Recommendations For K12 Math Exercises Based On Formula Features Extraction And Neural Cognitive Diagnosis

With the popularization of online education and big data,more and more students can enjoy the dividends of the Internet and get personalized and adaptive education for better devel-opment.In the field of intelligent education,personalized exercises recommendation for students has always been an significant research topic.Traditional methods implemented by collaborative filtering are similar to product recommendation system,which regards students as customers and exercises as goods.However,this method ignores the students' proficiency on concepts,thus has poor interpretability.In this research context,this thesis proposes three approaches in progressive orders for K12 math exercises recommendation,providing educators with a feasible,effective,and inter-pretable scheme.The three approaches are:an exercise modeling method based on for-mula understanding,a student modeling method based on neural cognitive diagnosis,and an exercises recommendation framework based on neural cognitive diagnosis and knowledge hierarchy constraint.At the beginning,the exercises text needs to be labeled to provide references for recommen-dation.However,the current methods usually ignore the rich semantic information in the formulas,which are common in math exercises.Besides,since some knowledge concepts are compound words,it is hard to select the best candidate.Therefore,this thesis proposes an exercise modeling method based on formula understanding.It uses BiLSTM+CRF network to capture the contextual key information as a basic structure,then designs a formula un-derstanding layer to dig out the explicit and implicit features embedded in the formulas,and a post-processing layer to adjust the compound candidates' scores for better selection.By applying this method,the knowledge of exercises can be automatically extracted and labeled with great accuracy.After exercises modeled,the student model,which reflects the students' characteristics for proficiency needs to be established.In this step,the famous educational psychology the-ory:cognitive diagnosis is introduced as the base,which can infer the students' cognitive states through Q-matrix and exercises response logs.Traditional cognitive diagnosis models(DINA,MIRT,etc.)usually use manual-designed students-exercises interaction functions to simulate the complex process,which is subjective and inefficient.In addition,they are not capable of dealing with large-scale knowledge concepts,like K12 math education,with exponential computing scale issue.Therefore,this thesis proposes a neural cognitive diag-nosis method based on DINA assumption for student modeling,which can learn the complex interaction functions between students and exercises factors automatically with intelligence and effectiveness,and solve the dilemma of large-scale knowledge,since the dimension of input data has little effects on the training performance of the network structure.Further,DINA assumption is used in the network,making the network reliable and interpretable.Based on the above work,this thesis proposes an exercise recommendation framework based on neural cognitive diagnosis and knowledge hierarchy constraint.The framework uses the exercises modeling method to generate Q-matrix automatically,the students modeling method to diagnose their proficiency states,then uses the diagnosis model to predict the probabilities of scoring,combined with the knowledge hierarchy constraint strategy to ad-just the scores,finally give accurate recommendations.Overall,this study proposes three innovative approaches in the progress of K12 math exer-cises recommendation progress.The whole scheme effectively solves the dilemma encoun-tered in the recommendation process and been proved with great effectiveness compared to traditional ways.