The Geometry of Semantic Opposition: Why Distributional Embeddings Resist Abstract Antonym Detection

Abstract

Standard distributional word embeddings place antonym pairs such as love and hate in high-similarity positions because they appear in similar linguistic contexts. This paper investigates the geometric structure of semantic opposition in GloVe embeddings across multiple dimensionalities and corpus sources, documenting a systematic ceiling in abstract antonym detection that persists regardless of embedding configuration. Concretely grounded opposition pairs (hot/cold, fast/slow, bright/dark) achieve reliable geometric separation with cosine similarities around negative 0.70. Abstract conceptual opposition pairs (good/bad, true/false, free/trapped) do not, with cosine similarities clustering near zero. The paper characterises this ceiling precisely as a property of what distributional corpora encode: physical reality forces consistent distributional patterns across all corpora and languages, while cultural and conceptual distinctions reflect locally variable frameworks. This finding motivates the mathematical track of the research program, where logically forced structure provides a calibration reference for the language ceiling.