The Delta Spreadsheet

John F. Burrows introduced Delta, a simple measure of
authorial difference in his Busa Award lecture (2001), and
further elaborated upon it in three articles (2002a, 2002b, 2003).
In all of these discussions Burrows relies on an Excel
spreadsheet that helps to simplify and partially automate the
calculation of Delta. At the ALLC/ACH conference in
Gothenburg, David L. Hoover presented the results of further
tests of Delta on prose and discussed a more complex version
of Burrows's spreadsheet that takes the automation of the
calculation and the analysis of results much further (2004a),
and he has just published two articles that rely on such
spreadsheets (2004b, 2004c).
Given the burst of activity in authorship attribution circles
following the introduction of Delta, many researchers are
interested in using it on various projects. Unfortunately, even
Hoover's 2004 versions of the spreadsheet are rather daunting
in their complexity, and their macros are difficult to understand
because they do not include comments. Further, the researcher
must do substantial analytical work on raw word frequency
lists before they can be inserted in the spreadsheet for Delta
testing. Once the lists are produced, the frequencies must be
transformed into text percentages and a zero frequency record
must be inserted in the list for each text if any of the most
frequent words does not occur in that text. This is not a
significant problem for analyses using only a small number of
the most frequent words because nearly all of them will occur
in each text, but, as Hoover has shown (2004b, 2004c),
increasing the word list to the 700-800 most frequent often
improves the accuracy of the analysis, and many of the 800
most frequent words will normally fail to appear in one or more
of the texts. Manually adding zero records may be an acceptable
method in small analyses, but it would be an extremely
time-consuming and error-prone process if the 800 most
frequent words in a set of fifty or more texts were to be
analyzed.
Hoover's analyses also show that removing personal pronouns
and words for which a single text provides nearly all the
occurrences significantly improves Delta (and other kinds of
statistical analyses of authorship), and these are non-trivial
processes that are difficult enough to prevent some researchers
from trying out these techniques. The addition of the various
possibilities for Delta Prime introduced in Hoover's second article (2004c) makes for still greater complication, and seems
likely to prevent the interested humanist who is not an Excel
maven from further testing these innovative measures on new
corpora and from using them in real authorship attribution
problems.
My current project involves further elaboration of Hoover's
spreadsheets to automate more of the necessary processes.
Beginning with a version provided by Hoover that includes
explanatory comments on the macros by Marc LeBlanc of
Wheaton College (MA), I hope to produce a spreadsheet that
can accept as input a list of the authors and texts, the raw word
frequencies from the corpus as a whole and from the individual
primary and test texts. The complete analysis will be performed
within the spreadsheet itself. This will allow anyone who has
access to any of the myriad of software tools that can produce
ranked frequency lists to try out Delta and the various Delta
Primes without needing to have expertise in text analysis, Excel,
or Visual Basic. The project is currently under way, with the
various formulas for calculating Delta and the various Delta
Primes already added and the analytic work planned out and
in progress. Initial testing has begun to determine whether or
not the macros will operate with acceptable speed, and whether
the limitations of Excel will impact the number of frequent
words that can be analyzed. If performance proves too poor, I
intend to use other methods than Visual Basic and link them
as seamlessly as possible with the spreadsheet. By the time of
the conference, I expect to have a fully operational version to
demonstrate and distribute to anyone who is interested.
A secondary benefit of the current project is more wide ranging,
having to do with the question of how to balance using the good
tools for performing the analysis and manipulation of the word
frequency lists (certainly Visual Basic is not one of them!), and
providing a tool that is usable by the largest possible number
of users, even if those users are not particularly computer
literate. This has long been a question of serious interest to
software developers, and the relatively small scale of this project
may allow it to come to the fore in interesting ways. I hope to
benefit from the expertise of conference attendees in continuing
to develop and improve The Delta Spreadsheet.
Bibliography
Burrows, J.F. "'Delta': a measure of stylistic difference and a
guide to likely authorship." Literary and Linguistic Computing
17 (2002a): 267-287.
Burrows, J.F. "The Englishing of Juvenal: computational
stylistics and translated texts." Style 36 (2002b): 677-99.
Burrows, J.F. "Questions of Authorship: Attribution and
Beyond." Computers and the Humanities 37.1 (2003): 5-32.
Burrows, J.F. "Questions of Authorship: Attribution and
Beyond." Paper delivered at the Association for Computers and
the Humanities and Association for Literary and Linguistic
Computing, Joint International Conference. June 14, 2001.
Hoover, D.L. "Testing Burrows's Delta." Paper delivered at the
Association for Literary and Linguistic Computing and
Association for Computers and the Humanities, Joint
International Conference, Göteborg, Sweden. 2004a.
Hoover, D.L. "Testing Burrows's Delta." Literary and
Linguistic Computing 19.4 (2004b): 453-475.
Hoover, D.L. "Delta Prime?" Literary and Linguistic
Computing 19.4 (2004c): 477-495.