| With the development of artificial intelligence and big data technology,a large number of internet exam platforms have emerged,and online exams have become one of the main exam forms in the current information age.Traditional teaching evaluation not only increases the workload of teachers,but also fails to timely improve students’ programming abilities.Therefore,online detection of student learning outcomes has become a popular research direction.As of now,the automatic scoring technology for objective questions is relatively mature,and there is still no appropriate solution for evaluating procedural subjective questions.For students,the subjective step scoring of procedural questions can improve their hands-on practical skills,timely grasp programming knowledge,and for teachers,it can reasonably formulate classroom teaching plans,improve classroom teaching efficiency,greatly promoting the development of information technology teaching.Therefore,the study of procedural step scoring is of great significance.The difficulty of evaluating subjective questions in programs lies in the diversification of students’ subjective consciousness,which leads to the diversification of problem-solving methods and the ever-changing structure of programs.The original automatic evaluation technology for programs only focused on program results,without considering students’ answering ideas and programming steps.Therefore,there are issues with low accuracy and incomplete scoring strategies in program evaluation technology.Based on the original automatic evaluation method,this article has done the following work:1.Propose an automatic scoring method based on multi-level feature fusion,which mainly includes three steps: first,the first layer detects whether the student program results are correct,and then compares the test cases to detect the student program results;Secondly,a second level of matching is carried out for keywords,and student programs are uniformly standardized.Attention mechanisms are used to assign weights to different keywords,and the editing distance algorithm is used to match the similarity between student keywords and template keywords;Finally,the third layer evaluates the program structure,transforming the student program into an intermediate representation of an abstract syntax tree and eliminating redundancy in the abstract syntax tree.In order to reduce the time complexity of the algorithm,the Tree Kernel algorithm is used to extract subtrees and compare hash values to calculate the similarity between the student program structure and the template structure.This method has been validated through experimental assignments of C language students in our school to effectively solve the problem of automatic program grading,with an accuracy of 75%,and the results are closer to manual grading.2.Propose a method of transforming abstract syntax trees into graph structures and utilizing graph neural networks for feature extraction,in order to deeply explore the hidden syntactic and semantic features in programs.Although the program scoring method based on multi-level feature fusion conforms to the strategy of manual scoring and has a significant improvement in accuracy compared to traditional automatic scoring methods,there are problems such as incomplete manual feature extraction and high time complexity.Therefore,this part of the work is mainly divided into three steps: as abstract syntax trees are a special graph structure,the CBOW model is first used as a word vector generation algorithm,and the abstract syntax tree is transformed into a word vector through word embedding and positional encoding;Secondly,the generated vectors are input into the attention network mechanism of the graph to train feature vectors,learn the linear mapping weight matrix,aggregate the information of neighboring nodes,and obtain normalized attention coefficients;Finally,the obtained feature information is inputted into the twin neural network to calculate Loss,and the similarity between the source program and the template program is compared.Based on the scoring rules,the student program’s rating is obtained with an accuracy of 82.4%,which is closer to the teacher’s rating.3.Based on the framework of the multi-level feature based program automatic evaluation method and the research on graph attention networks,a C language online test automatic scoring system has been designed and implemented.The most important function of this system is daily testing,and the specific application scenarios are as follows: teachers release daily test exercises,students answer in the system,and can compile and run without jumping to the C language compiler.After running,the grades can be viewed,and the grades are graded step by step based on the traditional system’s 0or 100 points;Another function is online exams,where teachers publish online exam invitations,allowing students to participate in exams,and teachers can export student grades.The C language online test automatic scoring system meets the requirements of educational informatization for talent cultivation,while also reducing the workload of teachers,which is conducive to promoting teaching on a large scale.To sum up,this paper conducts in-depth research on program automatic scoring from data preprocessing(C language standardization),new automatic scoring model,design and implementation of online evaluation system,proposes a multi-level feature fusion program automatic scoring scheme,and integrates the new scheme into the program automatic evaluation system to achieve automatic scoring,and the scoring results are closer to manual scoring;At the same time,based on the multi-level scoring model,the abstract syntax tree of the intermediate representation of the program is further processed,and feature information is extracted through the graph neural network model.The scoring accuracy can reach about 85%.The research results of this article have promoted the development of program automatic evaluation research in the field of intelligent teaching,which is suitable for daily teaching. |
PDF Full Text Request