🍽️ Restaurant Reviews Analysis
"See what your customers feel, not just what they say."
An AI-powered system that transforms large volumes of restaurant reviews into actionable business insights using Aspect-Based Sentiment Analysis (ABSA) and AI-driven summarization.
Built as part of Samsung Innovation Campus – SIC 7 Team 7
📌 Problem Statement
Restaurants receive thousands of emotional and unstructured reviews across platforms. Managers often:
- Struggle to identify real issues in food, service, price, or delivery
- Spend hours manually reading reviews
- Miss recurring patterns and hidden insights
This project automates the entire process — extracting key aspects, detecting sentiment, and generating clear business summaries.
🚀 Project Overview
The system pipeline consists of:
- Aspect Extraction (Token-Level)
- Sentiment Classification (Aspect-Level)
- Review Aggregation
- AI-Powered Summarization
- Dashboard Visualization
Users can:
- Upload an Excel file
- Fetch reviews from an API
- Instantly receive:
- ✅ Top positives
- ❌ Top negatives
- 📈 Ratings & trends
- 📊 Overall performance score
🧠 Technical Architecture
1️⃣ Aspect-Based Sentiment Analysis (ABSA)
📚 Dataset
- SemEval 2014 Task 4 – Aspect-Based Sentiment Analysis
🔧 Preprocessing
- Data balancing using:
- Synthetic data
- Class weights
- Tokenization using DeBERTa-v3 tokenizer
- BIO tagging scheme for aspect extraction
BIO Labels:
- B – Beginning of aspect
- I – Inside aspect
- O – Outside aspect
2️⃣ Multitask Transformer Model
Core Model:
Two task-specific heads:
- 🏷️ BIO Tagging Head → Aspect extraction
- 💬 Sentiment Classification Head → Polarity detection
Both heads share the same encoder, improving learning efficiency and overall accuracy.
📊 Model Performance
- Strong convergence in training & validation loss
- High F1 scores for aspect extraction
- Balanced performance across:
- Positive
- Negative
- Neutral classes
Confusion matrix results show strong classification accuracy across all sentiment categories.
✍️ AI Summarization Model
Model used:
Capabilities
- Extracts aspects & classifies sentiment
- Generates:
- 🟢 Pros summary
- 🔴 Cons summary
- 🟡 Overall summary
- Aggregates frequent trends
- Cleans & deduplicates repetitive insights
Output is structured, concise, and business-ready.
📊 Business Value
- ⚡ Fast insight into strengths & weaknesses
- 🎯 Actionable priorities for improvement
- 📈 Data-driven marketing & operations decisions
- 📦 Scalable for large review volumes
- 🖥️ Easy-to-use Streamlit dashboard
🖥️ Deployment
Built using Streamlit Web Application
How to Use
- Upload Excel file containing restaurant reviews
- Select the column that contains review text
- Click Analyze Reviews
- View dashboard insights
- Download results as Excel
What It Analyzes
- 🍔 Food Quality
- 🚚 Time / Delivery
- 🏪 Place / Service
📦 Tech Stack
- Python
- PyTorch
- HuggingFace Transformers
- DeBERTa-v3-base
- Flan-T5-large
- Streamlit
- Pandas / NumPy
- Matplotlib / Seaborn
🔮 Future Improvements
- Automatic review scraping from Google & review platforms
- Financial metrics integration
- Multilingual support
- Cloud hosting deployment
- Mobile application version
👥 Team
SIC 7 Team 7
- Anas Ibrahim
- Mohamed Makram
- Maryam Abdelnabi
🎯 Conclusion
This system converts massive volumes of unstructured restaurant reviews into structured, meaningful, and actionable intelligence using state-of-the-art NLP models.
It bridges the gap between:
What customers say
and
What businesses need to improve.
