A Parameter-Efficient and Data-Centric Framework for Ancient Chinese Text Recognition and Layout Analysis

Abstract

This paper presents the system developed for the EvaHan 2026 shared task on Ancient Chinese OCR and Layout Analysis. Participating in the Closed Track, we propose a highly parameter-efficient, data-centric framework based on the Qwen2.5-VL-7B-Instruct multimodal large language model (MLLM). While the official baseline utilizes the same backbone architecture, our approach significantly outperforms it by integrating orientation-aware image preprocessing and expert-constrained adaptive prompt engineering. We employed Low-Rank Adaptation (LoRA) with a minimal rank configuration (Rank=16) to train three independent, task-specific adapters. Our system achieved exceptional results, recording an Overall score of 0.9703 and an F1-score of 97.19% on printed text recognition (Task A)—effectively halving the baseline’s Character Error Rate. On handwritten texts (Task C), we maintained a highly competitive 90.18% F1-score. Furthermore, our model achieved significant progress in layout analysis (Task B), surpassing the baseline’s Macro F1 by 172% (0.4162 vs. 0.1530) and mAP by 37%. These results underscore that embedding explicit document structure and semantic constraints into MLLMs is more effective than simply scaling model parameters.

Resources

Details