Embodiments of a dialog system that utilizes grammar-based labeling scheme to generate labeled sentences for use in training statistical models. During the process of training data development, a grammar is constructed manually based on the application domain or adapted from a general grammar rule. An annotation schema is created accordingly based on the application requirements, such as syntactic and semantic information. Such information is then included in the grammar specification. After the labeled grammar is constructed, a generation algorithm is then used to generate sentences for training various statistical models.