Negated Adjectives and Antonyms in Distributional Semantics: not similar?

We investigate the relation between negated adjectives and antonyms pairs in English (e.g., not cold vs. hot - cold) using Distributional Semantics. We build vector representations of a set of antonyms and their negations on the basis of their contexts of use, and compare the similarities of the negated adjectives to each of the adjective in their antonym pair. We find that in a distributional semantic model a negated adjective (e.g., not cold) is typically more similar to the adjective itself (cold) than to its antonym (hot). The effect is less strong for antonyms that share their lexical root (morphological; e.g., happy - unhappy). No difference is observed between simple and double negations (e.g., not happy, not unhappy), and contrary and contradictory antonyms (e.g., hot - cold, dead - alive). Our results provides insights on negated adjectives, and in general the type of similarity captured by Distributional Semantics.