Who did what to Whom? Language models and humans respond diversely to features affecting argument hierarchy construction