BERT-Based Global Semantic Refinement and Local Semantic Extraction for Distinguishing Urgent Posts in MOOC Forums

Abstract

With the continuous development of the Internet, Massive Open Online Courses (MOOCs) have become the choice of more and more learners. MOOCs platforms provide a forum for students and teachers to interact in a timely manner and solve problems in the student learning process. Due to the large number of posts in the forum and the limited energy of teachers, many urgent posts that require priority responses are not handled promptly. In this paper, we propose a MOOC post classification model based on BERT to help teachers quickly and accurately distinguish urgent posts from non-urgent posts. The working process of the model is as follows: First, we use a pre-trained BERT to encode the post text; then, we use a downstream network based on a double pointwise convolution to perform global semantic refinement for the BERT encoding, and use a downstream network based on CNNs to perform local semantic extraction on BERT encoding, and finally merges two semantic refinement sub-vectors to classify the MOOC post. Experiments show that the urgent post’s F1-scores of our model on the three sub-datasets (Groups A, B, and C) of the Stanford MOOC post dataset achieve 83.54%, 81.79%, and 79.58%, respectively, which improves other methods on group A by 0.24%, group B by 0.89%, and group C by 0.88%. Our code is available at https://github.com/3269233844/MOOC-classification.

Keywords

• MOOCs
• post
• BERT
• global semantic refinement
• local semantic extraction