Abstract: Aiming at the problem of poor classification effect caused by only using semantic-level features in fine-grained emotion analysis task, a fine-grained emotion analysis model based on word-level and semantic-level attention was proposed. When applied to SemEVAL-2014 Task4 dataset, the accuracy of this model is improved by 3.51% and 2.17% compared with AEN-BERT and BERT-SPC models, respectively. Compared with the model using Glove word vector, the accuracy of this model is improved by 3%-4%. The introduction of word-level features in fine-grained sentiment analysis tasks can further enrich text feature representation, and the weight of words related to sentiment analysis can be strengthened through the attention mechanism to improve the accuracy of fine-grained sentiment analysis.