Research On The Automatic Short Answer Grading Method

With the rapid development of artificial intelligence and big data technology,the intelligent education model with which as basis has gradually become the mainstream of educational information development and a research hotspot in academia.Automatic grading is an important part in the field of intelligent education,to which scholars have proposed many solutions,but those methods are generally considered not suitable for automatic short answer grading of subjective questions.At the same time,with the growing number of various online exams,the manual grading of short answers has become more and more burdensome.If this task can be done automatically by a computer,it will not only greatly save the workload of teachers' grading,but also solve the problem of inconsistent grading caused by the subjectivity of different reviewers.In addition,the report on the 19 th National Congress of the Communist Party of China also clearly stated: "Strengthening education is fundamental to our pursuit of national rejuvenation.We must give priority to education,further reform in education,speed up its modernization,and develop education that people are satisfied with." Therefore,this paper with focus on short answer grading can not only improve the existing automatic grading system,but also enrich the application scenarios of intelligent education and promote the rapid development of education modernization.After a review of the development of automatic grading studies at home and abroad,this paper firstly analyzes the characteristics of short answer data and the shortcomings of existing methods.Subsequently,considering the features including the shortness of answer text,the mass of repeated sentences,the loudness of noise,and proneness to segmentation error,the paper proposes automatic grading model and puts forth the GCN model-based automatic short answer grading method as the existing methods generally ignore the word co-occurrence information in thecorpora and the global interaction information between samples.Finally,various types of experiments are conducted on the Chinese data set and the public Sem Eval-2013 English data set.The main work of the thesis includes two aspects.Firstly,given the features including shortness of the answer text,the mass of repeated sentences,the loudness of noise,and proneness to segmentation errors,etc.,an automatic grading framework based on character-level RCNN model is constructed.To be specific,firstly build answer text corpus,illegal vocabulary,and concept table based on students' answers and the key points,and grade student answers with obvious characteristics through rule matching strategies.Then the model uses a single character string as the input sequence,which can avoid the error transmission problem caused by the inaccurate word segmentation,and can also obtain the deep semantic information of the sentence by using the distribution vector representation method of characters.Finally,based on the proposed character-level RCNN model,an automatic short answer grading system is designed and implemented.Compared with the experimental results of multiple classic text classification models,the model achieves better performance.Second,in order to learn the co-occurrence relationship of words in the corpus and the global interaction information between samples,an automatic short answer grading method based on the GCN model is proposed.Typical short answer automatic grading methods are usually based on machine learning and neural network models with the former overly dependent on hand-designed features,limited by scalability and higher costs,and the latter ignores the word co-occurrence relationship and the sample between the corpus Global interactive information.However,automatic short answer grading often requires global information to learn multiple expressions of the same meaning and the relationship between different expressions and grading labels,so this paper attempts to use two layers of GCN(Graph Convolutional Neural Network)to encodeheterogeneous text graph composed of all student answers.The heterogeneous text graphs have sentence,vocabulary and phrase(word /bigram)level nodes,and construct connected edges between them according to the inclusion relationship or co-occurrence relationship between the nodes,and use the sentence-level TF-IDF value or PMI values as the edge weights represent the strength of the association between two nodes.The automatic short answer grading experiment conducted on the Chinese data set and the public Semeval-2013 English data set verified the effectiveness of the model.