Neural Natural Language Inference And Explanation Generation

With the rapid development of neural networks and the deposit of human-annotated data,deep learning has obtained noticeable results on various natural language processing tasks.However,to assist machine to understand human language,natural language understanding is still a long-standing problem in natural language processing and artificial intelligence.As one of the basic natural language understanding tasks,natural language inference plays a vital role to assist machine to understand text knowledge.Natural Language Inference(NLI)aims to determine the logic relationships(i.e.,entailment,neutral and contradiction)between a pair of premise and hypothesis.Existing NLI models use alignment mechanism to capture the aligned parts(i.e.,the similar segments)in the premise-hypothesis pairs,which imply the relationship of entailment.These alignment-based models achieve promising improvement on NLI.However,aligned parts in the premise-hypothesis pairs can sometimes mislead the judgment of neutral relation.Hence,only considering aligned parts results in a biased perspective for inference.On the other hand,since the prediction of logic relationship in existing NLI models is short of interpretability,researchers turn to research on generating explanation for NLI.Explanatory natural language inference can not only help people assert the trust of logic relationship prediction,also reveal the drawback of existing methods.Existing NLI explanation generators based on encoder-decoder framework generate explanations for the premise-hypothesis pairs,which results in the interpretability of final logic relationship prediction.However,these discriminative generators usually generate explanations with correct evidence but incorrect logic semantic.It is due tothat the logic semantic is implicitly encoded in the premise-hypothesis pairs and difficult to model.To address these issues in existing NLI models and explanation generators,this thesis proposes the following approaches:· For the problem of neutral biased prediction in alignment-based NLI models,this thesis proposes the Multi-Perspective Inferrer(MPI),a novel NLI model with high interpretability.MPI forms the holistic view from multiple perspectives associated with the three logic relationships to make the final decision.All parts in the sentence pairs are captured unequally regarding different perspectives within MPI.Additionally,to ensure the perspectives obtained by MPI as expected,this thesis introduces an auxiliary loss as explicit supervision on the representation of each perspective.The proposed explicit supervision does not only achieve better performance on MPIs,but endows the obtained perspectives with high interpretability as well,which is absent in the previous routing-by-agreement-based studies of NLP.· For the problem of generating incorrect logic semantic in NLI explanation generation,this thesis proposes a deep generative model called Variational Explanation Generator(Variational EG)with a latent variable to model the logic semantic.Training with the guide of explicit logic semantic in target explanations,latent variable in Variational EG could capture the implicit logic semantic in premisehypothesis pairs effectively.Additionally,to tackle the problem of posterior collapse while training Variaztional EG,this thesis introduces a simple yet effective approach called Logic Supervision on the latent variable to force it to encode logic information.Experiments on explanation generation benchmark—e-SNLI demonstrate that the proposed Variational EG achieves significant improvement compared to previous studies and yields a state-of-the-art result.Furthermore,Experimental analysis of generated explanations demonstrates the effect of the latent variable.