something else you could do is train both a forward and backward model then reinforce when they both converge to the same overall text

Posters: Shoalst0ne. Shoalst0ne

3 likes recorded by the source corpus.