You can deploy this model and start earning money today!

niconielse/
nlp_custom

Name: nlp_custom
Availability: InStock

Views

No views yet

Vietnamese E-commerce Comment Sentiment Analysis

This project implements a machine learning pipeline for classifying Vietnamese e-commerce comments into 5 star ratings (1-5 stars) based on sentiment analysis.

Features

Text Preprocessing: Vietnamese text cleaning and normalization
Multiple Models: Naive Bayes, SVM, Random Forest, Neural Network, Logistic Regression
Ensemble Method: Combines multiple models for better performance
Evaluation: Micro-F1 score as primary metric
Prediction API: Easy-to-use functions for new comment prediction

Project Structure

d:/OlympicAI/code/
├── data_preprocessor.py    # Text preprocessing utilities
├── models.py              # Classification models
├── train_model.py         # Main training script
├── requirements.txt       # Dependencies
└── README.md             # This file

Installation

Install required packages:

pip install -r requirements.txt

Usage

1. Prepare Data

Your CSV file should have two columns:

comment: Vietnamese text comments
rating: Integer rating from 1-5

2. Train Models

python

from train_model import train_and_evaluate_models

# Train all models
models, results, preprocessor = train_and_evaluate_models('your_data.csv')

3. Make Predictions

python

from train_model import predict_sentiment

# Predict single comment
text = "Sản phẩm tuyệt vời, tôi rất hài lòng!"
rating, confidence = predict_sentiment(
    text, 
    'model_nb.pkl',
    'tfidf_vectorizer.pkl'
)
print(f"Predicted rating: {rating} stars")

Model Performance

The system trains multiple models and compares their performance using Micro-F1 score:

Naive Bayes: Fast, good baseline performance
SVM: Strong performance with linear kernel
Random Forest: Robust ensemble method
Neural Network: Deep learning approach
Logistic Regression: Simple yet effective
Ensemble: Combines all models for best results

Rating Scale

1 star: Very dissatisfied (Rất không hài lòng)
2 stars: Dissatisfied (Không hài lòng)
3 stars: Neutral (Trung bình)
4 stars: Satisfied (Hài lòng)
5 stars: Very satisfied (Rất hài lòng)

Evaluation Metric

Micro-F1 Score: Measures overall classification accuracy across all classes, calculated as:

Micro-F1 = 2 × Precision × Recall / (Precision + Recall)

Where Precision and Recall are computed globally by counting total true positives, false negatives, and false positives.

Files Generated

After training, the following files are created:

model_*.pkl: Trained model files
tfidf_vectorizer.pkl: Fitted TF-IDF vectorizer
confusion_matrix_*.png: Confusion matrix visualizations

Example Output

FINAL RESULTS SUMMARY
==================================================
NB           | Micro-F1: 0.8234 | Macro-F1: 0.8123 | Weighted-F1: 0.8201
SVM          | Micro-F1: 0.8456 | Macro-F1: 0.8334 | Weighted-F1: 0.8423
RF           | Micro-F1: 0.8123 | Macro-F1: 0.8001 | Weighted-F1: 0.8089
MLP          | Micro-F1: 0.8345 | Macro-F1: 0.8234 | Weighted-F1: 0.8312
LR           | Micro-F1: 0.8567 | Macro-F1: 0.8445 | Weighted-F1: 0.8534
ENSEMBLE     | Micro-F1: 0.8678 | Macro-F1: 0.8556 | Weighted-F1: 0.8645

Best model: ENSEMBLE with Micro-F1: 0.8678

niconielse/nlp_custom