Abstract: Online social networks have gradually become popular and popularization. A number of social networks such as microblog have formed a unique form of literary and emotional expression. Because the expression of microblog is different from the expression of traditional articles, the sentiment analysis research based on short-text machine learning has become more and more difficult. Aiming at the new features of Microblog short text language expression, we crawl and collect a large amount of non-emotionally labeled Microblog data, and build a Microblog short text corpus to create a global relationship graph between words and short texts. The BERT (Bidirectional Encoder Representations from Transformers) document embedding is used as the feature value of the graph node, and graph convolution is used for feature transfer and feature extraction between nodes. We manually annotate non-emotionally labeled Microblog data which sample from the whole Microblog short text corpus. A semi-supervised machine learning method combined with global relationship graph is proposed to improve the performance of sentiment classifier. Experiments show that by increasing the proportion of unmarked data, the method can better capture global features and improve the accuracy of sentiment classification. Comparative experiments are carried out on self-built artificial labeling data, COAE2014 data set and NLP&CC2014 data set. The experimental results show that the method has a good performance in accuracy and recall.