AN ATTENTION-BASED RECURRENT NEURAL NETWORK MODEL FOR IMPROVED ASPECT-BASED SENTIMENT ANALYSIS OF CUSTOMER REVIEWS

Abstract

Sentiment Analysis involves the process of identifying and classifying the underlying opinion or attitude expressed in a piece of text, typically categorizing it as positive, negative, or neutral. Aspect-based sentiment Analysis targets the extraction of sentiments directed toward specific aspects within a review, offering more granular insights than assigning a single sentiment to an entire text. Conventional methods assign a single sentiment label to an entire review but often fail when the text contains mixed opinions about different elements. Deep learning models emerged to address this gap by learning contextual dependencies within sequences, offering improved performance over rule-based approaches. However, despite their strengths, these models often struggle to retain relevant information over long distances, leading to issues such as diluted sentiment signals and reduced sensitivity to aspect-specific content. To overcome these shortcomings, attention mechanisms were integrated to help models focus on crucial sentiment-bearing tokens and better align input elements with aspect terms. However, they still face challenges such as; a limited understanding of complex multi-aspect sentiment where models often fail to accurately interpret multiple opinions spread across different parts of a review, lack of aspect-specific focus as many models treat all words uniformly missing critical phrases linked to specific product features and loss of context in long sequences. A true experimental research design was adopted in this study. The first objective was to compare four deep learning models namely: Recurrent Neural Networks (RNN), Convolutional Neural Network (CNN), Recursive Neural Network (RecNN), and Memory Network(MN). The performance of these models on the SemEval Task 4 dataset was compared without attention mechanisms and with attention mechanisms included. From the performance comparison, RNN integrated with the attention mechanism recorded the highest accuracy score of 84.7%. The second objective was to improve on the above performance index by integrating pretrained multi-head attention and dual attention mechanism into the RNN model, aiming at solving the reported limitations of deep learning models in ABSA. The third objective was to evaluate the performance of the improved model on the SemEval Task 4 and MAMS dataset. On the SemEval Task 4 and MAMS datasets, the improved model achieved accuracy scores of 89.1% and 88.5%, F1-scores of 87.8% and 86.2%, precision scores of 89.6% and 87.3%, and recall scores of 86.7% and 84.2%, respectively. These results indicate consistent improvements across all standard classification metrics compared to the baseline RNN + Hierarchical Attention model. On average across both datasets, the improved model recorded performance gains of 4.1% in accuracy, 3.8% in F1-score, 3.0% in precision, and 4.5% in recall, highlighting the effectiveness of the applied enhancements in improving sentiment classification. These findings demonstrate that the integration of multi-head attention and dual attention mechanisms significantly improved the classification of aspect-specific sentiments in aspect-based customer reviews. The improved model can be deployed in e-commerce platforms, customer feedback systems, and other domains that require detailed, aspect-level sentiment insights to improve products and decision-making.