I'm looking at using Google's Ngram to pick out the word that is the closest to an opposite of a given word by looking at the frequency and correlation over time of the two words. For example, I predicted and found large and small have similar frequencies and appear positively correlated, and so do huge and tiny. Nevertheless, a careful reading of the descriptions of words from a dictionary is still likely to be the best way of finding an opposite.
What suggestions does everyone have for refining this approach? What concerns of potential pitfalls to this approach are there?