Organizing Knowledge : Speaking the Same Language
Helenius, Teemu (2026)
2026
Tietojohtamisen DI-ohjelma - Master's Programme in Information and Knowledge Management
Johtamisen ja talouden tiedekunta - Faculty of Management and Business
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2026-07-31
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-202607318663
https://urn.fi/URN:NBN:fi:tuni-202607318663
Tiivistelmä
Organizations coordinate action through shared vocabulary, but the meaning of that vocabulary is constituted through ongoing practice rather than fixed in written definitions. Large language models, deployed increasingly for organizational knowledge management, are trained on a fixed corpus and represent a synchronic snapshot of language at a point in time. This thesis investigates whether such models can participate in the maintenance of organizational meaning, and what limits their reliable use in practice.
A theoretical framework is developed, synthesizing three dimensions along which organizational meaning is constituted: a temporal dimension, describing whether a term's meaning is treated as fixed or continuously revised; a structural dimension, describing how abstract a term's representation is and how many senses it folds together; and a distributional dimension, describing how widely a term's governing code is shared beyond the group that produced it. Drawing on Peircean pragmatism, Saussure's synchronic/diachronic distinction, organizational semiotics, and computational linguistic evidence of semantic drift, the thesis argues that the gap between organizational and machine-represented meaning is a structural property of synchronic systems operating in diachronic environments, not a limitation correctable by better models or more careful prompting.
An empirical study, following a semiotic methodology adapted from Mingers and Willcocks' four-phase framework, examines an AI classification product tested against an enterprise knowledge base. Investigating the case revealed that no metadata category in the knowledge base carried an explicit definition, and that organizational vocabulary divides into a globally standardized layer, decodable by an outside agent, and a locally negotiated layer, decodable only through participation in organizational practice. Only the first layer could be brought into classification testing at all. Testing this layer against an LLM classifier showed that agreement with an independently constructed human baseline tracked how structurally open a category's definition was, and that adding explicit structural constraints reduced this variation directly, independent of whether a category's broader organizational meaning was otherwise resolved.
These findings are read as a direct empirical encounter with the theoretical boundary described above, rather than as a limitation of the research design: the categories that would have tested the gap most directly could not be evaluated from outside the organization at all. The thesis proposes that organizational knowledge management should pair categories with explicit statements of purpose and distributional status, and route locally negotiated categories away from automated classification by design.
A theoretical framework is developed, synthesizing three dimensions along which organizational meaning is constituted: a temporal dimension, describing whether a term's meaning is treated as fixed or continuously revised; a structural dimension, describing how abstract a term's representation is and how many senses it folds together; and a distributional dimension, describing how widely a term's governing code is shared beyond the group that produced it. Drawing on Peircean pragmatism, Saussure's synchronic/diachronic distinction, organizational semiotics, and computational linguistic evidence of semantic drift, the thesis argues that the gap between organizational and machine-represented meaning is a structural property of synchronic systems operating in diachronic environments, not a limitation correctable by better models or more careful prompting.
An empirical study, following a semiotic methodology adapted from Mingers and Willcocks' four-phase framework, examines an AI classification product tested against an enterprise knowledge base. Investigating the case revealed that no metadata category in the knowledge base carried an explicit definition, and that organizational vocabulary divides into a globally standardized layer, decodable by an outside agent, and a locally negotiated layer, decodable only through participation in organizational practice. Only the first layer could be brought into classification testing at all. Testing this layer against an LLM classifier showed that agreement with an independently constructed human baseline tracked how structurally open a category's definition was, and that adding explicit structural constraints reduced this variation directly, independent of whether a category's broader organizational meaning was otherwise resolved.
These findings are read as a direct empirical encounter with the theoretical boundary described above, rather than as a limitation of the research design: the categories that would have tested the gap most directly could not be evaluated from outside the organization at all. The thesis proposes that organizational knowledge management should pair categories with explicit statements of purpose and distributional status, and route locally negotiated categories away from automated classification by design.