One of key problems in implementing a dynamic interface between human and agents is how to do semantic mapping from natural language questions to OWL. The paper views the task as a two-class classification problem. A pair of question variable and OWL element is a sample. Two classes of "Matched" and "Unmatched" explain two relations between the question variable and the OWL element in a given sample. Building appropriate semantic mapping is the same as classifying the sample to a "Matched" class by an effective machine learning method and a trained model. Two types of features of samples are selected. Syntactical features denote the syntactical structure of a given sample. Semantic features present multiple relations between the question variable and the OWL element in one sample. Preliminary experimental results show that the sum precision of the learning-based model is better than that of the constraints-based method.