Research On Distributed Knowledge Representation Learning Over Large Knowledge Graphs

With the proliferation of knowledge graphs,it has been applied to various research fields,such as knowledge base completion,question and answer systems,etc.Knowledge Representation Learning(KRL)has been playing an essential role in many AI applications.Similar to the Word2 Vec method,the graph embedding is an effective method to represent entities and relations in knowledge graphs.KRL greatly improves the computability of the knowledge graphs by using rich mathematical expressions,and some previous KRL embedding models achieve good performance.However,most of the research work is currently scattered,and the underlying implementation is not uniform,and the researchers of the model focus on the study of model performance while ignoring the scale of the dataset.Two main issues of existing KRL embedding techniques have not been well addressed yet.One is that the size of input datasets processed by these embedding models is typically not large enough to accommodate large-scale real-world knowledge graphs;the other issue is that lacking a unified framework to integrate current KRL models to facilitate the realization of embeddings for various applications.We propose DKRL,which is a distributed KRL training framework that can incorporate different KRL models in the translational category using a unified algorithm template.In DKRL,a set of primitive interface functions is defined to be implemented by various knowledge embedding models to form a unified algorithm template for distributed KRL.Meanwhile,we propose a parameter-server-based algorithm framework PSDKRL,which divides the training data into relatively uniform data blocks to avoid the computational imbalance between the computing nodes due to different scales.The workers exchange parameters with the parameter server.Each worker internally maintains a common data area for parameter acquisition and transmission.The parameter retrieval thread is responsible for putting the data requested from the parameter server into the buffer,and the update thread passes the updated parameter to the parameter server.In addition,under the premise of ensuring the accuracy of the model,based on the distributed environment,it provides support for learning to realize large-scale knowledge graph data training.The multi-threaded training design is adopted inside the computing node to maximize the training efficiency while ensuring the accuracy of the data.The effectiveness and efficiency of our framework have been verified by extensive experiments on both benchmark and real-world knowledge graphs,which show that our approach can outperform the existing ones by a large margin.