Amazon Review Opinion Search Engine

Project Overview

Semantic opinion search engine for Amazon reviews — 48% MAP improvement using SBERT over Boolean baseline. 210K+ reviews

Project Structure

amazon_search_engine/
├── Codes/
│   ├── baseline_boolean.py            # Boolean search implementation (Tests 1-3)
│   ├── advanced_language_model.py           # First advanced method
│   ├── advanced_SBERT.py           # Second advanced method
│   ├── inverted_index.py      # Inverted index builder/handler
│   └── queries.json           # Query definitions for testing
├── Data/
│   ├── data.pkl               # Processed dataset
│   ├── reviews_segment.pkl    # Original Amazon reviews dataset
│   ├── inverted_index.pkl     # Pre-built inverted index
│   ├── lm.pkl                 # Logistic regression model (trained sentiment classifier)
│   ├── positive_words.txt     # Positive opinion lexicon
│   ├── negative_words.txt     # Negative opinion lexicon
│   ├── negation_words.txt     # Negation words list
│   └── intensifier_words.txt  # Intensifier words list
├── Outputs/
│   ├── Baseline/ {aspect_words}_test{number}.txt
│   ├── AdvancedMethod1/ {aspect}_test4.txt
│   └── AdvancedMethod2/ {aspect}_test4.txt
├
├── requirements.txt           # Python dependencies
└── Readme.md                  # This file

📁 Data Folder

The large reviews_segment.pkl and data.pkl files are not included in this repository due to their size. To run the full pipeline, place these files inside the Data/ directory before executing the scripts.

Data/
├── data.pkl
├── reviews_segment.pkl

Python Dependencies

See requirements.txt for complete list. Key dependencies include:

Step 1: Navigate to Project Directory

cd amazon_search_engine

Step 2: Create Virtual Environment if not created

Create virtual environment

python -m venv venv

Activate virtual environment if not activated

On Windows:

venv\Scripts\activate

On macOS/Linux:

source venv/bin/activate

Step 3: Install Dependencies

pip install -r requirements.txt

Step 4: Download NLTK Data (if needed)

nltk.download('wordnet') nltk.download('omw-1.4') nltk.download('punkt') nltk.download('stopwords') nltk.download('punkt_tab') nltk.download('averaged_perceptron_tagger')

Dataset Setup

Place the dataset files in the Data/ directory:

• reviews_segment.pkl: Main review dataset (210,761 reviews)
• data.pkl: Precomputed embeddings (for Advanced Method 2)
• Opinion lexicons: Positive/negative word lists
• Cached indices: Pre-built inverted index for faster retrieval (optional)

Usage

Running the Baseline Method (Tests 1-3)

python Codes/baseline_boolean.py

This script:

• Loads queries from queries.json
• Implements Boolean search for Tests 1-3
• Generates output files in Outputs/...

Test Types:

• Test 1: Aspect
• Test 2: Aspect AND Opinion
• Test 3: Aspect OR Opinion
• Test 4: Connotation

Running Advanced Method 1

python Codes/advanced_language_model.py

Running Advanced Method 2

python Codes/advanced_SBERT.py

Query Format

Queries are defined in queries.json:

Output Format

Output Files

All output files are saved in Outputs/ directory. {aspect_words}_test{number}.txt

Examples:

• audio_quality_test1.txt (Baseline Test 1)
• audio_quality_test2.txt (Baseline Test 2)
• audio_quality_test3.txt (Baseline Test 3)
• audio_quality_test4.txt (Advanced method)

File Content

Each line contains a single review ID:

Evaluation Queries

The project is evaluated on these 5 queries:

1. audio quality: poor
2. wifi signal: strong
3. mouse button: click problem
4. gps map: useful
5. image quality: sharp

All queries are defined in queries.json.

Acknowledgments

• Dataset: Amazon Review Dataset (Course provided)
• Opinion Lexicon: Hu and Liu (KDD 2004)

📊 Results

Method Performance Comparison

Method	Avg Precision	Avg Retrieved	Precision Improvement
Baseline (Boolean)	0.3914	1,278	—
Advanced Method 1 (LM)	0.5503	582	+40.6%
Advanced Method 2 (SBERT)	0.578	216	+47.7%

83% reduction in retrieved documents while improving precision — evaluated across 5 real-world opinion queries on 210,000+ Amazon reviews.

Key outcomes:

• 48% relative improvement in Average Precision (0.39 → 0.58)
• 83% reduction in retrieved documents while maintaining relevance
• Evaluated across 5 real-world opinion queries on 210,000+ Amazon reviews

Me

Elias Arellano Campos