Learning Entailment Relations by Global Graph Structure Optimization

Abstract

We propose a global inference algorithm for learning entailment relations between predicates. We first define a graph structure over predicates that represents entailment relations as directededges. Then, we use a global transitivity constraint on the graph to learn the optimal set of edges, formulating the optimization problem as an Integer Linear Program. We motivate thisgraph with an application that provides a hierarchical presentation for a set of propositions that focus on a target concept, and show that our global algorithm improves performance by more than 10% over baseline algorithms.