| Knowledge graph completion(KGC)research aims to use available knowledge to fill the missing information in the knowledge graph,which can improve the completeness of the knowledge graph and provide richer and more reliable knowledge support for the application of the knowledge graph in information retrieval,question answering system,recommendation systems and other fields.Traditional KGC usually rely on a large amount of labeled data to predict the missing parts of the knowledge graph.However,knowledge graphs usually contain a large number of long-tail relations,i.e.,most relationships only have a small number of labeled samples,and data labeling requires high labor and time costs in real-world applications.Unlike the traditional KGC,the few-shot knowledge graph completion(FKGC)task can effectively infer the missing information with a small number of samples,which has wider application value and research significance.FKGC studies usually contain two modules: the neighbor encoding module and the few-shot relation learning module.The neighbor encoding module is used to obtain better quality entity representations by aggregating entity neighbor information.The few-shot relation learning module is used to learn the corresponding relation representation based on a small number of reference triples under the same relation.Although the existing FKGC methods have achieved some success in two public datasets,they still have the following challenges: Firstly,existing studies only consider the local neighbor information of entities,ignoring the rich semantics contained in long-distance neighbors;Secondly,existing studies represent the relational features as a fixed vector,ignoring the uncertainty in the semantics of few-shot relations;Thirdly,existing studies aggregate the entire neighbor entities without removing the neighbor information that are irrelevant to the current task;Finally,existing studies only consider the common relational semantics among reference triples in the reference set,ignoring the diverse relational semantics of the same relation in different contexts.To solve the above problems,this thesis proposes the following work.(1)The research on FKGC incorporating global-local neighbor encoder and gaussian mixture model.For the problem of only considering local neighbors,this thesis proposes a global-local neighbor encoding module to simultaneously capture graph-level neighbor features and task-level neighbor features.The global encoder captures graph-level neighbor features based on relation paths,while the local encoder uses the task-aware attention mechanism to capture task-level neighbor features.Subsequently,the cascaded strategy is applied to fully integrate the two levels of neighbor features and obtain highquality entity representations.For the problem of uncertainty in the semantics of few-shot relations,this thesis designs an adaptive gaussian mixture model to model few-shot relations,which can adapt to different queries by dynamically adjusting component weights.The experimental results of link prediction on NELL-One and Wiki-One datasets show that the performance of the proposed model has significantly improved compared to the benchmark models.(2)The research on FKGC basing on box embedding.For the problem of irrelevant neighbor features,this thesis designs a neighbor encoder based on the box space,which captures relevant neighbor features by taking the intersection of entity boxes to filter out irrelevant features,and applies an attention mechanism to aggregate neighbor information.For the problem of only considering the semantics of common relations in the reference set,this thesis introduces a dual attention network.This network uses a position attention encoder to capture the contextual relational semantics between head and tail entities in the reference triples,and a channel attention encoder to capture the common relational semantics among the reference triples.In addition,this thesis applies a meta-learning framework to optimize the model which makes the model adapt new relations quickly.This thesis conducts link prediction experiments on two standard datasets,NELL-One and Wiki-One,and the experimental results show that the model achieve state-of-the-art performance in the task of FKGC. |
