Research Assistant
- Organization
- University of Koblenz
- Location
- Koblenz, Germany
- Duration
- 6 mos
Highlights
- Weakly supervised fake news detection with Mean Teacher, VAT, and pseudo-labelling
- Pseudo-labelling reached 0.731 test accuracy against 0.687 for a supervised BiLSTM baseline
Research role focused on robust NLP methods for fake news classification under limited-label and noisy-data conditions.
Work and responsibilities
- Weakly Supervised Fake News Detection: Researched and implemented weakly supervised methods (Mean Teacher, VAT, Pseudo-Labelling) with noise strategies to improve robustness under limited-label conditions.
- Transformer-Based Analysis: Conducted an empirical study of transformer architectures, attention mechanisms, and seq-to-seq / encoder-decoder models for fake news classification.
Methods and tools
Methods
- Mean Teacher
- Virtual Adversarial Training
- Pseudo-labelling
- Noise strategies
NLP
- Transformers
- Attention mechanisms
- Seq-to-seq models
- Encoder-decoder architectures