An explicit multi-aspect sentiment analysis using bidirectional encoder representation transformers

Abstract

Aspect-based sentiment analysis (ABSA) is a vital natural language processing (NLP) task that extends traditional sentiment analysis by identifying sentiments associated with specific aspects within a sentence. Although recent deep learning (DL) models have improved ABSA performance, most concentrate on extracting only a single aspect and predicting sentence-level polarity, thereby overlooking multiple explicit aspects and their individual sentiments. This reveals a research gap in fine-grained sentiment detection. The main problem addressed in this work is the absence of robust models capable of extracting multiple explicit aspects and accurately predicting aspect-wise polarity. To address this, the primary objective was to design a model supporting multi-aspect extraction and sentiment classification. Accordingly, this work proposes explicit aspect-aware bidirectional encoder representations from transformers (EA-BERT), a BERT-based architecture enhanced with a multi-head attention (MHA) mechanism to capture aspect-specific contextual features and predict their corresponding polarities. The model was trained and evaluated on the SemEval 2014 restaurant (Res2014) and laptop (Lap2014) datasets. EA-BERT achieved accuracies of 90.86% and 90.23% and macro-F-scores of 91.89% and 90.28%, respectively. Experimental comparisons demonstrate that EA-BERT consistently outperforms several recent ABSA models, confirming its effectiveness for multi-aspect sentiment classification.