Bi-GRU Urgent Classification for MOOC Discussion Forums Based on BERT

Abstract

MOOC provides a forum for communication between students and teachers and is a platform for students to articulate their problems in the education process. Because of the large number of students and the small number of teachers, distinguishing posts that require rapid intervention is challenging. Therefore, there is a need to automatically classify the posts in the forum and distinguish urgent posts requiring a quick response. This classification is mainly based on the numerical representation of the text included in the students’ posts. BERT is a model for representing words according to the context in which they are used. These context-dependent representations can capture many syntactic and semantic properties of words under diverse linguistic contexts. This paper aims to utilize such representation as a pre-trained embedding layer for a broader task, and then it is fine-tuned to define the urgent post. This embedding layer passes on the classification model consisting of a multi-layer bi-directional Gated Recurrent Unit (GRU). Also, different types of embedding layers have been experimented with other deep neural network structures for comparison. Despite the simplicity of the proposed structure, using BERT as an embedding layer achieves urgent posts classification with a weighted F-score of 91.9%, 91.0%, and 90.0% on the Stanford MOOC Posts dataset. Distinguishing urgent posts with high accuracy helps teachers prioritize their responses and answer learner questions promptly. Also, this helps reducing dropout rates and improving completion rates.