Can we explain natural language inference decisions taken with neural networks? Inference rules in distributed representations