Aspect-level sentiment classification determines the emotional polarity from sentence towards a specific aspect word. When using the global attention mechanism to calculate the attention scores of attribute words and other words of the sentence, the model will focus on words that are not related to the attribute words, and pay insufficient attention to long-distance dependent words and negative words, and cannot detect side-by-side relationships or phrases. To solve these problems, this paper proposed a semantic enhancement model based on the self-attention mechanism(SRSAM). The model first used the bidirectional long short-term memory model to obtain the text encoding, and then used the self-attention mechanism to calculate the multiple semantic encodings of the text encoding. Finally, it used the attribute words and semantic coding to interact to determine the emotional polarity of the attribute words in the sentence. The experiment on the SemEval 2014 dataset show that, since the model can find long-distance dependence and negative words, it has a certain detection effect on the parallel relationship and the phrase, and the classification accuracy is 0.6%~1.5% higher than the basic model.

[2020-11-05] Printed Article /doc/2019.07.0259/v2