Sentiment knowledge-induced neural network for aspect-level sentiment analysis

Aspect-based sentiment analysis has been a popular topic in natural language processing in recent years that aims to determine the sentiment polarity of a specific aspect in one context. However, most existing models only focus on feature extraction and ignore the significant role of words with sentiment tendency (e.g. good, terrible), which results in low classification accuracy. In this paper, a sentiment knowledge-based bidirectional encoder representation from transformers (SK-BERT) is proposed to overcome this shortcoming. To introduce sentiment knowledge, SK-BERT first integrates sentiment knowledge words into independent sequences and then encodes the sequence and context into static and dynamic vectors with the BERT pretrained models, respectively. All vectors are sent to the sentiment centre to generate different dimension representations for classification. We evaluate our model on three widely used datasets. Experimental results show that the proposed SK-BERT model outperforms other state-of-the-art models. Furthermore, visualization experiments are implemented to prove the rationality of SK-BERT.