🎭 Sentiment Analysis Dashboard

An AI-powered sentiment analysis dashboard built with DistilBERT fine-tuned on Amazon Product Reviews. Achieves 90%+ accuracy — a +22% improvement over the TextBlob baseline.

---

📋 Table of Contents

• Overview
• Key Features
• Model Performance
• Architecture
• Installation
• Usage
• Project Structure
• Dataset
• Model Training
• Results
• Technologies
• Disclaimer
• Future Improvements
• Contributing
• License
• Contact

---

🎯 Overview

This project builds an end-to-end NLP pipeline for sentiment analysis of product reviews. Using DistilBERT (a lighter, faster version of BERT), the model classifies reviews as Positive or Negative with high accuracy while also providing aspect-based sentiment analysis to identify sentiment toward specific product attributes.

Use Cases:

• E-commerce review analysis
• Brand monitoring
• Customer feedback processing
• Product quality assessment
• Business intelligence

---

✨ Key Features

🤖 High-Accuracy NLP Model

• 90%+ accuracy using fine-tuned DistilBERT
• Trained on 10,000 Amazon product reviews
• Optimized classification threshold
• Fast inference (~100ms per review)

🔍 Aspect-Based Analysis

• Detects sentiment for 6 product aspects:
  • 🏗️ Quality
  • 💰 Price
  • 🚚 Shipping
  • 👥 Customer Service
  • ⚡ Performance
  • 🎨 Design
• Visual radar chart for aspect scores
• Aspect mention detection

📁 Batch Processing

• Upload CSV files with multiple reviews
• Process hundreds of reviews at once
• Export results to CSV
• Interactive visualizations

💻 Interactive Dashboard

• Real-time sentiment prediction
• Gauge chart for probability scores
• Confidence metrics
• Demo reviews included

---

📊 Model Performance

Comparison Table

Model	Accuracy	F1 Score	Type
TextBlob	68.00%	0.68	Rule-based
DistilBERT	90%+	0.90+	Fine-tuned Transformer
Improvement	+22%+	+0.22+	-

Test Set Metrics

Metric	Score
Accuracy	90%+
Precision	90%+
Recall	90%+
F1-Score	90%+
AUC-ROC	0.96+

Error Analysis Insights

• Model excels at clear positive/negative reviews
• Handles mixed-sentiment reviews reasonably well
• Struggles with sarcasm and irony (known BERT limitation)
• Short reviews (< 10 words) have slightly lower accuracy

---

🏗️ Architecture

Model Pipeline

Raw Text ↓ Text Preprocessing (lowercase, remove HTML, URLs, special chars) ↓ DistilBERT Tokenizer (max_length=256, padding, truncation) ↓ DistilBERT Encoder (6 transformer layers, 768 hidden dims) ↓ Classification Head (Dense → Dropout → Dense) ↓ Softmax Output (Positive Probability, Negative Probability) ↓ Threshold Decision (default: 0.5, optimal: tuned) ↓ Sentiment Label + Confidence

Training Strategy

Fine-Tuning Configuration:

• All DistilBERT layers unfrozen
• AdamW optimizer with weight decay
• Linear warmup schedule (10% of steps)
• Gradient clipping (max_norm=1.0)
• Early stopping on validation accuracy

Hyperparameters:

CONFIG = {
    'model_name': 'distilbert-base-uncased',
    'max_length': 256,
    'batch_size': 16,
    'epochs': 3,
    'learning_rate': 2e-5,
    'weight_decay': 0.01,
    'warmup_ratio': 0.1
}

---

🚀 Installation

Prerequisites

• Python 3.8 or higher
• pip package manager
• 4GB RAM minimum
• GPU optional (CPU works fine)

Step 1: Clone Repository

git clone https://github.com/ADHIRAJ994/sentiment-analysis-dashboard.git
cd sentiment-analysis-dashboard

Step 2: Create Virtual Environment

# Windows
python -m venv venv
venv\Scripts\activate

# Mac/Linux
python3 -m venv venv
source venv/bin/activate

Step 3: Install Dependencies

pip install -r requirements.txt

Step 4: Download NLTK Data

import nltk
nltk.download('stopwords')
nltk.download('punkt')
nltk.download('vader_lexicon')

Step 5: Download Model Files

Download best_model.pt and tokenizer/ from releases and place in models/ folder.

