BKSBhupender Kumar Saini
Research focusCV

Sep 2020 – Feb 2021/NLP Research

Weakly Supervised Fake News Detection

  • NLP Research
  • University of Koblenz and Landau
Partner
University of Koblenz and Landau
Period
Sep 2020 – Feb 2021
Focus
NLP Research

Key outcomePseudo-labelling reached 0.731 ± 0.01 test accuracy, outperforming a supervised BiLSTM baseline at 0.687 ± 0.02

Paper: False Article Detection with Weakly Supervised LearningGitHub repository: fake_news_detection

Context

Fake news classification often suffers from limited high-quality labeled data, while large amounts of unlabeled news content remain available across political, health, and general news domains.

Inputs

  • FakeNewsNet and FakeHealth labeled datasets
  • NELA unlabeled news samples
  • Article text with true, fake, and unlabeled labels

Approach

  • Implemented weakly supervised learning pipelines for fake news detection using Mean Teacher, Virtual Adversarial Training, and Pseudo-Labelling
  • Built preprocessing workflows, experiment folds, training scripts, and evaluation reports for comparing semi-supervised methods
  • Explored noise strategies such as synonym noise and dropout to improve model robustness under limited-label settings

Outcomes

  • Pseudo-labelling reached 0.731 ± 0.01 test accuracy, outperforming a supervised BiLSTM baseline at 0.687 ± 0.02
  • Comparative evaluation of Mean Teacher, VAT, and Pseudo-Labelling using precision, recall, F1-score, and loss analysis
  • Foundation for later master thesis work on semi-supervised transformer training for text classification
  • A model can produce confident predictions even when its learning signal is incomplete or noisy; reliable AI requires attention to how the data was created and which uncertainties remain hidden

Related on the shelf