Research On Construction Technology Of Question Bank Based On Knowledge Graph

With the development of the Internet and artificial intelligence technology,traditional industries are constantly being transformed and upgraded by informatization technology.The state vigorously supports the development of informatization education.The huge demand for online education under the epidemic has highlighted the importance of education informatization.The question bank website provides a large number of test questions for users to choose,which can effectively help users learn and consolidate the knowledge they have learned.However,on the one hand,in the process of updating the question bank,it is necessary to manually judge the knowledge points of the test questions for classification,which is labor intensive and has the influence of subjective factors.On the other hand,existing question banks just divide test questions according to large chapter knowledge points,which leads to the repeated recommendation of similar test questions in the test question recommendation process,and reduces the learning efficiency.In response to the above problems,this project draws on the idea of constructing knowledge graphs.First,the named entity recognition of the test question text is performed to obtain the conceptual entity in the test question,and then the semantic relationship between the entities is obtained through the entity relationship extraction to complete the automatic extraction of the knowledge points of the test question.In the existing named entity method based on deep learning,the word embedding part adopts static word embedding,and the generated word vector cannot represent the polysemy of a word,which leads to incomplete extraction of subsequent features that play a key role in entity recognition,which reduces the accuracy rate of entity recognition.In addition,in the process of feature extraction,some local features are ignored when extracting global context features,which makes the recognition effect worse.To solve the above problems,this thesis proposes a named entity recognition method based on BERT(Bidirectional Encoder Representation from Transformers).Firstly,the BERT model is used to train on large-scale unlabeled data in an unsupervised manner to extract more grammatical and semantic features to solve the shortcomings of word vectors representing the same meaning in different contexts.Then the bidirectional long short-term memory network(Bi LSTM)and iterative expanded convolutional neural network(IDCNN)are used to complete the lower-level feature extraction task to make up for the shortcomings of the Bi LSTM model that ignores some local features,and then the attention mechanism is used to extract the key features for classification.Finally,the conditional random field is used to decode the output of the attention layer to complete entity recognition.Experiments have proved that the model proposed in this thesis can effectively improve the accuracy and recall rate of entity recognition.Existing entity relationship extraction methods based on dependency syntax analysis only consider the subject-predicate and verb-object dependencies,resulting in poor extraction results.In this thesis,on the basis of the subject-predicate and verb-object relationship,three other relationship types are added to improve the accuracy of entity relationship extraction.In addition,because syntactic analysis cannot cover all the characteristics of entity relationships,and some relationships are represented by special characters,which cannot be identified through syntactic analysis.In response to this problem,this thesis proposes an entity relationship extraction method that combines syntax and domain knowledge features.The main idea is to treat the words that reflect the relationship in the text as entities,and then identify the entity and the words that reflect the entity relationship through named entity recognition,then the relation extraction is achieved through matching rules,and the extraction results are combined with the entity relationship extraction results based on dependency syntax analysis to make up for the shortcomings of the low recall rate of entity relationship extraction based on a single syntactic feature.Experiments show that the method proposed in this thesis can effectively improve the recall rate of entity relationship extraction.On the basis of entity recognition and entity relationship extraction,the knowledge points examined in the test questions are obtained,and then the correlation between the test questions based on the knowledge points is mined,and this correlation degree is quantified to obtain a test question network with the test questions as the nodes and the relationship degree as the edge.Finally,save the test question network and test question texts into the neo4 j graph database to complete the construction of the question bank and lay the foundation for subsequent teacher questioning and test question recommendation.