(414f) Chatgmp: Introducing Virtual Tutors in Chemical Engineering Education

The overwhelming attention that ChatGPT, based on the GPT-3 model series [1] and developed by OpenAI, has gathered recently, is just one of many signs that the “Artificial Intelligence Revolution” is upon us. Also in education, artificial intelligence (AI) is rapidly becoming an integral part of digital platforms. It is in fact used across various applications, such as generating video captions in software like Microsoft Teams^®, to develop smart content, grade assignments and personalize the learning of each student based on their individual profile. This is motivated by the fact that this technology has the potential to answer the needs of both educators and learners. For example, Lu et al. introduced RadarMath [2], an intelligent tutoring system to support personalized learning for math education. Another example of AI used in education can be provided by Carnegie Learning with 'MATHiaU' [3], a tool that uses cognitive science and AI methods to create personalized lesson plans and tutoring that adjust based on students’ performance, and it also provides real-time feedback.

Among other benefits, AI provides the opportunity for educators to replace repetitive tasks with stimulant new experiences. This work aims to investigate the use of AI in this context, as a powerful tool to automate tasks and generate answers from previously collected data. The Department of Chemical and Biochemical Engineering at the Technical University of Denmark (DTU) offers a course in Good Manufacturing Practice (GMP). As part of the course, students are required to take part in an audit exercise, where they take on the role of an auditor investigating a company, represented by the teachers, about their good manufacturing practice. During the audit, students can ask questions regarding available documents and practices adopted by the company, which is necessary information to evaluate whether the company in question could be a valuable business partner or if detected non-conformities are too severe. The teachers’ role is to give both accurate, in case the documents are available, or vague responses, leaving the students to reflect upon the company’s behavior. After a few years of participating in the aforementioned audit exercise, the teachers have agreed to test the replacement of the physical teacher with an AI-powered virtual assistant. The hypothesis is that students will be more engaged by the gamified nature of the virtual assistant and will not be afraid to ask for any follow-up or further clarification of the given answers. Finally, the students will have the opportunity to practice more with the virtual tutor before the actual audit exercise, hopefully strengthening their learning.

This work presents the early stages of this experiment, aiming to develop ChatGMP, a digital audit tool to represent the fictional company and therefore replace the teachers. In practice, the virtual tutor consists of a pre-trained question-answering generative model (Transformer-based [4]), which is fine-tuned on domain-specific material. This ensures that the model contains both the general knowledge, given by the original pre-trained model, as well as the specific GMP material, derived from fine-tuning. In order to enable the students to interact with the virtual tutor, we build an interactive GUI.

The aim of this initiative is twofold: (i) to improve chemical engineering students’ learning experience and increase engagement by providing stimulating material and gamification, and (ii) to allow substantial time saving for the teachers. A survey conducted shows that students seem to be interested in the option to have the possibility to perform the audit exercises with a virtual tutor. Moreover, quantitative interviews were conducted with a small number of volunteers to test the model and they reveal a positive perception of the generated answers by the students.

References

[1] Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J. D., Dhariwal, P., ... & Amodei, D. (2020). Language models are few-shot learners. Advances in neural information processing systems, 33, 1877-1901.

[2] Lu, Y., Pian, Y., Chen, P., Meng, Q., & Cao, Y. (2021). RadarMath: An Intelligent Tutoring System for Math Education. Proceedings of the AAAI Conference on Artificial Intelligence, 35(18), 16087-16090. https://ojs.aaai.org/index.php/AAAI/article/view/18020.

[3] Carnegie Learning, MATHiaU, available at: https://www.carnegielearning.com/solutions/math/mathiau/, Accessed: 21-03-2023.

[4] Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., ... & Polosukhin, I. (2017). Attention is all you need. Advances in neural information processing systems, 30.