Logo image
Back
Semantic shift as scholarly record: a framework for fuzziness-aware detection and provenance-aware documentation across domains and time
Dissertation   Open access

Semantic shift as scholarly record: a framework for fuzziness-aware detection and provenance-aware documentation across domains and time

Hyung Wook Choi
Doctor of Philosophy (Ph.D.), Drexel University
Sep 2026
DOI:
https://doi.org/10.17918/00011533
pdf
Choi_Hyung-Wook_20264.56 MBDownloadView
PDF Open Access Open Access (License Unspecified)

Abstract

Language evolves continuously, and the meanings of words shift as communities of practice adapt to cultural, technological, and institutional changes. This dissertation addresses a persistent gap in existing computational approaches to semantic shift: while embedding-based methods can detect when and how word meanings diverge, they rarely preserve detected shifts as structured, reusable scholarly records. The result is that semantic change is routinely identified but seldom documented in a form that supports reproducibility, longitudinal reanalysis, or integration with knowledge organization systems. To address this gap, the dissertation proposes a two-layered framework. The first layer develops computational methods for detecting semantic drift across time and domains. Building on temporally aligned Word2Vec and contextual BERT representations, the framework quantifies drift through three complementary indicators---positional change, neighborhood overlap, and fuzziness score---and models semantic evolution as gradual and potentially overlapping rather than binary. A seed-based context anchor, grounded in historically attested dictionary definitions, ensures that measured displacement reflects genuine semantic change relative to a stable, interpretable reference state. The second layer addresses documentation. Detected shifts are first abstracted into a purpose-built Versioned Semantic Shift Ontology (VERSO), which represents semantic change as domain-level conceptual objects independent of specific pipelines or file formats. These VERSO records are then embedded within a PROV-O-compliant provenance layer that links semantic states through versioning and derivation relations, preserving the full lineage of conceptual change and enabling auditing, reanalysis, and cumulative extension. An interactive visualization system, TRACE, renders VERSO--PROV-O records as navigable sense genealogies that expose branching, convergence, and provenance on demand. The framework is validated through four case studies. The analysis of network (1816--2025) traces multi-domain semantic generalization across physical, organizational, social, and computational registers. The analysis of cloud documents a focused cross-domain extension from meteorological to computing discourse, characterized by a transitional fuzziness peak followed by sense consolidation. A third case study applies the framework to a cluster of stigma-related terms in addiction discourse---junkie, addict, substance abuse, and user---demonstrating that corpus-derived drift measurements can be linked to externally dated policy events within the archival architecture, and that the same terminological cluster follows structurally distinct trajectories across registers. A fourth case study disaggregates this cluster to examine junkie as a case of partial semantic shift, in which the drug-referential sense is progressively attenuated while the bleached modifier-bearing sense expands, producing concurrent sense coexistence that neither substitution nor accumulation models adequately capture. The dissertation makes three contributions: to computational linguistics, through fuzziness-aware methods for domain-sensitive detection of gradual and overlapping semantic change; to ontology engineering, through a two-stage versioning mechanism that separates semantic abstraction from provenance tracking; and to information science, through a reproducible infrastructure that treats semantic shift as an archival object with long-term scholarly value.

Metrics

1 Record Views

Details

Logo image