Syntax-aware Natural Language Inference with Graph Matching Networks

The task of entailment judgment aims to determine whether a hypothesis is true (entailment), false (contradiction), or undetermined (neutral) given a premise. While previous methods strike successful in several benchmarks and even exceed the human baseline, recent researches show that it remains arguable if the methods learn the statistical bias in the datasets. In this paper, we propose the syntax-aware Natural Language Inference (SynNLI) model, which utilizes graph matching networks to obtain syntax-guided contextualized representation while aligning the premise and the hypothesis accordingly. We show that the proposed method outperforms multiple baseline models on MNLI develop set, and visualize the model internal behavior.