Summary
LLMs such as GPT-4 can be used to automate the creation of lecture slides from course material.
Abstract
In this paper, we explore the automation of script-to-slide generation using three Generative Pretrained Transformer (GPT) models. We created a training set of script-slide pairs from a single presentation. Next, we created three GPT models using this training data. We used custom fine-tuning for GPT-3 and few-shot learning for GPT-3.5 and GPT-4. These three models used the on-sample training scripts to generate test slides. To assess the relative performance of each model, we compared the generated slides against the training slides and scripts. GPT-4 outperformed the other models in terms of accuracy, similarity, and relevance. However, it underperformed the other models on runtime performance and cost efficiency. This research demonstrates the potential of using GPT models to automate lecture slide generation, highlighting areas for further investigation and improvement.
Resources
My final project for Creating AI-Enabled Systems at Johns Hopkins University