Sentiment Analysis (NLP)

Customer review sentiment analysis with Python and NLP.
The project uses a synthetic review dataset (positive, neutral, negative), applies text preprocessing (cleaning, tokenization, stopwords removal, lemmatization), converts text to TF-IDF features, and trains classifiers (Naive Bayes, Logistic Regression, Random Forest).
The best model is selected based on macro F1-score, and results are visualized with confusion matrix, word clouds, and top TF-IDF features.

---

Features

• Generate synthetic review dataset
• Text preprocessing:
  • lowercasing, URL & punctuation removal
  • stopwords filtering
  • lemmatization
• TF-IDF vectorization (unigrams + bigrams)
• Models: Multinomial Naive Bayes, Logistic Regression, Random Forest
• Evaluation: accuracy, precision, recall, F1-score
• Visuals: confusion matrix, word clouds, top features per class
• Saved artifacts: best model + vectorizer (joblib), metrics JSON

---

Project Structure

sentiment-analysis-nlp/
├─ README.md
├─ LICENSE
├─ requirements.txt
├─ data/
│  └─ generate_reviews.py
├─ src/
│  ├─ train_nlp.py
│  └─ utils.py
└─ outputs/
   └─ figures & reports (auto-created)

---

Setup

python -m venv .venv
# Windows:
.venv\Scripts\activate
# macOS/Linux:
source .venv/bin/activate
pip install -r requirements.txt

---

Generate Synthetic Reviews

python data/generate_reviews.py --n 8000 --seed 42 --out data/reviews.csv

---

Train & Evaluate

python src/train_nlp.py --input data/reviews.csv --outdir outputs --test-size 0.2 --seed 42

Outputs

• metrics.json – per-model scores & best model
• classification_report.txt
• confusion_matrix.png
• wordcloud_positive.png, wordcloud_negative.png
• top_features.txt
• best_model.joblib, vectorizer.joblib

---

Example Results

Confusion Matrix

Best model performance across classes:

---

Word Cloud (Positive Reviews)

---

Word Cloud (Negative Reviews)

---

Top Features

File: outputs/top_features.txt
Shows top discriminative words/phrases learned by the classifier for each class.

---

Data Schema

column	description
review_id	unique id
text	raw review text
label	sentiment {negative, neutral, positive}