Research On Knowledge Representation Learning Based On Path Molding

In the era of big data,there will be massive amounts of structured,semi-structured and unstructured data on the Internet.How to manipulate these massive amounts of data and use the useful information therein has received widespread attention.Knowledge Graph(KG)expresses the semantic information of various structured data in a computer-understandable way,which helps us better organize and manage data.However,existing knowledge graph is incomplete.It greatly limits the application of Knowledge Graph.Therefore,it is necessary to explore the implicit information to realize knowledge reasoning and Knowledge Graph completing.A Knowledge Graph is a complex semantic network composed of interrelated knowledge,in which an entity is regarded as a node of the network,and a relation is regarded as an edge of the network.Knowledge representation learning is to learn and calculate the entities and relations semantic connections.This technology effectively improves the performance of knowledge acquisition,representation and reasoning.Therefore,this paper do further research on the basis of knowledge representation learning.The specific contents are as follows:(1)Aiming at the individual locality problem of different KGs,a dynamic margin(DM)translation principle is proposed in this paper,and by introducing this principle into the classical TransE model.Initially,different model parameters are set for different KGs during the training process.Secondly,according to the increase of the number of trainings,the margin of the loss function is dynamically optimized,which effectively improves the individual locality problem of different KGs,thus improving the training efficiency.(2)In fact,there are abundant semantic information in multi-step relations between entity pairs.PTransE takes into account multi-step relations,and thereby achieves significant improvement in the tasks of entity prediction and relation prediction.However,in the case that there are many multi-step relations between a pair of entities,PTransE doesn't make any distinguish between them.In this paper,dynamic factors are added to paths representation process to achieve flexible transformation between relations and paths embedding.Thus,this method can effectively learning similar paths to alleviate the above problem.(3)The existing path-based knowledge representation learning models only consider two or three steps path.In order to better reflect the complex reasoning in KGs,long and short term memory(LSTM)structure is used to combine multiple relations in a path to increase the length of the embedded path and obtain more semantic information between entity pairs.Moreover,the improved path-constraint resource allocation(PCRA)algorithm is used in the training process to calculate the reliability of each path.