Static And Dynamic Knowledge Graph Completion Based On Hierarchical Chinese Data

As a downstream task of extraction and an upstream task of inference and processing,knowledge graph completion is an important link in the field of natural language processing.At present,some progress has been made in knowledge graph completion,but the analysis effect of hierarchical data in data concentration is difficult to meet the requirements of knowledge graph construction.In this paper,a method of constructing hierarchical Chinese datasets is proposed by analyzing the hierarchical nature of Chinese datasets,which improves the static completion model and introduces time information to the dynamic completion model to realize the completion of knowledge graph.The main research contents of this paper are as follows:(1)A Chinese data subset construction method based on the hierarchy of multiple coupling relations of datasets is proposed.The existing open source English dataset has two major problems through the hierarchical analysis of dataset structure.First,there are a large number of meaningless triples in the dataset,and the existence of these triples will improve the accuracy of model testing.Secondly,there is the problem of unbalanced triplet data quantity in the dataset.According to the above two problems,this paper puts forward four kinds of coupling relation and a method of building data subset,the method is based on two gradation index,measuring data of the above four kinds of coupling relation has good extraction effect,and build two Chinese data subset.Experimental results show that the effect of Chinese data subset on the model is better than the existing open source English dataset.At the same time,by improving the maximum amount of data difference in the constructed Chinese dataset,and then testing on the model,it is concluded that the greater the amount of data difference,the worse the effect of the model.(2)A coupling relation recognition model based on the coordination of spatial rotation and reflection parameters is proposed.For the static knowledge graph completion task,through analyzing the data processing methods of the existing model,it is found that the model does not extract the four coupling relations well.Therefore,on the basis of ATT model,this paper introduces spatial rotation parameters and reflection parameters,and makes them map the corresponding coupling relation in space respectively.In this way,the model can recognize the hierarchical coupling relations,and then carry out spatial mapping through hyperbolic-tangent space to obtain the semantic and distribution of spatial relations.Then,the improved ATTm model is compared with the existing model,and experiments are carried out on the open source English dataset and the constructed Chinese data subset respectively.The results show that the improved ATTm model is better than other models.At the same time,the maximum data difference was compared to further verify the above conclusion.(3)A dynamic knowledge graph completion model of relation pair difference weight matrix is proposed.Extending triples to quads by introducing time information into datasets and models.Specifically,for the dataset,the external text information is linked to the existing entities in the dataset,and the relevant semantic information including time is extracted from the external text information,so that the dataset is based on the corresponding time information of each entity as the criterion of whether the update on time is effective.For the model,this paper proposes that the relative pair error values that follow and violate the time order become smaller and larger respectively,and the result of the relative pair error is combined with the ATTm model in the form of weight matrix.The improved ATTm-d model is compared with the existing dynamic completion model,and the Chinese and English datasets are compared.It is found that the introduction of time information removes the triplet data that does not meet the time meaning,and the accuracy of the model is further improved,which can have a better recognition effect on the hierarchy of datasets.