🏛️ 🚀 Deconstructing Fake News Narratives to Enhance Media Literacy 🚀 🏛️

A Generative AI–powered tool that leverages Natural Language Processing(NLP), Sentiment Analysis, and pretrained large language models(LLMs) for a variety of use cases to detect potential misinformation in news articles, explain why they may be risky (via disinformation tactic detection), and help users make better-informed decisions—improving media literacy.

📄 Table of Contents

• 🎯 Project Overview
• ✨ Key Features
• 🏗️ Architecture & Workflow
• 📊 Data & Model Training
• 🛠️ Core Components
  • 🧠 1. Fake vs. Real News Classifier
  • 🔍 2. Vector Search & Similarity Analysis
  • 🚨 3. Disinformation Tactic Identification
  • ✅ 4. Fact-Checking Integration
  • 🤖 5. User Explanation Layer (Generative AI)
• 🛡️ Tech Stack
• ⚙️ Installation & Setup
• 📈 Evaluation & Metrics

---

🎯 Project Overview

This repository implements a GenAI tool for:

• 🚩 Flagging potential misinformation in news articles.
• 🧩 Explaining risk factors via detection of disinformation tactics.
• 🔗 Augmenting user judgment through similarity comparisons with known fake/real articles and fact-check results.
• ✍️ Synthesizing findings into a concise, human-readable summary using a generative model.

The ultimate goal is to empower users to recognize misleading narratives and improve their media literacy in an interactive, understandable way.

---

✨ Key Features

• 🔎 Supervised Fake/Real News Classification
  • Fine-tuned transformer-based classifier on labeled dataset (titles & texts).
• ⚡ Retrieval-Augmented Similarity Search
  • Vector embeddings stored in MongoDB Atlas Vector Search for semantic similarity against known fake/real articles.
• 🚨 Disinformation Tactic Detection
  • Rule-based (regex, heuristics) and NLP-based methods (syntax/entity analysis, sentiment) to identify sensationalism, vagueness, appeals to emotion, loaded language, conspiracy markers, etc.
• ✅ Fact-Checking API Integration
  • Query Google Fact Check Tools API for provided claims or URLs to fetch existing verifications.
• 🤖 Generative Explanation Layer
  • Uses a distilled Llama-based model (DeepSeek-R1-Distill-Llama-8B-GGUF) to combine classifier verdict, similarity insights, tactic detections, and fact-check findings into a clear user-facing summary.
• ☁️ Scalable Training & Deployment
  • Trained on Google Vertex AI (textembedding-005 for embeddings, Vertex AI Training for fine-tuning).
• 🗄️ Caching & Vector DB
  • MongoDB Atlas for vector storage and caching embeddings without storage limits.
• 🌐 Interactive Interface
  • Simple web UI for users to input an article title, key claim, or URL and receive analysis.

---

🏗️ Architecture & Workflow

1. 📝 User Input
   • Accepts: article text, title, key claim, or URL.
2. 🧹 Preprocessing
   • Clean text (remove special characters, lowercase), tokenize, remove stopwords.
   • Generate sentence/document embeddings via Vertex AI textembedding-005 API.
3. 🧠 Fake/Real Classifier
   • Fine-tuned transformer (e.g., BERT-mini or similar) on embeddings + text features.
   • Outputs label (“Fake” / “Real”) with confidence score.
4. 🔍 Vector Similarity Search
   • Query MongoDB Atlas Vector Search with the new embedding.
   • Retrieve top-K most similar known Fake articles and top-K Real articles.
   • Compute similarity statistics (e.g., “80% of top-10 neighbors are Fake with avg sim = 0.78”).
5. 🚨 Disinformation Tactic Identification
   • Rule-based: regex patterns (ALL CAPS, clickbait phrases).
   • Lexicon-based: emotional/conspiracy term lists.
   • NLP-based: Google Natural Language API for syntax, entity, sentiment analysis.
6. ✅ Fact-Checking Integration
   • Query Google Fact Check Tools API with user-provided claim(s) or URL.
   • Fetch matches: claim, rating (False, Misleading, True), publisher, date.
7. 🤖 Generative Explanation
   • Collate:
     • Classifier result + confidence.
     • Similarity summary (top neighbors breakdown).
     • Detected tactics list with notes.
     • Fact-check findings.
   • Prompt distilled generative model (DeepSeek-R1-Distill-Llama-8B-GGUF) with structured input.
   • Produce a clear summary: reasons to be cautious/confident, what to watch for, suggestions for further verification.
8. 📤 Output to User
   • Display label, confidence score, similarity insights, tactic detections, fact-check results, and the generative summary in a user-friendly format.

---

📊 Data & Model Training

• 📥 Dataset: KaggleHub datasets containing labeled Real vs. Fake news articles (~95K samples).
  • Fields: title, text, label (0 = Fake, 1 = Real), optionally metadata (date, source).
• 🧹 Preprocessing:
  • Remove HTML tags, special characters; lowercase; tokenize; remove stopwords.
  • Optionally filter very short or irrelevant articles.
• 📈 Embeddings:
  • Use Vertex AI textembedding-005 API to generate vector representations for text/title.
  • Store embeddings in MongoDB Atlas for both training analysis and inference similarity search.
• ⚙️ Training Setup:
  • Split dataset: train/validation/test (e.g., 70/15/15).
  • Fine-tuned a transformer (e.g. BERT-mini) on text tokens or embeddings.
  • Track metrics: accuracy, precision, recall, F1 on validation/test.
• 🔧 Hyperparameters & Infrastructure:
  • Vertex AI Training for scalable GPU training.
  • Early stopping based on validation loss.
  • Save best checkpoints; export model for serving.
• 📊 Results (Example Placeholder):
  • Achieved ~84% accuracy on held-out test set.

