Deductive Closure Training of Language Models for Coherence, Accuracy, and Updatability F i n d i n g s - A C L