The KAIA research series (Papers 1-3) documented a consistent empirical finding: geometric semantic architectures built on distributional word embeddings work reliably for concrete physical semantic dimensions (temperature, speed, luminosity, scale, social exchange) and fail reliably for abstract human conceptual dimensions (moral evaluation, epistemic truth, existential state, freedom, social belonging). Prior papers attributed this failure to the distributional training objective. This paper argues that the explanation of the training objective is incomplete and that the deeper cause lies in a fundamental distinction between two classes of semantic opposition: physically forced universals and culturally contingent constructs. Physical opposites are forced by reality. Hot and cold cannot co-occur in the same object at the same scale simultaneously. This is true in every human language, every culture, every corpus of text. The distributional geometry of these concepts is clean because the underlying physics is clean. Abstract opposites are constructed by human conceptual frameworks. Good and bad, true and false, free and trapped are not properties of physical reality but categories that humans impose on experience. These categories vary across languages, cultures, and philosophical traditions. The distributional geometry of these concepts is entangled because the underlying conceptual structure is entangled in the language that encodes it. This distinction has direct implications for what geometric semantic architecture can and cannot be expected to do, and for what a system capable of reasoning about human conceptual categories would actually require.
Tiffney Bare (Fri,) studied this question.