MIXI researchers Matt Curinga and Aaron Hung team up with Matt’s sister, Dr. Rebecca Curinga from CUNY to present their research on translanguaging with GPT-3 at AAAL 2025.
Paper Abstract. This study investigated the hypothesis that ChatGPT can work as an effective bilingual “tutor” for multilingual learners. Since its release, ChatGPT has captured both public and academic interest and concern with its ability to create naturalistic language. The current version, GPT-4o, is both multilingual and multimodal (i.e., it reads and generates images; works in speech-to-text and text-to-speech modes). It is trained on dozens of natural languages, and “fluent” in more than 20. We are interested in the ways GPT-4 operates across languages, not as software that supports dozens of monolingual versions or for machine translation, but as a multilingual “speaker” that employs multiple languages to communicate.
Research in translanguaging suggests that multilingual students learn more effectively when they draw on their full linguistic resources when engaging with academic content. This study uses a translanguaging and raciolinguistics approach to interrogate a series of conversations between ChatGPT and the researchers. These dialogs were generated via the real-time “voice-chat” mode available on the mobile app where the user speaks (rather than types) their prompts, and the software responds through audio rather than writing. ChatGPT creates text transcripts used for further analysis. Researchers framed conversations around a hypothetical case of a multilingual learner preparing for the New York State English Regents Exam. This assessment, specifically, has been noted as an obstacle for high school graduation and further study for multilingual students. Bilingual conversations were created using identical initial conditions in English-Spanish and English-Mandarin Chinese. We report on the specific prompts used to elicit bilingual, translanguaging behavior from ChatGPT. Analysis attends to the ways that ChatGPT’s language mirrors and diverges from expert bilingual teachers, in terms of language structure, functional uses, and cultural appropriateness across languages. Findings discuss the circumstances where ChatGPT can be an effective tutor, despite differences from human tutors’ language.
Where:
Sheraton Denver Downtown Hotel
1550 Court Place
Denver, CO 80202-5107
United States
When: Friday, March 25, 2025, 10:15 AM - 11:15 AM [add to gcal]