Detecting Misinformation, Bias & Hostile Rhetoric in Political Communication Using NLP and Deep Learning
This project analyzes and flags manipulative language in political discourse—spanning speeches, tweets, and media articles. Using advanced Natural Language Processing (NLP) and Transformer-based Deep Learning models, it classifies political statements for truthfulness, detects manipulative or hostile rhetoric, analyzes media bias, and extracts key rhetorical themes.
✨ Real-time, multi-angle detection pipeline revealing how and why political speech may be misleading, emotionally charged, or biased.
- 🔎 Fact-Check Claims: Classify political statements as true, false, or misleading.
- 💬 Rhetoric Detection: Identify manipulation tactics (e.g., emotional appeal, nationalism, repetition).
- 🧠 Hostile Speech Analysis: Detect personal attacks and hate speech in tweets.
- 📰 Media Bias Framing: Analyze partisan leanings and framing in political news.
- 🎙 Theme Extraction: Uncover core themes and ideological messaging in speeches.
🎙 SPEECH 1 - MANIPULATION ANALYSIS REPORT
📊 Overall Sentiment: 😡 Negative (-0.46)
⚠️ Fact-Check: Likely False Statement Detected
🎭 Detected Tactics:
✅ Repetition | ✅ Emotional Appeal | ✅ Personal Attacks | ✅ Nationalism | ✅ Misinformation
🧠 Verdict: The speech contains manipulative or misleading content with emotionally charged, divisive language.
| Dataset | Description | Use Case |
|---|---|---|
| LIAR | 12.8K fact-checked political statements | Fake news classification |
| Trump Speeches | 45+ official transcripts | Rhetorical + sentiment analysis |
| Trump Tweets | 5.6K tweets labeled for hostility | Hate speech / insult detection |
| All The News | 140K+ news articles (2016–2022) | Media bias & framing analysis |
- Named Entity Recognition (NER)
- Tokenization, Lemmatization
- Sentiment & Emotion Analysis
- TF-IDF & Hate Speech Lexicons
- Entity Co-occurrence Mapping
- Logistic Regression, SVM, XGBoost, Random Forest
- Topic Modeling: LDA for ideological themes and topics
- BERT / BERT-Large: Context-aware classification
- DistilBERT, RoBERTa: Light & interpretable models
- BiLSTM: Sequential pattern recognition
- Hybrid Architectures (e.g., BERT + BiLSTM)
| Task | Best Model | Score / Insight |
|---|---|---|
| ✅ Fake News Detection | Logistic Regression | 88.8% Accuracy |
| 🚨 Insult Detection | BERT + BiLSTM | 93.6% F1 Score |
| 📰 Media Bias Detection | Co-occurrence Analysis | Based on framing patterns (heuristic) |
| 🎙 Theme Extraction | LDA + Sentiment | Topics: Witch Hunt, Fake News, Elections |
| 💬 Deep Rhetoric Class. | DistilBERT (Trial) | 88.8% Accuracy |
Natural Language Understanding, Transformer Fine-Tuning, Text Classification, Sentiment Analysis, Rhetoric Detection, Topic Modeling, Explainable AI, Data Visualization, and Model Interpretation (SHAP).
- Multi-language political discourse analysis 🌍
- Real-time Twitter/News feed monitoring 🔁
- Dashboard with dynamic visualizations 📊