As much valuable domain knowledge is hidden in enterprises' text repositories (e.g., email archives, digital libraries, etc.), it is desirable to develop effective knowledge management tools to process this unstructured data so as to extract domain knowledge for business decision making. Ontology-based semantic annotation of documents is one of the promising ways for knowledge discovery from text repositories. Existing semantic annotation methods usually require many labeled training examples before they can effectively operate, and this bottleneck holds back the widely applications of these semantic annotation methods. In this paper, we propose a semi-supervised semantic annotation method, self-teaching SVM-struct, which uses fewer labeled examples to improve the annotating performance. The key of the self-teaching method is how to identify the reliably predicted examples for retraining. Two novel confidence measures are developed to estimate prediction confidence. The experimental results show that the prediction performance of our self-teaching semantic annotation method is promising.