NLP Sentiment Analysis Pipeline

An end-to-end Machine Learning pipeline for binary sentiment classification, utilizing an advanced Ensemble Voting strategy and N-gram TF-IDF vectorization.

Engineering Goal: Maximize classification accuracy on movie reviews by combining the strengths of linear models (Logistic Regression, SVM) and tree-based models (Random Forest) via soft voting.

Academic Context: Final Project for the "Introduction to Data Science" course at Institut Teknologi Sepuluh Nopember (ITS), Indonesia (2025). Ranked via a private Kaggle competition.

---

🏗 Architecture & Methodology

This project implements a robust pipeline designed for high-dimensional text data. Unlike simple baselines, it uses advanced feature extraction and model stacking.

Key Steps:

1. Advanced Preprocessing (src/preprocess.py):
   • Noise Removal: Regex-based cleaning to strip URLs, user handles (@user), and non-alphabetic characters.
   • Normalization: Lowercasing and whitespace trimming.
2. Feature Extraction:
   • TF-IDF Vectorization: configured with N-grams (1, 3) to capture phrase-level context (e.g., "not good").
   • Optimization: Uses sublinear_tf scaling and strict document frequency limits (min_df=3, max_df=0.7) to reduce feature noise.
3. Ensemble Modeling (src/model.py):
   • Voting Classifier (Soft Vote): Combines probability predictions from three distinct models:
     • Logistic Regression: For linear separability and speed.
     • Random Forest: To capture non-linear relationships.
     • SVM (RBF/Linear): For high-dimensional margin maximization.
4. Evaluation:
   • 5-Fold Cross-Validation to ensure model stability.
   • Metrics: Accuracy, Precision, Recall, and F1-Score.

---

📂 Project Structure

├── data/
│   ├── train.csv          # Labelled training dataset
│   └── test.csv           # Unlabelled testing dataset
├── notebooks/
│   └── eda.ipynb          # Exploratory Analysis (Class balance, Text length distribution)
├── src/
│   ├── preprocess.py      # Regex cleaning & TF-IDF N-gram vectorizer
│   ├── model.py           # Ensemble Voting Classifier definition
│   └── predict.py         # Inference logic for generating submissions
├── outputs/
│   └── submission.csv     # Final predictions generated by the pipeline
├── main.py                # Entry point orchestrating the full flow
└── requirements.txt       # Python dependencies

---

🚀 Getting Started

Prerequisites

• Python 3.8+
• Key Libraries: scikit-learn, pandas, nltk, textblob

Installation

git clone https://github.com/noecrn/NLP-Sentiment-Analysis-Pipeline.git
cd NLP-Sentiment-Analysis-Pipeline
pip install -r requirements.txt

Usage

Run the full pipeline (Clean Vectorize Train Ensemble Predict):

python main.py

This script will:

1. Load and clean the training data.
2. Train the Voting Classifier and print Cross-Validation accuracy.
3. Generate predictions for test.csv and save them to outputs/submission.csv.

---

📊 Performance

• Vectorization: TF-IDF (Unigrams + Bigrams + Trigrams)
• Validation Strategy: 5-Fold Cross-Validation
• Target Metric: Accuracy

---

👤 Author

• Noé Cornu - Engineering Student @EPITA - GitHub | LinkedIn
• Baptiste Rio - Engineering Student @ EPITA