Few-Shot Knowledge Graph Completion By Subgraph Reasoning

Knowledge graphs play a very important role in the current society.Many tasks based on natural language,such as recommendation systems,information extraction,reading comprehension,etc.,rely on the construction of knowledge graphs.Although there are a large number of triple facts in the knowledge base,the knowledge graph still faces the problem of lack of knowledge.Knowledge graph completion is a hot topic of current knowledge graph research.It aims to infer new facts(triplets)through existing entities and relationships.Knowledge representation is the main method to solve the completion of the knowledge graph,mainly by projecting the entities and relations in the graph into a continuous vector space,and predicting the missing relations through the trained vectors.Most of the existing knowledge representation methods require enough data and train all relationships together.In real world,there are a large number of long-tail relationships in the knowledge base,and with the dynamic evolution of the knowledge base,new types of relationships will appear in the knowledge graph.In order to solve the above problems,scholars have proposed a small sample knowledge graph completion problem.The current research on few-shot knowledge graph completion is based on entity representation and local first-order graph structure,and does not consider the path rule logic around the triplet.The subgraph structure around the triplet can well preserve the Path information of target triplet.In order to be able to save the information of the subgraph structure in the knowledge completion of the small sample scenario,this paper has made the following improvements:(1)A few-shot knowledge graph completion method based on adaptive metric matching is proposed.We first obtain the subgraph representations of the support set and query set for each task separately,use an adaptive matching network to learn the matching attention coefficients of the positive and negative samples of the query set and the positive samples of the support set,and finally integrate the various samples of the support set to get The prototype means that the matching score between the prototype vector and the query sample vector is calculated.(2)A few-shot knowledge graph completion method based on subgraph relationship learning is proposed.The main idea of method is: save the path information around the triples in the form of local subgraphs,use the relation graph neural network to extract the subgraph features of each relation,and average the sample features in the same task in the support set.Prototype features,calculate the similarity score by splicing the sample features with the prototype features,and finally use the MAML gradient optimization algorithm to learn the public initialization of each task to adapt to the new relationship.