Social Media Dataset

Dataset Overview

• Data Source: Crawled over 1 million Twitter posts.
• Data Filtering:
  • Used a pretrained VGG19 model to filter out non‑emoji images; 95% of data were initially screened.
  • Manual filtering retained text‑image multimodal data, removing approximately 40% of the data.
• Data Processing:
  • Extracted emojis from text using regular expressions.
  • Obtained embedded text from emoji packs using the PaddleOCR platform and manual correction.
• Data Modalities: Includes four modalities.

Dataset Status

• The dataset will be open‑sourced after the paper is accepted.