Recent neural headline generation models have shown great results, but aregenerally trained on very large datasets. We focus our efforts on improvingheadline quality on smaller datasets by the means of pretraining. We proposenew methods that enable pre-training all the parameters of the model andutilize all available text, resulting in improvements by up to 32.4% relativein perplexity and 2.84 points in ROUGE.
展开▼