Exploring the Efficacy of Twi-Language Interfaced ChatGPT as a Self-Learning Support Tool for Junior High School Students in Rural Communities.

Abstract

Research shows that differences between the language of instruction and the primary language of students affect learning negatively. Large language models have been found to be useful in teaching and learning. Unfortunately, most large language models are not built on low resource languages like the prevalent local language in Ghana, Twi. To allow those that are most comfortable interacting in low resource languages benefit from large language models in their educational journey, this research seeks to answer the question, "Does a Twi language interface on ChatGPT make it a good learning support tool for junior high school students in rural communities in Ghana?” The main objectives of this research were to develop a Twi-interfaced ChatGPT that facilitates interaction in Twi and assess the students’ acceptance of various interaction modalities. Students were divided into three groups that interacted with two systems each of varied modalities and system configurations. Post-interaction, a questionnaire was administered to each group to deduce interaction and language preferences as well as evaluate the developed system from a user standpoint. It was found that students preferred to type as opposed to interacting in speech. They found it more efficient and enjoyable. Students also preferred to study in Twi with ChatGPT via the human translator as opposed to the Voice mode feature built into the system. In comparing their interaction in Twi with ChatGPT via the human translator to the developed system, students found the human system more efficient. This proves that the concept could work in the near future as speech recognition systems and text-to-speech system approach human level accuracy.