A Comparative Study of Arabic Sentiment Swap Models for AraSentEval 2026

Abstract

Sentiment swap is a controlled text generation task that rewrites a sentence by inverting its sentiment polarity while preserving semantic content and fluency. In this paper, we present our system for AraSentEval 2026 Subtask 2 on Arabic sentiment swap, a particularly challenging problem due to Arabic’s rich morphology and dialectal variation. We investigate multiple modeling paradigms, including encoder–decoder and multilingual approaches, and propose an enhanced system that combines targeted data augmentation and ensemble learning. Specifically, we augment underrepresented dialectal patterns to improve robustness and ensemble two Arabic-focused sequence-to-sequence models, AraBART and AraT5v2. Experiments are conducted on the MA’aks parallel dataset under fine-tuned settings. Our system ranked first in AraSentEval 2026 Subtask 2, achieving a BLEU score of 43.0, chrF of 65.36, and sentiment preservation accuracy of 0.7554. The results demonstrate that dialect-aware augmentation together with model ensembling substantially improves sentiment-controlled generation in Arabic and establishes strong baselines for future research in low-resource sentiment manipulation. Keywords: Arabic NLP, sentiment swap, style transfer, AraSentEval, text generation

Resources

Details