Aspect-level sentiment classification with emotional keywords attention network

Aspect-level sentiment classification (ASC) aims to uncover the sentiment polarity expressed towards specific entities in market analysis.Despite its popularity, two fundamental obstacles persist: 1) manual annotation is notoriously labour-intensive, resulting in data-scarce regimes where models struggle to acquire sufficient knowledge; 2) accurate sentiment inference demands rich prior semantic knowledge that is difficult to acquire even when large corpora are available.To address these challenges, we first automatically identify emotion-bearing keywords for each aspect and summarise their distributional properties.We then introduce a convolutional neural network enhanced with an attention mechanism that explicitly highlights these keywords through an emotion-keyword mask, thereby guiding the model to focus on sentiment-relevant context.Extensive experiments on standard SemEval benchmarks show that the proposed mask consistently improves performance and that our overall approach achieves competitive results against recent strong baselines.