The emotional state of athletes is crucial to the level of competition and training, especially during the competition, the emotional fluctuations of athletes have a direct impact on the performance of the competition. Recognizing the emotions of athletes through their social media data can detect the fluctuations in their emotions in a timely manner, which can be used to adjust the training plan or intervene early to avoid affecting the competition results. Using machine learning and deep learning techniques, artificial intelligence can analyze multiple data forms such as text, voice, image and video to identify and understand the emotional state of athletes. In this paper, we propose a multi-granularity sentence sentiment analysis method, which constructs the whole sentence sentiment analysis model DABLSTM-L1 by fusing sentiment word vectors and high-level semantic features, and constructs the interactive attention sentence sentiment recognition model Att-CNN-BLSTM to extract the interactive sentiment features of the whole sentence and local sentences, and finally fuses the whole sentence sentiment analysis model DABLSTM-L1, local sentence sentiment analysis model and interactive attention sentence sentiment classification model Att-CNN-BLSTM to predict the sentiment polarity of text. The experimental results show that multi-granularity sentence sentiment analysis can effectively identify the emotional state of athletes.