Improving statistical parsing by linguistic regularization