Research And Implementation Of Information Credibility Evaluation Method Based On Knowledge Representation Learning

With the advent of the Big Data era,5G communication and other new communication technologies,as well as online social platforms and other forms of information dissemination have evolved,resulting in an explosive growth in the speed and scale of information dissemination.It is undeniable that massive network information provides convenient for human daily life.However,the false information hidden not only brings trouble to people's life,but also to national information security.Therefore,it is of great significance to research the credibility evaluation of network information.In the current network environment,information data presents the characteristics of multi-source heterogeneity and low density of effective information,which make it difficult to collect and utilize information effectively.Previous evaluation methods cannot mine hidden information among large scale data,and need quantities of manpower and time cost.Therefore,how to use information data effectively is the key to solve the problem of network information credibility evaluation.In order to solve the above problem,the paper researches the network information credibility evaluation method in Big Data environment.The main work contents are:(1)Under the recent network environment,information data presents the characteristics of large volume and disordered structure,which puts forward higher requirements for collection and utilization of data.This paper utilizes the different representations of structured and unstructured data to extract the entities and relationships contained in the information data,and stores them in the form of triples in the knowledge graph,so as to facilitate the integration and mining of complex association relationships between entities.Aiming at the problem that the template-based relationship extraction method is limited by template coverage,we extract relationships and obtain the knowledge graph based on deep learning algorithm.(2)In view of the low density of real information in the network environment,this paper proposes an information credibility evaluation method based on knowledge representation learning.Based on the translation invariance of vector and the assumption that the information triples satisfy the vector triangle rule,the model maps relationships and entities into vector space.Considering the weak ability of traditional knowledge representation learning to deal with complex relationships and strong dependence on super parameters,the negative sampling method based on entity classification is adopted to reduce the randomness and non-negativity of corrupted triples.Furthermore,through the adaptive adjustment mechanism of super parameters and the monitoring of the lose function,a large learning rate is adopted in the early stage of parameter updating process to make the model converge quickly,and then the parameters are reduced to make the model converge accurately.The performance on large real knowledge bases shows that the average ranking of prediction results is improved by about 4%,and the accuracy of prediction results in the top ten is improved by more than 10%.(3)Aiming at the limitation of traditional translation embedding methods which only refers to the direct relationship,the model comprehensively considers the direct relationship and indirect relationship between entities through single path aggregation algorithm based on neural network and multi-path aggregation algorithm based on neighbor.The advantage of the long short-term memory neural network in dealing with the sequences which have semantic dependency can be used to alleviate the gradient vanishing problem of the recurrent neural network in dealing with the long-term sequences,and better represent the semantic relationship between entities.(4)According to the market demand and business function demand of automated information credibility evaluation by enterprise users and ordinary users,the paper designs and implements an automated credibility evaluation system.The system can be divided into knowledge graph building module,information credibility evaluation module and evaluation result visualization module.First,it takes the structured information data and unstructured text information data as the target,uses crawler technology to collect,stores in the form of triples.And then it constructs energy model based on knowledge representation learning algorithm and artificial neural network,updates parameters through adaptive mechanism.Finally,it provides a visual interface to show the results of the relationships to be evaluated and the system prediction respectively.The system can be applied to not only information credibility evaluation,but also in knowledge completion and database cleaning fields.