Abstract: Topical shifts—differences between the topic distribution in training and test data—can substantially reduce the performance of non-topical text classification and regression models. In this ...