Doc2GraphFormer: Bridging Structured Graph Learning with Transformer Attention for Efficient Document Understanding

doi:10.1007/978-3-032-04627-7_29

Doc2GraphFormer: Bridging Structured Graph Learning with Transformer Attention for Efficient Document Understanding

2026-01-01

0

PRE

AI

DOI：10.1007/978-3-032-04627-7_29

Original

Original request for help

Save

Abstract

En 中文

Structured document understanding requires capturing both semantic relationships and spatial structures within visually rich documents. While Graph Neural Networks (GNNs) effectively model structured dependencies, they struggle with long-range dependencies and global context awareness, which are strengths of transformers. In this work, we propose Doc2GraphFormer, a hybrid framework that integrates graph-based structured parsing with multi-head attention mechanisms, evolving from GraphSAGE-style message passing to a more expressive graph-enhanced transformer architecture. Additionally, we introduce a novel modality fusion strategy, inspired by LayoutLMv3, to enhance the integration of textual, visual, and structural embeddings. We evaluate Doc2GraphFormer on FUNSD, XFUND, demonstrating its superior performance in semantic entity recognition and relation extraction. Our findings highlight the potential of combining graph reasoning with selfattention for efficient document understanding. Our code will be made available on: https://github.com/biswassanket/doc2graphformer

Keywords:

Multimodal Document Understanding

Graph Neural Networks

Key Information Extraction

Entity Detection and Linking

Graphformer

Doc2GraphFormer: Bridging Structured Graph Learning with Transformer Attention for Efficient Document Understanding

Abstract

Journal

Organization