---

💻 Usage

Run Web Dashboard

streamlit run app.py

Access at: http://localhost:8501

Python API

import torch
from transformers import DistilBertTokenizer, DistilBertForSequenceClassification

# Load model
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
tokenizer = DistilBertTokenizer.from_pretrained('models/tokenizer')
model = DistilBertForSequenceClassification.from_pretrained(
    'distilbert-base-uncased',
    num_labels=2
)
checkpoint = torch.load('models/best_model.pt', map_location=device)
model.load_state_dict(checkpoint['model_state_dict'])
model = model.to(device)
model.eval()

# Predict
def predict(text):
    encoding = tokenizer.encode_plus(
        text,
        max_length=256,
        padding='max_length',
        truncation=True,
        return_tensors='pt'
    )
    
    with torch.no_grad():
        outputs = model(
            input_ids=encoding['input_ids'].to(device),
            attention_mask=encoding['attention_mask'].to(device)
        )
        probs = torch.softmax(outputs.logits, dim=1)
        pred = torch.argmax(outputs.logits, dim=1)
    
    sentiment = 'POSITIVE' if pred.item() == 1 else 'NEGATIVE'
    confidence = probs[0][pred.item()].item()
    return sentiment, confidence

# Example
sentiment, confidence = predict("This product is absolutely amazing!")
print(f"Sentiment: {sentiment} ({confidence*100:.1f}% confidence)")

Batch Processing via CSV

import pandas as pd

# Create CSV with reviews
df = pd.DataFrame({
    'review': [
        'Great product, highly recommend!',
        'Terrible quality, waste of money.',
        'Good value for the price.'
    ]
})
df.to_csv('reviews.csv', index=False)

# Upload to dashboard and analyze

Train from Scratch

# Run notebooks in order
jupyter notebook notebooks/01_eda.ipynb
jupyter notebook notebooks/02_bert_model.ipynb
jupyter notebook notebooks/03_finetuning.ipynb
jupyter notebook notebooks/04_deployment.ipynb

---

📁 Project Structure

Sentiment Analysis Dashboard/

│ ├── 📁 data/ │ ├── train_processed.csv # Preprocessed training data │ └── test_processed.csv # Preprocessed test data │

├── 📁 models/ │ ├── best_model.pt # Fine-tuned DistilBERT weights │ ├── model_config.json # Model configuration │ └── tokenizer/ # Saved tokenizer files │ ├── config.json │ ├── tokenizer.json │ ├── tokenizer_config.json │ └── vocab.txt │

├── 📁 results/ │ ├── plots/ │ │ ├── class_distribution.png │ │ ├── wordclouds.png │ │ ├── top_words.png │ │ ├── training_history.png │ │ ├── confusion_matrix.png │ │ ├── roc_curve.png │ │ ├── model_comparison.png │ │ ├── aspect_sentiment_heatmap.png │ │ ├── aspect_statistics.png │ │ ├── error_analysis.png │ │ └── threshold_analysis.png │ └── metrics/ │ ├── training_history.csv │ ├── test_results.json │ └── final_results.json │

├── 📁 notebooks/ │ ├── 01_eda.ipynb # Exploratory Data Analysis │ ├── 02_bert_model.ipynb # BERT model training │ ├── 03_finetuning.ipynb # Fine-tuning & optimization │ └── 04_deployment.ipynb # Deployment preparation │

├── 📄 app.py # Streamlit dashboard ├── 📄 requirements.txt # Python dependencies ├── 📄 config.json # Project configuration ├── 📄 .gitignore # Git ignore rules └── 📄 README.md # This file

---

📊 Dataset

Source

• Dataset: Amazon Polarity Reviews
• Platform: Hugging Face Datasets
• Link: amazon_polarity

Statistics

Split	Positive	Negative	Total
Train	5,000	5,000	10,000
Val	850	850	1,700
Test	1,000	1,000	2,000

Preprocessing Steps

1. Convert to lowercase
2. Remove HTML tags
3. Remove URLs
4. Remove special characters
5. Remove extra whitespace
6. Truncate to max_length=256 tokens

---

🎓 Model Training

Phase 1: Baseline (Week 1)

