Abstract
Natural Language Inference (NLI) is a fundamental and challenging task inNatural Language Processing (NLP). Most existing methods only apply one-passinference process on a mixed matching feature, which is a concatenation ofdifferent matching features between a premise and a hypothesis. In this paper,we propose a new model called Multi-turn Inference Matching Network (MIMN) toperform multi-turn inference on different matching features. In each turn, themodel focuses on one particular matching feature instead of the mixed matchingfeature. To enhance the interaction between different matching features, amemory component is employed to store the history inference information. Theinference of each turn is performed on the current matching feature and thememory. We conduct experiments on three different NLI datasets. Theexperimental results show that our model outperforms or achieves thestate-of-the-art performance on all the three datasets.