---

🛠️ Core Components

🧠 1. Fake vs. Real News Classifier

• Model: Fine-tuned transformer-based classifier (BERT-mini).
• Inputs: Title & body text (optionally combined features).
• Output: Label (“Fake” / “Real”) + confidence score.
• Implementation Notes:
  • Use Hugging Face Transformers, PyTorch or TensorFlow.
  • Save artifacts (tokenizer, checkpoint) for inference.
  • Write test cases to verify inference outputs.

🔍 2. Vector Search & Similarity Analysis

• Embedding Generation: Vertex AI textembedding-005 API.
• Vector DB: MongoDB Atlas Vector Search.
• Workflow:
  1. On training: generate embeddings for all labeled articles; store in collection with metadata (label, title, snippet).
  2. On inference: generate embedding for input; query top-K similar vectors.
  3. Retrieve similarity scores & labels of neighbors; compute summary (e.g., “Top-5: 4 Fake (avg sim 0.82), 1 Real (avg sim 0.65)”).
• Benefits:
  • Contextualizes classifier output.
  • Helps when classifier confidence is low.

🚨 3. Disinformation Tactic Identification

• Rule-based Patterns:
  • ALL CAPS titles 🔠, excessive punctuation (!!!, ???).
  • Clickbait phrases (“You won’t believe”, “Shocking truth”).
  • Vague sourcing (“Experts say” without naming).
• Lexicon-based:
  • Emotional word lists (anger, fear, excitement).
  • Conspiracy markers (“hidden truth”, “they don’t want you to know”).
• NLP-based (Google Natural Language API):
  • Syntax analysis: identify manipulative sentence structures.
  • Entity analysis: detect vague vs. credible source mentions.
  • Sentiment analysis: high emotion without facts.
• Output:
  • List of detected tactics, each with a brief explanation.
  • Example:
    • 🔥 Sensationalism: Title in ALL CAPS, emotionally charged wording.
    • ❓ Vague Sourcing: Uses “experts say” without naming a source.

✅ 4. Fact-Checking Integration

• API: Google Fact Check Tools API.
• Workflow:
  • Extract key claim(s) or URL from user input.
  • Query API to get matched fact-check entries.
  • Parse results: claim text, rating (False, Misleading, True), publisher, date.
• Output:
  • Summary list/table of fact-check findings:
    • 📰 Claim: “[quoted claim]”
      • 🔖 Rating: False
      • 🏢 Publisher: Snopes
      • 📅 Date: Jan 15, 2024
• Benefit:
  • Provides authoritative verification or debunking references.

🤖 5. User Explanation Layer (Generative AI)

• Model: DeepSeek-R1-Distill-Llama-8B-GGUF (distilled/quantized).
• Prompt Design:
  • Structure input with labeled sections:

    News Title: [title]
    Classifier Prediction: [Fake/Real] (confidence: X%)
    Similarity Analysis: [“Top 5 neighbors: 4 Fake (avg sim 0.82), 1 Real (avg sim 0.65)”]
    Detected Tactics: [List with notes]
    Fact-Check Results: [List of claims & ratings]
    

  • Example prompt:

    “Analyze the following news assessment and provide a concise summary for the user, highlighting key reasons for concern or confidence, and advising what to look out for:

    News Title: ...
    Classifier Prediction: ...
    Similar Articles Found:
      Label: Fake, Similarity: ...
      Label: Real, Similarity: ...
    Fact-Check API Results:
      Claim: ..., Rating: ..., Publisher: ...
    Detected Disinformation Tactics: ...
    ```”
    

• Output:
  • A friendly summary paragraph or bullet points:
    • Why the article may be unreliable or likely genuine.
    • Which tactics reduce credibility.
    • How similarity to known Fake/Real informs trust.
    • Relevant fact-check outcomes.
    • Suggestions for further verification (e.g., check original source, corroborate with reputable outlets).
• Deployment:
  • Local inference runtime (e.g., llama.cpp) or lightweight server.
  • Cache prompt templates; optimize for latency.

---

🛡️ Tech Stack

• Languages & Frameworks:
  • Python, PyTorch/TensorFlow, Hugging Face Transformers.
  • FastAPI for backend server deployment.
• Cloud & APIs:
  • Google Vertex AI: text embeddings, model training, endpoints.
  • Google Natural Language API: syntax, sentiment, entity analysis.
  • Google Fact Check Tools API.
• Database:
  • MongoDB Atlas with Vector Search for embedding storage & similarity queries.
• Generative Model Inference:
  • Local inference of DeepSeek-R1-Distill-Llama-8B-GGUF via llama.cpp.
• DevOps & Deployment:
  • Docker: containerize API, model servers.
  • Kubernetes or Google Cloud Run for scalable deployment (optional).
  • CI/CD: GitHub Actions for linting, testing, deployment pipelines.
• Utilities:
  • Python scripts for preprocessing, embedding generation, batch inference.
  • Logging, monitoring (inference latency, error rates).
• Frontend:
  • Simple React or Streamlit interface for user input and results display.

---

⚙️ Installation & Setup

🔄 Clone the Repository

git clone https://github.com/JiaJun-dotcom/deconstruct-fake-news.git
cd deconstruct-fake-news