Fine-grained sentiment classification based on semantic extension of target word

The current existing fine-grained sentiment analysis method usually extract the context of the sentence while ignoring the semantic representation of the target words. We extent the target words in comments as the additional input parameters to the deep learning model in this paper. And the influence of the number of extended words on the model's performance is also discussed thoroughly during the experimenting process. Main procedures of our proposed fine-grained sentiment classification method can be described as: 1) firstly, target words are expanded by using the semantic distance of the word embedding, which used as the key information; 2) bidirectional LSTM neural network is used to extract the semantic information afterwards; 3) additionally, the attention mechanism is employed to learn the sentiment weight distribution of the target words among the text automatically. The experiments conducted on the SemEval 2014 Task 4 corpus showed that the proposed method outperforms the other LSTM model.