The study asks whether large language models can generate personalized quiz questions that help deaf and hard of hearing (DHH) students learn from video. To answer it, the researchers designed and refined a prototype over several rounds, then tested it in user studies with undergraduate DHH students. The prototype generated 2 kinds of questions. Visual Questions point to video segments where visual information could be misread, and Emotion Questions mark moments where earlier DHH learners showed, through emotional responses, that they were having trouble with the video.
- LLM-generated questions showed significant potential for personalized learning among DHH students using video-based material.
- Students' interaction with the AI system still ran into difficulties.
- The authors trace those difficulties to an accessibility gap between text-based AI prompts and the sign languages that DHH students use as a first or native language.
The authors conclude that LLM-based educational technology has to be designed with language diversity and culture in mind. A tool built around written prompts can sit poorly with learners whose primary language is signed, and they argue that this mismatch needs attention if such tools are going to serve DHH students well.
