Appraisal expression recognition is very important in sentiment analysis. Because of the lack of labeled corpus, most former works in appraisal expression recognition are focused on construction of rules and templates manually. In order to reduce the training work of labeling corpus and further mining information of unlabeled corpus, a new algorithm based on co-training was proposed, which mainly used massive unlabeled corpus and only a small number of labeled corpus. The proposed algorithm was based on Tri-training and combined Support Vector Machine (SVM), Maximum Entropy (MaxEnt) and Conditional Random Field (CRF) to build a new approach for candidate appraisal expression classification. By comparing the Tri-training based algorithm with the former single classifier based algorithms, the former can effectively improve the performance of appraisal expression recognition in subjective sentences.