Research And Implementation Of Automatic Essay Scoring Model

The automatic essay scoring technology can automatically analyze and score the essay,which has become one of the hot research issues in the application of natural language processing technology in the education field.Since it was proposed in 1966,automatic essay scoring technology has been successfully applied in large essay examinations such as Educational Testing Service（ETS）and Chinese College English Test（CET）,and has played a good supporting role in composition scoring.Automatic essay scoring not only saves the cost of manpower and material resources,but also greatly improves the fairness and accuracy of essay scoring.Compared with the traditional machine learning methods,the neural network method based on deep learning has achieved better performance in the essay scoring task.However,essay scoring is a complex human behavior,which requires comprehensive evaluation of essay from different levels,such as essay theme,words,rhetoric and so on.Therefore,the performance of a single neural network model is often not ideal.Secondly,the deep-seated neural network model has a large number of parameters,which requires more computing resources in the process of model training.In addition,although pre-trained word vectors have achieved good performance in many tasks,their performance is not ideal in automatic essay scoring.To solve the above problems,this paper first studies the impact of deep and shallow semantic features on the performance of essay scoring;Secondly,an end-to-end lightweight automatic essay scoring model is studied;Thirdly,the semantic deviation caused by pre-trained word vector is deeply discussed;Finally,on the basis of summarizing the advantages of the first three parts,this paper proposes an automatic essay scoring method based on heterogeneous network fusion.The main research content of this article includes the following four aspects:（1）Aiming at the current automatic essay scoring method that separates the deep and shallow semantic features,and ignores the impact of multilevel semantic fusion on essay scoring,this paper proposes a neural network model based on multilevel semantic features.Firstly,the convolution neural network is used to capture the local semantic features,and the hybrid neural network is used to capture the global semantic features,so as to obtain the semantic features of the essay from the deep level;Secondly,the topic level features are obtained by using the topic vector at the text level,and the shallow linguistic features such as grammatical errors which are difficult to be mined by the neural network model are constructed at the same time;Finally,the essay is scored automatically by feature fusion.The experimental results verify the effectiveness of the algorithm in automatic essay scoring tasks.（2）Aiming at the inefficiency of the current automatic essay scoring method based on deep learning and the limitations of Feature Engineering,this paper proposes a lightweight automatic essay scoring method based on attention word embedding network.This method adopts an end-to-end training method,does not include any feature engineering,has few parameters and is easy to train.The experimental results show that the model can effectively score the essay automatically,and the efficiency of the model is significantly improved.（3）It is found that because the training corpus of pre-trained word vector is quite different from the essay corpus in semantic expression and language style,the use of pre-trained word vector will bring the problem of semantic deviation.Therefore,this paper proposes an automatic essay scoring model based on hybrid word vector,which includes pre-trained word embedding and self-training word embedding at the same time.Experimental results show that this model can effectively alleviate this semantic deviation problem.（4）Aiming at the problem that the current automatic essay scoring methods lack the fusion of different structure neural networks and ignore the complementarity of the composition semantics extracted by different structure networks,based on the advantages of the first three methods,this paper proposes an automatic essay scoring method based on heterogeneous network fusion,including convolution neural network,recurrent neural network and self attention network.In addition,through experiments on pre-trained word vectors with different scale structures,the effects of different pre-trained word vectors and self-training word vectors on the performance of automatic essay scoring are analyzed.