Leveraging Sentence Similarity in Language Generation: Improving Beam Search using Range Voting