Additional Information Aware Knowledge Graph Representation Learning

Google put forward the concept of knowledge graph,how to apply knowledge graph to various intelligent tasks has become an important research topic.Inspired by representation learning,knowledge representation learning has become an new tool for representing knowledge graphs,which can better apply the semantic information of knowledge graphs to tasks such as reasoning.Various knowledge representation learning models have been proposed.However,most of the current knowledge representation learning models only consider the information of the triplet itself when training the model,only learn the singlestep path of the triplet,and do not consider the rich information contained in the knowledge graph such as path information in the large-scale knowledge graph.From the above perspective,we study the knowledge representation learning considering additional information.The main research contents are as follows:We propose an adaptive knowledge representation learning model named Adaptive-PTransE that considers additional path information,which optimizes the PTransE model from three aspects: path selection,score function and loss function.First of all,for the path selection,we only sample the paths within two steps,avoiding the increase of extra noise from too long paths;In measuring the reliability of the path,we use the PCRA algorithm proposed by the PTransE model,and only selects the path with a reliability larger than ? to avoid too many useless paths from affecting the performance of the model.Secondly,for the setting of the score function,we propose a weight coefficient ? to measure the importance of triples and additional paths.By changing the value of ? to adapt to different knowledge graphs,the score function is more flexible.Thirdly,for the setting of loss function,on one hand,we adopt a loss function based on negative sampling,which solves the problem that the PTransE model can only train a pair of positive and negative samples at the same time.Finally,a fixed distance threshold is used in the loss function of the PTransE model to maximize the distance between positive and negative examples.We use a variable distance threshold to adapt to the best distance in the training process,so that the model can get better training.We conduct entity prediction experiments on the FB15 k and WN18 datasets.The results show that the Adaptive-PTransE model is 1%?3% higher than the PTransE model on MR and Hits@10.Aiming at the problem of representing the three relational modes of symmetry/antisymmetric,opposite,and combination at the same time,we propose a rotation-based knowledge representation learning model named Path-RotatE that considers additional paths.First of all,we put forward the RotatE model,a knowledge representation learning model based on rotation,which can represent three combined relationship models well.Secondly,we propose a new rotation combination method to improve the RotatE model.Thirdly,in the calculation of path reliability,the PCRA algorithm proposed by PTransE only measures the reliability of the path based on the amount of resources flowing from the head entity to the tail entity,ignoring the correlation between the path and the direct relation.We consider the correlation between the path and the direct relation.In this way,the reliability of the path is more in line with the actual situation.Finally,we conduct entity prediction experiments on datasets such as FB15 k,FB15-237,WN18 and WN18 RR.The results show that the PathRotatE model has a certain improvement in MR,MRR and Hits@N compared to RotatE,PTransE and other baseline models.