Arabic NLP Project

An Arabic NLP project that explores text classification, question answering, translation, retrieval-augmented generation, preprocessing strategies, traditional machine learning models, transformer models, and an interactive Streamlit/FastAPI interface.

The project focuses on Arabic educational question-answering data and compares how different preprocessing methods, feature representations, and models affect performance.

---

Interface Demo

---

Traditional ML Experiment Overview

This visualization shows all traditional model combinations tested in the project:

Dataset → Word Embedding / Feature Representation → Model → Macro F1

The plot summarizes how different Arabic preprocessing strategies, embedding methods, and classifiers affected the final Macro F1 score.

---

Project Tasks

Task	Description
Arabic Text Classification	Classifies Arabic questions into predefined categories
Arabic QA Generation	Generates answers for Arabic questions
Machine Translation	Translates Arabic question-answer data
RAG Question Answering	Uses retrieval-based context for answering questions
Model Comparison	Compares traditional ML, Seq2Seq, and transformer-based approaches
Interface Deployment	Runs a FastAPI backend with a Streamlit frontend

---

Main Experiments

1. Arabic Preprocessing

Several preprocessing versions were tested, including:

• Original raw Arabic text
• PyArabic preprocessing
• Hamza normalization
• Tashkeel handling
• Tatweel removal
• Punctuation-focused preprocessing
• Regex aggressive preprocessing

The goal was to test whether Arabic text normalization improves classification and generation performance.

---

2. Traditional Machine Learning Classification

Traditional ML models were trained using multiple feature representations:

Feature Type	Description
BoW	Bag-of-Words sparse word-count representation
TF-IDF	Term frequency-inverse document frequency
Word2Vec-CBOW	Dense word embeddings trained using CBOW
Word2Vec-SG	Dense word embeddings trained using Skip-Gram
FastText	Subword-aware word embeddings
BERT	Transformer-based embeddings
GPT	Transformer-based embeddings

Models compared:

• Linear SVM
• XGBoost
• Multinomial Naive Bayes

---

3. Transformer Classification

Transformer-based classification was tested to compare deep contextual embeddings against traditional ML methods.

---

4. Question Answering

The project also includes Arabic question-answering experiments using:

• Seq2Seq models
• Transformer-based generation models
• RAG-based retrieval and answering

---

System Architecture

Arabic Question
   ↓
Preprocessing
   ↓
Classification / QA / RAG Pipeline
   ↓
Model Inference
   ↓
Prediction or Generated Answer
   ↓
Streamlit Interface

---

Running the Project

Install dependencies:

pip install -r requirements.txt

Start the backend API:

uvicorn app_api:app --reload

Start the Streamlit interface:

streamlit run app_ui.py

---

Repository Structure

Arabic-NLP-Project/
│
├── app_api.py                          # FastAPI backend
├── app_ui.py                           # Streamlit frontend
├── requirements.txt                    # Dependencies
│
├── preprocessing.ipynb                 # Arabic preprocessing pipeline
├── classification_traditional_ml.ipynb # Traditional ML classification
├── classification_transformer.ipynb    # Transformer classification
├── machine_translation.ipynb           # Translation experiments
├── question_answering_seq2seq.ipynb    # Seq2Seq QA
├── question_answering_rag.ipynb        # RAG QA
│
├── AAFAQ_Dataset.csv                   # Original dataset
├── AAFAQ_Dataset_Translated.csv        # Translated dataset
│
├── classification plots/               # Classification visualizations
├── preprocessed datasets/              # Processed dataset versions
├── QA_seq2seq_outputs/                 # Seq2Seq QA outputs
├── QA_transformer_outputs_final/       # Transformer QA outputs
└── classification_seq2seq2_outputs/    # Classification outputs

---

Key Features

• Arabic text preprocessing and normalization
• Traditional ML model comparison
• Transformer-based classification
• Arabic question-answering generation
• Retrieval-augmented question answering
• Machine translation pipeline
• Streamlit user interface
• FastAPI inference backend
• Visual analysis of model performance

---

Why This Project Matters

Arabic NLP is challenging because Arabic has rich morphology, different writing forms, optional diacritics, spelling variation, and preprocessing sensitivity.

This project investigates how preprocessing choices and embedding methods affect Arabic NLP performance across multiple modeling approaches.

---

Limitations

• The dataset size limits generalization.
• Some generated answers may be semantically weak even when similarity metrics look acceptable.
• Traditional ML models depend heavily on preprocessing quality.
• Transformer models require more compute and careful fine-tuning.
• RAG performance depends on the quality of retrieved context.

---

Future Work

• Improve Arabic answer generation quality
• Add better transformer fine-tuning
• Add larger Arabic datasets
• Improve RAG retrieval quality
• Add model confidence explanations
• Dockerize the full application
• Deploy the interface publicly

---

Tech Stack

• Python
• PyTorch
• Hugging Face Transformers
• scikit-learn
• XGBoost
• FastAPI
• Streamlit
• Pandas
• NumPy
• PyArabic