Research On Exercise Understanding And Application Algorithms Based On Deep Learning

It's an urgent need to apply artificial intelligence technology to online education platforms in the field of modern education.The current status of online education shows that high-quality teachers are still scarce resources,which limits the large-scale popularization of online education;Meanwhile,the current national policy states: ‘In order to solve the educational need of society,artificial intelligence should be applied to online learning educational platform.'Automatically recommending similar exercises to students based on natural language processing technology is a core application in the field of artificial intelligence online education,which can help students master knowledge points.Designing a proper exercise similarity model has important practical significance and application value for improving the quality of exercise recommendation and teaching efficiency.The traditional similar exercise recognition model has two shortcomings:(1)They only obtain the hyperplane that separates dissimilar exercises,but do not focus on increasing distances between dissimilar exercises while reducing distances between similar exercises;(2)They only focus on the information that the text of exercise's question can bring,but ignore the role of the exercise's answer text.These shortcomings lead to inaccurate results of recommending similar exercises by the current models.Therefore,the research goal of this article is to establish a comprehensive vector space representation of exercises,design an accurate exercise similarity model,and provide application tools for finding similar exercises.This paper innovatively designs exercise similarity models that can improve the representation ability of the exercise and the recommendation effect of similar exercises for the above shortcomings.The experiments are based on the data set of Chinese mathematical exercises in a real online education system.The contributions are as follows.1)In order to map the exercises to the exercise representation space where the distance between similar exercises is close and the distance between dissimilar exercises is far away,the exercise similarity models SBERT-CLS and TBERT-CLS based on Siamese architecture and Triplet architecture are designed using the Siamese network architecture.The latter has a MAP score of 0.61,which is 0.23 higher than the bestperforming baseline model VSM(i.e.,a relative increase of 60.5%).2)In order to better capturing the relationship between the exercise question text and answer text,the exercise similarity models SBERT-QA and TBERT-QA are designed to support not only the matching of the two question texts and the answer text,but also the question text of one exercise and the answer text of the other exercise.The latter has a MAP score of 0.61,which is 0.04 higher than the TBERT-CLS model that only considers the question text similarity matching(i.e.,a relative increase of 6.6%),which proves the importance of the comprehensive consideration of the question text and the answer text for obtaining effective exercise representation.3)In order to further obtain a comprehensive text representation of the exercises,we use Text-CNN to perform pooling operation and design the exercise similarity models SBERT-CNN,TBERT-CNN,SBERT-QA-CNN and TBERT-QA-CNN.Among them,the MAP score of the TBERT-QA-CNN model is 0.66,which is 0.01 higher than the TBERTQA model without CNN pooling(i.e.,is an increase of 1.5%),and 0.28 higher than the best performing baseline model VSM(i.e.,a relative increase of 73.7%).This article visually analyzes the exercise representations obtained by the designed exercise similarity models,and compares the specific performance of each model in recommending similar exercises with actual recommendation cases,proving the role of the model we designed in improving the recommendation effect of exercises and verifying the practical significance and application value of the model.