Beijing Normal University at EvaHan 2026: Enhancing Ancient Chinese Character Recognition and Layout Analysis via VLM Fine-Tuning and Linguistic Post-Processing

Abstract

This paper describes the system submitted by the Beijing Normal University (BNU) team for the EvaHan 2026 shared task. We participated in Task A (Printed Text Recognition), Task B (Layout Element Analysis), and Task C (Handwritten Character Recognition). For text recognition (Tasks A and C), we proposed a hybrid pipeline combining supervised fine-tuning (SFT) of Vision-Language Models (VLMs) with a linguistic rule-based post-processing module. In the Open Track, we further explored the use of a general-purpose VLM to correct semantic errors while maintaining visual fidelity to ancient variant characters. For Task B, we adopted a method integrating a VLM with structured prompting strategies. Our system consistently surpassed the official baselines, achieving an F1 score of 94.53% in Task A and 91.33% in Task C, while demonstrating enhanced localization precision in Task B.

Resources

Details