Fake News Detection using BERT

Fabricated news now travels faster than any fact-checking desk can respond, yet most readers have no tool that gives a fast, explainable first read on a suspicious claim. Manual verification does not scale, and simple keyword filters are trivially gamed. Transformer language models changed this equation: BERT's pre-trained contextual representations let a classifier learn subtle linguistic cues of deception — hedging, sensationalism, unverifiable superlatives — from a few thousand labeled examples. This project turns that capability into a complete student build: bert-base-uncased fine-tuned on the LIAR benchmark of real PolitiFact statements, wrapped in a demo that scores any pasted claim and highlights the exact words that drove the verdict. Because the verdict is inspectable, the report and viva can discuss why a statement was flagged — and where the model is known to fail, such as satire and breaking-news topics.

1. LIAR statements are WordPiece-tokenized (max 128 tokens); the six truthfulness grades are kept for training and collapsed to a binary fake/real verdict at inference.
2. Speaker metadata (party, job, credit history counts) is encoded and concatenated with the BERT CLS embedding in a hybrid classification head.
3. bert-base-uncased is fine-tuned for about 5 epochs with AdamW; every epoch logs accuracy, precision, recall, F1 and the confusion matrix.
4. At inference, the demo app tokenizes the pasted statement, runs it through the fine-tuned model and renders the verdict with a confidence bar.
5. Gradient-based token influence scores color each word by how strongly it pushed the decision toward fake or real.
6. Batch mode reads a CSV of statements and writes per-statement verdicts and confidences to a report file.

Tech stack:

• Python 3.10, PyTorch, Hugging Face transformers (bert-base-uncased)
• scikit-learn (metrics, confusion matrix), NumPy, pandas
• Matplotlib (training curves, label distribution)
• Single-file HTML/CSS/JS demo app (analyzer, dataset explorer, training views)
• LIAR benchmark dataset (Wang, ACL 2017)
• Trained weights exported from the included fine-tuning run

• [Statement analyzer] Paste any claim and get a fake/real verdict with a confidence bar; token-influence highlighting colors the words that most pushed the decision.
• [BERT fine-tuning pipeline] bert-base-uncased with a dropout + linear classification head, fine-tuned with AdamW and a linear warmup schedule; all hyperparameters exposed in the notebook.
• [Hybrid text + metadata head] Speaker metadata from LIAR (party, job title, credit history) feeds an auxiliary input concatenated with the CLS embedding, following the paper's hybrid finding.
• [Dataset explorer view] Browse the LIAR label distribution (six truthfulness grades, train/valid/test splits) inside the demo app.
• [Training-curve view] Per-epoch F1 and train/validation loss plots that the notebook reproduces during the build.
• [Batch scoring mode] Score a CSV of statements into a verdict report for larger-scale analysis.
• [Full evaluation logging] Accuracy, precision, recall, F1 and the confusion matrix computed on the official LIAR test split — never pre-claimed.

• Complete source code (preprocessing, fine-tuning, hybrid head, inference, demo app)
• Jupyter training and evaluation notebook (buyer-run procedure: fine-tune, evaluate, inspect errors)
• Project report PDF (background, LIAR analysis, methodology, evaluation, error analysis)
• PPT presentation for final review
• Viva Q&A preparation document (transformers, attention, fine-tuning, LIAR, evaluation metrics)
• Setup guide (environment, dataset download, training, running the demo)

• The model classifies short political statements in English — the LIAR domain. Satire, breaking-news topics, non-English text and multimedia claims are outside its training distribution and it handles them poorly.
• A verdict is a statistical prediction, not a fact-check: low-confidence outputs must be read as model uncertainty and always passed to human review.
• Emotionally charged but true statements can be misclassified; the confusion matrix in the report documents the error patterns openly.
• Fine-tuning BERT needs a GPU (or long CPU patience) — the notebook documents expected runtimes and a smaller-sample quick-run mode.
• Speaker metadata helps only when present; pure-text mode drops the auxiliary input and scores slightly lower.