Fake News Detection using Transformers

Misinformation spreads faster than fact-checks can follow, and readers have no reliable way to judge an unfamiliar article in the moment. Automated detection is genuinely hard — style alone does not prove falsehood — which is exactly why this project is built with unusual honesty about its limits. A RoBERTa classifier fine-tuned on the ISOT dataset (44,898 labeled real and fake news articles) learns the stylistic fingerprints of fabricated news, and a transparent signal layer scores interpretable factors: sensational vocabulary density, emotional manipulation markers, attribution presence (named sources vs "experts say"), and headline–body consistency. The demo never claims certainty — it presents a risk assessment with the evidence, like a spell-checker for credibility. The report documents where the model fails (well-written disinformation, satire) as carefully as where it succeeds.

1. Article text is cleaned and split into headline and body; the publishing source is extracted when available.
2. RoBERTa-base, fine-tuned on ISOT, outputs a fake-vs-real probability from the full article text.
3. In parallel, six rule-based signals are computed: sensational word density, emotion-lexicon hits, named-source attribution count, source-list match, headline–body semantic similarity, and capitalization patterns.
4. The classifier probability and signal scores are combined with documented weights into a final risk band: likely real, uncertain, or likely fake.
5. The demo renders the verdict card with per-signal bars and highlights the contributing phrases in the article text.
6. Evaluation on ISOT's held-out split reports accuracy, precision/recall and F1 honestly, with an error analysis of satire and well-written fakes in the report.

Tech stack:

• Python 3.11, Hugging Face Transformers
• RoBERTa-base (roberta-base) fine-tuned on ISOT
• ISOT Fake News Dataset (44,898 articles)
• PyTorch (fine-tuning)
• scikit-learn (metrics)
• spaCy (linguistic features)
• Flask + HTML/CSS/JS (demo)

Dataset & model details

• Dataset: ISOT Fake News Dataset (Ahmed et al., 2017) — 44,898 full-text articles: 21,417 real (Reuters) and 23,481 fake, collected 2016–2017. The standard public benchmark for fake-news text classification. Source: the authors' release (widely mirrored, e.g. Kaggle "Fake News").
• Task: Binary classification — input = news article text (+ source), output = real/fake probability plus six interpretable signal scores.
• Model: RoBERTa-base (125M parameters) fine-tuned on ISOT; signal layer uses lexicon and pattern features documented in the report.
• Metrics: Accuracy, precision, recall and F1 on ISOT's held-out split — reported as design targets from the training run, never pre-claimed.

• [RoBERTa classifier] RoBERTa-base fine-tuned on the ISOT Fake News dataset distinguishes real from fabricated articles by learned stylistic patterns.
• [Explainable signal layer] Six transparent signals — sensationalism, emotional manipulation, attribution quality, source credibility, headline–body consistency, all-caps density — each scored and shown.
• [Verdict card] Every analysis renders as a risk card: likely-real / uncertain / likely-fake with confidence and the top contributing signals.
• [Signal highlighting] The demo highlights the exact words and phrases driving each signal inside the article text.
• [Source credibility check] The publishing domain is checked against a documented list of known unreliable outlets and satire sites.
• [Uncertainty honesty] Borderline articles get an "uncertain — verify manually" verdict with fact-checking guidance instead of a forced classification.
• [Live web demo] Paste any article or analyze the preloaded real and fake samples; the full signal breakdown computes instantly.

• Fine-tuned RoBERTa fake-news classifier
• Signal-analysis pipeline source (6 documented signals)
• Source credibility reference list (documented)
• Web demo with preloaded real and fake sample articles
• Evaluation notebook (accuracy, F1, error analysis on satire)
• Project report PDF (background, transformers, misinformation detection limits, results)
• PPT presentation for final review
• Viva Q&A preparation document (RoBERTa, ISOT, precision/recall, adversarial limits)

• Style-based detection cannot verify facts — a well-written false article can pass, and the report's error analysis documents this openly.
• ISOT articles are from 2016–2017; news style and topics have shifted, so the demo includes a dataset-age discussion.
• Satire sites are flagged by the source list, not understood as humor — satire is an explicit known-failure case.
• F1 is a design target until the training run executes; the report records the achieved figure.
• The tool assists readers; it is not a content-moderation system and must not be used to auto-censor.