Research On Predicting The Difficulty Of Test Paper Based On Title Attribute And Text

Since the selection examination has been widely concerned by the society,the research on the propositional quality of the test questions has a very important significance,and the adjustment of the difficulty of the test questions is the key to improve the quality of the proposition.Taking the mathematics test questions as an example,this paper analyzes the data of eight sets of tests for an important selection test in a province and city for four consecutive years,and estimates the difficulty parameters of the test questions by using the project response theory.Considering that different tests are composed of different subjects,the difficulty of different test estimates is not comparable.Through the equivalent experimental design,all tests are combined to establish a matrix of questions,and the scores on all different test forms are converted to the same score.On the scale,Achieve the goal of unified evaluation.In this paper,18 subject attributes are separated from the test questions by artificial coding.Starting from the four dimensions of base-development,problem type,knowledge,and ability,nine topic attributes that have a significant impact on the difficulty of the test questions have been selected through T tests for 160 topics.They are base-development,problem type,algorithm,sequence,arithmetic solution,data processing,abstract summary,analysis and problem solving,reasoning and innovation.Based on the extracted topic attributes,33 text word vectors,66 text word vectors,and combination of title attributes and text word vectors as independent variables,a multiple linear regression model was proposed.In order to reduce the risk of overfitting,the support vector regression model was further proposed and the penalty function was used.To reduce the complexity of the model has solved the problem of overfitting.The empirical results show that the vector SVM model based on topic attributes and 33 text words is superior.It has a smaller mean square error on the test set and the model fits well.In addition,because the difficulty itself is a vague concept,the difficulty of the questions is divided into four levels from easy to difficult,and the difficulty level is classified by the support vector machine for the title and answer text and the title attribute respectively.The results show that the prediction accuracy of the support vector machine model based on the topic attributes is higher in the two prediction models,which is 75%.It also confirmed the feasibility of pre-testing difficulty levels with support vector machines.