Few-Shot Knowledge Graph Completion Based On Meta-Learning

Knowledge graphs are prevalent in many fields,such as information retrieval and data mining.Large knowledge graphs often suffer from the problem of incompleteness,i.e.,missing potential triples.Completing the knowledge graph can improve the user experience of knowledge-graph-related applications.Traditional knowledge graph representation learning methods require a large amount of training data,so they have difficulty completing the few-shot relations in knowledge graphs.We proposed meta-learning-based few-shot knowledge graph completion methods,which can predict potential triples of relations with a few reference triples.The main work and contributions of this thesis are as follows:(1)We proposed a meta-attention model Meta A for few-shot knowledge graph completion.Meta A utilizes an attention neighbor encoder to capture the most relevant attribute features of entities to the uncompleted relations from the entity’s neighbor relations and entities by constructing the dynamic attention from entities to their neighbors,which enhances the entity embeddings.Meta A computes the attention of query triples and reference triples,as well as the attention between reference triples through an attention matching network,then performs simultaneous updates and multi-step matching of their embedding representations to compute the validity of query triples accurately.Experimental results on benchmark datasets demonstrate that Meta A can improve the model performance by obtaining relevant information from external data through the attention mechanism.(2)We proposed a meta pattern learning model Meta P for few-shot knowledge graph completion.Meta P extracts pattern information directly from triples by a pattern learner incorporating a residual update mechanism and then calculates the validity of the triple patterns by a pattern matcher incorporating a validity balance mechanism,which reduces the reliance of the model on external data.The residual update mechanism helps Meta P retain the original information of pre-trained entity embeddings during the optimization process and obtain better entity embeddings.The validity balance mechanism can reduce the interference to the model caused by the candidate triples that are difficult to identify whether they are valid or invalid.It also improves the accuracy of the model in calculating the validity of triples.Experimental results on the benchmark datasets demonstrate the effectiveness of these two mechanisms and the robustness of Meta P.