ikraam05/dziri-sentiment-analysis

Sentiment Analysis on Algerian Dialect using DziriBERT # Algerian Dialect Sentiment Analysis using DziriBERT This project focuses on sentiment analysis for Algerian dialect (Darija) tweets using the Twifil dataset and DziriBERT transformer model. --- # Project Objectives - Analyze Algerian dialect tweets - Perform dialect-specific preprocessing - Fine-tune DziriBERT for sentiment classification - Compare balancing strategies: - SMOTE - ADASYN - back transaction - Focal Loss - class weighting - focal loss + class weighting - focal loss + back transaction # Dataset Dataset used: Twifil Dataset Contains: - Algerian dialect tweets - Arabic - French - Arabizi - Code-switching # Model Base model: DziriBERT # Evaluation Metrics -Accuracy -F1-score -F1-macro -Recall -G-Mean -Confusion Matrix # Main Results Model F1-Macro Baseline --> 0.68 SMOTE --> 0.64 Focal Loss --> 0.68 # Main observation: Advanced balancing methods did not significantly improve performance due to the linguistic complexity of Algerian dialect and ambiguity of Neutral tweets. # Technologies Used -Transformers -HuggingFace -Scikit-learn -PyTorch -Matplotlib -Seaborn # Author Ikram Yadel --- # 📄 requirements.txt ```txt transformers datasets torch scikit-learn imbalanced-learn pandas numpy matplotlib seaborn wordcloud emoji