With the development of information technology and social media, user sentiment analysis tends to play an increasingly important role in public opinion monitoring, information prediction and product evaluation. However, collecting sufficient manual sentiment labels in supervised learning is still difficult and costly, and unsupervised learning is lack of label guidance. Therefore, a semi-supervised sentiment analysis model based on sociological theory is established in this paper, which is mainly divided into two parts: label addition and emotion analysis. First, a UR-S (User Relationship using Social relations) model was built, which was inspired by sentiment consistency and emotional contagion. Then a TRS-SAT (Text Relationship Strength using Social relations, user Attribute and Text similarities) model based on UR-S model and add labels was established. Finally, the TRS-SAT model and CNN (convolutional neural network) were combined to construct SA-SRS-CNN (Sentiment Analysis using Social Relationship Strength and Convolutional Neural Network) model. The model uses CNN to mine the deep connection between the feature set and the emotional labels to improve the emotional performance. Experiments show that the accuracy, recall, and the F value of the proposed model increased by 11.40%, 5.90% and 8.65%, respectively compared with SVM, and increased 4.12%, 4.17%, and 4.14%, respectively compared with CNN, which suggests that the model is innovative and practical and can provide a good theoretical basis for public opinion analysis.