Research On Open-world Knowledge Reasoning Based On Graph Neural Networks

Knowledge graphs highly summarize the connections between real-world things in the form of triplets,and play the role of knowledge storage and intelligent computing in artificial intelligence.However,knowledge graphs often have a large amount of missing data which can affect the application in downstream tasks.Knowledge reasoning can automatically infer potential information from the existing knowledge in a knowledge graph,and discard the complication of manual completion and update.Traditional knowledge reasoning methods are oriented to the closed world and cannot predict emerging entities or relations.Graph neural networks can effectively learn graph structural features without relying completely on the representation of nodes and edges.Therefore,this thesis considers using graph neural networks to design methods capable of processing open-world data.The main work of this thesis is as follows:(1)The problem of few-shot is widespread in the knowledge graph,so an inductive link prediction model using the graph neural network and meta-learning is proposed for the few-shot datasets whose entities are unseen.This model can be naturally extended to knowledge graphs containing unseen entities,and can still work effectively on the fewshot data.Specifically,the data are firstly divided into four tasks based on relation types,and a subgraph structure is constructed for each triplet in each task;then the data features in the subgraph are learned,and the meta-learning framework is used to train each task in turn.The relations are taken as the learnable parameters of the graph neural network,and the nodes and edges are updated by message passing;finally,the triplets are scored based on the information in the subgraph structures.The experimental results show that the model has an average improvement of 3.58% on the metric Hits@10.(2)The open-world knowledge graph contains both unseen entities and unseen relations,especially when it involves privacy issues in fields such as public security,the data cannot be obtained in advance.Therefore,a knowledge extrapolation model for joint entity type is proposed,in which the features of the knowledge graph are further extracted by relational correlation graphs.It can be generalized to knowledge graphs containing unseen data.During the training,some entities and relations are regarded as unseen,and the pre-trained model is used to initialize the seen entities,relations and entity types in the knowledge graph;then unseen entities and relations are represented using seen components through two message passing modules,and the graph neural network is used to update the features in the knowledge graph;finally,the final entity and relation representations are obtained through the meta-learning training mechanism.(3)Applying the proposed open-world knowledge reasoning methods to the specific scenario,a question and answer system based on knowledge reasoning is designed.Users can query the information in the knowledge graphs by inputting triplets or questions,and the system infers the unknown knowledge through the knowledge reasoning models.