Research And Application Of Knowledge Graph For Inquiry-based Experimental Courses

E-learning has been widely used in teaching at home and aboard for its convenience,easily for sharing and novel forms of learning.For Chinese STEM field,students can improve their experiment ability through performing AR experiments,which is a typical form of E-learning.However,AR experiments lack combination of STEM knowledge at present,so we design a STEM knowledge graph to assist teachers and students to finish teaching and learning knowledge through AR experiments,and help students with autonomous learning.The main research of the dissertation is as follows:(1)In terms of knowledge extraction,the effects of named entity recognition and relation extraction are improved by feature extraction.Firstly,the knowledge dictionary,indicator words and keywords are obtained through word frequency,which can be important features of CRF models.Moreover,the indicator words and keywords are classified based on the named entities for improving the recall rate of named entity recognition.Secondly,based on the statistical results of the distance between knowledge entity and key components of the sentences,position is also used as one of the features of multi-classifier.Hence,the enhanced features are used in relation extraction.Experiment results show that the feature extraction methods proposed are effective for the Chinese STEM field,the precision and recall rate increase as well.(2)In terms of knowledge graph completion,an improved negative triplet sampling method is proposed,which can improve the knowledge representation learning results of the Trans H model.Firstly,adding knowledge for the STEM knowledge graph form crowdsourced experimental teaching resources.Secondly,translation models are used to do link prediction of the knowledge graph.To be specific,a negative entity sampling method based on Bernoulli is proposed to calculate the probability of different replacement ranges of entities,then we design a quality evaluation function to choose high-quality negative triplets for solving the zero loss problem.In the STEM knowledge graph,experiment results demonstrate that the proposed methods outperform the traditional knowledge representation learning models in link prediction.(3)In terms of personal knowledge recommendation based on the paths of knowledge graph,the path extraction and weight operation methods are proposed in proving and exploratory experiments respectively.The paths of knowledge graph are extracted by the knowledge points related to the experiment and experiment scores of students,then put the paths into the Bi-LSTM model for mining semantic as well as order information.In proving experiments,the Softmax function is used to calculate the weight of the paths,which represent the students' demand level of knowledge.Besides,in exploratory experiments,implementing dynamic knowledge recommendation with different experiment steps.Experiment results represent that the proposed methods can improve n DCG rate of the knowledge recommendation.