Khondo: A Multimodal Benchmark for Document Packet Splitting of Bangla Forms

2026-07-23

AI Summary

This work addresses the challenge of automatic segmentation and original page-order reconstruction in government document bundles for low-resource languages, focusing on Bengali. It introduces the first multimodal benchmark specifically designed for Bengali document bundle segmentation, encompassing five stitching configurations across 14 administrative domains, with annotations for page boundaries, domain types, and ground-truth page sequences. Notably, the benchmark explicitly frames page-order reconstruction as a core challenge and enables zero-shot evaluation by leveraging multimodal large language models (MLLMs) with combined visual and textual prompts to assess clustering and sequencing capabilities under varying stitching schemes. Experiments reveal that while current models effectively cluster pages from the same source, they struggle significantly with recovering correct page order—particularly for Bengali compared to English—highlighting page-order reconstruction as a critical bottleneck.

Abstract

Document packets, multiple documents concatenated into a single file, are common in government and administrative workflows, yet splitting them into their constituent documents is difficult, especially for low-resource languages. We introduce Khondo (Bangla for split/segment), the first benchmark for document packet splitting on Bangladeshi government forms. Unlike prior English and OCR-text-based datasets, Khondo is bilingual (Bangla--English) and vision-native; where models operate directly on page images. It spans five concatenation schemes, from sequential to fully shuffled, across 14 administrative domains, with ground-truth boundaries, domain types, and page order. Zero-shot evaluation of MLLMs shows they cluster pages into their source documents fairly well but struggle in restoring the original page order once shuffled. To isolate what drives this difficulty, we run two controlled analyses, varying the prompt instruction and then the packet language. Both primarily affect ordering rather than clustering: (a) explicit page-order instructions are necessary but insufficient, and (b) English packets are ordered more reliably than Bangla, making page arrangement the dominant challenge and language a secondary but consistent factor. Khondo establishes page-order reconstruction as a key open problem in vision-based, low-resource document understanding, and provides a controlled benchmark for measuring progress toward solving it. Our dataset and code is available at https://huggingface.co/datasets/Mausul/khondo

Problem

• document packet splitting
• page-order reconstruction
• low-resource languages
• vision-based document understanding
• Bangla forms

Innovation

• document packet splitting
• vision-native benchmark
• low-resource languages
• page-order reconstruction
• multimodal document understanding