Abstract

Aspect-Based Sentiment Analysis (ABSA) is an opinion mining method that finds fine-grained opinions, which are aspect terms and their associated sentiment polarities in user-generated text. Current methods are characterized by high computational cost, lack of sensitivity to subwords and error propagation between sequential subtasks. It is suggested in this paper that a new lightweight and end-to-end ABSA architecture, which includes three closely linked components (i) a Context-Augmented Subword Embedding (ii) mechanism to augment static embeddings with contextual bias, polarity modifier cues, and domain-relevant reinforcement, (iii) a Hierarchical Multi-Scale Convolutional Neural Network (HMSCNN) to learn a wider range of local, intermediate, and long-range sentiment dependencies, using multi-dilated parallel convolutions. The extensive experiments with the SemEval-2014 Restaurant and Laptop benchmark datasets have shown that the proposed framework will run to achieve a Span-level Aspect Term Extraction (ATE) F1-score of 83.6%, Attention-based Sentiment Classification (ASC) accuracy of 86.4% with Macro-F1 of 84.1% and Joint F1-score of 76.2%, outperforming Convolutional Neural Networks-Recurrent Neural Networks (CNN-RNN) and recurrent baselines by 4.4%–13.3%. The additive contribution of each component is confirmed with aid of ablation studies. These findings indicate that the suggested framework can provide a good compromise between representational richness and computational efficiency and can be used in real-time and resource-constrained fine-grained sentiment analysis tasks.