Labeling text data is quite time-consuming but essential for automatic text classification. Especially, manually creating multiple labels for each document may become impractical when a very large amount of data is needed for training multi-label text classifiers. To minimize the human-labeling efforts, we propose a novel multi-label active learning appproach which can reduce the required labeled data without sacrificing the classification accuracy. Traditional active learning algorithms can only handle single-label problems, that is, each data is restricted to have one label. Our approach takes into account the multi-label information, and aims to label data which can optimize the expected loss reduction. Specifically, the model loss is approximated by the size of version space, and we optimize the reduction rate of the size of version space with Support Vector Machines (SVM). Furthermore, we design an effective method to predict possible labels for each unlabeled data point, and approximate the expected loss by summing up losses on all labels according to the most confident result of label prediction. Experiments on seven real-world data sets (all are publicly available) demonstrate that our approach can obtain promising classification result with much fewer labeled data than state-of-the-art methods.