Abstract: The original Tri-training algorithm classifies the labeled data by the method of random sampling,forming three training sets for three classifiers.There is an phenomenon that the number of different categories may have huge differences between the exiting labeled data sets in this training sets formed by random sampling three classifiers, which may lead the categories of training sets into imbalance, and influence the accuracy of classifier.By employing a method of classification sampling to replace the random sampling, Tri-training algorithm was improved and a classification model was established. Classification experiment were performed on HIT question set and expanded question set. The results were compared with those of original Tri-training algorithm on the same data sets, which indicates that the new algorithm has good adaptability, and the accuracy of the algorithm is improved. With the increase of training set and the number of unlabeled samples, the generalization ability and the accuracy of the classifier are improved.