SnappFood Reviews Sentiment Analysis using Bidirectional LSTM

A deep learning project for binary sentiment analysis of Persian user reviews from the SnappFood dataset. It applies Persian text normalization and tokenization before training a Bidirectional LSTM classifier.

---

Features

• Persian text normalization using Hazm
• Tokenization with TensorFlow/Keras
• Bidirectional LSTM architecture
• Early Stopping
• ReduceLROnPlateau learning rate scheduling
• Best model checkpointing
• Saved tokenizer for inference
• Performance visualization

---

Dataset

This project uses the Cleaned SnappFood Persian Sentiment Analysis Dataset available on Kaggle:

Kaggle - Cleaned SnappFood Persian Sentiment Analysis Dataset

The dataset contains approximately 66K Persian user reviews before preprocessing.

Each review is labeled as:

• Positive
• Negative

---

Dataset Split

The dataset was divided using stratified sampling:

Split	Percentage
Training	72.25%
Validation	12.75%
Testing	15%

Stratification was applied to preserve the class distribution.

---

Data Preprocessing

Before training, the dataset undergoes several preprocessing steps.

Text Normalization

Persian text is normalized using Hazm, including:

• Character normalization
• Standardizing Persian/Arabic characters
• Removing unnecessary spaces
• Cleaning formatting inconsistencies

Although the dataset was pre-cleaned, Hazm normalization was applied to ensure consistent Persian character representation during training and inference.

Tokenization

The cleaned text is converted into integer sequences using the Keras Tokenizer.

Configuration:

• Vocabulary Size: 15,000
• OOV token enabled

Sequence Padding

Since reviews have different lengths, sequences are padded to a fixed size.

• Maximum sequence length: 40
• Post-padding
• Post-truncation

---

Model Architecture

Input Text
     │
Hazm Normalization
     │
Keras Tokenizer
     │
Padding (Length=40)
     │
Embedding
(Vocabulary=15000, Dimension=128)
     │
Bidirectional LSTM (128, dropout=0.2)
     │
Bidirectional LSTM (64, dropout=0.2)
     │
Dropout (0.5)
     │
Dense (64, ReLU)
     │
Softmax Output (2 classes)

---

Hyperparameters

Parameter	Value
Vocabulary Size	15,000
Maximum Sequence Length	40
Embedding Dimension	128
BiLSTM Units	128, 64
Dense Units	64
Dropout	0.5
Batch Size	64
Learning Rate	0.001
Epochs	10
Optimizer	Adam

---

Training Strategy

The model was trained using several techniques to improve convergence and reduce overfitting.

Early Stopping

• Monitor: Validation Loss
• Patience: 5

Learning Rate Scheduling

ReduceLROnPlateau automatically decreases the learning rate when validation performance stops improving.

Model Checkpoint

The best-performing model is automatically saved according to validation loss.

---

Results

Test Performance

Metric	Score
Accuracy	86%
Macro Precision	86%
Macro Recall	86%
Macro F1-score	86%

---

Classification Report

The model achieved consistent performance across both sentiment classes, showing that it learned meaningful sentiment patterns from Persian reviews.

Class	Precision	Recall	F1-score	Support
Negative	0.84	0.88	0.86	4,845
Positive	0.88	0.84	0.86	5,051
Accuracy			0.86	9,896
Macro Avg	0.86	0.86	0.86	9,896
Weighted Avg	0.86	0.86	0.86	9,896

Observations

• The model achieved an overall test accuracy of 86%.
• Performance is well balanced across both sentiment classes, with identical F1-scores (0.86) for Positive and Negative reviews.
• The Negative class obtained a slightly higher recall (0.88), indicating that the model successfully identifies most negative reviews.
• The Positive class achieved a higher precision (0.88), meaning positive predictions are highly reliable.
• The similar macro and weighted averages suggest that the model generalizes consistently across the dataset without a strong bias toward either class.

---

Training Curves

The repository includes:

• Training Accuracy
• Validation Accuracy
• Training Loss
• Validation Loss

These plots illustrate the convergence behavior of the model throughout training.

---

Learning Rate Schedule

The repository also includes the learning rate schedule generated during training.

---

Confusion Matrix

The confusion matrix provides a detailed view of prediction performance on the test set.

---

Training Log

Training history is available, including:

Epoch	Train Accuracy	Train Loss	Learning Rate	Validation Accuracy	Validation Loss
0	83.21%	0.3761	0.0010	85.83%	0.3278
1	87.78%	0.2946	0.0010	86.01%	0.3387
2	89.58%	0.2551	0.0010	85.44%	0.3503
3	90.92%	0.2233	0.0010	85.50%	0.4016
4	93.01%	0.1755	0.0005	85.16%	0.4534
5	93.88%	0.1551	0.0005	84.64%	0.4979

---

Example Predictions

Confidence	Prediction	Persian Review
99.25%	Positive	کیفیت غذا عالی بود و خیلی خوشمزه بود
90.26%	Negative	غذا سرد رسید و اصلا کیفیت خوبی نداشت
95.87%	Positive	ارسال خیلی سریع بود و از سفارش راضی هستم
95.40%	Negative	بدترین تجربه‌ای بود که داشتم، دوباره سفارش نمی‌دهم
74.36%	Negative	غذا معمولی بود، نه خوب نه بد

---

Project Structure

├── data/
│   └── cleaned_snappfood.csv
├── models/
│   ├── tokenizer.pkl
│   └── best_model.keras
├── results/
│   ├── classification_report.txt
│   ├── confusion_matrix.png
│   ├── learning_rate_schedule.png
│   ├── training_curves.png
│   └── training_log.csv
├── config.py
├── data_preprocessing.py
├── inference.py
├── main.py
├── models.py
├── train.py
├── predict.py
├── visualization.py
├── requirements.txt
└── README.md

---

Author

Sara Kaveh

GitHub: https://github.com/sara-kaveh