A Code Automatic Evaluation Method For Students'Program Based On Code Format And Semantic Features

With the development of computer science and technology and the continuous ma-turity of online education,there are more and more computer courses online and offline,and the number of students participating in various computer courses ranges from thou-sands to millions.Although the continuous development of high and new technology has improved the ability of schools or educational platforms to provide content to large-scale students,it is still difficult to provide students with timely and accurate assessment and feedback(including correction of homework and course feedback).At present,pro-viding personalized feedback is the main manifestation of high-quality education,but for small face-to-face classrooms,the cost is already very high,and for large-scale class-rooms,the cost is prohibitively high.Therefore,how to automatically provide timely and accurate feedback is the core challenge of large-scale computer science education.This thesis designs a student code automatic evaluation model based on code for-mat and semantic features to realize the automation of code correction.When evaluating the student code,the student's code format,algorithm ideas and whether the program is correct are all important evaluation indicators.Therefore,this thesis summarizes the characteristics of student codes and designs an evaluation model for these characteris-tics.At the same time,based on the realization of the code evaluation model,this thesis optimizes for the problem of slow model training.The main research content of this paper is as follows:This thesis designs an automatic code evaluation model for stu-dents' code algorithm ideas and code correctness.The model contains the following three important modules:(1)Code embedding module based on code semantics.This module uses the NCC method for code embedding training,effectively transforming the semantic features of the code into a vector sequence.(2)Code feature extraction mod-ule based on attention mechanism.This module uses the hidden attention mechanism to deeply extract the code features in the code vector sequence,and provides a more excel-lent code feature representation for the scoring model.(3)Code correctness detection module based on black box testing.This module uses a black box test method to de-tect the correctness of the student's code and provides the correctness guarantee for the scoring model.Next,this thesis designs a selection function based on the self-attention mechanism to solve the problem of high training overhead due to the long input se-quence length in the code evaluation model.At the same time,this thesis designs a new CNN structure to replace the RNN structure in the code evaluation model to achieve the purpose of reducing model parameters and optimizing training time.Finally,this thesis designs an automatic evaluation model for student code format,and designs a module to convert codes into pictures.This module can convert code text into pictures according to specific rules to highlight the format features of the code.After experimental verification,the code automatic evaluation model for student code algorithm ideas and code correctness proposed in this thesis has reached an eval-uation accuracy rate of 83.1%under the real data set,and the three important modules have played a very important role.The optimization scheme designed in this thesis with high cost of true model training can reduce the model parameters by 53.1%,and reduce the model training time by 30%-45%,while only reducing the model accuracy rate by less than 5%.The automatic evaluation model designed in this thesis for the student code format achieves an evaluation accuracy of 82.2%under the real data set.