MambaDocNet: Scalable multilingual document understanding for complex script structures

Akkshita Trivedi

doi:10.1016/j.patcog.2026.115093

MambaDocNet: Scalable multilingual document understanding for complex script structures

2026-10-08

0

PRE

AI

DOI：10.1016/j.patcog.2026.115093

Original

Original request for help

Save

Abstract

En 中文

Indian scripts such as Hindi, Bangla, and Odia exhibit complex visual compositions with dense ligatures, stacked modifiers, and nonlinear character combinations. These properties make document understanding particularly challenging for Transformer-based models, which face scalability issues due to quadratic attention complexity and weak generalization to tasks involving Indic textual structures. To overcome these limitations, we introduce MambaDocNet, a unified and efficient framework for multilingual document entity detection and segmentation. The model leverages selective State Space Models (SSMs) within a Mamba encoder to capture long-range dependencies efficiently while maintaining linear computational complexity. Spatially aware embeddings preserve document layout information, and lightweight task-specific decoders enable fine-grained text region detection and segmentation. Extensive experiments on benchmark datasets demonstrate statistically comparable results with many of the text, layout, and image-based baselines. MambaDocNet achieves F1-scores of 93.8% on FUNSD, 97.7% on CORD, and 97.2% on SROIE, and remains competitive with the state-of-the-art results on the IndicScript dataset with 67.9% (Odia), 68.1% (Bangla), and 70.4% (Hindi). MambaDocNet remains scalable while generalizing well across detection and segmentation tasks on Indic script documents.

Keywords:

Document understanding

State space models (SSM)

Multilingual learning

MAMBA

Document image

MambaDocNet: Scalable multilingual document understanding for complex script structures

Abstract

Journal

Organization

Cited Papers