Research On Knowledge Graph Representation Learning Based On Fusion Of Multi-source Information

Knowledge graph(KG)organizes world knowledge in the form of triplet,which contains useful information.The purpose of representation learning(RL)is to learn the representation of entities and relations,so as to realize the application in specific task.Many methods have been proposed,there are some problems.On the one hand,limited by the limited structured data,the data is relatively sparse;on the other hand,most of the methods have limited expression ability and do not make full use of the potential information.The text description information of entity is introduced into RL.By embedding the related words of entities,the embedding based on text information is generated.The triple structure representation of KG is learned,and the two representations are trained jointly.Two forms of text data are considered.In the first case,the entity is used as the subject to describe the text,and the entity is generated by weighting the words in the text.Based on the embedding of the text information,different weights are set for the description words according to the relationship.In the second case,the entity name appears in the form of words or phrases,and the vector merging structure information of entity name is embedded in the corpus.The results show that the proposed model has better performance in link prediction and triplet classification tasks.The model is composed of encoder and decoder.The encoder generates new embedding for each entity by aggregating neighbor embedding and relationship aware global graph structure,and the weight of neighbor information is determined by graph attention network.By combining with Trans E or Dist Mult,the encoder can be optimized without other parameters.Two kinds of convolutional neural network decoders are constructed.By using multiple filters of different shapes or generating filters according to the relationship,more interactions between the source entity and the relationship embedding are captured while retaining the features of the encoder,so as to realize the matching calculation with the target entity.The results show that the model based on graph structure information achieves better results in each task.The role of each part of the model is proved by ablation experiments.Figure 28;Table 16;Reference 52...