Reviews: Defending Against Neural Fake News

Detailed comments: Contribution 1 [Significance: High]: New approach for learning and generating multi-field documents i.e. documents where not only a "body" of text is to be generated, but which also contain other fields such as domain, date, authors and headline. The approach is a language modelling approach where these fields are used as context. The approach consists in ordering the fields into two subsets during the training in an attempt to address issues such as expensive marginalization or dealing with a large number of field orderings during inference time. The approach seems novel to me as I am not personally aware of previous work that approaches this problem in this way. Question: It wasn't clear to me if some of the fields have some restricted domains from which they can be sampled.