Learning Behavior And Sentiment Analysis In MOOCs

The outbreak Coronavirus pneumonia in this spring is a big test for the online education of our country.Different kinds of schools carry out types of online teaching during the infection prevention and control,achieving the goal of no suspension of classes.This shows the necessity and importance of further development of online education.As a form of online education,massive open online courses（MOOCs）are widely used and have boomed in recent years because of advantages such as non-geographical and no time limitation.However,most courses have high dropout rates and learners are prone to have poor performance due to the unconstrained learning environment.Previous studies demonstrate that learning behaviors and sentiment are important factors influencing learning performance.Thus,learning behavior and sentiment analysis are key aspects to reduce dropout rate and improve learning performance in the online learning environment,which have become focus of current research works.Although there have been many studies,the following problems still exist in the applications of learning behavior and sentiment analysis.Firstly,for dropout behavior prediction,learning behavior discrepancy leads to a wide range of fluctuation of prediction results,which may result in low prediction results.Secondly,there lacks of causal analysis between learning behaivors and performance.Most studies have focused on correlation analysis between them,resulting in unreliable conclusions and uneffective decision support.Thirdly,for sentiment analysis,supervised methods rely on a large amount of labeled data.Constructing large-scale labeled datasets is very laborious and time consuming.Besides,there exist a large amount of unlabeled data that have not been fully utilized.To address the problems mentioned above,this dissertation studies the learning behavior and sentiment analysis in online learning environment,aiming at accurate identification of at-risk learners and their sentiment,and making causal analysis.The main works and contributions of this dissertation are summarized as follows:（1）A dropout behavior prediction approach based on web log is proposed.Based on the learning behavior feature extraction from log records,this approach proposes a novel hybrid model combining decision tree and extreme learning machine.The key of this model is to map decision tree to extreme learning machine,to realize feature selection,fast training and dropout behavior prediction simultaneously.Experimental results demonstrate that our approach not only improves the accuracy and F₁-score,but also requires no iteration training.（2）A causal analysis approach between learning behaviors and learning performance is proposed.Expert knowledge is utilized to construct the initial network.On this basis,structure learning which learns network from data is utilized to construct the final causal Bayesian network.Then we make reasoning based on the constructed causal network,to mine the behavior patterns with good and poor learning performance.Experimental results demonstrate that our approach not only improves the accuracy of learning performance prediction,but also makes causal analysis and proposes five kinds of intervention suggestions.（3）A co-training semi-supervised learning model for sentiment polarity identification of course review is proposed.The model encodes texts into vectors from views of word embedding and charater-based embedding,taking advantage of static word vectors generated by large-scale corpus and dynamic word vectors generated by domain corpus.To select high confident samples,a double-check strategy for sample selection is proposed.To address the classification imbalance and promote the identification performance,a mixed loss function based on semi-supervised learning is proposed,both considering the labeled data with asymmetric and unlabeled data.Experimental results demonstrate that our model not only improves the accuracy and F₁-score of sentiment polarity identification with only a small amount of labeled data,but also has higher stability.（4）An unsupervised rule-based aspect and opinion extraction approach for course review is proposed.Based on part-of-speech tagging and dependency parsing,this approach proposes a series of rules for extracting aspects and opinions according to the part-of-speech tagging pattern and dependency relationships between words.Furthermore,the aspects and opinions in candidate set generated by rules are filtered.Moreover,the sentiment polarity of opinion is identified to analyze the reasons for the positive and negative reviews from the fine-grained level.Experimental results demonstrate that our approach not only improves the precision and recall of aspects and opinions extraction without labeled data,but also makes analysis of the aspects and opinions from teaching,educational resources and online learning platform.