Automatic Evaluation Of Summative Texts Of University Courses Based On Deep Learning

Summative texts are usually written descriptions by students to summarize the course content and learning outcomes.At present,summative texts evaluation methods for courses are often incorporated into grade evaluations for non-exam courses in universities,which is more effective than the traditional evaluation methods that directly answer simple questions such as multiple-choice questions and right/wrong questions.In recent years,under the boom of large-scale online learning and assessment system,although there have been a lot of research works to improve the performance and effectiveness of the system,the automatic evaluation of university course summative texts is still a very important problem to be solved.According to the Xinhua dictionary and Baidu encyclopedia,composition refers to writing an essay or a writing exercise for students,so the summative texts of university course can be classified as composition.With the rapid development of computer technology,AES(Automated Essay Scoring)has been further applied.The analysis and evaluation of essays are the main elements of this area,which has lower labor cost and higher grading efficiency than teachers directly correcting essays.At present,the commonly used model for constructing and evaluating AES systems are still a combination of neural network models and hand-crafted features;however,the highly variable nature of university courses summative texts types and the fact that university courses summative texts are usually longer unstructured documents with higher requirements for contextual semantic information pose challenges for building automatic evaluation models for summative texts.Therefore,to address the above issues,this thesis is based on the BERT(Bidirectional Encoder Representation from Transformers)model and optimizes it to make AES better applied to the field of automatic evaluation of summative texts for university courses.The thesis mainly includes the following three contents:(1)Because the BERT model is a huge and complex hierarchical model,the gradient may disappear during the pre-training task.Using the Real Former method to modify the training method of the BERT pre-training model;Aiming at the problem that the training corpus of the basic BERT pre-training model has nothing to do with the research field of this paper,collect the corpus of professional fields to re-train the pre-training model;The Mean-Max-Pooled method is used to convert the word vector of BERT model into sentence vector to obtain more text semantic information.Finally,the optimized BERT model is obtained to complete the automatic scoring task of summative texts of university courses.The accuracy of the model reached 79.17%,which is 12.5% higher than the basic model.(2)An evaluation index system proposed for the summative texts of university courses,which contains three features: keywords,key sentences,and topic words.Firstly,the text is extracted with feature vectors using an optimized BERT model and multiple features are fused.Secondly,the text is trained using a text classification method.Finally,the task of automatic scoring of university courses summative texts are completed,and the accuracy of experimental results using this method is improved by 3.22% compared with the basic model.(3)Using the optimized BERTmodel obtained based on the previous research,an automatic evaluation system of university courses summary texts are successfully built and applied to the actual environment to effectively evaluate the courses summary texts,which verify the effectiveness of the above research technology.