An end-to-end NLP system that analyses customer reviews across multiple banks — classifying sentiment, predicting star ratings (1–5), and identifying key complaint themes — deployed as a live Flask web application.

Banks receive thousands of customer reviews and complaints daily. Reading each one manually is impossible at scale. This project builds an automated NLP pipeline that:

• Classifies reviews as positive, negative, or neutral
• Predicts the star rating (1–5) from raw review text alone
• Identifies recurring complaint themes (service, waiting time, fees, etc.)
• Performs intent analysis to understand what customers actually want

This enables banks to prioritise critical complaints, monitor satisfaction trends, and improve customer service — all without manual effort.

• Customer Service Teams can auto-triage high-priority complaints
• Bank Managers get data-driven insights into which services are failing
• Marketing Teams can identify what customers love and amplify it
• Scales to handle thousands of reviews per minute once deployed

• ~500 customer reviews and complaints across multiple banks
• Multi-class target: Star ratings 1 through 5
• Real-world noisy text — abbreviations, typos, mixed case, punctuation

Bank-Reviews-Complaints-Analysis/
├── results/
│   ├── 2020-06-21 16_52_49-Window.png
│   ├── 2020-06-21 16_53_18-Window.png
│   ├── 2020-06-21 16_53_57-Window.png
│   └── 2020-06-21 16_58_01-Window.png
├── static/
│   └── styles.css
├── templates/
│   ├── home.html
│   └── result.html
├── Bank Reviews-Complaints Analysis.ipynb   # Full EDA + Modelling
├── app.py                                   # Flask application
├── nlp_model.pkl                            # Trained classification model
├── transform.pkl                            # Label encoder / transformer
├── tfidf_transform.pkl                      # Fitted TF-IDF vectorizer
├── Dockerfile
├── Procfile
├── requirements.txt
└── README.md

Extensive NLP feature engineering was applied to extract signal from raw text:

• Basic text stats: word count, character count, sentence count
• Linguistic features: uppercase words, lowercase words, title-case words
• Punctuation density: count of punctuation marks as a noise/emotion proxy
• Stopword analysis: stopword count to identify filler-heavy reviews
• Custom pattern features:
  • Words ending with specific suffixes (e.g., et)
  • Words starting with specific prefixes (e.g., no) — a strong negative signal
  • Word frequency dictionaries per review

df["count_punctuations"] = df["Reviews"].apply(
    lambda x: len([c for c in str(x) if c in string.punctuation])
)
df["count_words_upper"] = df["Reviews"].apply(
    lambda x: len([w for w in str(x).split() if w.isupper()])
)
df['words_start_with_no'] = df['Reviews'].apply(
    lambda x: len([w for w in x.lower().split() if w.startswith('no')])
)

• Applied TF-IDF (Term Frequency–Inverse Document Frequency) to convert review text into numerical feature vectors
• Serialized the fitted vectorizer (tfidf_transform.pkl) so the same transformation is applied at inference time

• Trained a multi-class classification model to predict star ratings directly from review text
• Serialized trained model as nlp_model.pkl for production inference

• Built a Flask web application for real-time star rating prediction
• User inputs any review text → model predicts a rating from 1 to 5
• Containerized using Docker for portability
• Originally deployed on Heroku (migrating to Render — see below)

Key findings from the analysis:

• Reviews with high punctuation counts (e.g., multiple ! or ?) are strongly correlated with 1-star ratings
• Words starting with "no" (no service, not helpful, nowhere) are the top negative signal
• Uppercase word density is a proxy for customer frustration
• Most complaints cluster around: long wait times, poor customer service, hidden fees, and account issues
• 4 and 5-star reviews consistently mention: staff helpfulness, speed, digital banking ease

# Clone the repository
git clone https://github.com/vicky60629/Bank-Reviews-Complaints-Analysis.git
cd Bank-Reviews-Complaints-Analysis

# Install dependencies
pip install -r requirements.txt

# Run the Flask app
python app.py

Open http://localhost:5000 in your browser.

docker build -t bank-reviews-app .
docker run -p 5000:5000 bank-reviews-app

1. Open the web app in your browser
2. Type or paste any bank review text into the input box
3. Click "Predict"
4. Get an instant star rating prediction (1–5) based on your review

Example Input:

"The staff was extremely rude and I waited 45 minutes just to speak to someone. Totally unacceptable."

Predicted Output: ⭐ 1 Star

Home Screen	Prediction Result

What worked well:

• TF-IDF combined with engineered text features gave a strong signal beyond just word frequencies
• Custom features like "words starting with no" proved surprisingly effective as negative sentiment indicators
• End-to-end deployment validated the full ML lifecycle — not just model training

Future enhancements:

• Replace TF-IDF with BERT / sentence-transformers for deeper semantic understanding
• Add aspect-based sentiment analysis (e.g., separate scores for service, speed, fees)
• Build real-time complaint dashboard using Streamlit or Plotly Dash
• Train on larger datasets (e.g., Trustpilot, Google Reviews scrape)
• Add multilingual support for Indian regional languages
• Integrate with MLflow for experiment tracking

Vicky Gupta — Data Engineering Analyst @ Accenture (4.5 years) | Aspiring Data Scientist

Passionate about NLP and building end-to-end ML systems that solve real business problems. Experienced in PySpark, ETL pipelines, and deploying ML models to production.

🔗 LinkedIn | GitHub

📧 [email protected]

This project is licensed under the MIT License — see the LICENSE file for details.

⭐ If you found this project useful, please star the repository — it helps others discover it!