Data-Driven Semantic Axes ICA Reveals That Independent Opposition Dimensions in Word Embeddings Do Not Match Human Conceptual Categories

We apply Independent Component Analysis (ICA) to 190 WordNet antonym difference vectors in a 50-dimensional GloVe space and discover that the statistically independent semantic opposition dimensions in distributional embeddings do not align with the axes that human conceptual intuition predicts. ICA finds specificity (specific/general), causality (cause/effect), directness (explicit/implicit), and change (static/dynamic) as more geometrically independent than the traditional valence (love/hate), moral (good/bad), and truth (true/false) axes. ICA-discovered axes reduce mean axis entanglement from 0.210 to 0.123 (42% improvement in Experiment 23) and maximum pairwise similarity from 0.744 to 0.430. The geometry-first approach reveals a genuinely new working axis: directness (explicit/implicit), which achieves bidirectional top-3 opposition hits that the original design never produced. Experiment 24 confirms that the architecture-level benchmarks (B2, B3, B5, B8b) are stable at 80%, 80%, 70%, and 100%, regardless of axis design; these properties belong to the GloVe embedding geometry, not the axis labeling. The pole-swap antonym detection ceiling, by contrast, persists across both axis designs, confirming it is a property of the operation rather than the axis definition.