A Research For Automatic Scoring Approach And Interpretability Of Subjective Tests Based On Deep Learning

Manual marking is still the main marking method in the current educational environment,but it will also increase the workload of teachers and make it difficult to test the learning effect of students.The automatic scoring of subjective questions can be aimed at solving the above problems,that is,the workload of teachers’ marking papers is greatly reduced,and students’ papers are graded and analyzed in real time,so as to better urge teachers to adjust teaching strategies and improve teaching work.In addition,considering that the educational scene needs to achieve man-machine trust,and the research on automatic scoring of subjective questions is set up to assist teachers’ teaching and students’ learning,this means that if the scores can be reasonably explained,the persuasiveness of the results can be enhanced.Therefore,this paper will further explore the interpretability of automatic scoring of subjective questions on the basis of automatic scoring of subjective questions.Aiming at the problems of polysemy,sparse features and less contextual information in automatic scoring of subjective questions,this paper proposes an automatic scoring method of subjective questions based on deep learning(ACMB),which fully considers that BERT model can alleviate the problems of polysemy in automatic scoring of subjective questions,and uses adaptive convolutional neural network to extract local features of subjective questions to solve the problems of sparse features and less contextual information in subjective questions.At the same time,the adaptive convolution generation network solves the disadvantage that the existing convolution neural network uses the same filter weight for different input subjective text data in each iterative training period.In addition,in order to solve the problem of poor interpretability of deep learning,this paper uses the interpretability analysis framework based on SHAP attribution analysis,and combines local interpretation and global interpretation to interpretably analyze the scoring results of ACMB method.The experimental results on ASAP and ASAP-ZH data sets show that compared with other baseline models,the ACMB method proposed in this paper is superior to previous studies in all data subsets.At the same time,SHAP can also effectively extract the scoring points of texts and make interpretable analysis on ACMB method,thus further enhancing users’ trust in ACMB method.