A multimodal Bangla meme dataset for hate speech, sentiment, and sarcasm analysis

This repository contains a curated multimodal Bangla meme dataset designed for research on hate speech detection, affective analysis, and pragmatic language understanding in low-resource settings. The dataset comprises 5,126 publicly available Bangla memes, each consisting of an image and embedded Bangla text commonly found in online social media discourse.

The memes were manually collected from multiple publicly accessible social media platforms, including Facebook, Instagram, Reddit, and similar online sources. All images were processed using the Python Imaging Library (Pillow), and embedded Bangla text was extracted using EasyOCR. The extracted text was manually reviewed for accuracy and translated into English to facilitate cross-lingual and multilingual research.

Each meme was independently annotated by three native Bangla-speaking annotators for five binary attributes: sarcasm, humor, vulgarity, violence, and hate speech. A majority-voting strategy was applied, where a label was assigned as positive if at least two annotators agreed. This annotation process ensures label consistency while acknowledging the subjective nature of meme interpretation.

In addition to meme-level annotations, the dataset includes a curated Bangla hate-related lexicon. Hate-associated terms were extracted from the annotated text and refined through manual inspection. The final lexicon contains 441 unique entries, categorized into four semantic classes: vulgar, insult, violent, and hate-related words. This resource enables lexicon-assisted and hybrid approaches to hate speech detection.

The dataset is organized into image files, structured CSV annotation files, and lexicon resources, enabling seamless use in multimodal machine learning experiments. It supports research tasks such as multimodal hate speech detection, sarcasm and humor analysis, affective content modeling, cross-lingual learning, and low-resource Bangla language processing.