In relation classification, previous work focused on either whole sentence or key words, meeting problems when sentence contains noise or key words are extracted falsely. In this paper, we propose coarse and fine-grained networks for relation classification, which combine sentence and key words together to be more robust. Then, we propose a word selection network under shortest dependency path (SDP) supervision to select key words automatically instead of pre-processed key words and attention, which guides word selection network to a better feature space. A novel opposite loss is also proposed by pushing useful information in unselected words back to selected ones. In SemEval-2010 Task 8, results show that under the same features, proposed method outperforms state-of-the-art methods for relation classification.