Improved systems and methods for generating training data for classification models are disclosed. In an example, a training application accesses two fragments of text. The application represents each fragment of text as a parse thicket. The parse thickets jointly represent syntactic and discourse information. From the parse thickets, the application generalizes the text by identifying common entities or common rhetorical relations between parse thickets. The generalized text is added to a training data set, thereby increasing the coverage of the training set.