Research On Knowledge Graph Construction Method For Sparse Labeling

The construction of knowledge graph is an important direction of current research.In the process of graph construction,entity and relation knowledge need to be learned and mined from a large number of labeled natural language texts.However,due to the diversity of data sources and labor costs,sparse data labeling is a common problem in the process of knowledge graph construction,which increases the difficulty of knowledge acquisition and limits the construction of knowledge graph.Therefore,this dissertation focuses on the knowledge graph construction method for sparse labeling.The main research work is as follows.Firstly,there is a lack of augmentation methods for text semantic features in few-shot named entity recognition based on data augmentation,which limits the ability to learn and represent entity information.Therefore,a few-shot named entity recognition method based on collaborative graph attention network is proposed.This method realizes the full mining and representation of entity feature information by designing a sentence feature encoder based on collaborative graph attention network.Additionally,to promote the information interaction between the support set and the query set,an entity classifier based on match processing is proposed.The attention weight makes the support set adapt to different query instances,thereby improving the classification and prediction ability of the entity.Secondly,because the contextual feature information of the relation can enhance the understanding of the relation semantics in the sentence,and the traditional relation extraction model ignores the learning of these contextual information,resulting in the problem of insufficient relation feature mining.Therefore,a few-shot relation extraction method based on multi-level embedding representation is proposed.This method uses multi-level embedding of sentences to represent local and global text sequence information,to mine deep relational semantic features.In order to alleviate the influence of sample feature deviation,a support prototype extraction method based on query prototype attention is proposed.This method uses query prototype-level attention to guide the extraction of support prototypes to achieve accurate classification of query relations.Thirdly,aiming at the problem that important multi-hop features are lost in the process of information transmission in the node feature representation learning of knowledge graph completion,a graph structure encoding method based on gated graph sub-hop convolutional neural network is proposed.This method uses the graph structure representation method based on the sub-hop matrix to directly transmit the multi-hop neighbor information to the coding node,and uses the sub-hop gate mechanism to filter the information,thereby reducing noise redundancy while introducing multi-hop information.Additionally,in the link prediction part,a structural feature decoder based on high-dimensional structure analysis weights is proposed,which uses high-dimensional structure analysis weights to score candidate triples to complete the knowledge graph.Finally,aiming at the problem that the traditional few-shot knowledge graph completion model cannot fully learn the correlation information between the same relation triples,a few-shot knowledge graph completion method based on enhanced relation semantic mapping is proposed.This method uses the mapping matrix to map the entity features and the graph structure features of triples into the relation embedding space,to enhance the perception of relational semantics and better understand the association information between different entities under the same relation.Additionally,in order to enhance the semantic expression of the prototype to the relation,a prototype network based on attention convolution is proposed to extract reliable relation prototype representation,which is used to predict unlabeled triples with similar relation features.In addition,comparative experiments and effectiveness experiments are conducted on public datasets for the few-shot named entity recognition model,few-shot relation extraction model,knowledge graph completion model,and few-shot knowledge graph completion model proposed in this dissertation.