SyncDoc: a synergistic evolution framework for structured document understanding

Abstract

Structured document understanding requires models to jointly exploit textual, layout, and visual information, but existing discriminative models such as LayoutLMv3 often over-rely on complete OCR text during fine-tuning, limiting their robustness when processing low-quality scans, irregular layouts, or incomplete text. To mitigate these issues, we develop SyncDoc, a synergistic evolution framework built upon LayoutLMv3, which introduces three complementary modules: a Pixel-level Compensation Residual (PCR) module to enhance fine-grained visual structures, a Dynamic Masking Strategy (DMS) to reduce dependence on textual semantics, and a Deep Cross-modal Orientation (DCO) module to improve consistency between textual and visual representations. Experiments on benchmark datasets FUNSD and CORD show that SyncDoc consistently outperforms the LayoutLMv3 baseline, achieving entity-level F1 scores of 91.93 and 97.53, respectively. Ablation studies confirm that each module contributes to performance gains, demonstrating that visual structure enhancement, dynamic textual perturbation, and cross-modal alignment are complementary. SyncDoc effectively reduces the model’s over-dependence on OCR text while maintaining high accuracy on complex forms and receipt documents. Additional OCR token masking experiments further show that SyncDoc maintains stronger robustness on the evaluated datasets when textual inputs are partially missing.