Hateful Memes Benchmark for Multimodal Detection

Facebook AI created this benchmark dataset to measure progress in multimodal reasoning for hate speech detection. The dataset pairs images with text to form memes, each labeled for hateful content. It was published in 2020 and last updated on the platform in December 2022.

Use Cases

• Train a multimodal classifier to predict the 'hateful' label from combined image and text meme features.
• Benchmark model performance on the challenge task of detecting hate speech where meaning depends on both visual and textual context.
• Analyze the relationship between specific visual elements and textual phrases that constitute hateful content.
• Fine-tune vision-language transformers like CLIP or ViLT using the provided image-text pairs and binary labels.

Strengths

• Dataset was created and published by Facebook AI researchers to drive progress in a defined challenge.
• Focuses on a complex, real-world task requiring joint reasoning over image and text modalities.

Limitations

• Exact size, row count, and specific column names are not provided in the available metadata.
• The dataset's geographic and demographic bias is unknown, which could limit model generalizability.
• Data is from 2020 and may not reflect current meme trends or evolving hate speech patterns.

Provenance

Source

Facebook AI (Meta).

Collection Method

Created for the Hateful Memes Challenge benchmark; specific collection methodology is detailed in the associated research paper.

Time Range

null

Freshness

Last updated on the platform in 2022, but the underlying data is from the 2020 challenge.

Geography

Region tag suggests 'us', but specific spatial coverage is not detailed.