• Rule-based sentiment with TextBlob
• Result: 68% accuracy
• Used as benchmark for improvement

Phase 2: Fine-Tuning (Week 2)

• Loaded pre-trained DistilBERT
• Added classification head
• Fine-tuned for 3 epochs
• Result: 88-92% accuracy

Phase 3: Optimization (Week 3)

• Threshold tuning
• Error analysis
• Aspect-based sentiment extraction
• Result: 90%+ optimized accuracy

Training Details

Epoch 1/3: Train Loss: ~0.35 | Train Acc: ~85% Val Loss: ~0.28 | Val Acc: ~88% Epoch 2/3: Train Loss: ~0.22 | Train Acc: ~91% Val Loss: ~0.24 | Val Acc: ~90% Epoch 3/3: Train Loss: ~0.15 | Train Acc: ~94% Val Loss: ~0.26 | Val Acc: ~91%

---

📈 Results & Visualizations

Training History

• Loss decreases consistently across epochs
• Validation accuracy improves from 88% to 91%+
• No significant overfitting observed

Confusion Matrix Insights

        Predicted
    Negative  Positive

Actual Neg ~450 ~50 Pos ~50 ~450

Aspect Analysis Results

Aspect	Coverage	Positive Rate
Quality	High	~65%
Price	High	~55%
Shipping	Medium	~70%
Service	Medium	~60%
Performance	Medium	~68%
Design	Low	~72%

Key Findings

1. Price aspect has lowest positive rate (most complaints)
2. Shipping aspect has highest positive rate
3. Quality is most frequently mentioned aspect
4. High confidence predictions (>90%) are almost always correct

---

🛠️ Technologies Used

NLP & Deep Learning

• PyTorch 2.0.1 - Deep learning framework
• Transformers 4.35.0 - DistilBERT model
• NLTK 3.8.1 - Text preprocessing
• TextBlob 0.17.1 - Baseline model

Data Science

• NumPy 1.24.3 - Numerical computing
• Pandas 2.1.1 - Data manipulation
• scikit-learn 1.3.0 - Metrics and utilities

Visualization

• Matplotlib 3.7.2 - Static plots
• Seaborn 0.13.0 - Statistical plots
• Plotly 5.17.0 - Interactive charts
• WordCloud 1.9.2 - Word cloud generation

Web Development

• Streamlit 1.28.1 - Web dashboard

Development Tools

• Jupyter Notebook - Development
• Git & GitHub - Version control

---

⚠️ Disclaimer

Limitations

1. Domain Specificity

   • Trained on Amazon product reviews
   • May not generalize to other domains (e.g., movie reviews, news)

2. Language

   • English only
   • May struggle with slang or informal language

3. Sarcasm

   • Known weakness of transformer models
   • Sarcastic reviews may be misclassified

4. Context Window

   • Reviews truncated at 256 tokens
   • Very long reviews may lose context

Responsible Use

• Not suitable for critical business decisions without human review
• Always validate predictions on your specific domain
• Consider bias in training data

---

🚀 Future Improvements

Short-term (1-3 months)

• Multi-class sentiment (5-star rating)
• Multi-language support
• Real-time Twitter/social media analysis
• Sarcasm detection module

Medium-term (3-6 months)

• Deploy as REST API (FastAPI)
• Database integration for history
• User authentication
• Analytics dashboard

Long-term (6-12 months)

• Fine-tune on domain-specific data
• Active learning pipeline
• Mobile application
• Integration with e-commerce platforms

---

📄 License

This project is licensed under the MIT License. MIT License Copyright (c) 2026 Adhiraj Chakravorty Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions: The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software. THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

---

📞 Contact

Adhiraj Chakravorty

• 📧 Email: [youradhi20@gmail.com]
• 💼 LinkedIn: https://www.linkedin.com/in/adhiraj-chakravorty-788685344/
• 🐱 GitHub: @ADHIRAJ994

Project Links

• 🔗 Repository: github.com/ADHIRAJ994/sentiment-analysis-dashboard
• 🌐 Live Demo: https://sentiment-analysis-y5kp2cxntkssecxadpvfa5.streamlit.app/

---

---

🎊 Thank you for checking out this project!

⬆ Back to Top