Leveraging large language models for analogy generation and extraction

Sehgal, Shradha

Permalink

https://hdl.handle.net/2142/124321 Copy

Description

Title

Leveraging large language models for analogy generation and extraction

Author(s)

Sehgal, Shradha

Issue Date

2024-04-25

Director of Research (if dissertation) or Advisor (if thesis)

Zhai, ChengXiang

Department of Study

Computer Science

Discipline

Computer Science

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

M.S.

Degree Level

Thesis

Keyword(s)

• Large Language Models
• Analogies
• Education
• Information Extraction
• Natural Language Processing

Abstract

Analogies draw parallels between distinct entities based on their shared characteristics. They make abstract and complex ideas more accessible and relatable and are therefore used for creative writing, communication, and scientific innovation. Particularly valuable in the educational domain, analogies can be used to explain complex concepts to students creatively and uniquely. By linking educational concepts to other familiar topics, students often have better retention, understanding, and reasoning of complex topics. However, generating analogies is a complicated task due to the nuanced understanding required to draw meaningful parallels between concepts. Thus, analogies are most often created by domain experts with a deep understanding of topics. With the recent advent of large language models and their textual representative power, it becomes interesting to study their applications to analogies, both for automated generation and identification. In this thesis, we explore the use of large language models for automatically generating and extracting educational analogies. We study various prompting techniques to create text-based scientific analogies pertaining to high school concepts. We note the shortcomings of existing data sources and propose a new large-scale dataset consisting of over 3k target concepts and their analogies. Beyond generation, we study how we can automatically extract analogies from a large corpus of text documents. This shows promise to mine a vast database of analogies from the web, that can be used for informing and educating students. Finally, we propose a pipeline to generate multimodal analogies by leveraging structural scientific concepts, thereby offering a text and visual representation of analogies. Our research contributes to the educational web platform, Analego (https://timan.cs.illinois.edu/analego), where students can explore our vast collection of analogies to enhance their learning.

Graduation Semester

2024-05

Type of Resource

Thesis

Copyright and License Information

Copyright 2024 Shradha Sehgal

Owning Collections