Hierarchical prosody structure generation is a key component for a speech synthesis system. One major feature of the prosody of Mandarin Chinese speech flow is prosodic phrase grouping. In this paper three methods are proposed to predict prosodic phrase. The first is a statistic probability model, which efficiently combines the local POS and word length information. Experiments show by choosing appropriate threshold the model can reach a high precision and high recall ratio. Secondly we use the decision tree learning algorithm combined with the pause rules of Chinese empty words to predict prosodic phrase boundary in unrestricted Chinese text. The experiments show that the approach improves overall performance. Another is an SVM-based method. The precision and recall ratio are improved after using SVM classifier.