🇩🇿 Darija Sentiment Analysis — Algerian Darija Sentiment Classifier

This project is an end-to-end Sentiment Analysis system for Algerian Darija (الدارجة الجزائرية)
using a custom dataset of 105,000 labeled samples split into train, validation, and test.

The goal is to build a complete pipeline that cleans raw Algerian text, visualizes the data,
trains a deep learning model, and provides a real-time prediction API connected to a React interface.

---

🧹 Data Preprocessing

A full preprocessing pipeline is implemented to clean Algerian Darija comments.

✔ Includes:

• Removing URLs, mentions, emojis, punctuation, and non-Arabic characters
• Normalizing Arabic letters and removing diacritics
• Lowercasing text
• Removing repeated characters (e.g., "راااااائع" → "رائع")
• Removing Arabic + Algerian stopwords
• Tokenizing text with Keras Tokenizer (20k vocab)
• Padding sequences to a fixed length
• Saving cleaned datasets and tokenizer for training

The preprocessing script generates:

• cleaned train/val/test datasets
• tokenizer.pkl
• text statistics

---

📊 Data Visualization (Before & After Preprocessing)

Two analysis scripts visualize the dataset and compare text before and after cleaning.

Visualizations include:

• WordCloud for raw data
• WordCloud for cleaned data
• Class distribution (0 / 1 / 2)
• Text length distribution
• Sample comparisons before/after cleaning

These plots help understand the dataset and verify that preprocessing improves consistency.

---

🧠 Model Architecture (CNN + Word2Vec)

The model used is a deep 1D Convolutional Neural Network with pretrained Word2Vec embeddings.

Final Architecture:

• Embedding Layer
  • Word2Vec vectors (200 dimensions)
  • Trainable
• SpatialDropout1D (0.2)
• Conv1D Layer (128 filters, kernel=3, ReLU, Same padding, L2 regularization + BatchNorm + Dropout 0.2)
• Conv1D Layer (128 filters, kernel=5, ReLU, Same padding, L2 regularization + BatchNorm + Dropout 0.2)
• GlobalMaxPooling1D
• Dense Layer (64 units, ReLU + BatchNorm + Dropout 0.2)
• Dense Layer (32 units, ReLU + BatchNorm + Dropout 0.2)
• Output Layer (3 units, Softmax)

📌 Training:

• Optimizer: Adam
• Loss: Categorical Crossentropy
• Callbacks: EarlyStopping, ReduceLROnPlateau
• Accuracy & loss curve plots saved during training

---

🌐 Backend API (Flask)

A lightweight REST API is provided to serve real-time sentiment predictions.

Server Features

• Loads trained sentiment_cnn_model.h5 and tokenizer.pkl
• Cleans input text using the same preprocessing pipeline as during training
• Returns both predicted class and confidence score
• Handles single or batch predictions

🖥️ Frontend (React)

The frontend is built with React and provides a user-friendly interface for interacting with the Flask backend.

Frontend Features

• Input box for Algerian Darija comments
• Predict button
• Live display of predicted sentiment and confidence
• Character & word counters
• Mobile responsive layout
• Frontend Repository

Frontend Repository

https://github.com/Ynsaid/Darija_Sentiment_Analysis_Frontend

Test live website

https://darija-sentiment-analysis-frontend-1.onrender.com/