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

We propose a method for natural language generation, choosing the most\nrepresentative output rather than the most likely output. By viewing the\nlanguage generation process from the voting theory perspective, we define\nrepresentativeness using range voting and a similarity measure. The proposed\nmethod can be applied when generating from any probabilistic language model,\nincluding n-gram models and neural network models. We evaluate different\nsimilarity measures on an image captioning task and a machine translation task,\nand show that our method generates longer and more diverse sentences, providing\na solution to the common problem of short outputs being preferred over longer\nand more informative ones. The generated sentences obtain higher BLEU scores,\nparticularly when the beam size is large. We also perform a human evaluation on\nboth tasks and find that the outputs generated using our method are rated\nhigher.\n