Unified Binary Aspect-Based Sentiment Analysis with Transformer Models: A Cross-Domain Evaluation of Models

Abstract

Aspect-Based Sentiment Analysis (ABSA) is a Fine-grained Sentiment Classification task that aims to detect the sentiment polarities expressed towards specific aspects in the text. ABSA has also become prevalent in applications of product feedback mining, brand monitoring and recommendation systems due to the growth of user-generated feedback spread on social media posts. The conventional forms of deep learning, like ATAE-LSTM and TD-LSTM, have worked sufficiently but have constraints because of shallow context perception and the inability to provide long-term dependencies. The examination performs and benchmarks the utilization of current transformer-based structures, Bidirectional Encoder Representations from Transformers (BERT), Robustly Optimized BERT Pretraining Approach (RoBERTa)and Decoding-enhanced BERT with Disentangled Attention (DeBERTa) to binary ABSA classification, i.e., distinguishing between positive and negative sentiments but ignoring neutral labels. The approach will be to encode aspect terms by using special tokens and fine-tuning models using benchmark data such as SemEval (Laptop & Restaurant) as well as the Multi-modal Aspect-based Sentiment Analysis (MAMS) dataset. In addition, a comparative literature survey is made among the recent models, including multimodal advanced structures, which confuse text and visual information. Benchmarking experimental testing (SemEval-Laptop, SemEval-Restaurant, and MAMS) demonstrates that DeBERTa-base had an accuracy of 92.45 and an F1-score of 92.44, compared with RoBERTa-base with 91.89 accuracy and BERT-base with 91.05 accuracy.