Online Course Recommendation System Based On Stance Detection

With the popularization of network terminals and the development of internet technology,the internet has become one of the main ways for people to acquire knowledge,and the arrival of the pandemic has ignited the spark of online learning.Online learning,which breaks the time and space constraints of traditional education,has played a huge teaching role during the pandemic and gradually become a popular learning mode.In order to facilitate learners to quickly select from massive learning resources,online course recommendation systems have emerged.In the development history of online learning platforms so far,a large amount of data has also been generated,and learners’ comments have always been considered to contain a lot of valuable information.The current recommendation system uses sentiment analysis to mine learners’ interests from comments,but sentiment analysis does not have the specificity of courses,and the emotional orientation expressed under different course stances may not necessarily match.Therefore,this paper uses a stance detection model to analyze learners’ comments on courses,mine learners’ interests,and use graph neural networks to learn the historical interaction information between learners and courses and the correlation information between courses themselves for recommendation work.The current stance detection models that integrate external knowledge in stance detection tasks concatenate the learned representations of text and knowledge and feed them into a classifier to complete the stance detection task.The method of directly concatenating text and knowledge representations does not compute the interaction features between the representations,and has weak modeling ability for the interaction information between text and knowledge.It cannot measure their similarity,which may lead to the problem of inaccurate latent common space representation.The current online course recommendation methods focus on using historical interaction data to model learners and courses.However,the real-world interaction data is often sparse,and using the correlation information between courses can effectively alleviate the sparsity.The historical interaction graph and the related course graph are two different structured graphs,and there are some correlations and differences between different graphs.Using multiple graph data can cover a wider range of learners and courses,and combining the two graph data can better reflect the multidimensional relationship between learners and courses.How to effectively combine the two graphs to model courses is a problem that needs to be solved.Therefore,the specific work of this paper is as follows:In order to model the interaction information between text and knowledge,we propose a stance detection method based on Prompt and external knowledge joint.Specifically,this method embeds the external knowledge and stance text directly into the model,and uses the Prompt learning method and the multi-head self-attention mechanism of the masked language model to learn the interaction features between text and external knowledge,to learn the representation on the latent common space.At the same time,it combines the sentiment dictionary to expand the Prompt label word set,so that the model can automatically capture the emotional features existing in the text when predicting the results.The experimental results show that this method achieves the best comprehensive performance,and has a better learning effect on the opposing stance.In order to combine the historical interaction graph and the related course graph to model courses,we propose a graph neural network rating prediction model based on related course embedding.Specifically,we first construct the related course graph by calculating the similarity between courses and their labels,then design a method to capture the interaction information between historical learners and ratings from the historical interaction graph,and extract the related course information from the related course graph when modeling courses,and combine the two kinds of information to complete the course modeling.Finally,we combine the learner representation and course representation to complete the rating prediction for courses.The experimental results show that this method can learn richer course representations and have smaller error in rating prediction for courses.Based on the above research,a front-end and back-end separated online course recommendation system has been built using technologies such as Vue,Spring Boot,and Py Torch to achieve course recommendation.In the comment analysis module of the system,the proposed stance detection model is used to analyze the comment information as text for analysis,and learner information and course information are added as external knowledge to the model.The model mines and updates learner information based on the provided data.In the system’s personalized recommendation module,a recall module is used to form a set of candidate courses,and then the proposed course rating prediction model is used to predict ratings for the courses in the candidate course set,and finally,the courses are recommended based on the ratings obtained through sorting.