Figure 6

The score assigned to increasing numbers of reviews drawn from the tagged positive and negative sets. For each sentiment dictionary we show mean sentiment and 1 standard deviation over 100 samples for each distribution of reviews in Panels A-F. For comparison we compute the fraction of the distributions that overlap in Panel G. At the single review level for each sentiment dictionary this simple performance statistic (fraction of distribution overlap) ranks the OL dictionary in first place, the MPQA, LIWC, and labMT dictionaries in a second place tie, WK in fifth, and ANEW far behind. All dictionaries require on the order of 1,000 words to achieve 95% classification accuracy.