Research On Natural Language Inference Algorithm Infusing Knowledge Graph

The main task of natural language inference is to make machine infer the implicative relationship between premise and hypothesis under the condition of given text pair.It is a fundamental and important task in natural language processing,and has a wide application prospect.It has been proved that the introduction of knowledge graph may be critical for a task of entailment.However,the proposed models do not fully consider the noise in the usually large knowledge graphs;encoding these subgraphs using graph neural networks may loss some important information.In view of the above problems,this paper proposes the strategy for constructing subgraph.By filtering the path between premise and hypothesis,the strategy can optimize the information of structured data.In order to maximize the knowledge information in the subgraph,this paper constructs a graph-encoding network which based on the capsule to encode the extracted knowledge subgraph.The main research work and innovations are as follows:1.Aiming at the problem of noise in the process of subgraph construction,this paper propose the subgraph construction strategy based on optimal path information.The weight information and path length information are combined to filter the knowledge subgraph,so as to reduce the number of unmatched entities in the graph and increase the context relevance of the subgraph.In order to further obtain the knowledge information in the subgraph,the subgraph composed of premise and hypothesis is encoded by graph convolutional networks,and the obtained fixed length vector is integrated into the text-based inference model for training.2.Aiming at the problem of information loss caused by the encoding process by traditional graph neural network,this paper proposes a knowledge enhanced natural language inference model(Caps-KGEIM)based on capsule network.Through capsule unit,the scalar output features of graph convolution neural network are transformed into vector form,so as to further enhances structural knowledge information mentioned in subgraph.The final model contains a common text-based inference model and a graph-based inference model.It enriches the external knowledge information in the text-based inference model.3.This paper uses neo4 j to manage knowledge graph data,builds neural network model with deep learning framework of python,and evaluate our approach on CNLI and RITE datasets respectively.The results show that the subgraph construction method in this paper can get purer knowledge subgraph than the fixed-hop method.Compared with several classical baseline models,the proposed model has different range of improvement.