Hate Speech Classification — Model Comparison

Comparative study of 8 classifiers on a hate-speech dataset — from classical ML baselines to sequence models and transformer-based sentiment.

Accompanied by a small Flask demo for interactive prediction.

Stack: Python · scikit-learn · TensorFlow / Keras · Hugging Face · Flask · Jupyter

---

Models compared

#	Model	Notebook
1	Logistic Regression	LogisticRegression.ipynb
2	Multinomial Naive Bayes	MultinomialNaiveBayes.ipynb
3	Support Vector Machine	SVM.ipynb
4	Random Forest	RandomForest.ipynb
5	Gradient Boosting	GradientBoosting.ipynb
6	XGBoost	XGBoost.ipynb
7	LSTM (deep sequence model)	LSTM.ipynb
8	Transformer + VADER (sentiment-augmented)	Transformer(VADER).ipynb

Full write-up: Report.pdf

---

Results

Add assets/results.png — a matplotlib bar chart of accuracy / F1 for all 8 models, generated from your notebooks.

Fill this table with the numbers from your notebooks / Report.pdf:

Model	Accuracy	Precision	Recall	F1
Logistic Regression	–	–	–	–
Multinomial NB	–	–	–	–
SVM	–	–	–	–
Random Forest	–	–	–	–
Gradient Boosting	–	–	–	–
XGBoost	–	–	–	–
LSTM	–	–	–	–
Transformer + VADER	–	–	–	–

Winner: (fill in — which model performed best, and why you think so)

---

Dataset

• File: hate.csv
• Source: (add original source — Kaggle URL or paper reference)
• Size: (rows)
• Class balance: (class distribution — hate vs. non-hate)

Preprocessing pipeline

Documented in DataCleaning and EDA.ipynb:

1. Lowercase + strip URLs / mentions / hashtags
2. Tokenize
3. Remove stopwords
4. Lemmatize
5. Vectorize
   • Classical models → TF-IDF
   • LSTM → learned embeddings
   • Transformer → pretrained embeddings + VADER sentiment features

Train/test split

See Train-Test Split.ipynb — stratified 80/20 split with a fixed random seed for reproducibility across all notebooks.

---

Setup

git clone https://github.com/prayas17/hate-speech-classification.git
cd hate-speech-classification
python -m venv .venv
# Windows
.venv\Scripts\activate
# macOS / Linux
source .venv/bin/activate

pip install -r requirements.txt
jupyter lab

Run notebooks in this order:

1. DataCleaning and EDA.ipynb
2. Train-Test Split.ipynb
3. Any model notebook (they're independent from here on)

Flask demo

A small web app that takes text input and returns a prediction using the best-performing model.

cd Flask
python app.py
# → open http://localhost:5000

---

Repo layout

.
├── DataCleaning and EDA.ipynb
├── Train-Test Split.ipynb
├── LogisticRegression.ipynb
├── MultinomialNaiveBayes.ipynb
├── SVM.ipynb
├── RandomForest.ipynb
├── GradientBoosting.ipynb
├── XGBoost.ipynb
├── LSTM.ipynb
├── Transformer(VADER).ipynb
├── hate.csv
├── Report.pdf
├── requirements.txt
├── Flask/                # Flask demo app
└── assets/
    └── results.png       # comparison chart

License

MIT