BTMD: Bengali Trimodal Misinformation Dataset

Dataset Summary

BTMD (Bengali Trimodal Misinformation Dataset) is a large-scale, manually annotated multimodal dataset for Bengali (Bangla) fake news detection. It contains 6,058 carefully curated news instances spanning three modalities — text, image, and video — collected from diverse Bengali online sources.

The dataset supports both binary classification (Real vs. Fake) and fine-grained misinformation classification (Real, Misinformation, Satire, Clickbait), making it a comprehensive benchmark for evaluating multimodal misinformation detection models in a low-resource language setting.

BTMD covers 11 news categories and was constructed through a rigorous multi-stage pipeline involving crowdsourced data collection, source verification, expert annotation, and inter-annotator agreement analysis.

Supported Tasks

Task	Description	Labels
Binary Classification	Distinguish real from fake news	Real, Fake
Fine-grained Classification	Classify the type of misinformation	Real, Misinformation, Satire, Clickbait

Languages

• Bengali (Bangla) — ISO 639-1: bn

Dataset Structure

Data Fields

Field	Type	Description
id	string	Unique identifier (UUID) for each news instance
heading	string	News headline (may be null for image/video-only instances)
text	string	Full article text content (may be null)
image_path	string	Relative path to the associated image file(s). Multiple images are semicolon-separated (e.g., images/img1.png;images/img2.png). May be null for text-only or video-only instances.
video_path	string	Relative path to the associated video file (may be null)
label	string	Binary label: Real or Fake
multi_category	string	Fine-grained label: Real, Misinformation, Satire, or Clickbait
source_category	string	Source platform (e.g., Facebook, YouTube, Website, Newspaper)
category	string	News domain category (e.g., Politics, Health, Science)

Data Splits

The dataset is provided as a single unsplit file with 6,058 instances. The original paper employs 5-fold cross-validation for evaluation. Users are free to define their own splits as needed.

Split	Instances
Full Dataset	6,058

Modality Availability

Each instance contains at least one modality (text, image, or video). The textual modality includes news headlines and article content, while the visual modality consists of images associated with each news item.

Modality	Instances
Has Text (heading or article)	4,457
Has Image	4,289
Has Video	1,383

Label Definitions

Label	Definition
Real	Factually accurate content whose textual claims and associated visual or video content can be verified using credible sources.
Misinformation	False or misleading content presented as factual information, including fabricated claims, manipulated information, or authentic media shared in a misleading context.
Satire	Content created primarily for humor, irony, or social commentary that may be mistaken for genuine news when shared without its original context.
Clickbait	Content containing sensational or misleading headlines, thumbnails, or descriptions intended to attract attention while inaccurately representing the underlying information.

Dataset Statistics

Binary Class Distribution

Binary Class	Count	Percentage
Real	3,029	50.0%
Fake	3,029	50.0%
Total	6,058	100.0%

Fine-grained Category Distribution

Category	Count	Percentage
Real	3,029	50.0%
Misinformation	1,859	30.7%
Satire	594	9.8%
Clickbait	576	9.5%

News Domain Distribution

Category	Total	Real	Fake	Misinfo.	Satire	Clickbait
Politics	879	409	470	327	131	12
Technology	679	348	331	84	164	83
Entertainment	636	345	291	200	55	36
Sports	613	313	300	214	73	13
International	610	302	308	230	36	42
Science	579	286	293	121	0	172
Environment	551	291	260	170	0	90
Religion	527	264	263	154	99	10
Health	402	201	201	101	0	100
Education	362	161	201	167	34	0
Miscellaneous	220	109	111	91	2	18

Modality Co-occurrence

Modality Combination	Count	Percentage	Real	Fake
Text + Image	2,735	45.15%	1,363	1,372
Image Only	1,473	24.31%	744	729
Text + Video	1,176	19.41%	601	575
Text Only	467	7.71%	219	248
Video Only	126	2.08%	72	54
Text + Image + Video	81	1.33%	30	51

Source Distribution

Source	Total	Real	Fake
Facebook	3,627	1,558	2,069
YouTube	831	228	603
Website	589	498	91
Newspaper	532	476	56
News Channel	274	266	8
Miscellaneous	97	0	97
TikTok	55	0	55
Twitter	37	3	34
Instagram	14	0	14
Blog	2	0	2

Text and Video Statistics

Statistic	Word Count	Video Duration (s)
Minimum	1	3.87
Maximum	2,738	300.12
Mean	156.80	87.47
Median	59	72.62
Std. Dev.	211.91	69.79
Q1	11	26.68
Q3	247	128.35

Data Collection

Collection Pipeline

BTMD was constructed through a structured multi-stage pipeline:

1. Crowdsourced Acquisition: A custom Telegram bot collected news content (text, images, videos, URLs) from volunteers across multiple undergraduate and graduate programs at the American International University-Bangladesh (AIUB).

