Can neural networks understand logical entailment?