Abstract: How to leverage the abundant unlabeled data with a few labeled training examples to construct a strong classification system is a focus issue. Both semi-supervised learning and active learning attempt to exploit the unlabeled data to improve the recognition rate of supervised learning algorithms and minimize the cost of data labeling. This paper proposed two approaches,Entropy Priority Sampling (EPS) and Simple Disagreement Sampling (SDS),to select samples in active learning , which are applied into the Tri-Training algorithm as Tri-EPS and Tri-SDS methods. Several experiments with these approaches on the UCI, remote sensing image and environmental audio datasets are carried out in order to illustrate the results of the proposed methods and compare their performance with that of Tri-Training algorithm. Experimental results show that the active learning combined with semi-supervised learning can effectively improve the performance.