Design And Implementation Of Relation Extraction System Based On Meta Learning

Relation extraction is an important task in information extraction technology,and it is also a key part in the construction of knowledge graphs.This technology has great commercial value and application demand.However,the existing relation extraction systems generally face the challenge of two major problems: the cold-start data demand problem and incremental data processing problem.Recently,more and more researchers have begun to pay attention to these problems.The few-shot relation extraction has achieved some good results in responding to the problem of cold-start data demand problem.In one way,the existing methods of the few-shot relation extraction are based on samples.It will cause the relation modeling deviation with very few samples and reduce the performance of the model.In another way,the existing continual relation extraction methods are rare,and suffering from the catastrophic forgetting problem and order-sensitivity problem.This thesis proposes a few-shot relation extraction algorithm and a continual relation extraction algorithm.The contributions of this thesis mainly include:(1)A relation extraction framework based on meta-learning and multi-view classifier integration is proposed.The initialization parameters of classifiers are trained by meta-learning method,so that it can be applied to the task of relation extraction with few samples.(2)A method of modeling semantic representation of relation using external knowledge base is proposed.The multi-view training labels are obtained by this method,and the meta-training task sequence is reconstructed,which improves the performance of meta-learning few-shot relation extraction model.(3)A meta-learning continual relation extraction framework is proposed,which combines model interaction information and task semantic information for negative sampling,and uses instance representation to filter memory instances.The mechanism improves the ability of the model to resist catastrophic forgetting problem and ordersensitivity problem.The method obtains the best performance on the testing datasets.(4)Through a detailed investigation of the actual application scenarios of the knowledge graph construction project,the data processing,relation extraction and multi-person collaboration functions that the relationship extraction system should meet are analyzed,and the ease of use and maintainability of the system are considered.On this basis,a detailed design scheme of a relation extraction system is proposed and the system is implemented.To sum up,this thesis proposes a few-shot relation extraction algorithm based on knowledge enhancement and a continual relation extraction algorithm based on relation awareness.The design and implementation of a relation extraction system meets the actual needs.