NAT-MTT: Noise-Aware Multi-Task Transformers for Cross-Domain Aspect-Based Sentiment Analysis

This paper proposes a noise-aware multi-task transformer framework that jointly performs aspect extraction (AE) and aspect sentiment classification (ASC) using a shared BERT/RoBERTa encoder with dual task-specific heads. Robustness is promoted through a systematic noise-aware training (NAT) strategy that injects controlled synthetic perturbations (spelling errors, word dropout, synonym replacement, slang) according to a curriculum schedule, mixing clean and noisy instances in each batch. Experiments on SemEval-2014 (Restaurants, Laptops) and large-scale Amazon (Electronics, Apparel) and Yelp (Food) reviews demonstrate consistent gains over strong single-task, multi-task, and cross-domain baselines in in-domain, cross-domain, and multi-domain settings. On the SemEval Rest 14 dataset, the proposed model achieves improvements of +2.6 F1 (AE) and +2.9% accuracy (ASC) on the Rest14 benchmark over the strongest baseline. with maximum gains of +3.1 F1 (AE) and +3.3% (ASC) over the strongest cross-domain baseline (BGCA), reduces noise-induced performance degradation by up to 42% (NAT contribution vs. identical model without NAT), and improves cross-domain transfer with minimal additional parameters and a 22% training-time overhead. Ablation and error analyses show that multi-task learning and NAT are both critical to robustness, particularly under high noise and domain shift. These findings indicate that jointly learning multiple ABSA subtasks with noise-aware augmentation is an effective and efficient route to deployable, real-world ABSA systems.