We have three competent reviews all of which recommend acceptance. In my opinion the main room for improvement involves empirical evaluation. As exact calculation of the log loss is intractable a downstream task evaluation would be nice. But the paper seems acceptable with the current results.