A parallel neural network structure for sentiment classification of MOOCs discussion forums

Yi Gao, Xia Sun, Xin Wang, Shou Xi Guo, Jun Feng

Abstract: Forum posts in Massive Open Online Courses (MOOCs) support an important way for online learners to interact with each other and with instructors. Instructors explore the sentiment from posts in MOOCs to detect learners’ trending opinions towards the course so that they can improve MOOCs. However, it is unrealistic to expect instructors to adequately track learners’ sentiment under the large number of messages exchanged on the forums. Fortunately, sentiment classification can automatically analyze learners’ emotion on the course of MOOCs from posts. Traditional classifiers based on machine learning algorithm, which often depend on human-designed features and have data sparsity problem. In contrast to traditional approaches, we develop a novel neural network model called parallel neural network (PNNs) for sentiment classification of MOOCs discussion forum to alleviate the aforementioned problems. In our model, we design a parallel neural network structure to replace the popular serial neural network structure so that PNNs can preserve the validity of features as far as possible when neural network model training. Meanwhile, we also introduce Self-attention mechanism that automatically identifies which features play key roles in sentiment classification to obtain the important components in posts. We experiment on a public MOOCs dataset and two common sentiment classification datasets, and achieve a good performance. That means PNNs is a substantially reliable classification model for identifying the sentiment polarity of posts. The study has great potential application value on the platform of large scale courses, which can help instructors to gain the emotional tendency of learners for the course content in real time, so that timely intervention to support learning and may reduce the dropout rates.