I think it should actually give you better results. It's monotonic in kl divergence, and does a much better job of taking into account how common the feature is rather than just how different it is. You no longer need to do things like throwing out phrases that appear less than x times if you use it.