Research On The Problem Difficulty Prediction Based On Deep Learning

With the popularization of the Internet,the number of people in online education has reached a new high,and more and more unlabeled problem data have been uploaded to the Internet,making it difficult for student users to distinguish the difficulty of the problem.In order to enable users to better choose their own suitable difficulty of the problem,save the time of users to screen the problem,and help users to carry out personalized learning,the prediction of difficulty of the problem has become an urgent problem to be solved.Problem difficulty prediction is one of the important problems in the field of educational data mining.The existing problem difficulty prediction models are based on the estimation of the problem difficulty by professionals,or mining the relevant characteristic information from a large number of users’ questions.However,such a method is greatly affected by human subjective factors,and its efficiency is relatively low.Therefore,a new model is needed to accurately and quickly predict the difficulty of the problem.For the prediction of difficulty of the problem,this paper will use the methods of data augmentation and neural network language model in natural language processing to analyze and use the text information of the problem to predict the difficulty of the problem.Aiming at the problems existing in the difficulty prediction,a method based on natural language processing is adopted to predict the difficulty of the problem.Firstly,the neural network model based on attention mechanism is used to enhance the problem text data set.Second,multi-tasking learning is used to improve the performance of difficulty prediction.The main work of this paper is as follows:(1)Through the web crawler technology,different data sets are obtained from four different online question judging systems,and these data are preprocessed,and the difficulty of the problems is marked and classified according to the user’s question information.(2)A data augmentation model DASA(Data Augmentation based on Self-Attention)based on the attention mechanism is proposed for the task of problem difficulty prediction.The data augmentation model uses textual random masking and self-attention mechanism.In the experiment,by comparing DASA with the data augmentation methods commonly used in natural language processing on multiple problem difficulty data sets,the experiment proves that the data set enhanced by DASA has a significant improvement in the accuracy of the model.(3)Because the data sets of the problems in a certain discipline are relatively small,and the content distribution of the problems in different online learning platforms varies greatly.In order to reduce the required data set information,further improve the accuracy of difficulty prediction and reduce the training time.In this paper,we apply the multi-task learning model to the BERT(Bidirectional Encoder Representation from Transformers)model,and propose a MTBERT(Multi-task BERT)problem difficulty prediction model.The features from different data sets are shared by the MTBERT model to improve the generalization ability of the model.Experiments were carried out in Leetcode,Zoj and other real data sets,and the traditional neural network was compared with the original Bert model,thus verifying the effectiveness of the method.Finally,the experiment shows that the techniques of data enhancement and Neural network language model can effectively improve the accuracy of difficulty prediction and the efficiency of model training.