Sentiment Analyzer — Neural Network

A neural network built from scratch in PyTorch that reads movie reviews and classifies them as positive or negative. Trained on 25,000 real IMDB reviews, achieving 85.08% accuracy on unseen data.

---

Project Overview

This project takes raw text as input, processes it through an embedding-based neural network, and outputs a sentiment prediction with a confidence score.

"This movie was absolutely incredible!" → POSITIVE 😊 (96.3% confidence)
"What a waste of time. Terrible plot."  → NEGATIVE 😞 (94.1% confidence)

---

How It Works

Text can't be fed directly into a neural network — computers only understand numbers. The model processes text through this pipeline:

Raw Text
   ↓
Tokenize         → ["this", "movie", "was", "great"]
   ↓
Vocabulary       → [11, 27, 8, 84]      (each word gets a unique ID)
   ↓
Embedding        → [[0.2, 0.8, ...], ...]  (each ID → 64-number meaning vector)
   ↓
Mean Pooling     → average all word vectors into one review vector
   ↓
Neural Network   → 0.96  →  POSITIVE

The embedding layer is the key insight — words with similar meanings (like "great" and "amazing") end up with similar vectors, which is how the model understands language.

---

Architecture

Layer	Details
Embedding	10,000 vocab × 64 dimensions = 640,000 parameters
Mean Pooling	Averages 200 word vectors → single 64-dim review vector
Linear (fc1)	64 → 128 hidden units
ReLU	Non-linear activation function
Dropout (0.3)	Randomly disables 30% of neurons to prevent overfitting
Linear (fc2)	128 → 1 output
Sigmoid	Squishes output to 0–1 (probability of positive sentiment)
Total Parameters	648,449

---

Training Results

Trained for 5 epochs on 25,000 IMDB reviews with a batch size of 64.

Epoch	Train Loss	Train Acc	Test Loss	Test Acc
1	0.5748	70.02%	0.4595	79.02%
2	0.3744	83.75%	0.3846	83.11%
3	0.3024	87.54%	0.3609	84.46%
4	0.2597	89.73%	0.3536	85.03%
5	0.2299	91.20%	0.3593	85.08%

Note: The gap between train (91.2%) and test (85.08%) accuracy is called overfitting — the model begins memorizing training reviews rather than purely learning patterns. The dropout layer slows this down but doesn't eliminate it entirely.

---

Project Structure

Sentiment-Analyzer/
├── sentiment.py        # Main model, training, and prediction code
└── README.md

---

How to Run

# Create a virtual environment
python3 -m venv .venv
source .venv/bin/activate

# Install dependencies
pip install torch datasets numpy matplotlib

# Run the model
python3 sentiment.py

---

Tech Stack

• Python 3.12
• PyTorch — neural network definition and training
• Hugging Face datasets — IMDB dataset loading
• NumPy — numerical operations
• Matplotlib — training visualization

---

Example Predictions

Review: This movie was absolutely incredible. The acting was superb...
Result: POSITIVE 😊 (confidence: ~96%)

Review: What a waste of time. The plot made no sense and the acting was terrible...
Result: NEGATIVE 😞 (confidence: ~94%)

Review: It was okay I guess, nothing special but not horrible either...
Result: (near 50% — ambiguous, as expected)

---

Concepts Covered

• Text tokenization and vocabulary building
• Word embeddings
• Mean pooling
• Binary cross-entropy loss
• Backpropagation and gradient descent
• Dropout regularization
• Overfitting vs generalization

---

Next Steps

• LSTM / GRU — recurrent layers that process word order, not just averages
• Pre-trained embeddings — use GloVe or Word2Vec instead of training from scratch
• Transformer (BERT) — fine-tune a pre-trained model for 92%+ accuracy
• Larger vocabulary — increase from 10,000 to 50,000 words