From Methodologies to Metrics: A Review of Aspect-Based Sentiment Analysis Approaches

Abstract: ABSA (Aspect-Based Sentiment Analysis) has been developed as a fine-grained sentiment analysis tool, which finds the sentiment towards a particular aspect, enabling more accurate sentiment mining in a variety of domains. Over the past decade ABSA research has transcended lexicon-driven and traditional machine learning methodology using deep learning and transformer-based pre-trained language models to generative large language models. Nevertheless, underlying issues remain: implicit aspect extraction, low cross-domain and cross-lingual robustness, dataset imbalance, and interpretability concerns of complex neural networks. In addition, the rapid scaling of ABSA subtasks has led to some fragmentation in methodological advances in earlier investigations. By methodically reviewing the development of methodological paradigms, benchmark datasets, and evaluation approaches, this review has offered a systematic and rigorous assessment of the literature on ABSA. Unlike previous reviews, the study adopts a holistic, task-aware view and makes a direct connection between ABSA subtasks and the accompanying modeling methodologies. The review explores new research directions such as explainable ABSA, meta-based learning frameworks, multilingual and low-resource modeling, and large language model integration, thus providing a structure toward the road to developing more resilient, interpretable, and generalizable ABSA systems.