Research On Extracting Concept Relations In Primary Mathematic

With the development of Internet technology,online education has changed the learning style of students.The emergence of online education platforms,on the one hand,make it easier for students to access online educational resources.On the other hand,they also cause “information overload” and “knowledge trek”.A complete knowledge system not only effectively integrates online educational resources,but also helps to characterize the learning status of students.The learning path planned according to the knowledge system can alleviate the negative impact of “knowledge trek”.The relation-extraction for concepts is one of the main tasks of knowledge system construction.At present,the most efficient relationextraction approaches are mainly supervised algorithms.However,such methods suffer from low quality of text,scarcity of corpus,difficulty in labeling data,and low efficiency of feature engineering.Therefore,this paper studies the relationextraction algorithm between concepts based on the encyclopedic corpus and the distant supervised methods.The main contributions are as follows:1.Propose attention mechanism based on relation representation.The CNN model and the RNN model can only extract undirected relation,while the relationship between a pair of concepts is directed.This paper proposes the attention mechanism based on relation representation,which enable the model focus on the directed relation between a pair of concepts.2.Propose GCLSTM model based on GCN and LSTM.Considering that the GCN model is difficult to capture directional information,and the LSTM is difficult to extract local related information between adjacent words,this paper proposes a GCLSTM model by combining the advantages of GCN and LSTM models.First,it use GCN's gate linear unit to extract multipoints information related to the pair of concepts.Then,it weights the multipoints information through the attention mechanism based on the relation representation.Finally,it uses the bidirectional LSTM to integrate weighted multipoints information and extract relation.3.Propose BTRE model based on Transformer.Considering that the GCLSTM model is difficult to extract longdependent local information and high complexity,this paper proposes BTRE model which integrates both Transformer architecture and attention mechanism.This model can extract longdependent directed information between words with low model complexity.It is suitable for directed relationship extraction.4.Design and implement the concepts relationextraction system.This system can perform data processing,online data annotation,model training and visualization,demonstrate knowledge graph,and support the team members working at same time.