Computational Analogies in the Era of Large Language Models

Attend a PhD defense presentation exploring the intersection of analogical reasoning and large language models in computational linguistics. Discover how this research addresses the fundamental gap in understanding how current language models perform analogical reasoning tasks, despite their impressive capabilities in text generation and classification. Learn about a comprehensive six-part framework that advances both theoretical understanding and practical evaluation of analogies in LLMs, including a complexity-grounded taxonomy of analogies that reveals the essential role of knowledge-enhanced approaches for proportional and long-text analogies. Examine the analysis of student-generated analogies in biochemistry education, investigating how hand-engineered features and LLM-generated embeddings distinguish strong from weak analogies in educational contexts. Explore linguistic probing techniques that reveal the relationship between LLMs' syntactic-semantic encoding capabilities and their sentence-level analogy performance. Understand the development of knowledge-enhanced methods specifically designed for challenging proportional analogies, along with a generation pipeline for realistic, long-text analogies that addresses limitations in existing datasets. Gain insights into Graph Neural Network-based evaluation methods and a systematic taxonomy of abstraction levels that distinguishes between related phenomena in computer science literature. This defense presents implications for both cognitive modeling and practical natural language processing applications, establishing a comprehensive framework for understanding, evaluating, and improving analogical reasoning capabilities in the era of large language models.