Tracking the Criminal of Fake News Based on a Unified Embedding
summary
Proposes a unified embedding that maps texts to points on a unit sphere and trains with a triplet loss so same-author texts sit closer than others, and adds dynamic triplet selection to speed learning. Implements on a Chinese dataset collected from eight sites — claimed about 130 articles from 3600 authors — and reports higher attribution, verification and clustering scores versus n-gram/CNN and SVM baselines, with gains growing as author count increases. Runs on Tesla P40, TensorFlow/Keras.
AI-written summary · about
chapters (11)
- 0:00Start
- 0:21Introduction
- 5:10Potential Victims
- 7:44Writing Styles
- 13:24Triplet Loss Function
- 16:35Selection Method
- 17:37Dynamics Selection
- 18:31Experiment Procedures and Results
- 19:12Experiment Environment
- 20:38Authorship Verification
- 21:50Summary
related talks
45mDeep Learning for Natural Language Processing
38mVideo Killed the Text Star: OSINT Approach
26mMaximizing Multimodal: Exploring the search frontier of text-to-image models
34mApplying Machine Learning Techniques to Detect and Understand Fake News
22mMaking Natural Language Processing Robust to Sociolinguistic Variation- 34mApplying Machine Learning Techniques to Detect and Understand Fake News
28m"Embeddings: How Computers Learned to Read"
45mApplying the four step "Embed, Encode, Attend, Predict" framework to predict document similarity