Recently, the relationship between automated and human evaluation of topic
models has been called into question. Method developers have staked the
efficacy of new topic model variants on automated measures, and their failure
to approximate human preferences places these models on uncer