I think a lower amount of training data for Haskell might be a reason.