Augmenting dissertation mentorship through multi-modal generative AI: adapting language and visuals to diverse learning styles

Abstract

Purpose Generative artificial intelligence (GAI) triggered an unprecedented disruption, with studies exploring opportunities and concerns about language capabilities of language models (Generative Pre-trained Transformer 3/4, Large Language Model Meta AI, Gemini etc.) in education. This study aims to present a multi-modal GAI dissertation mentor prototype, followed by an experiment within a classroom. The AI mentor was developed based on Gardner's multiple intelligences theory, where the AI agent can adapt to diverse learning styles like visual, kinesthetic and reading and/or writing. Design/methodology/approach The system comprises a multi-modal retrieval augmented generation (RAG) backend and web-based frontend interface, allowing educators to create a knowledgebase of academic resources with text, tables and images. For the experiment lecture notes, e-books and presentations pertaining to a research design unit within an undergraduate degree were used. The solution can also extract and interpret relevant images within the educator's documents. The AI mentor was evaluated in a vocational college, with learners querying about filling the college's research proposal form, methodology assistance and prompts linking theory to their context. Findings Results for the AI-driven experiment confirm that increased accessibility, a reliable and/or verified data corpus and the AI mentor ability to explain with language adapted to each learning style are the three main features that distinguish the specialized prototype from generic chatbots like ChatGPT or Gemini. 51% of learners sitting for the pilot session preferred the kinesthetic mode, a finding in line with the college's vocational structure. Originality/value While human interaction with the lecturer and/or mentor remains critical, the proposed AI solution augments mentorship provision beyond classroom hours and to a level of detail, which is difficult to achieve in traditional classroom or mentor meeting settings. While prior multi-modal RAG systems such as MuRAG and SAM-RAG exist, our work is distinct in applying a unified text-table-image pipeline to an educational tutor and combining this with learning style-based conditioning.

Ähnliche Arbeiten