Research On Open-World Knowledge Graph Completion Based On Meta-Learning

Knowledge graph structures complex knowledge in the form of networks and is used in many downstream tasks,such as Web search and intelligent recommendation systems.However,most knowledge graphs commonly suffer from missing relationships and incomplete data,and the purpose of the knowledge graph complementation task is to improve the completeness of the knowledge graph and thus its performance in downstream tasks.With the high speed and dynamic development of real-world knowledge,new entities and relationships appear constantly.Most existing embedding-based knowledge graph complementation methods assume that the knowledge graph is fixed,and when new entities and relationships appear,the model needs to be retrained,which will bring huge costs.At the same time,long-tailed relations with lower frequency in the knowledge graph also involve entities with long-tailed distribution,while most current knowledge graph complementation methods that solve the problem of few samples only consider longtailed relations and ignore the existence of long-tailed entities.To address the above two problems,the main research works and contributions of this paper are as follows.1.This paper studies and analyzes the knowledge graph completion techniques,considers how to deal with the knowledge graph complementation problem under the open world setting,constructs links between visible entities and visible entities,and links between invisible entities and visible entities in the open world knowledge graph,and studies the meta-learning methods for solving the few-sample problem,and applies the meta-learning methods to solve the open world knowledge graph completion task the problem of long-tailed relations and long-tailed entities exists.2.This paper proposes a Reptile Meta-Learning Based Open-World Knowledge Graph Completion(RML)model,the model mainly contains description encoder,triadic scorer and meta-learner,in which the description encoder uses text descriptions of entities and relations to generate embedding representations,contextual encoding of text descriptions by Bi-LSTM,and then feature extraction using local panning feature of CNN,which can link invisible entities to the knowledge graph.Considering that the head and tail entities under different task relation semantics should have different embedding representations,the task relation specific aggregation method is used to realize the interaction between head and tail entities and task relations to obtain the embedding representation with richer semantic information.To improve the model prediction capability,the triad scorer uses the Trans E function to evaluate the rationality of the triads.The meta-learner designed based on Reptile can quickly update the model parameters to adapt to new tasks by a small sample of relations,solving the problem of long-tailed relations and long-tailed entities that exist in the knowledge graph.3.Based on RML,this paper also proposes the Multi-Attention Meta-Learning Based Open-World Knowledge Graph Completion(MAKGC)model.Taking tail entity prediction as an example,the model fuses the following multiple semantic interaction information: semantic interaction information between head entities and task relations,semantic interaction information between candidate tail entities and task relations,and semantic interaction information between candidate tail entities and head entities,and fuses these information based on the word-level attention mechanism,so that the head and tail entities have more accurate representations.In addition,by introducing the self-attention mechanism,the model can further focus on the key semantic information and remove the noise,so as to mine the deeper semantic association information.4.In this paper,according to the problem setting of Reptile meta-learning method,a new dataset DBP50 ktext is constructed for the meta-learning open world knowledge graph complementation task,and comparison experiments and ablation experiments are conducted on the dataset DBP50 ktext and other two datasets WDtext and DBPtext,and the experimental results show that the two models proposed in this paper perform well in each evaluation index performs better than the three sets of knowledge graph complementation algorithms selected in this paper,and achieves more accurate and efficient prediction.