Research And Application Of Chinese Composition Scoring Based On Deep Learning

With the continuous upsurge of Chinese learning around the world,research on Automated Essay Scoring(AES)for Chinese has attracted increasing number of researchers' interest.The research goal of AES is to assign a proper score for an essay written by a student automatically,which is an important application of Natural Language Processing(NLP)in the field of education.AES can not only reduce the influence of subjective of human raters,but also decrease workload of human raters.At present,the research and development of AES system mainly focus on English essay,thus theoretical and practical research on Chinese AES has not been well developed.This thesis mainly focuses on solving the theoretical and practical problems existing in Chinese AES,and develops an AES system based on deep learning methods.The main research contents are listed as follows:(1)the classical Support Vector Model(SVM)has succeeded in many NLP tasks.This study uses regression-base SVM method to build a stable and effective baseline model for Chinese AES.(2)An AES model based on multiple linear regression has been constructed in this work by using various features extracted from an essay.Our experimental results show that the performance regression-based model is significantly better than that of the baseline model.(3)A LSTM(Long Short-Term Memory)-based AES model has been proposed in this thesis,which employs a CNN(Convolutional Neural Network)layer to capture local semantic information.(4)BERT,one of the most influential Pre-trained Language Models(PLMs),has made the state-of-the-art results in NLP community.A BERTbased AES model has been proposed by fine-tuning BERT for our task.Innovation of this thesis includes that we tentatively use varied approaches for Chinese AES study and build up a practical AES system.In order to improve performance of AES model,we introduce the latest PLM based on full word coverage Chinese Bert into the our task.Our results suggest that BERT-based Chinese AES model outperform other approaches,whose QWK(Quadratic Weighted Kappa)coefficient reaches 0.71.This thesis analyzes existing AES studies and uses the latest QWK coefficient as the unified evaluation method of the model.On the basis of the above research,this thesis develops an automatic scoring system for Chinese AES for Chinese learners.