That's a valid concern, given the paper makes clear that the effect over the polite/impolite scale seems to be model dependent (it finds the reverse correlation of earlier studies on even older models).