Research And System Realization Of Evaluation Algorithm For Short Text Subjective Questions

At present,there are various types of examinations,whether in schools or in professional title evaluation.In fact,examinations are a process of testing and selection,which can test the true level of candidates and provide accurate basis for talent selection and evaluation in relevant fields.With continuous development of computer and artificial intelligence technology,gradually changed the traditional mode of examination,online examination system become the important trend of the application of,in a lot of automated test system has been applied in the test,could finish the process of examination and grading based on computer efficiently,compared with the traditional exam mode is more efficient,cheaper,can get a more accurate assessment results.At present,in the online examination system,the automatic scoring technology is mostly used in the objective question scoring.This kind of technology is relatively mature and can get the results of scoring efficiently.However,there is still a lack of the application of automatic scoring technology in the subjective question scoring.There are many kinds of such questions,such as composition or other forms of short answer questions.The answering of subjective questions is open to a certain extent,and many questions do not have fixed answers,so the complexity of automatic grading is high.Nowadays,automatic grading cannot be achieved by relying on computers,but manual grading is still needed.Marking,therefore,people need to be in a very short period of time to read and understand an exam of expression,and then use their own standard to evaluate the degree of an exam to master knowledge,this way there is an obvious shortage problem,increase the grading teachers’ work pressure,reduce the efficiency of the score,and the accuracy of grading results often can not meet the higher requirements.This thesis focuses on the subjective topics of short texts with standard answers,such as short answer questions and noun explanation questions,and studies the automatic scoring algorithm of this kind of subjective questions,so as to achieve automatic scoring,which can play an auxiliary and double-check role in manual grading.In this way,the subjective consciousness of the reviewer can be relatively eliminated,and the accuracy and fairness of the scoring can be further improved.Focusing on the automatic scoring of short essay subjective questions,this thesis divides the score points of the standard answers of short essay subjective questions into fullmatch and non-fullmatch.(1)For the score points marked Full Match,regular matching is conducted between the score points and candidates’ answers;If the regular match is successful,the candidate’s answer will be subtracted from the corresponding fullmatch answer and the score value will be obtained.Otherwise,the original candidate’s answer will be retained and the score value will not be obtained.(2)The calculation of non-fullmatch score points can be divided into three steps.In the first step,features are extracted from the answers of candidates and non-fullmatch score points,and then they are spliced.The spliced vector value is input into the fully connected neural network to calculate the similarity value,and then the score A is calculated according to the similarity value.The second step is to extract the key words of non-fullmatch score points and conduct two-way matching with the answers of candidates to calculate the score B;In the third step,score A and score B are respectively weighted and added together to obtain the final score of candidates in the non-fullmatch section.The fullmatch score and the non-fullmatch score are added together to create the total score for the candidate.The main work of this thesis is as follows:(1)In this thesis,with the characteristics of the joining together of ideas,by extracting based on BERT generate other vector,based on the improved jaccard similarity coefficient and the entropy characteristics,based on dependency syntactic analysis and part-of-speech tagging feature vector and five parts based on semantic role labeling similarity characteristics and splicing,eventually form a long one-dimensional vector of length 1619.(2)According to the different methods of feature extraction,this thesis designs four kinds of network models.By selecting different hyperparameters for each model for training,and adopting the idea of data enhancement,the fully connected network model which is more suitable for automatic grading of short text subjective questions is finally selected.(3)Based on Text Rank algorithm,a simple and efficient keyword extraction tool is implemented for the subsequent model establishment.(4)Design and implement an interactive Web-side subjective question scoring review system with Django,and prove the usability and efficiency of the system.In this thesis,the essay the subjective topic to study the automatic grading model,and successfully applied to the system,in the process of grade a quandary,automatic grading can rise to review,the role of reference,can be eliminated relatively certain subjective consciousness,further improve the accuracy and impartiality of the score,the better able to ensure the fairness of the exam,has certain practical significance.