Conceptual change is difficult to observe because names often remain stable while meanings move. A familiar term can gradually acquire new applications, relations and political implications without changing its lexical form. Conversely, a concept can receive a new name while preserving much of its earlier mechanism. Measuring conceptual change through word frequency alone therefore tells us relatively little. Even sophisticated computational models can mistake changes in topic, syntax or corpus composition for changes in meaning. The more useful question is multidimensional: what exactly has moved? A canonical definition may change while public usage remains attached to the previous formulation. A concept may expand into another discipline while keeping its central distinction intact. Its network of neighbouring ideas may shift. A translation may preserve function while changing genealogy. Another concept may move statistically closer until the two are treated as interchangeable by a retrieval system even though specialists still regard them as distinct. Computational representations are valuable because they reveal distributional patterns that are difficult to see manually. Diachronic embeddings and contextualized models can compare usages across time and corpora. But a vector cannot by itself decide what a concept means. Koselleck’s conceptual history reminds us that meanings are embedded in historical oppositions, institutions and expectations. Kuhn similarly shows that conceptual change can accompany transformations in the structures through which a field perceives problems. Computational evidence therefore needs to be combined with version records, occurrences, relations, source concentration and human interpretation. A further difficulty is hardening. A formulation repeated extensively across indexed corpora may become computationally dominant even after scholars revise it. Search engines continue surfacing the older version; language models reproduce it because it has greater statistical mass. Semantic change and computational memory then move at different speeds. The most interesting evidence may lie precisely in that divergence. Instead of one universal drift score, concepts need profiles showing several dimensions of movement. Definition, context, relations, time, governance and distribution can each change independently. Such a profile does not eliminate interpretation; it organizes evidence for interpretation. Meaning is not simply a point migrating through vector space. It is a maintained relation among words, uses, institutions, histories and the communities that continue to recognize a concept as the same—or decide that it has become something else.
Giulianelli, M., Del Tredici, M. and Fernández, R. (2020) ‘Analysing Lexical Semantic Change with Contextualised Word Representations’, ACL.
Hamilton, W. L., Leskovec, J. and Jurafsky, D. (2016) ‘Diachronic Word Embeddings Reveal Statistical Laws of Semantic Change’, ACL.
Koselleck, R. (1979) Vergangene Zukunft.
Kuhn, T. S. (1962) The Structure of Scientific Revolutions.
Kutuzov, A. et al. (2018) ‘Diachronic Word Embeddings and Semantic Shifts: A Survey’, COLING.
