How Smart is BERT? Evaluating the Language Model's Commonsense Knowledge