Improving Zero-Shot Translation by Disentangling Positional Information