Abstract: Automatic abstracting is an important direction in the area of natural language processing. The purpose of this technique is to explore the mechanism of acquiring and abstracting information from natural language texts, and then the programs which can automatically write abstracts will improve the efficiency of information retrieval and spread. A new abstracting method based on text multilevel dependency structure is presented in the paper. The new method is neither superficial as the mechanic method, nor limited as the understanding method. The formal description of the text multilevel dependency structure is given and it is proved that the text multilevel dependency structure is very suitable for automatic abstracting. Also presented are the methods of text structures recognition, reduction, and abstract generation from the compressed structure. The experiments show the expected results obtained, and the feasibility and advantage of the new abstracting method is validated.