Federated Fine-Tuning of DziriBERT for Algerian Dialect Sentiment Analysis

This study examines federated learning for sentiment analysis in Algerian Arabic and code-switched social media content using transformer-based language models. A corpus of 13,260 user-generated comments collected via the X platform API with balanced positive and negative labels is preprocessed, stratified into train, validation, and test splits, and divided into ten highly non-IID clients using a Dirichlet procedure. We first fine-tune DziriBERT centrally as a solid baseline, obtaining roughly 86.35% accuracy and 86.35 macro-F1 on the held-out test set. We then implement FedAvg, FedProx, and a semisupervised FL variant with only 20% labeled clients using DziriBERT, and evaluate multilingual DistilBERT with FedAvg as an alternative backbone. FedAvg achieves 84.77% accuracy with DziriBERT, FedProx improves to 85.41%, semi-supervised FL achieves 80.69% with limited labels, and DistilBERT FedAvg reaches 81.37% accuracy. All results are based on a single random seed and partition.