Attention-based contextual local and global features for urgent posts classification in MOOCs discussion forums

Abstract

The Massive Open Online Courses (MOOCs) platform offers communication channels for students to share concerns about the educational process. Due to the large number of students compared to the instructors’, it is challenging to identify urgent forum posts that require attention and prompt response from the instructor. This paper presents an innovative automated classification model called the “Attention Based on Contextual Local and Global Features (AT-CX-LGF)” classifier to identify MOOCs’ urgent posts. It can aid instructors in managing many posts and prioritizing their responses, allowing them to respond more quickly to student questions and reduce dropout rates while increasing completion rates. The suggested model obtains word embedding to represent the context information using BERT (Bidirectional Encoder Representation from Transformer). It depends on several phases. First, it extracts local and semantic (or global) contextual features using multi-layer CNN and Bi-LSTM. Then, two attention layers parallelly identify the most significant local and global features. After that, the outputs of the attention layers are concatenated and normalized. Finally, fully connected, and sigmoid layers are used for the classification process. On three groups (A, B, C) gathered from the Stanford MOOC Posts dataset, the AT-CX-LGF classifier obtained urgent posts recall of 87%, 87.1%, and 90.6% with 5.5%, 2.4%, and 7.5% improvements over the most recent algorithms, respectively. Furthermore, the model outperformed the state-of-the-art method in the weighted F1-score with handling the concept drift of the dataset.

Keywords

• Multi-head self-attention
• MOOCs
• Natural language processing
• Deep learning
• BERT