Research And Application Of Few-Shot Relational Extraction Technology For Educational Knowledge Graph

The educational knowledge graph is a structured data representation that uses knowledge points from various subjects as nodes and the relationships between these points,such as inclusion,dependency,and association,as edges.It can be used to support various applications in the field of education,such as intelligent recommendations,personalized learning,and knowledge management.However,due to the involvement of multiple disciplines and levels in the field of education,constructing a complete and high-quality educational knowledge graph requires a significant amount of manual annotation and maintenance work,which is a very time-consuming and labor-intensive process.Therefore,how to effectively extract relationships with less supervised data has become an important and challenging research topic.This paper focuses on few-shot relationship extraction technology in the educational knowledge graph,exploring issues such as insufficient robustness,incomplete feature representation,and poor generalization ability caused by sparse data in existing methods.The main research contents are as follows:(1)A few-shot relation extraction method based on How Net with dual attention mechanism is proposed to address the noise problem existing in the current methods.The method decomposes the semantic meaning of entities in the text sequence and calculates the semantic weight of the entity and instance using a dual attention mechanism to improve the model’s robustness.Specifically,the method uses the How Net semantic network to divide entities into multiple sub-words,and the first layer of the attention mechanism selects the sub-word that best matches the contextual context to alleviate the ambiguity problem of entities in different contexts.The second layer of the attention mechanism is used to encode different instances that extract prototype features for different relationship types to mitigate the impact of noisy data.The method performed well in the Chinese dataset Fin Re and achieved good results in few-shot relationship prediction tasks of varying difficulty compared to other methods.(2)A few-shot relation extraction method based on multi-view graph attention network is proposed to address the problems of insufficient feature representation and poor generalization ability existing in the current methods,which can mine hidden graph structural features from text and fully utilize these features to improve the accuracy of relationship prediction.Specifically,the method takes each character in the text sequence as a node and uses a Gaussian graph generator to construct edges between nodes from multiple perspectives to represent hidden relationships between text.From different perspectives,the model uses different graph attention networks to learn the relationship weights between each node,enabling it to automatically select the correct relationship.Then,the weight information and feature information are input into a graph convolutional network for feature fusion to obtain more context information and improve generalization ability.The method was evaluated on the FewRel1.0 and FewRel2.0 datasets and achieved good results.(3)This paper designs and implements a knowledge extraction visualization system for educational knowledge graphs.Users can upload their data and perform a small amount of annotation to train their own customized knowledge extraction models.The system provides the above two different algorithms for modeling,and users can choose appropriate algorithms for knowledge extraction according to the needs of different subject areas.In addition,the system adds a quality evaluation module for the extraction results.Users can use this module to evaluate the accuracy of the extraction results and retrain the model with higher-quality results to improve its accuracy.