Abstract

This paper describes our system for the EvaHan 2026 shared task. We design and experiment with data-centric strategies across three subtasks: printed text OCR (Task A), layout element analysis (Task B), and handwritten text OCR (Task C). Our approach employs systematic data augmentation using 17 transformation strategies, comprehensive manual annotation refinement for layout analysis, and style transfer augmentation for handwritten texts. We use pre-trained Qwen2.5-VL-7B-Instruct with LoRA fine-tuning as the base model. According to the evaluation metrics adopted by the organizers, our system achieves 27.5% and 4.5% CER reduction over official baselines for Tasks A and C respectively. Manual annotation refinement for Task B achieves 205% improvement in Micro F1 and 258% improvement in Macro F1 on the validation set, demonstrating that annotation quality is the primary bottleneck for layout analysis in closed-modality settings.

Resources

Details