BaFCo, a new benchmark dataset for Bangla form comprehension, was introduced on July 6, 2026, by a team of researchers including Abu Tyeb Azad and ten others. This initiative aims to enhance document comprehension capabilities for Multimodal Large Language Models (MLLMs), particularly for low-resource languages like Bangla.
Challenges in Document Comprehension for Bangla
Document comprehension remains a significant challenge for MLLMs due to the limited availability of high-quality annotated data for languages such as Bangla. With the growing adoption of these systems in real-world applications, the need for comprehensive datasets has never been more critical.
The BaFCo dataset addresses this gap by providing a meticulously curated collection of 200 multi-page complex Bangladeshi government forms. These forms span various sectors, including agriculture, education, banking, and land management, showcasing the diversity and complexity of the documents.
Key Features of BaFCo Dataset
To ensure a thorough understanding of the structural and contextual intricacies of the forms, BaFCo employs a detailed annotation schema. This schema includes 26 types of form entities, alongside a coarse form entity set comprising 5 types. Such granularity is essential for effective Document Layout Analysis (DLA) and Key Information Extraction (KIE).
- 200 multi-page forms from various sectors
- 26 fine-grained entity types for detailed analysis
- 5 coarse entity types for broader categorization
Evaluation of Multimodal Large Language Models
The research team evaluated several leading MLLMs, including ChatGPT, Gemini, Claude, Qwen, and Kimi. The evaluation utilized both zero-shot and chain-of-thought prompting under varying reasoning conditions. Results indicated notable limitations in the current models' abilities to accurately localize highly granular form entities.
This research was accepted at the 19th European Conference on Computer Vision (ECCV), held in 2026, highlighting its significance in advancing the field of artificial intelligence and document understanding.
🤖 This article was rewritten by Feed and Figures' editorial AI from a report originally published by arXiv NLP. Facts and quotes are preserved from the original; the rewrite focuses on clarity and structure. For the unedited original, see the source link below.
