Automatic citation metadata extraction is an important aspect of digital library development. The previous methods which using hidden Markov models to extract citation metadata mostly need label many training data manually. To save the high cost of labeling training data manually, this paper describes a method for citation metadata extraction using hidden Markov models. This method use unlabeled data (plain texts which we want to extract metadata) as training data. The results of experiment show that our method has good performance in precision and recall.