Enhancing Aspect Sentiment Analysis through Local and Global Context Fusion

Aspect sentiment analysis, a task in text classification, predicts sentiment polarity regarding specific aspects within text. Recent research typically emphasizes global context modeling using attention mechanisms and external semantic knowledge. However, balancing local and global context is vital as overemphasizing global context increases model complexity. To address this, we propose the “Local and Global Feature Fusion Network model,” using a multi-head attention mechanism. Our approach involves initial context encoding using bidirectional gated recurrent units (GRUs). We employ semantic distance-based masking to filter less relevant words, creating local context representations. Additionally, a multi-head Aspect-aware attention network independently extracts features from local and global contexts. Pretrained BERT models enhance performance. Experiments on Twitter, Laptop, and Restaurant datasets, evaluated using Accuracy and Fl-score, demonstrate that our model outperforms other aspect-based sentiment classification algorithms with a more compact parameter size.