Infinity-Parser2: A New GNN-Based Approach to Document Parsing

Researchers have introduced Infinity-Parser2, a novel document parsing model that uses graph neural networks (GNNs) to better capture the structural relationships in documents. Unlike traditional OCR-based or transformer-only approaches, Infinity-Parser2 models documents as graphs, where nodes represent text regions and edges represent spatial and semantic relationships.

The model shows particular strength in parsing complex multi-column layouts, tables, and mixed-format documents. By learning from graph representations, the model can preserve document hierarchy and logical reading order more effectively than sequential approaches.

Technical Details

Infinity-Parser2 builds on previous work by incorporating a GNN backbone that processes the document graph in multiple stages. First, a layout analysis module extracts text blocks and their positions. Then, a GNN refines relationships between blocks, predicting which ones belong to the same logical section or reading flow.

Key innovations include:

• Edge-aware attention: The model uses attention mechanisms that consider both spatial distance and content similarity between text blocks.
• Hierarchical decoding: Output is generated in a tree-like structure to preserve section hierarchy.
• End-to-end training: The entire pipeline is trained jointly on labeled document datasets.

Performance and Implications

In benchmarks on public document parsing datasets, Infinity-Parser2 achieves state-of-the-art results on complex layouts, particularly in F1 scores for structure recovery. The authors note that the model generalizes well to unseen document formats, suggesting potential for real-world use in digitizing historical documents or processing diverse business reports.

For developers working on document understanding pipelines, this work offers a promising direction: moving from flat OCR outputs to structured, graph-based representations that capture document semantics. The code is expected to be open-sourced, though no release date has been announced.

The paper is available on arXiv under the ID 2607.07836.