Aspect-Based Sentiment Classification for E-Commerce Reviews

Project conducted: December 2024 (was not uploaded at the time)

Building a machine learning pipeline to classify customer sentiments across multiple product aspects in the e-commerce domain.

---

Introduction

This project aims to extract structured insights from unstructured customer reviews collected from Hasaki.vn. By leveraging NLP and machine learning, it automatically identifies six key product aspects:

• Store
• Service
• Packaging
• Price
• Quality
• Others

Each review is classified by both aspect and sentiment polarity (Positive/Negative/Neutral), supporting customer experience analysis at a granular level.

---

Dataset

Data Source

Data was crawled using Selenium from product pages and review sections on Hasaki.vn.

Labeling Process

A hybrid labeling approach was used:

• Gemini API: for automated tagging of aspects and sentiments.
• Manual verification: to improve reliability.

Each review may be linked to multiple aspects, each with a corresponding sentiment label.

---

Tools & Technologies

• Languages: Python
• Libraries: pandas, numpy, scikit-learn, gensim, tensorflow, keras
• Embedding: Word2Vec
• Model persistence: joblib, pickle
• Development: Google Colab (please update file paths if running locally)

---

Project Workflow

1. Data Crawling

• File: crawl_comment.ipynb
• Crawls product IDs and corresponding customer reviews
• Saves raw review data to data/data_crawl.xlsx

2. Labeling

• File: code_label_gemini_api.ipynb
• Uses Gemini API to assign:
  • Aspect tags: e.g., Service, Packaging...
  • Sentiment: Positive, Negative, Neutral
• Outputs structured file data/data_label.xlsx with one-hot encoded aspect columns

3. Data Preprocessing

• File: data_preprocessing.ipynb
• Includes:
  • Lowercasing, punctuation/special character removal
  • Emoji replacement
  • Tokenization using Vietnamese rules
  • Stopword removal using a curated .txt list
• Output: cleaned review sequences in data/data_preprocess.xlsx

Example wordcloud:

4. Embedding & Tokenization

• Trains a custom Word2Vec model
• Tokenizes input sequences using Keras Tokenizer
• Produces:
  • word2vec_sentiment.model (embedding)
  • tokenizer.pkl (used for training neural networks)

5. Model Training

• Folder model_code/ contains 6 training notebooks (1 per aspect)
• Models trained:
  • Logistic Regression
  • SVM
  • Random Forest
  • Neural Network
• Best-performing model selected per aspect
• Trained .joblib models stored in model_file/

6. Evaluation

• Performed with:
  • Accuracy
  • Precision / Recall / F1-score
  • Confusion Matrix
  • Classification Report
  • ROC Curve

Example for aspect_quality:

Summary all:

Aspect	Model	Accuracy
Service	Neural Network	92.3%
Store	Random Forest	90.1%
Packaging	Random Forest	89.72%
Others	Random Forest	88.23%
Price	Neural Network	85.4%

---

Result

Result from running main.ipynb with a Vietnamese review input

---

How to Use

Option 1: Full Pipeline

1. Run the following in order (on Colab):
   • crawl_comment.ipynb
   • code_label_gemini_api.ipynb
   • data_preprocessing.ipynb
   • Any training notebook in model_code/
   • main.ipynb

Option 2: Pre-trained Models

1. Download:

   • embedding_model_file/
   • model_file/

2. Update paths in main.ipynb

3. Input a Vietnamese review → returns aspect-level sentiment

---

Learning Outcomes

• Applied aspect-based sentiment analysis to real-world e-commerce data
• Trained and compared multiple classification models
• Built a functional inference pipeline ready for deployment
• Practiced crawling, labeling, NLP preprocessing, model evaluation, and modular code organization