Figure 3. Bipartite map of grape varieties (circles, left; coloured by derived VarietyFamily, sized by review count) and the tasting descriptors (diamonds, right; coloured by sensory category, sized by total association) that characterise them. An edge links a variety to a descriptor when that term's mean per-wine TF-IDF weight across the variety's reviews is ≥ 0.02; only the top 12 descriptors per variety are drawn. Varieties are included at ≥ 300 reviews (generic blends excluded). n = 40 varieties, 36 descriptors, 461 edges. Each descriptor's vertical position is the TF-IDF-weighted mean of its varieties' positions, so the vocabulary self-sorts: dark-fruit/cedar/tannin terms gravitate to the Cabernet–Syrah cluster while red-cherry/earth terms gravitate to Pinot — rejecting the falsifier of a uniform shared vocabulary. Descriptors derived by controlled-lexicon extraction (SPEC §4.6); negation is not modelled. Source: Wine Reviews (zynicide, Kaggle), Wine Enthusiast; snapshot 2019-09-20.