BnPostDuo

Description

BnPostDuo is a Bengali multimodal social media dataset of 4,708 Facebook image posts across eight content categories: Educational (593), Entertainment (597), Marketing (581), Motivational (600), News (598), Political (598), Religious (551), and Social Awareness (590). Posts were manually collected from publicly accessible Facebook pages and groups during 2025–2026. An initial balanced collection of 4,800 posts (600 per category) was reduced to 4,708 after perceptual-hash (pHash) duplicate screening, manual visual inspection, and removal of 92 confirmed duplicates. Each record links the user-written Bengali caption with text embedded in the post image. Image text was extracted with EasyOCR (raw_ocr) and then manually reviewed against the original images to correct recognition errors and retain the meaningful visible text (main_ocr). Category labels were assigned by human annotators. A stratified subset of 480 posts (60 per category) was independently relabelled by three annotators with the original labels hidden, giving a Fleiss' kappa of 0.8905 (411 unanimous, 69 two-to-one decisions). Files: - BnPostDuo_combined.csv: all 4,708 records (image_id given as category/filename, e.g. marketing/1.jpg) - educational.csv, entertainment.csv, marketing.csv, motivational.csv, news.csv, political.csv, religious.csv, socialawareness.csv: one file per category (image_id given as filename only) Fields: image_id, caption, raw_ocr, main_ocr, phash (64-bit perceptual hash, hex), image_width, image_height (pixels), label. The original Facebook images are not redistributed, out of respect for content creators' rights; only text and image metadata are released. The dataset supports Bengali social media topic classification from captions, OCR text, or both; analysis of OCR errors (raw vs. curated text); and few-shot or zero-shot evaluation of LLMs and VLMs. Processing scripts and documentation: https://github.com/Azu2Bro/BnPostDuo. For academic and research use.

Steps to reproduce

1. Collect Bengali image posts from public Facebook pages/groups for the eight categories; save each image and record its caption under a unique image_id. 2. Run EasyOCR (Bengali) on each image with no preprocessing; store the output as raw_ocr. 3. Manually review each image against raw_ocr; transcribe and correct the main visible text into main_ocr, omitting decorative or unrelated text. 4. Compute a pHash, width, and height for each image; inspect hash-matched images and remove confirmed duplicates. 5. Assign category labels; validate them on a stratified 480-record subset with three blinded annotators and compute Fleiss' kappa. 6. Export eight per-category CSVs and one combined CSV.

Institutions

• Chittagong University of Engineering & Technology

  Chittagong, Chittagong

Categories

Licence