Small but Significant: On the Promise of Small Language Models for Accessible AIED
Project Overview
The document explores the significant role of generative AI, especially large language models (LLMs) like GPT, in education, addressing both their widespread adoption and the challenges posed by their resource-intensive nature which can hinder access for underfunded institutions. It advocates for the use of smaller language models (SLMs), such as Phi-2, which demand fewer resources while still delivering effective educational solutions. The authors highlight the potential of SLMs to improve accessibility and equity within the educational landscape, particularly in the realm of knowledge component discovery, a vital aspect of artificial intelligence in education (AIED). Overall, the findings underscore the promise of generative AI in enhancing educational outcomes while also recognizing the need for more accessible models to ensure that all institutions can benefit from these advancements.
Key Applications
Knowledge Component (KC) Discovery using Phi-2
Context: Educational settings with limited resources, targeting instructors and students
Implementation: Phi-2 was used to measure question similarity and apply clustering algorithms to identify KCs from multiple-choice questions.
Outcomes: The approach outperformed both instructional experts and GPT-4o in predicting student performance based on generated KCs.
Challenges: Dependence on the quality of the training data and potential limitations in the models' performance compared to larger models.
Implementation Barriers
Resource Constraints
Educational institutions often face budget limitations, technical infrastructure, and privacy requirements that hinder the adoption of AI technologies.
Proposed Solutions: Utilizing small language models (SLMs) like Phi-2, which require less computational power and can be deployed locally, thus reducing costs and ensuring privacy.
Project Team
Yumou Wei
Researcher
Paulo Carvalho
Researcher
John Stamper
Researcher
Contact Information
For information about the paper, please contact the authors.
Authors: Yumou Wei, Paulo Carvalho, John Stamper
Source Publication: View Original PaperLink opens in a new window
Project Contact: Dr. Jianhua Yang
LLM Model Version: gpt-4o-mini-2024-07-18
Analysis Provider: Openai