Year: 2026
Type: dissertation
Abstract: Morality is a fundamental domain of human cognition and plays a critical role in maintaining social order. Moral norms determine what counts as right and wrong, yet such judgments vary across time, cultures, and individuals. Understanding how morality evolves is fundamental to unraveling the dynamic nature of human cognition and societies. Language provides a crucial window into this process. Language enables humans to express their moral concerns, establish societal norms, and transmit values across generations. Given the intimate relationship between language and morality, this dissertation develops computational methodologies to study moral change through large-scale and systematic analyses of natural language. Advances in artificial intelligence (AI) and natural language processing (NLP) have made it possible to investigate morality automatically and at an unprecedented scale, far beyond what traditional lab-based experimental or observational methods can achieve. However, existing computational approaches to moral inference often overlook the evolving nature of morality. They tend to focus on isolated time periods or specific cultural contexts and lack clear connections to theoretical frameworks in moral psychology. This dissertation addresses these limitations by proposing a theoretically grounded computational paradigm for understanding human morality and the various ways in which it can change. The dissertation is organized into five main chapters. Chapter 2 provides a comprehensive review of both empirical and computational approaches to the study of human morality. By synthesizing psychological theories of moral change over time and across cultures, I demonstrate how current computational approaches to moral inference fall short of capturing moral dynamics. Chapter 3 introduces a cognitively grounded computational framework that leverages human word association to estimate moral perception toward concepts. I discuss the superiority of this framework against language models employed in prior studies of moral inference. Chapter 4 explores the under-examined phenomenon of moralization---the process through which concepts acquire moral significance over time. Building on the framework from Chapter 3, I extend it diachronically and demonstrate how the combination of human word association networks, large-scale historical text corpora, and graph neural networks can provide better insights into moralization than existing computational methods. Chapter 5 focuses on the evolution of the moral lexicon itself, examining how established methods from computational semantic change inference reveal metaphor's critical role in shaping moral language over historical periods. Chapter 6 expands the scope beyond English-speaking cultures to examine cross-cultural variation in moral norms. In particular, I investigate how large language models represent moral values across cultures and propose strategies to mitigate cultural biases in these systems. By studying the intersection of language, morality, and computational modeling, this dissertation offers a new paradigm for investigating moral change at scale. It provides novel insights into the origins and dynamics of moral language while contributing to the development of more reliable, ethically informed, culturally inclusive, and theoretically grounded AI systems for future studies of morality and ethics.
Source: TSpace (University of Toronto)
Author: Ramezani, Aida
Institution: University of Toronto
Language: English
Subfield: Cognitive Neuroscience
Field: Neuroscience
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7203957033
