PolyMeme: Fine-Grained Internet Meme Sensing.

Area of Science:

• Computer Science
• Social Media Analysis
• Artificial Intelligence

Background:

• Internet memes are a multimodal digital content format popular on social media.
• Automatic meme detection is crucial for tracking trends and harmful content spread.
• Existing datasets lack diversity in meme formats, styles, and content.

Purpose of the Study:

• To introduce the PolyMeme dataset, a diverse collection of approximately 27,000 memes across four categories.
• To address the limitations of existing datasets in capturing meme variety.
• To develop and evaluate deep learning models for accurate meme detection.

Main Methods:

• Collected approximately 27,000 memes from Reddit, categorizing them.
• Manually labeled a portion of the dataset for training and validation.
• Trained deep learning networks (ResNet, ViT) using the PolyMeme dataset and other image datasets for meme detection.

Main Results:

• Deep learning models trained on PolyMeme achieved an estimated error rate of 7.35% for classification.
• Meme detection models demonstrated high accuracy, reaching 98% on the test set.
• The inclusion of regular images with text did not significantly improve meme detection performance.

Conclusions:

• The PolyMeme dataset enhances meme detection capabilities by accounting for diverse meme formats.
• Accurate automatic meme detection is feasible with advanced deep learning techniques and comprehensive datasets.
• This work contributes to better understanding and mitigating the spread of online misinformation and harmful content through memes.