Phrase boundary location provides an important information for bracketing and tagging the phrase automatically．This paper describes an experimental model for the automatic prediction of the phrase boundary location．It consists of three processing stages：first，automatically identify the phrase boundaries using statistics from treebank；then，post—tune the results using local tuning rules generated by an error—driven the machine learning method；at last，refine the results of the last two stages with the overall tuning rules summarized by man．Experimental results on a corpus of 1 434 sentences demonstrate a high rate of the success for predicting the phrase boundary(96.33％correct the prediction for the close testing and 94.54％correct the prediction for open testing)．