Final Hate Speech Classification Project

This project builds a machine learning pipeline for detecting hate, offensive, and neutral speech in short-form text. It includes dataset processing, cleaning, balancing, TF-IDF vectorization, model training, and an example inference step.

---

📁 Project Structure

.
├── main.py
├── hate_speech.csv
├── cleaned_hate_dataset.csv        # (generated)
└── hate_speech_model.pkl           # (generated)

---

🔍 Features

• Cleans and normalizes tweet text
• Removes stopwords, URLs, mentions, emojis, and special chars
• Handles class imbalance via oversampling
• Trains a neural network classifier (MLP)
• Uses TF-IDF embeddings with n-grams
• Saves trained model + vectorizer as artifacts
• Contains inference demo for quick testing

---

🧠 Model Architecture

Algorithm: MLPClassifier (sklearn)

Key Params

• hidden_layer_sizes=(100,)
• activation="relu"
• solver="adam"
• max_iter=200
• early_stopping=True

---

🗂 Dataset Details

Classes:

• 0 → Hate
• 1 → Offensive
• 2 → Neutral

The script balances the dataset by oversampling under-represented labels before training.

---

🧹 Text Cleaning Pipeline

The clean_text() function applies:

• Lowercasing
• URL + username removal
• Hashtag stripping
• Punctuation removal
• Stopword filtering
• Token reconstruction

Example input:

"@user Check out this link! http://site.com #Wow"

Output:

"Neutral"

---

🚀 Running the Project

1. Install dependencies

pip install -r requirements.txt

(If NLTK stopwords aren't installed, script will download them.)

2. Execute training

python main.py

This will:

• Load dataset
• Clean + balance
• Train model
• Save artifacts
• Display metrics

---

📦 Output Artifacts

After training completes:

• cleaned_hate_dataset.csv
• hate_speech_model.pkl

The .pkl file stores:

{
  "vectorizer": TfidfVectorizer,
  "model": MLPClassifier
}

---

🧪 Example Prediction

The script ends by running inference on sample inputs such as:

"I hate you so much!"
"Check out my new blog post"
"Had a great time at the concert!"

Outputs include predicted class labels.

---

🧰 Requirements

Recommended environment:

• Python 3.8+
• scikit-learn
• imbalanced-learn
• pandas
• nltk

---

📬 Notes

• Designed for academic + research use
• Not production-hardened
• Requires dataset hate_speech.csv to exist locally

---