BART Model's Corruption Methods for Multi-Sentence Sequences

When pre-training a model on a document, a common strategy is to intentionally alter the input text and task the model with restoring the original. Which of the following alteration techniques is uniquely dependent on the input text containing more than one sentence?

When preparing text data to train a language model, various 'corruption' techniques are used to alter the original input, which the model then learns to restore. Some of these techniques operate on the word or token level, while others operate on the sentence level. Match each corruption technique described below with the structural requirement of the input text.

Analyzing Text Corruption Strategies