Automatic Scoring Of Chinese Subjective Questions Based On Word-level Deep Neural Network

The traditional written test is still the most widely used form of investigation in modern teaching,and the types of test questions can be divided into two types: objective questions（such as true and false questions,fill-in-the-blank questions,and multiple-choice questions）and subjective questions（such as composition and short answer questions）.Among them,because the answer to the objective question is uniquely determined,and there are few knowledge points to be examined,the scoring is relatively simple;in contrast,the answer to the subjective question is more open,and more knowledge points are examined,and the scoring is relatively difficult.At present,manual scoring is still mainly used.However,the use of the original manual scoring method is highly susceptible to subjective factors（such as whether the students’ handwriting is beautiful and clear,whether the paper is clean and tidy,and whether the teacher is in a good state at the moment,etc.）.There is a positive correlation between the number of examinee and the size of the test.Therefore,the research on the automatic scoring method for Chinese subjective questions has important practical significance and application value.This paper comprehensively compares the research status of automatic scoring methods for Chinese and English subjective questions at home and abroad,and analyzes the characteristics of the Chinese subjective question data set and some shortcomings of the existing automatic scoring methods.Aiming at the phenomenon that Chinese word segmentation tends to cause errors to propagate layer by layer in the network,a "word-level CNN question classification method with residual connections" is proposed.Aiming at the problem of low matching degree of answers,this paper proposes an automatic scoring method for Chinese subjective questions based on Siamese and ESIM networks.Finally,on the self-built Chinese subjective question question classification data set and automatic scoring data set,the comparison experiments and ablation experiments of the two types of models are carried out respectively.The main work and innovations of this paper are as follows:（1）Aiming at the lack of Chinese datasets,this paper collects and organizes "Computer Network final exam questions for undergraduates majoring in software engineering in a university over the years" and "The real questions of the Computer Network grade examination over the years".Finally getting the Chinese subjective question classification data set containing 3948 questions.As each question has 1 standard answer and2 student answers,thus obtaining a Chinese subjective question automatic scoring data set containing 7896 student answer data.（2）Aiming at the phenomenon that Chinese word segmentation tends to cause errors to propagate layer by layer in the network,this paper proposes a "word-level CNN question classification method with residual connections".Given a question to be classified,the input layer of the model first preprocesses it,that is,constructs a dictionary containing all the words in the question classification data set.Then,the dictionary subscript corresponding to each word in the sentence is input into the embedding layer of the model,and a 300-dimensional word vector representation is obtained by training with the Word2 vec tool.Next,the word vector representation is input to the convolutional layer,FNN layer,residual connection,and pooling layer of the question classification model in this paper in turn,and the final semantic feature representation of the question to be classified is extracted.Finally,the prediction layer of the model uses the Softmax function and the argmax function to output the class label of the question.The model uses a cross-entropy loss function to measure the difference between the true class label and the predicted class label,and uses error back-propagation and gradient descent algorithms to optimize the model parameters.Through multiple sets of comparative experiments and ablation experiments,the validity of the question classification model in this paper is analyzed and demonstrated.（3）Aiming at the problem of low matching degree of answers,this paper proposes an "Automatic Scoring Method for Chinese Subjective Questions Based on Siamese and ESIM Networks".The input of the scoring model is the word vector concatenation representation of the question,the question category label,and the answer（student answer or standard answer）.Then,the embedding representation of the two answers（student answer and standard answer）is encoded separately using the Siamese-CNN model.Next,the shallow semantic feature representation extracted by CNN is input to the ESIM network,and successively passes through the BiLSTM encoding layer-> Attention interaction layer-> BiLSTM synthesis layer-> Average and maximum pooling layer,the output is a fixed-length vector containing the overall semantic relevance of the two answers.Finally,the prediction layer of the model uses the Softmax function and the argmax function to convert the final vector representation into class labels for scoring intervals.Through multiple sets of comparative experiments and ablation experiments,the effectiveness of the automatic scoring model for Chinese subjective questions in this paper is analyzed and demonstrated.To sum up,the automatic scoring method for Chinese subjective questions based on the word-level deep neural network model in this paper can help reduce the teaching burden of teachers,improve teaching efficiency,and at the same time avoid the phenomenon of unfair and unjust scoring caused by teachers’ subjective factors.Thereby improving the accuracy of automatic scoring of subjective questions.