Abstract: A new method, which combines the corpus approach with the machine learning approach, is put forward in this paper to identify simple, non recursion noun phrases （BaseNP）. Firstly, all different part of speech （POS） strings （BaseNP rules） which are corresponding to BaseNP are extracted from the training corpus tagged with POS and the boundary of each BaseNP. By means of training and based on linguistics knowledge, some BaseNP rules which have lower precision and have no linguistics sense apparently are deleted. Secondly, the remaining BaseNP rules are employed to identify BaseNP in new sentences. In the process, a heuristic algorithm of longest match, which is combined with the machine learning method of inductive decision trees to consult contexts, is applied. Experiments show that this new method results in higher precision and recall precision.