Fine-Grained Sentiment Analysis Approach on Customer Reviews Based on Aspect-Level Emotion Detection

Abstract

In the era of digital platforms, customer reviews constitute a vital resource for understanding user sentiment and perception toward products and services. Traditional sentiment analysis methods predominantly operate at the document or sentence level, often missing fine-grained emotional cues tied to specific product or service aspects. To address this limitation, this study proposes a novel Fine-Grained Sentiment Analysis (FGSA) framework that performs aspect-level sentiment classification using a joint learning approach. The proposed model employs a hybrid deep learning architecture that integrates transformer-based contextual encoders with Bidirectional Long Short-Term Memory (Bi-LSTM) layers. This design allows the model to capture both rich contextual semantics and sequential dependencies a combination that has not been widely adopted in existing FGSA research. Additionally, we introduce a new annotated dataset of 5,000 customer reviews spanning multiple domains (electronics, food and beverages, and general services), enabling robust training and evaluation. Experimental results show that the model outperforms standard baselines, achieving an F1-score of 82.0% for aspect extraction and an accuracy of 79.8% for sentiment classification. Further analysis reveals consistent patterns, such as positive sentiments linked to design and quality, and negative sentiments associated with customer service and delivery. These insights highlight the practical value of aspect-level sentiment modelling. The key contribution of this work is the integration of a transformer-Bi-LSTM joint architecture for aspect-based sentiment analysis, supported by a domain-diverse benchmark dataset. This framework enhances the interpretability and granularity of sentiment insights and sets a foundation for future research in multilingual and multimodal contexts.