Summary: A question classification algorithm based on semantic kernel function is proposed. This algorithm constructs Support Vector Machine (SVM) kernel function based on the grammatical structure of the question. Firstly, the given question is parsed into syntactical structural tree, and then sub-trees of syntactical tree are used to represent the question. Secondly, features are extracted from three aspects of the question: lexical, syntactical and semantic, to form a richer feature space. Thirdly, the kernel function is constructed based on syntactical structural tree of the question. Finally, using the potential semantic indexing method and the lexical, grammatical and semantic features of the question, the feature space is mapped into a more efficient space by the semantic kernel. The experimental results on the TREC dataset show that the classification accuracy can be improved by lexical, grammatical, and semantic enhancement.