2. Source Verification: Each submitted news item was independently verified by trained annotators through cross-referencing with credible sources, including established news outlets, official reports, fact-checking organizations, and trusted online platforms.

3. Annotation: Five trained annotators with domain expertise labeled each instance using a custom cross-platform annotation tool. Annotations include binary labels, fine-grained categories, and annotator confidence scores.

4. Quality Assurance: Instances with confidence scores ≤ 70% were re-annotated under blind review by all five annotators. Final labels were determined by majority voting (≥ 3/5 agreement).

Source Diversity

Data was collected from diverse online sources including social media platforms (Facebook, YouTube, TikTok, Twitter, Instagram), news portals, newspapers, news channels, blogs, and online forums.

Annotation Tool

The custom annotation platform is publicly available at: https://github.com/Faysal1000/fake-news-annotation-tool

Annotation Quality

Inter-Annotator Agreement

Agreement was evaluated on a stratified 500-instance subset using Cohen's κ (pairwise) and Fleiss' κ (five-rater), under a blind re-labeling protocol.

Annotation Scheme	Class	Cohen's κ	Fleiss' κ
Binary	Overall	0.848	0.848
Multi-class	Real	0.825	0.825
Multi-class	Misinformation	0.814	0.814
Multi-class	Satire	0.842	0.842
Multi-class	Clickbait	0.802	0.802
Multi-class	Overall	0.822	0.822

All κ values indicate almost perfect agreement (κ > 0.80) according to the Landis and Koch benchmark.

Usage

Loading the Dataset

from datasets import load_dataset

# Load the full dataset
dataset = load_dataset("Faysal4200/BTMD")

# Access a sample
sample = dataset["train"][0]
print(sample["heading"])       # News headline
print(sample["label"])         # Binary label: Real or Fake
print(sample["multi_category"])# Fine-grained label
print(sample["image_path"])    # Path to image file (if available)
print(sample["video_path"])    # Path to video file (if available)

Loading with Images

from datasets import load_dataset
from huggingface_hub import hf_hub_download
from PIL import Image

dataset = load_dataset("Faysal4200/BTMD")

# Load image(s) for a sample
# Note: image_path may contain multiple semicolon-separated paths
sample = dataset["train"][0]
if sample["image_path"]:
    image_paths = sample["image_path"].split(";")
    for img_path in image_paths:
        img = Image.open(hf_hub_download(
            repo_id="Faysal4200/BTMD",
            filename=img_path.strip(),
            repo_type="dataset"
        ))
        img.show()

5-Fold Cross-Validation (as in the original paper)

from datasets import load_dataset
from sklearn.model_selection import StratifiedKFold

dataset = load_dataset("Faysal4200/BTMD", split="train")

skf = StratifiedKFold(n_splits=5, shuffle=True, random_state=42)
labels = dataset["label"]

for fold, (train_idx, val_idx) in enumerate(skf.split(range(len(dataset)), labels)):
    train_set = dataset.select(train_idx)
    val_set = dataset.select(val_idx)
    print(f"Fold {fold+1}: Train={len(train_set)}, Val={len(val_set)}")

Citation

If you use this dataset in your research, please cite:

@article{ahmmed2025mome,
  title     = {MoME-BanglaFake: A Mixture-of-Modality-Experts Framework for Bengali Fake News Detection Using a Novel Trimodal Dataset},
  author    = {Ahmmed, Faysal and Rafsan, Resadus Salehin and Akther, Airin and Mansib, Muhtadi and Esika, Ainea Esrat and Mridha, F. M.},
  journal   = {International Journal of Information Management Data Insights},
  year      = {2026},
  note      = {Paper under review. Citation details will be updated upon publication.}
}

Note: The full citation with volume, pages, and DOI will be updated once the paper is published in International Journal of Information Management Data Insights.

License

This dataset is released under the MIT License.

Contact

For questions or issues regarding this dataset, please contact:

• Faysal Ahmmed — 22-47069-1@student.aiub.edu or faysalahmmed4200@gmail.com
• F. M. Mridha — firoz.mridha@aiub.edu

Affiliation: Department of Computer Science, American International University-Bangladesh (AIUB), Dhaka-1229, Bangladesh.