Multilingual AIs: Using AI to Produce Teaching and Learning Materials

• In Event: Poster Session 10
  In Poster Session: Multilingual AIs: Using AI to Produce Teaching and Learning Materials

Proposal

Relevance
In line with the CIES theme for 2025, we propose a discussion on the potential of AI to assist in multilingual contexts, where the language students encounter in their classrooms is different from the native language of their environment. We are examining how machine learning can assess a student’s repertoire and produce mixed language, differentiated materials. We are designing a platform that can assess and expand students’ content understanding by drawing from all their linguistic knowledge, allowing K-12 students to access concepts without linguistic barriers.
With the rise of big data and AI, learning analytics have become more precise. The ability to analyze data at high granularity allows accurate prediction of the bounds of student knowledge. Simultaneously, multiple natural language processing tools have been developed that allow users to communicate with machines effectively. Although some teachers worry that AI might negatively impact student learning by acting as a crutch, researchers are finding that AI paired with professional development can provide learning gains for students who are transitioning between languages of instruction.
Theory
Thus far, little effort has been made to combine AI and natural language processing in education. We are interested in how these tools can assess the entirety of a student’s linguistic repertoire and expand it. Using a translanguaging framework (Garcia, 2009), it should be possible to develop language tools that allow teachers to translate teaching-learning materials (TLM) to the local language, including low-resourced languages (Zhu et al, 2023), while differentiating TLM based on each student’s repertoire. Low-resourced languages in the context of AI and natural language processing are defined by a lack of availability of linguistic material for training natural language processing and machine learning models. However, by leveraging the emergent abilities of scaled language models, like in-context learning (Zhu et al., 2023), we can be LLMs can be fine-tuned specifically for low-resourced languages and their unique linguistic characteristics. This can help improve translation accuracy and relevance for these languages. In essence, we propose an AI that can integrate multiple languages in one material. We are particularly interested in the effects of this tool in contexts where the local language is the lanaugue of instruction and TLM is that language are underresourced.
Findings
In our presentation, we will discuss the potential for AI to help classroom teachers differentiate materials and communicate complex ideas in K-12 classrooms with linguistic barriers. We will hold an open discussion on applications of learning analytics to assess student repertoires in real time using a translanguaging framework. Attendees of our session should leave with a deeper understanding of the potential for AI to offload linguistic work for EFL teachers and lower barriers to entry for content subjects such as history and science, which are more reliant on complicated subject specific lexicon. We hope to explore the ways AI can reduce the translanguaging load for our students, rather than solely act as a language acquisition tool.

Authors

• Jessica Mejia, Graduate School of Education, University of Pennsylvania

• Gabriel Ares, Graduate School of Education, University of Pennsylvania