You get this when you complete the test: > The score is computed by taking the average semantic distance between each of the words. These distances are compute

You get this when you complete the test: > The score is computed by taking the average semantic distance between each of the words. These distances are computed by measuring how often the words are used together in similar contexts. It sounds like they are looking for bigrams (or n-grams), playing Markov chain golf. Sort of a neat idea for this application. It would also explain sensitivity to ordering the words. > ...so if you rearrange words properly you could get much better results As a personal score, but that would be a source of error in measurement, I suppose.