HinglishMemeX / README.md

pushkarsharma

Update README.md

5737f9d verified 5 months ago

preview code

raw

history blame contribute delete

4.06 kB

HinglishMemeX

A Code-Mixed Multimodal Dataset for Misinformation and Satire Detection in Indian Memes

Overview

HinglishMemeX is a curated multimodal dataset consisting of 1,370 Indian social-media memes that combine images + Hinglish (Hindi+English code-mixed) text. Each meme is paired with:

OCR‑extracted Hinglish text
A distilled English factual claim
A supporting evidence URL (from a trusted fact-checking source)
A veracity label: real, fake, satire, or partially_true

This dataset is designed for misinformation detection, satire identification, multimodal classification, and retrieval‑augmented fact verification.

Dataset Structure

HinglishMemeX/
  ├── images/
  │     ├── 000001.jpg
  │     ├── 000002.jpg
  │     └── ...
  ├── metadata.csv
  ├── README.md

Each metadata entry contains:

id: unique ID for the meme
image: path or URL to the meme image
ocr_text_hinglish: OCR text extracted from the meme (Hinglish)
claim_en: distilled factual English claim
evidence_url: link to fact-checking source
label: one of real, fake, satire, partially_true
source: origin (AltNews, Factly, BOOMLive, satire pages, etc.)
split: train / validation / test

Tasks Supported

Multimodal misinformation detection (4-way classification)
Satire detection
Claim verification via external evidence
OCR-based text understanding in code-mixed Hinglish
Retrieval-Augmented Generation (RAG) over evidence URLs

Dataset Statistics

Total memes: 1,370
Classes: Real, Fake, Satire, Partially True
Splits:
- Train: ~70%
- Validation: ~10%
- Test: ~20%

Data Collection & Curation

Images collected from public fact-checking portals (AltNews, BOOMLive, Factly) and popular social media satire pages.
Hinglish text was extracted using EasyOCR and Google Vision API, followed by light manual correction.
Claims were distilled into short English factual statements.
Evidence URLs were added for transparency and retrieval.
Double annotator labeling with adjudication for disagreements.

Benchmarks & Baselines

Baseline experiments were conducted using:

CLIP ViT-L/14 (vision-only)
CLIP + IndicBERT (late fusion)
Cross-attention dual encoders (deep fusion)

Evaluation metrics: macro-F1, accuracy, per-class F1.

Satire and partially-true memes are particularly challenging due to semantic overlap.

Ethical Considerations

Contains politically sensitive content; models trained on this dataset may inherit biases.
Some memes may include misinformation or sensitive themes—handle responsibly.
Recommended to use the dataset for research only.
Provide confidence scores and retrieved evidence when deploying models.

⚠️ Limitations

Focused on Indian context → may not generalize globally.
Natural OCR errors remain in some samples.
Subjective boundaries between satire and partially-true content.

Loading the Dataset (Hugging Face)

from datasets import load_dataset

dataset = load_dataset("pushkarsharma/HinglishMemeX")

def add_path(example):
    example["image"] = f"images/{example['id']}.jpg"
    return example

dataset = dataset.map(add_path)

License

Please refer to the LICENSE file for dataset licensing details. Choose a license such as CC BY 4.0 or CC BY-SA 4.0 depending on your redistribution permissions.

Citation

If you use HinglishMemeX, please cite:

@misc{hinglishmemex2025,
  title = {HinglishMemeX: A Code-Mixed Multimodal Dataset for Misinformation and Satire Detection in Indian Memes},
  author = {Sharma, Pushkar},
  year = {2025},
  institution = {Indian Institute of Technology Patna}
}

Contact

Maintainer: Pushkar Sharma
Email: pushkarrokhel@gmail.com

Thank you for using HinglishMemeX!

Datasets:

pushkarsharma
/

HinglishMemeX