Research On Key Technologies Of Autonomous Modeling And Reasoning Based On Knowledge Graph Combining Temporal Information

The fourth revolution of industry means that we are about to enter an era of intelligence.People use a series of information technologies to promote industrial transformation,and traditional industries will undergo revolutionary changes.However,the intelligence of industrial information is still in the research stage,and there are certain difficulties,mainly because industrial data is heterogeneous and multi-source,and most of them are unstructured data.With the rise of knowledgedriven concepts,knowledge graphs have begun to enter our daily life,whether it is using search engines such as Google and Bing,or using intelligent assistants such as IBM Watson,Siri and Cortana,in fact,people are constantly interacting with knowledge every day.Atlas to interact.At present,knowledge graphs are widely used by technology giants,which proves the feasibility of using knowledge graphs to store complex data,and based on this to do knowledge reasoning to obtain hidden knowledge.Knowledge graphs as a kind of large-scale semantic integration and interactive operation New technologies have aroused great public attention and research interest,and construction methods and inference algorithms have emerged in endlessly.But most static knowledge graphs are not enough to describe dynamic processes and knowledge that evolves over time,and the knowledge graph inference completion model proposed for these static knowledge graphs also ignores the importance of time information for inference algorithms.Existing knowledge map topics can be divided into two categories: general knowledge maps and professional knowledge maps.General knowledge maps contain multiple association information between real-world entities,and professional knowledge maps contain association relationships between various term entities.Regardless of the type,most of them are static knowledge graphs,so most of the knowledge inference algorithms derived from these static knowledge graphs do not consider time information.In applications,whether knowledge graphs are used to assist error detection,or Using knowledge graphs to model expert experience to provide references for optimizing industrial processes,most of them are dynamic processes.Static knowledge graphs are not enough to characterize such industrial knowledge.Ignoring the importance of time information will greatly limit The availability of knowledge graphs and the accuracy of inference algorithms.Based on the above considerations,the main research content and completed work of this article include the construction and reasoning of knowledge graph,as follows:1)Construction of knowledge graphFirst completed the crawling and cleaning of power grid news knowledge,realized the entity extraction based on the attention mechanism on the grid dataset and the experience knowledge of grid experts,and used the results of the entity extraction to build a custom dictionary,based on the dictionary and based on The method of syntactic analysis is combined to realize the relationship extraction,and finally the knowledge map of the small power grid industry is completed by combining the time information.In terms of knowledge graph reasoning,because most existing knowledge graph reasoning models only learn from the fact that time is unknown,and ignore the useful time information in the knowledge graph,for this point,this article considers relational time-aware embedding and entities Time perception is embedded in two ways.2)Relational time-aware embeddingConsidering the relationship-based time perception method and the entity-based time information perception method,a large number of predecessor reasoning models are summarized,and the feasibility of time information to improve the ability of reasoning algorithms is demonstrated.Considering time information in relation embedding,the basic Trans series model is improved based on this strategy.By decomposing the time stamp into a sequence composed of temporal markers,the recurrent neural network is then used to learn the time-aware representation of the relationship type,and The relationship time perception obtained as a key item is included in the scoring function.At the same time,in order to further improve the prediction accuracy of the model,we have established different hyperplanes for different relationships to overcome the difference between the head entity relationship and the time information.Confusion between tail entities.The experimental results on the ICEWS18 data set show that no matter which index,the relationship fusion time information can improve the model performance.3)Entity time-aware embeddingConsidering time information in the entity embedding,the entity embedding is defined as a function,which accepts an entity and a time node as input,and provides the entity with a hidden representation under the time node.Based on this strategy,the Dist Mult model was improved,and good performance was achieved on the data set in ICEWS2014 and GDELT.