A Latent Discriminative Model for Compositional Entailment Relation Recognition using Natural Logic

Recognizing semantic relations between sentences, such as entailment and contradiction, is a challenging task that requires detailed analysis of the interaction between diverse linguistic phenomena. In this paper, we propose a latent discriminative model that unifies a statistical framework and a theory of Natural Logic to capture complex interactions between linguistic phenomena. The proposed approach jointly models alignments, their local semantic relations, and a sentence-level semantic relation, and has hidden variables including alignment edits between sentences and their semantic relations, only requires sentences pairs annotated with sentence-level semantic relations as training data to learn appropriate alignments. In evaluation on a dataset including diverse linguistic phenomena, our proposed method achieved a competitive results on alignment prediction, and significant improvements on a sentence-level semantic relation recognition task compared to an alignment supervised model. Our analysis did not provide evidence that directly learning alignments and their labels using gold standard alignments contributed to semantic relation recognition performance and instead suggests that they can be detrimental to performance if used in a manner that prevents the learning of globally optimal alignments.