BaFCo introduces a benchmark for complex Bangla form comprehension

Researchers have introduced BaFCo, a new benchmark dataset designed to address the scarcity of high-quality annotated data for low-resource languages like Bangla in document comprehension tasks. The dataset focuses on Document Layout Analysis (DLA) and Key Information Extraction (KIE) by curating 200 multi-page complex Bangladeshi government forms from sectors such as agriculture, education, banking, and land management.

• BaFCo includes a fine-grained annotation schema with 26 types of form entities and a separate coarse entity set of 5 types to capture structural complexity.
• The study evaluates latest Multimodal Large Language Models (MLLMs) from ChatGPT, Gemini, Claude, Qwen, and Kimi series using zero-shot and chain-of-thought prompts.
• Results highlight limitations in current MLLMs' ability to comprehend Bangla forms, particularly in accurately localizing highly granular form entities.

The dataset and code are.