Prompting large language models to power educational chatbots

The recent rise in both popularity and performance of large language models has garnered considerable interest regarding their applicability to education. Technologies like ChatGPT, which can engage in human-like dialog, have already disrupted educational practices given their ability to answer a wide array of questions. Nevertheless, integrating these technologies into learning contexts faces both technological and pedagogical challenges, such as providing appropriate user interfaces and configuring interactions to ensure that conversations stay on topic. To better understand the potential large language models have to power educational chatbots, we propose an architecture to support educational chatbots that can be powered by these models. Using this architecture, we created a chatbot interface that was integrated into a web application aimed at teaching software engineering best practices. The application was then used to conduct a case study comprising a controlled experiment with 26 university software engineering students. Half of the students interacted with a version of the application equipped with the chatbot, while the other half completed the same lesson without the chatbot. While the results of our quantitative analysis did not identify significant differences between conditions, qualitative insights suggest that learners appreciated the chatbot. These results could serve as a starting point to optimize strategies for integrating large language models into pedagogical scenarios