Predicting the Need for Urgent Instructor Intervention in MOOC Environments

Abstract

In recent years, massive open online courses (MOOCs) have become universal knowledge resources and arguably one of the most exciting innovations in e-learning environments. MOOC platforms comprise numerous courses covering a wide range of subjects and domains. Thousands of learners around the world enrol on these online platforms to satisfy their learning needs (mostly) free of charge. However, the retention rates of MOOC courses (i.e., those who successfully complete a course of study) are low (around 10% on average); dropout rates tend to be very high (around 90%). The principal channel via which MOOC learners can communicate their difficulties with the learning content and ask for assistance from instructors is by posting in a dedicated MOOC forum. Importantly, in the case of learners who are suffering from burnout or stress, some of these posts require urgent intervention.
Given the above, urgent instructor intervention regarding learner requests for assistance via posts made on MOOC forums has become an important topic for research among researchers. Timely intervention by MOOC instructors may mitigate dropout issues and make the difference between a learner dropping out or staying on a course. However, due to the typically extremely high learner-to-instructor ratio in MOOCs and the often-huge numbers of posts on forums, while truly urgent posts are rare, managing them can be very challenging –– if not sometimes impossible. Instructors can find it challenging to monitor all existing posts and identify which posts require immediate intervention to help learners, encourage retention, and reduce the current high dropout rates.
The main objective of this research project, therefore, was thus to mine and analyse learners’ MOOC posts as a fundamental step towards understanding their need for instructor intervention. To achieve this, the researcher proposed and built comprehensive classification models to predict the need for instructor intervention. The ultimate goal is to help instructors by guiding them to posts, topics, and learners that require immediate interventions.
Given the above research aim the researcher conducted different experiments to fill the gap in literature based on different platform datasets (the FutureLearn platform and the Stanford MOOCPosts dataset) in terms of the former, three MOOC corpora were prepared: two of them gold-standard MOOC corpora to identify urgent posts, annotated by selected experts in the field; the third is a corpus detailing learner dropout. Based in these datasets, different architectures and classification models based on traditional machine learning, and deep learning approaches were proposed.
In this thesis, the task of determining the need for instructor intervention was tackled from three perspectives: (i) identifying relevant posts, (ii) identifying relevant topics, and (iii) identifying relevant learners. Posts written by learners were classified into two categories: (i) (urgent) intervention and (ii) (non-urgent) intervention. Also, learners were classified into: (i) requiring instructor intervention (at risk of dropout) and (ii) no need for instructor intervention (completer).
In identifying posts, two experiments were used to contribute to this field. The first is a novel classifier based on a deep learning model that integrates novel MOOC post dimensions such as numerical data in addition to textual data; this represents a novel contribution to the literature as all available models at the time of writing were based on text-only. The results demonstrate that the combined, multidimensional features model proposed in this project is more effective than the text-only model. The second contribution relates to creating various simple and hybrid deep learning models by applying plug & play techniques with different types of inputs (word-based or word-character-based) and different ways of representing target input words as vector representations of a particular word. According to the experimental findings, employing Bidirectional Encoder Representations from Transformers (BERT) for word embedding rather than word2vec as the former is more effective at the intervention task than the latter across all models. Interestingly, adding word-character inputs with BERT does not improve performance as it does for word2vec. Additionally, on the task of identifying topics, this is the first time in the literature that specific language terms to identify the need for urgent intervention in MOOCs were obtained. This was achieved by analysing learner MOOC posts using latent Dirichlet allocation (LDA) and offers a visualisation tool for instructors or learners that may assist them and improve instructor intervention. In addition, this thesis contributes to the literature by creating mechanisms for identifying MOOC learners who may need instructor intervention in a new context, i.e., by using their historical online forum posts as a multi-input approach for other deep learning architectures and Transformer models. The findings demonstrate that using the Transformer model is more effective at identifying MOOC learners who require instructor intervention.
Next, the thesis sought to expand its methodology to identify posts that relate to learner behaviour, which is also a novel contribution, by proposing a novel priority model to identify the urgency of intervention building based on learner histories. This model can classify learners into three groups: low risk, mid risk, and high risk. The results show that the completion rates of high-risk learners are very low, which confirms the importance of this model. Next, as MOOC data in terms of urgent posts tend to be highly unbalanced, the thesis contributes by examining various data balancing methods to spot situations in which MOOC posts urgently require instructor assistance. This included developing learner and instructor models to assist instructors to respond to urgent MOOCs posts. The results show that models with undersampling can predict the most urgent cases; 3x augmentation + undersampling usually attains the best performance. Finally, for the first time, this thesis contributes to the literature by applying text classification explainability (eXplainable Artificial Intelligence (XAI)) to an instructor intervention model, demonstrating how using a reliable predictor in combination with XAI and colour-coded visualisation could be utilised to assist instructors in deciding when posts require urgent intervention, as well as supporting annotators to create high-quality, gold-standard datasets to determine posts cases where urgent intervention is required.