Testing Burrows's Delta

Delta, a simple measure of the difference between two texts, has been proposed by John F. Burrows as a tool in authorship attribution problems, particularly in large ‘open’ problems in which conventional methods of attribution are not able to limit the claimants effectively. This paper tests Delta's effectiveness and accuracy, and shows that it works nearly as well on prose as it does on poetry. It also shows that much larger numbers of frequent words are even more accurate than the 150 that Burrows tested. Automated methods that allow for tests on large numbers of differently selected words show that removing personal pronouns and words for which a single text supplies most of the occurrences greatly increases the accuracy of Delta tests. Further tests suggest that large changes in Delta and Delta z-scores from the likeliest to the second likeliest author typically characterize correct attributions, that differences in point of view among the texts are more significant than differences in nationality, and that combining several texts for each author in the primary set reduces the effect of intra-author variability. Although Delta occasionally produces errors in attribution with characteristics that would normally lead to a great deal of confidence, the results presented here confirm its usefulness in the preliminary stages of authorship attribution problems.