DISSERTATION DFENSE
 

Author : Amarakoon Mudiyanselage Thilini Ishanka Wijesiriwardene
Advisor: Dr. Amit Sheth
Date: Oct 16th, 2025
Time: 10:00 am
Place: AI Institute Seminar room
Join Zoom Meeting: https://sc-edu.zoom.us/j/86209851277?pwd=frrKtfwOKKs4EHZWkly1pfahO0z7n6…
Meeting ID: 862 0985 1277
Passcode: 309819

Abstract

Analogy-making is central to human cognition, requiring the integration of abstract reasoning, pattern recognition, and background knowledge. Despite significant advances in language modeling, the capacity of current methods to accurately identify, model, and evaluate analogies remains fundamentally underexplored.

Analogies are central to human cognition, enabling individuals to perceive deep similarities between superficially different situations. Effective analogy-making requires integrating knowledge about the external world with abstract reasoning and pattern recognition capabilities. While current language models (LMs), trained on massive textual corpora using autoregressive or masked objectives, achieve impressive performance across Natural Language Processing (NLP) tasks such as text generation, summarization, and classification, their capacity for analogical reasoning remains poorly understood. Three factors contribute to this gap: the inherent complexity of analogy-making, the scarcity of suitable evaluation data, and the absence of systematic frameworks for quantifying analogy complexity. This dissertation bridges this gap by advancing both the theoretical understanding of analogies in LMs and the practical tools needed to benchmark and improve their analogical capabilities.

This work makes six interconnected contributions to computational analogy research. First, we introduce a complexity-grounded taxonomy of analogies and develop evaluation methods that assess Large Language Models (LLMs) across this spectrum, revealing that knowledge-enhanced approaches are essential for proportional and long-text analogies. Second, we analyze student-generated analogies in Biochemistry, demonstrating how both hand-engineered features and LLM-generated embeddings contribute to distinguishing strong from weak analogies in educational contexts. Third, through linguistic probing techniques, we investigate the relationship between LLMs' syntactic-semantic encoding capabilities and their performance on sentence-level analogies. Fourth, we propose knowledge-enhanced methods specifically designed to address the challenging proportional analogies identified in our taxonomy. Fifth, we develop a generation pipeline for realistic long-text analogies addressing the limitations of existing overly-clean datasets, and benchmark state-of-the-art LLMs while exploring Graph Neural Network-based complementary evaluation methods. Sixth, recognizing that analogy research requires distinguishing between related phenomena of abstraction, we present a systematic taxonomy of abstraction levels, addressing the lack of consistent operational definitions in the Computer Science literature.

Together, these contributions establish a comprehensive framework for understanding, evaluating, and improving analogical reasoning in the era of large language models, with implications for both cognitive modeling and practical NLP applications.