Authors: Mrs. Bhumika Alte, Satyam Mali, Yashraj Mhase, Kishor Hirgal
Abstract: Aspect-Based Sentiment Analysis (ABSA) provides a fine-grained approach to understanding opin-ions by extracting aspect–opinion–sentiment relation-ships from text. It is particularly valuable in domains such as product reviews, customer services, banking, and social media, where identifying specific strengths and weaknesses is essential. The subtask of Aspect-based Sentiment Triplet Extraction (ASTE) extends ABSA by simultaneously identifying aspect terms, corresponding opinion expressions, and their sentiment polarities. This work proposes an improved ABSA framework that integrates pre-trained language models (PLMs) with a pruned syntactic encoding mechanism to efficiently capture both local and global contextual dependencies. Additionally, a dynamic encoding strategy is introduced to overcome the limitations of traditional local encod-ing, which often fails to capture long-range relation-ships between aspects and opinions. The combination of syntactic pruning and dynamic encoding enhances the association between aspect and opinion terms, leading to more accurate sentiment classification. Experimental evaluations on benchmark ABSA datasets are expected to demonstrate that the pro-posed model achieves higher accuracy and robustness compared to existing methods. This approach effec-tively combines syntactic structure and contextual un-derstanding, improving interpretability and performance in aspect-level sentiment prediction tasks.
DOI: http://doi.org/10.5281/zenodo.20787176
