Figure 3. Undocumented, general-English headwords and near-synonyms can be acquired experimentally.

(A) The distribution over the inferred accuracies of the annotators validating harvested synonyms. (B) The true positive rate (blue) and false discovery rate (red) of the validation process as a function of the posterior probability of annotation accuracy. Diagnostic statistics were computed using known and random pairings. (C) The Receiver-Operator-Characteristic curve for the statistical model of the validation process, computed using known and random pairings. (D) The distribution over the posterior log-odds in favor of annotation accuracy for the novel synonym-headword pairings, annotated with exemplar pairings (rejected in red and accepted in blue). (E) The distributions over semantic similarity scores for the true negative (red), true positive (green), and novel synonym pairs (blue). (F) Bootstrapped (10,000 re-samples) distributions over the average semantic similarity scores for each group of pairings, computed using the data depicted in (E).