A Simple Fusion of Self-Attention and BERT for Aspect-Level Sentiment Classification

Aspect-level sentiment classification is aimed to identify the sentiment polarity of a user with respect to certain aspects (features) of a product, and fine-grained sentiment judgements can provide more accurate recommendation results for a wide range of users. Existing deep learning methods have achieved good performance on aspect-level sentiment classification tasks. However, some approaches ignore the modelling of comment semantics or key aspect terms, resulting in ineffective integration of the overall semantics of user comments and the specific sentiment attributes of aspect terms. In this paper, we propose a novel aspect-level sentiment classification approach which uses the pre-trained model to learn high-level semantic features of the input text, employing the self-attention mechanism to capture aspect word feature representations of text sequences. Meanwhile, we introduce dynamic sentiment weight in the model architecture to guide the model to effectively fuse the high-level features of the input sequences and aspect words during the training process, thus obtaining the final feature representations of the input sequences. Experimental results on 2 datasets demonstrate that our proposed method significantly outperforms other baseline methods, proving its effectiveness and superiority.