This project analyzes Amazon product reviews and classifies them as positive or negative using Natural Language Processing.
It compares traditional machine learning approaches with a transformer-based model, DistilBERT, to understand how different NLP methods perform on review sentiment classification.
Online reviews contain valuable customer feedback, but reading thousands of reviews manually is not practical.
This project uses NLP to automatically identify review sentiment and evaluate model performance using clear metrics and visual analysis.
| Area | Description |
|---|---|
| Text Cleaning | Preprocessed Amazon review text for modeling |
| Exploratory Analysis | Analyzed review length and text distribution |
| TF-IDF Features | Converted review text into numerical features |
| PCA / SVD Projection | Visualized feature separation in 2D |
| Classical ML | Compared traditional machine learning models |
| DistilBERT | Used transformer-based sentiment classification |
| Evaluation | Reviewed performance using ROC-AUC and classification metrics |
Shows how long the reviews are after preprocessing. This helps understand text length patterns before training.
2D PCA/SVD projection of TF-IDF features, showing how sentiment classes overlap in the feature space.
DistilBERT achieved strong classification performance with an ROC-AUC score of 0.9867.
| Category | Tools / Methods |
|---|---|
| Language | Python |
| Environment | Jupyter Notebook / Google Colab |
| Data Processing | Pandas, NumPy |
| Visualization | Matplotlib, Seaborn |
| Feature Engineering | TF-IDF, SVD / PCA |
| Machine Learning | Scikit-learn |
| Transformer Model | DistilBERT |
| Task | Binary Sentiment Classification |
| Metric / Check | Purpose |
|---|---|
| Accuracy | Measures overall classification correctness |
| Precision | Checks how many predicted positives are correct |
| Recall | Checks how many actual positives are captured |
| F1-score | Balances precision and recall |
| ROC-AUC | Measures class separation quality |
| Feature Projection | Helps visualize sentiment overlap |
| Review Length Analysis | Understands dataset characteristics |
I worked on this project as an NLP and model evaluation contributor.
My work focused on:
- cleaning and preprocessing Amazon review text
- analyzing review length distribution
- building TF-IDF feature representations
- visualizing feature space using PCA/SVD
- training and comparing sentiment classification models
- evaluating DistilBERT performance using ROC-AUC
- documenting results through plots and metric summaries
Make sure you have:
- Python installed
- Jupyter Notebook or Google Colab
- Required Python libraries installed
-
Clone the repository.
git clone https://github.com/SHREENITHI-TV/Amazon-Reviews-Sentiment-Analysis.git -
Open the notebook.
Amazon_Reviews.ipynb -
Install dependencies if needed.
pip install pandas numpy matplotlib seaborn scikit-learn transformers torch -
Run the notebook cells in order.
-
Review preprocessing, visualizations, model results, and ROC curve.
This project demonstrates practical experience with:
- NLP text preprocessing
- sentiment classification
- feature engineering with TF-IDF
- dimensionality reduction using PCA/SVD
- classical ML model comparison
- transformer-based NLP modeling
- model evaluation and visualization
- metric-driven validation
- Build a simple Streamlit demo