Multi-head attention model for aspect level sentiment analysis

Aspect level sentiment classification task requires topical polarity classification for different description aspect. There is a polysemy in the same vocabulary, and the emotional polarity is different for different objects. Word embedding can capture semantic information but cannot adapt to the polysemy. Attention mechanism has achieved good performance in the above tasks; however, it is only able to get the degree of association between words and unable to get detailed descriptions. In this paper, the ELMOs model is used to adjust the polysemy of the word. The Transformer model is used to extract the features with the highest degree of relevance to the target object for emotional polarity classification. Our work contribution is to overcome the polysemy interference, and use the attention mechanism to model the network relationship between words, so that the model can extract important classification features according to different target words. Experiments on laptop and restaurant datasets demonstrate that our approach achieves a new state-of-the-art performance on a few benchmarks.