You make a great point. Presumably biasing the training sets to those built from corrected sources would help, but what about using n-grams to throw out poor e

You make a great point. Presumably biasing the training sets to those built from corrected sources would help, but what about using n-grams to throw out poor examples of ungrammatical sentences? It wouldn't be perfect, but some kind of score based on a much lower prevlence of n-grams from the generated sentence vs the original might indicate acceptable cases and those where they were fairly similar might indicate the generated sentence could in fact be valid (and thus should be discarded)