With the universal use of GPS and rapid increase of spatial Web objects, spatial keyword query has been widely used in Location-Based Services (LBS). Most of the existing spatial keyword query processing models only support location proximity and strict text matching which makes the semantically related objects cannot be provided to users and even may lead to the empty answer problem. In addition, the current index structures (such as IR-tree, Quadtree) cannot process numerical attributes which are usually contained in the descriptive information associated to the spatial objects. To deal with these problems, this paper proposes a spatial keyword query method that can support semantic approximate query processing. Firstly, the user original query is expanded by Conditional Generative Adversarial Nets (CGAN) method to generate a series of query keywords that are semantically related to the original query keywords. And then, a hybrid index structure called AIR-tree is built to facilitate the query matching, which can support the text semantic matching and process numerical attributes with Skyline method. Experimental analysis and results demonstrate that the proposed method achieves higher execution efficiency and better user satisfaction compared with the state-of-the-art methods.