Research On Automatic Grading Algorithm For Essay Questions Based On Yolo V4+Word2Vec

In the field of education,the grading of essay questions based on natural language processing(NLP)has attracted a lot of researchers.Research on automatic scoring of question and answer based on natural language processing is mostly oriented to electronic documents,while the current homework and test questions are mainly based on paper-based documents that far from practical application.It has not yet been a complete solution of how to apply the automatic scoring method to the evaluation of paper objects which has become the key and difficult problem in the application of natural language processing in the education industry.Under the background mentioned above,this research starts with the teacher’s examination paper correcting work,observes and investigates the teacher’s correcting action.According to the requirements in the process of examination paper correcting and the characteristics of computer work,a three-stage automatic grading system framework of paper examination paper was determined,which includes detection firstly,then identification and grading at last.Then,took the subject examination of System Modeling and Simulation as the research object,the test papers are designed,made and collected.Finally,a set of automatic scoring system for paper papers was realized by combining target detection algorithm,OCR API and automatic scoring algorithm for essay questions.For the application scenarios of this research,the existing Yolo v4 object detection algorithm was used to build the test paper data set firstly whose preprocessing of target detection and the video frame detection method was proved,thus realize the target detection of the name column and topic content of the test paper.Secondly the Tencent Cloud OCR API was used to realize handwriting recognition of the object detection results;Thirdly Word2 Vec was used to train a word vector model with a mixed corpus containing Wikipedia,domestic news,and subject professional text,and use word and sentence vectors as text features to calculates similarity between texts,and formulates scoring rules for essay questions.finally,regular expressions,multi-threading and other methods was used to connect Yolo v4,OCR,and semantic similarity scoring algorithms.The experiment showed that: Yolo-v4 network that improved the image data enhancement method,obtained 68.77% mAP50 scores and 49.48% mAP75 scores.The "vertical translation" video detection scheme was used to improve the detection accuracy of the test paper content by 21.25%.The MSE of the Word2 vec similarity scoring algorithm and expert scoring are all 0.81 and below,and the error is less than 10%.It can meet the requirements of correcting in the test of the integrated algorithm.This resarch has successfully realized the prototype system of automatic marking for essay questions of examination paper.The research results have provided theoretical basis and method reference for the intelligent marking of paper objects.The prototype of the exam paper automatic scoring system was achieved by the use of the object detection,OCR,question and answer automatic scoring theories.The realization of this prototype system proves that it is feasible in theory and technology to use these three algorithms to achieve automatic grading of paper papers,which has important application value for reducing teachers’ workload and a certain reference value for future research.