A Lightweight Approach for Sentiment Analysis Using Data Augmentation, Aspect-Based, and Priority-Based Strategies

This Sentiment analysis, which is a vital task in NLP, strives to identify the emotional tone contained in text-based data. Despite the fact that deep learning architectures such as BERT have set high performance, they tend to be computationally expensive and complicated by class imbalance issues in data. In this study, we introduce a light yet powerful method for improving sentiment analysis performance by a hybrid strategy combining data augmentation, aspect-based, and priority-based methods. Data augmentation techniques such as synonym replacement and back-translation are utilized to counteract class imbalance and increase dataset diversity. In addition, aspect-based sentiment analysis (ABSA) is added specially for the restaurant review dataset to achieve more fine-grained and context-sensitive sentiment classification. In contrast to previous BERT-based hybrid models, we rely on a Decision-Based Recurrent Neural Network (D-RNN) only, which retains high accuracy at the cost of less computational overhead. Experimental verification on a combined dataset that includes Twitter tweets, Amazon product reviews, and restaurant reviews proves the efficacy of the introduced technique. Comparative analysis with techniques like LSTM, Bi-LSTM, CNN, and GRU further establishes the superiority and generalization ability of the introduced method. This work opens the door to effective sentiment analysis in low-resource settings without sacrificing model performance.