📰 French Satire Detection with Machine Learning

This project tackles the challenge of automatically detecting satirical vs. informative news headlines in French. It leverages a pipeline of custom web scraping, data cleaning, and machine learning, culminating in a strong logistic regression baseline, with future extensions toward deep learning and LLMs.

---

🚀 Project Structure

• 1_data_collection.ipynb
  Scrapes headlines from 4 French news sources (2 satirical, 2 informative). Headlines are saved with source-based labels.

• 2_data_cleaning_and_merging.ipynb
  Applies cleaning rules to remove irrelevant, misleading, or malformed titles. Duplicates and non-French entries are also filtered out. Results are merged into a single dataset for modeling.

• 3_logistic_regression_baseline.ipynb
  Builds a TF-IDF + Logistic Regression baseline classifier. Includes detailed evaluation and interpretability (feature importance, confusion matrix, top tokens by class).

• models/
  Contains the serialized final pipeline (logistic_pipeline.joblib).

• data/
  Folder structure:

  • 1_raw/ contains unfiltered scraped data
  • 2_cleaned/ contains cleaned per-source datasets
  • cleaned_dataset.csv is the final ready-to-use dataset

---

📊 Current Dataset Summary

• Raw data: 10,164 headlines
• Cleaned data: 9,682 headlines
  Balanced between satirical (Le Gorafi, Nordpresse) and informative (Le Figaro, Libération)

---

🧠 Current Model Performance

A logistic regression model trained on TF-IDF features achieves the following:

• Accuracy: ~82%
• Precision/Recall (balanced) across satire/informative classes
• Clear interpretability of the most discriminating tokens

---

🛠️ TODOs & Next Steps

📌 Data Cleaning Improvements

• Outlier detection for overly long titles
  → visualize length distribution via mean + std / median + IQR
• Better token-level cleaning: lowercase, stopword removal, tokenization, lemmatization (for traditional NLP models)

🧼 Refactoring Ideas

• Refactor scrape() for modularity
  → isolate per-site logic (URL & parsing)
• Replace per-source CSVs in data_collection with raw .txt (1 title/line)
  → only convert to CSV after cleaning + labeling

🧪 Modeling Enhancements

• Train a Recurrent Neural Network (RNN/LSTM) with word embeddings
• Fine-tune a pre-trained French LLM (e.g. CamemBERT, FlauBERT)
  → Use HuggingFace transformers for tokenization + classification head

---

💡 Why This Project?

Satirical content is intentionally close to real news — making it both fascinating and challenging to detect. This project simulates a real-world scenario for text classification and is a great demonstration of:

• Applied NLP pre-processing
• Feature engineering & interpretability
• Multi-source data wrangling
• And soon, deep learning model fine-tuning

---

📎 Author

Created by Luca Matarelli — ML enthusiast with a passion for language and social context in AI.

---