ClimateMemes

Code and dataset for the paper: What Media Frames Reveal About Stance: A Dataset and Study about Memes in Climate Change Discourse

This repository provides the CLIMATEMEMES dataset and code for evaluating multimodal models on stance and media-frame detection in climate-change memes.

Table of Contents

• Dataset

• Installation

• Code Overview

• Usage

• Citation

• License

Dataset

The CLIMATEMEMES dataset includes 1,184 climate-change memes collected from 47 subreddits (2016–2024). Each meme is annotated with:

• Stance: Convinced, Skeptical, or Neither

• Media Frame: Real, Hoax, Cause, Impact, Allocation, Propriety, Adequacy, Prospect

See paper for definitions.

Installation

This project uses Poetry for dependency management.

git clone https://github.com/mainlp/ClimateMemes.git
cd ClimateMemes
pip install poetry
poetry install

Code Overview

The code/ folder includes scripts for running different multimodal models:

LLaVA models

• llava.py — main script for LLaVA experiments

• llava_backbone.py — backbone architecture for LLaVA

• llava_embedding.py — embedding module for LLaVA

Molmo models

• molmo.py — main script for Molmo experiments

• molmo_backbone.py — backbone architecture for Molmo

• molmo_embedding.py — embedding module for Molmo

These scripts allow evaluation of model performance under different input configurations, including image-only, text-only, and combined inputs.

Ensure that the dataset directory structure is preserved when running the scripts.

Citation

If you use this repository, please cite the following paper:

@misc{zhou2025mediaframesrevealstance,
      title={What Media Frames Reveal About Stance: A Dataset and Study about Memes in Climate Change Discourse}, 
      author={Shijia Zhou and Siyao Peng and Simon Luebke and Jörg Haßler and Mario Haim and Saif M. Mohammad and Barbara Plank},
      year={2025},
      eprint={2505.16592},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2505.16592}, 

License

This project is licensed under the MIT License — see the LICENSE file for details.