Research On Fact Verification Method Based On Graph Neural Network

With the accelerated evolution of information warfare to intelligent warfare,the combat space of modern warfare continues to expand,and the composition of forces is more diverse.A large amount of situation information and intelligence data are difficult to distinguish,which greatly restricts the efficiency of combat command and decision-making.How to optimize the military information service system so that it can effectively and accurately analyze the acquired military information,distinguish the real battlefield situation and understand the battlefield situation from incomplete information,has become an urgent problem to be solved.This thesis aims at achieving accurate information services for military information systems,focusing on the task of fact verification in the field of natural language processing,which aims to use natural language technology to retrieve relevant evidence from the knowledge base and verify the given statement.We focus on the theoretical problems and technical difficulties in the current fact verification system,and conduct in-depth research to further improve the accuracy of the fact verification service.In general,our contributions in this thesis can be summarized as follows:(1)We propose an entity-graph based reasoning method for fact verification.This thesis innovatively proposes an entity graph to extract the key semantic features of evidence and model the semantic relationships in the evidence.This method uses the named entity recognization to extract key evidence entities from the evidence,and builds a graph structure with entities as nodes based on the position relationship.Then we conduct evidence reasoning on the entity graph through the node update mechanism based on graph attention neural network,which can help mine fine-grained features and semantic relationship features.In this thesis,a large number of experiments have been conducted on the public benchmark dataset for fact verification,i.e.,FEVER.We use the metrics such as label accuracy and FEVER score to evaluate the performance of the model.The experimental results demonstrates the superiority of our method for fact verification task.In addition,compared to the baselines,our model shows better performance in the scenario that require multiple evidences to verfiy the claim.(2)We propose a hierarchical reasoning-based heterogeneous graph neural network for fact verification.This thesis proposes to use an modified graph attention neural network mechanism to model the relationship of different-granularity semantic units,including entities,sentences and contexts.In order to better model the human information reasoning process,this method constructs a heterogeneous graph structure composed of entities,sentences and context nodes,and uses a hierarchical feature update strategie to realize the propagation of reasoning information from fine-grained to coarse-grained nodes.Finally,we construct different inference paths and selectively aggregate the prediction results of different inference paths.Experimental results show that this method has a greatly performance improvement in single evidence and multiple evidence scenarios compared with the baseline models.In addition,this method can provide a clear semantic reasoning path to verify the claim,which shows a good interpretability.(3)We propose a graph-based contrastive learning method for fact verification.Aiming at difficult samples with different labels but similar semantics,this thesis proposes to levearge the contrastive learning tasks to learn their discriminative representations,and alleviate the feature oversmoothing problem in the fact verification model based on deep graph neural networks.Based on the entity graph,this method leverages an unsupervised graph contrast task to train the graph convolutional encoder,and use a supervised case contrast task to train the text encoder,which can push samples of the same category in the vector space closer,as well as pushed different types of samples further.Experimental results on FEVER demonstrates the superiority of this method compared to previous models.In addition,when the number of training samples decreases,the method can maintain a relatively stable performance,indicating that it has better robustness.