Aspect-Based Sentiment Analysis for MOOC Reviews

Overview

This repository contains the implementation of an Aspect-Based Sentiment Analysis (ABSA) system for analyzing reviews of Massive Open Online Courses (MOOCs). Developed as part of the CS4082 Machine Learning course under the supervision of Dr. Naila Marir, this project aims to extract and classify sentiments related to specific aspects of MOOCs using a dataset of 1.45 million Coursera reviews. The system leverages traditional machine learning models (e.g., Logistic Regression, SVM, Decision Tree) and natural language processing (NLP) techniques, with results visualized through an interactive Streamlit dashboard.

Features

• Data Preprocessing: Cleaning, tokenization, and lemmatization of review text using SpaCy.
• Aspect Extraction: Identification of key themes using Dependency Parsing, RAKE, and Word2Vec.
• Sentiment Classification: Rule-based (TextBlob) and supervised classification (Logistic Regression, SVM, BERT, etc.).
• Visualization: Interactive Streamlit dashboard featuring aspect filters, pie charts, bar charts, and sample comments.
• Model Performance: High accuracy (Decision Tree: 94.28%, F1-Score: 94.40%) and AUC-ROC (Logistic Regression: 98.45%).

Dataset

The project utilizes a publicly available dataset of 1.45 million Coursera course reviews, containing:

• Reviews: Textual feedback from learners.
• Reviewers: Names of reviewers.
• Date_reviews: Submission dates of reviews.
• Rating: Numeric ratings provided by reviewers.
• Course_id: Unique identifiers for courses.

Repository Structure

├── data/
│   └── Coursera_Reviews.csv    # Link to Dataset
├── src/
│   ├── preprocessing.py        # Data cleaning and preprocessing script
│   ├── aspect_extraction.py    # Aspect extraction using SpaCy, RAKE, Word2Vec
│   ├── sentiment_analysis.py   # Sentiment classification (rule-based and supervised)
│   └── Streamlit.py            # Streamlit dashboard implementation
├
├── requirements.txt            # Python dependencies
├── README.md                   # This file
└── LICENSE                     # License file

Installation

Follow these steps to set up the project locally:

1. Clone the Repository:

   git clone https://github.com/your-username/mooc-absa.git
   cd mooc-absa

2. Create a Virtual Environment and Install Dependencies:

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

3. Prepare the Dataset:
   Ensure cleaned_reviews.csv is placed in the data/ directory.

---

Usage

Run the following commands to use the system:

1. Preprocess the Data
   Clean and prepare the dataset:

   python src/preprocessing.py

2. Extract Aspects
   Identify key aspects from the reviews:

   python src/aspect_extraction.py

3. Perform Sentiment Analysis
   Classify sentiments for extracted aspects:

   python src/sentiment_analysis.py

4. Launch the Streamlit Dashboard
   Visualize results interactively:

   streamlit run src/dashboard.py

   The dashboard provides:

   • Aspect filtering.
   • Sentiment distribution (pie chart).
   • Aspect frequency (bar chart).
   • Sample comments for each sentiment category.

Model Performance

The table below summarizes the performance of the evaluated models:

Model	Accuracy (%)	F1-Score (%)	AUC-ROC (%)
Decision Tree	94.28	94.40	93.22
Logistic Regression	92.47	92.90	98.45
SVM	91.91	92.30	N/A
Naive Bayes	80.24	83.12	93.04
Linear Discriminant	70.63	75.20	85.40
Quadratic Discriminant	48.07	60.05	69.53

Requirements

• Python: 3.8 or higher
• Key Libraries: spacy, textblob, sklearn, gensim, streamlit, pandas, numpy, matplotlib, transformers
• Full list available in requirements.txt

Future Work

• Integrate live feedback processing for real-time analysis.
• Utilize contextual embeddings (e.g., BERT) for enhanced feature extraction.
• Expand dashboard with additional interactive visualizations.
• Support non-English reviews for broader applicability.

Limitations

• High computational cost due to the large dataset.
• Potential bias in reviews (e.g., skewed toward strongly positive/negative sentiments).
• Limited to English reviews, reducing generalizability.

Contributors

• Araa Almarhabi (S20106395)
• Albatool Moathen (S21107416)
• Howyna Ahmed (S20106603)
• Fahad Dubush (S22107768)

License

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

Acknowledgments

We express gratitude to Dr. Naila Marir for her guidance and the Coursera platform for providing the dataset used in this project.