Multimodal Classification of Moderated Online Pro-Eating Disorder Content
summary
A deep learning pipeline labels moderated pro‑eating‑disorder content on Tumblr by combining image features from AlexNet and word2vec tag embeddings, a human‑annotated seed of 6,000 posts, and training on a crawl of 877,000 Tumblr photo posts where only 569 removed entries are verifiable; a multimodal deep neural network outperforms text‑only and image‑only SVMs and achieves high recall (reported ~85%), showing a viable first‑pass and online‑learning tool for moderation though limited by small ground truth and compute tradeoffs.
AI-written summary · about
chapters (18)
- 0:00Introduction
- 0:31Social Media
- 0:54Deviant Content
- 2:05Deviant Health
- 3:05Why Deviant Health
- 4:27Contributions
- 5:34Methodology
- 6:26Data
- 7:34Human Reading Cast
- 7:56Rating Task
- 8:53Multimodal Features
- 10:11Data Breakdown
- 10:55Results
- 11:58Conclusion
- 12:17First Pass
- 12:29Online Learning
- 13:53Questions
- 14:23Tradeoff
related talks
32mAnalyzing social media for suicide risk using natural language processing
33mMachine Learning to moderate ads
09mSocial Media Image Analysis for Public Health
32mDecoding Emotions from Art to Memes
37mA Practical Guide to Robust Multi modal Machine Learning in Education
23mLeveraging Generative AI and Transformers in Healthcare
26mMaximizing Multimodal: Exploring the search frontier of text-to-image models
26mModern NLP for Proactive Harmful Content Moderation