Efficient and Lightweight Deep Learning Framework for Aspect-Based Sentiment Analysis

Aspect-Based Sentiment Analysis (ABSA) has emerged as a critical task in Natural Language Processing, enabling fine-grained opinion mining across domains such as e-commerce, customer service, and social media. This paper proposes a lightweight BiLSTM + Attention architecture designed to jointly perform Aspect Term Extraction (ATE) and Aspect Sentiment Classification (ASC) with high interpretability and low computational overhead. The model integrates (1) bidirectional context capture through BiLSTM, (2) token-level relevance scoring via an attention mechanism tailored to aspect terms, and (3) multi-task learning to enhance both ATE and ASC performance. Preprocessing involves BIO tagging and aspect-sentiment pairing for structured supervision. Evaluations on SemEval and MAMS datasets yield an accuracy of 81% and a macro F1-score of 81.09%, while maintaining a compact model size of 10.11 MB, demonstrating suitability for edge deployment. Confusion matrix analysis confirms robustness across sentiment classes, and attention heatmaps offer transparency in predictions. Compared to transformer-based models like BERT, our architecture achieves competitive performance with significantly reduced resource requirements. The model achieves over 80% accuracy on benchmark datasets with inference times under 100 ms and low battery usage, validating its edge-deployable design across real-world mobile and embedded systems.