This paper presents our system and findings for AraSentEval 2026 Subtask 1 on Arabic Dialect Sentiment Analysis. We propose an automated sentiment classification system grounded in advanced Natural Language Processing (NLP) techniques. The proposed approach leverages pre-trained Transformer-based architectures to categorize textual inputs into three sentiment polarities: positive, negative, and neutral. Initially, a text normalization procedure is applied to unify the orthographic and graphical variations characteristic of the Arabic language. This process is further complemented by repetition reduction techniques, which aim to mitigate textual noise and enhance the overall consistency of the data. Subsequently, the data are adapted to the requirements of the pre-trained models to ensure coherent tokenization. The processed texts are then encoded into numerical representations that serve as inputs during training and evaluation. Finally, we conduct a comprehensive benchmarking study of five Transformer-based architectures to assess their effectiveness. The best-performing experimental setup yielded remarkable results on the AraSentEval 2026 benchmark, achieving a micro-F1 score of 75.96% on the official test set.

This paper presents our system for the AraSentEval 2026 shared task on Arabic dialect sentiment analysis. We propose a multi-model ensemble combining AraBERTv2 and CAMeLBERT with supervised contrastive learning to improve sentiment classification. The system incorporates dialect-aware preprocessing, class-weighted cross-entropy loss with label smoothing, supervised contrastive loss for enhanced sentence representations, and rule-based post-processing for dialect-specific patterns. Our approach achieves a macro F1-score of 0.83 on the official test set, demonstrating the effectiveness of contrastive learning with pretrained Arabic language models for dialectal sentiment analysis.

This paper describes our system submitted to the AraSentEval 2026 Shared Task, Subtask 2: Arabic Sentiment Swap. The task requires rewriting Arabic sentences to invert their sentiment polarity while preserving the core meaning. We propose a multi-step pipeline approach that uses large language models (LLMs). Our method decomposes the sentiment inversion problem into three stages: (1) sentiment expression extraction, where the model identifies all sentiment-bearing words and phrases in the input sentence; (2) opposite expression generation, where each identified expression is replaced by its semantic opposite; and (3) sentence reconstruction, where the final output is assembled to ensure grammatical correctness and natural fluency. Our system achieves 74.3% sentiment style accuracy, 27.22 BLEU, and 55.04 chrF on the official test set.