Aspect-Based Sentiment Analysis: Evolution of Techniques and Pathways for Future Research

The importance of sophisticated sentiment analysis techniques capable of extracting intricate understanding of the sentiments and opinions of users has grown with the rise in textual data on the internet, both social media and e-commerce reviews. The traditional sentiment analysis that most often only operates on a document or sentence level often fails to provide the detail needed to understand a specific issue about a product or service. This is tackled by Aspect-Based Sentiment Analysis (ABSA), or Aspect-Based Opinion Mining (ABOM), that correctly finds textual aspects and determines the sentiment on each. Crowding out relevant words and classifying their moods is an important aspect of ABSA. The field has experienced substantial improvements in methodology such as rule-based systems with linguistic features, advanced deep learning models, and unsupervised machine learning (e.g. Latent Dirichlet Allocation, Latent Semantic Analysis). It is worth noting that BERT based architectures have done a commendable job and have achieved good F1 scores on numerous sentiment classification tasks. However, it is found that research has persistent problems, including overreliance on a few benchmark datasets such as SemEval, and a blatant absence of dataset and domain diversity. All these issues together restrict the strength and external applicability of the existing ABSA solutions to a variety of real-life problems. The paper presents an easy-to-understand overview of the present and prospective developments of ABSA through the review and comparison of recent studies within the discipline and assessing a wide range of models, datasets, measures, and issues.