Вспомогательные задачи предварительного обучения

AbstractExisting models for extractive summarization are usually trained from scratch with a cross-entropy loss, which does not explicitly capture the global context at the document level. In this paper, we …...