Research On Automatic Grading Method And Text Feature Selection Of Subjective Questions In History Test Paper

In recent years,the rapid development of artificial intelligence has made breakthroughs in many fields such as speech recognition,image processing,automatic translation,and automatic driving.With the rise of a new generation of artificial intelligence,it has also injected new vitality into education reform.At present,the main method of testing teaching results and students' learning in the education field is traditional paper-pencil testing.Knowledge inspection is divided into objective questions and subjective questions.The answers to objective questions are deterministic and the correction is relatively simple.The knowledge and skills are relatively small;the answers to the subjective questions are vague in terms of presentation,the correction is relatively complicated,and the knowledge and skills to be investigated are relatively large.Domestic scoring of subjective questions is mostly done by teachers' manual marking.The speed of marking is limited.Manual marking scores are easily affected by the teacher's subject knowledge,understanding of standard answers,psychology,and physiology.Therefore,the research on the automatic scoring technology of subjective questions has very important practical significance.Subjective questions can be divided into open-ended answer subjective questions and subjective questions with reference answers according to whether there are standard answers.The typical representative of open-answer subjective question scoring is composition scoring,also known as composition automatic scoring;subjective question scoring with reference answers is also called short-answer automatic scoring.At present,there are four main technical routes for automatic scoring of short answers: rule-based,similarity-based,machine learning,and deep learning.This article mainly uses deep learning to solve the problem of automatic scoring of short answers.This paper uses the student answer data of the new senior high school entrance orientation historical test in a certain city in the 2020-2021 school year as the research data,and mainly conducts two researches:(1)Research on the impact of different deep learning models and scoring model combinations on QWKappa and accuracy;(2)Integrate ordinal-logistic to construct an ordered scoring model,and compare it with the multi-category scoring model.The research found that:(1)Among the four deep learning models,the QWKappa and accuracy of Bi LSTM+Text CNN are better than those of the other three deep learning models,and they are consistent across the scoring models.It shows that using the method of model fusion,the scoring model constructed by fusing the convolutional neural network and the recurrent neural network is better than using the convolutional neural network or the recurrent neural network alone;(2)the scoring model and the deep learning model are interactive Function,in Bi LSTM+Text CNN and Text CNN,the scoring model of representative answers and single-choice questions has achieved the best results,but in LSTM and Bi LSTM,the scoring model of single-choice questions has achieved the best results.In addition,the scoring model that considers multiple choice questions is better than the scoring model based on student answers and the scoring model based on reference answers.This is manifested in the construction of multi-category scoring models and ordered scoring models;(3)in many cases.Based on the classification scoring model,the ordinal-logistic is used to construct an ordered scoring model.It is found that the ordered scoring model can better utilize the ordered information between the scores,making the machine score and the manual score more consistent,but in different The effect of improvement on subjective questions is different;(4)Although the QWKappa value performs better in each scoring model,the accuracy rate is generally low.This is mainly due to the limited training data size but the greater difficulty of the task.