Comparison of features for random corpus. For all pairs of features, the differences between F-measures for PHI and the differences between F-measures for non-PHI are significant at α = 0.05. The only exceptions are the difference of F-measures in non-PHI of lexical bigrams and POS information (marked by †), the difference in F-measures in PHI of MeSH and orthographic features (marked by ‡), and the difference in F-measures in non-PHI of MeSH and orthographic features (marked by •). Best F-measures are in bold.