Khondo: First Benchmark for Document Packet Splitting of Bangla Forms
Khondo has been established as the inaugural standard for splitting document packets, focusing on forms used by the Bangladeshi government. Document packets, which consist of several documents merged into a single file, are prevalent in administrative processes but pose challenges for separation, particularly in low-resource languages. In contrast to previous datasets that relied on English and OCR text, Khondo is both bilingual (Bangla–English) and vision-native, working directly with page images. It encompasses five different concatenation methods, ranging from sequential to completely shuffled, across 14 administrative sectors, complete with ground-truth boundaries, domain classifications, and page sequences. A zero-shot assessment of multimodal large language models (MLLMs) indicates they effectively group pages into their original documents but find it difficult to restore the original sequence after shuffling. Two controlled studies modifying prompt instructions and model architectures pinpoint this challenge.
Key facts
- Khondo is the first benchmark for document packet splitting on Bangladeshi government forms.
- Document packets are multiple documents concatenated into a single file.
- The benchmark is bilingual (Bangla–English) and vision-native.
- It covers five concatenation schemes across 14 administrative domains.
- Zero-shot evaluation of MLLMs shows good page clustering but poor order restoration.
- Two controlled analyses vary prompt instruction and model architecture.
- The dataset includes ground-truth boundaries, domain types, and page order.
- Prior datasets were English and OCR-text-based.
Entities
Locations
- Bangladesh