| International Journal of Applied Information Systems |
| Foundation of Computer Science (FCS), NY, USA |
| Volume 13 - Number 5 |
| Year of Publication: 2026 |
| Authors: Gabriel Ihuoma Lilian, Laud Charles Ochei, Uche Oghenekaro |
10.5120/ijaisdc016f874c3d
|
Gabriel Ihuoma Lilian, Laud Charles Ochei, Uche Oghenekaro . IntelliDocFormer: An Extended LayoutLMv3 Architecture with Graph-based Relational Reasoning and Document-Type-Aware Multimodal Fusion. International Journal of Applied Information Systems. 13, 5 ( Nov 2026), 68-77. DOI=10.5120/ijaisdc016f874c3d
The digitisation of finance, healthcare, legal services and academia has produced large volumes of unstructured and semi-structured documents, from invoices and forms to scientific papers and contracts. Extracting information from them requires systems that understand both textual content and visual layout. Transformer-based models such as LayoutLM, DocFormer and Donut process documents multimodally, but they reason weakly about relations between spatially separated yet semantically related elements, treat all document types uniformly, have used visual encoders designed for natural scenes, and are computationally demanding. This paper presents IntelliDocFormer, an extended LayoutLMv3-based architecture that adds a Document-Type Classification (DTC) module, graph-based relational reasoning with a Graph Attention Network over a sparse layout graph, a Vision Transformer with Feature Pyramid Network (ViT-FPN) visual encoder, and an adaptive multimodal fusion layer. The architecture is analysed against existing models and evaluated with the Architecture Trade-off Analysis Method (ATAM) using a résumé-screening scenario extended to a utility tree of seven quality-attribute scenarios covering accuracy, throughput, degraded input, long documents, modifiability, explainability and fairness. The analysis shows that the modular design addresses document heterogeneity and complex layouts, and it identifies the sensitivity points, trade-offs and risks of the design, including the retained quadratic cost of the backbone. The contribution is conceptual and architectural; an implementation and validation plan on DocBank, FUNSD, CORD and DocVQA is specified for future empirical work.