Parameter Learning for Log-supermodular Distributions