Performance Impact Caused by Hidden Bias of Training Data for Recognizing Textual Entailment