An Exercise in Non-Ideal Authorship Attribution

Much has been written about the appropriate conditions
for non-traditional authorship (e.g., Rudman
1997). Substantial known samples by the possible authors
of a disputed text should be available, they should
be controlled for genre, point of view, etc., and additional
similar texts should be available from other contemporary
authors to show that false positive identifications
are unlikely. Less attention has been paid to authorship
problems compelling enough to demand an attempt in
spite of less-than-ideal conditions. One such case is Maria
Ward’s Female Life Among the Mormons (1855),
one of four widely read books published in 1855-56 that
were largely responsible for the litany of anti-Mormon
literature of the late 19th and early 20th centuries, including
works by Arthur Conan Doyle, Zane Grey, Robert
Louis Stevenson, and Mark Twain. (The first edition
of Female Life sold at least 40,000 copies; within three
years it had been translated into Danish, French, German,
Hungarian, and Swedish (Worldcat).)
Female Life, published as an anonymous autobiographical
account, was early panned and labeled fiction (New
York Times 1855), and there is strong evidence that ‘Maria
Ward,’ who also claims to have ‘edited’ Austin Ward’s
The Husband in Utah (1857) is a pseudonym (the book
was later published as Male Life Among the Mormons; I
refer to it as Male Life below). Confusingly both Ward
books are sometimes categorized as fiction, and Mrs B.
G. Ferris is sometimes listed as a (second) author of Female
Life (Worldcat). Moreover, both Mr and Mrs Ferris
wrote non-fictional anti-Mormon books–his a history of
Utah and Mormonism, Utah and the Mormons (1854),
hers a memoir, The Mormons at Home (1856). Finally,
efforts to confirm Ferris’s authorship of Female Life led
Arrington and Haupt (1968: 253) to conclude that she
did not write it, but that her book may have provided the
basis for Male Life.
A Computational Analysis
The case of Female Life certainly qualifies as a non-ideal
authorship attribution problem. Each of the four authors
of interest apparently wrote only one book. All four
books are framed as nonfiction,
but some extraordinary
errors and the fact that ‘Maria Ward’ and ‘Austin Ward’
are likely pseudonyms suggests that their books are fiction, makes their gender doubtful, and suggests that they
may be two pseudonyms for one person. Additionally,
only the middle third of Mrs Ferris’s book deals with the
Mormons, and the attribution of Female Life to her is
tenuous and unexplained. The most important facts are
these: Given this less-than-ideal scenario, can we shed any light
on the authorship of Female Life? I begin with a group
of large, Mormon-related, first-person fiction and nonfiction
texts. With many candidate authors, any similarities
between Mrs Ferris and Maria Ward will be more significant.
In Fig. 1, only Kane and Stenhouse are represented
by pairs of texts, and Stenhouse’s texts fail to cluster. Yet
‘StenhouseMa’ is an excerpt from Tell it All that supposedly
relates Mary Burton’s story in her own words, so
that this is not a definite error. Adding Mrs Ferris and
Maria Ward (Fig. 2), does not support Ferris’s authorship
of Female Life: the two authors’ texts are widely separated,
but the close clustering of the Wards’ texts suggests
a single author. (In the captions, NP 70% indicates that
no personal pronouns are included and words are eliminated
if a single text accounts for more than 70% of their
occurrences. Pronouns are closely tied to the number and
gender of characters; words that are frequent because of
a single text are typically character names.) Large 1 and 3 Person Mormon and Non-
Mormon Texts
A larger set of texts with more pairs by single authors
allows me to train the method on similar texts, select
the most effective analysis, then insert Ward and Ferris
and re-test–compensating for the lack of primary author
training texts. Consider then 32 large Mormon and
non-Mormon texts (Fig. 3). Eleven of thirteen pairs by
known authors cluster together, capturing known authorship
very effectively, though most contain Mormon
non-Mormon texts. The two failures may not be errors:
Stenhouse’s texts are discussed above, and Bell claims
only to have ‘prepared’ Stephens’s Rebel Cousins. Adding
Maria Ward’s and the Ferrises’s texts (Mrs Ferris’s in
three sections, isolating the Mormon part), produces Fig.
4. The known texts cluster as before, and the Wards form
a tight cluster separate from the Ferrises, whose own
close clustering may suggest collaboration. (A smaller
set of Mormon-related texts with the dialogue deleted to
eliminate any effects of different proportions of dialogue
and narration, produces similar results.) Delta Analysis
Delta analysis is problematic here because of the complexity
of the case, but several analyses with different
configurations of primary and secondary authors never
strongly identify Mrs Ferris as the author of Female Life,
though a few identify her as the author of Male Life when
Maria Ward is not among the primary authors. When either
of the Wards is included among the primary authors,
however, he or she is regularly and strongly identified
as the author of the other Ward’s text. (For discussion of
Delta, see Burrows 2002 and Hoover 2004.)
T-Tests
It seems worthwhile to try another approach using ttests.
Each test assumes that two of our three main authors
(Maria and Austin Ward and Mrs Ferris) are different,
uses t-tests to create a set of words that strongly
differentiate them, and then uses these words to test the
third author with PCA (principal components analysis).
(The method is inspired by Burrows 1992.) This makes
sense, however, only if we know what to expect when
each set of assumptions is false, so I begin with several
sets of three known texts by one, two, or three authors.
In two of three tests involving three texts by one author,
the third text intermingles with one of the two texts distinguished
by the t-tested words. In the third test, however,
the three texts remain quite distinct. Thus, if Mrs
Ferris wrote all three texts and they cluster separately,
no conclusions can be drawn. With three texts by two
authors, t-tests that distinguish two texts by one author
and those that distinguish two texts by two authors produce
very different results. In the former case, all three
texts typically cluster separately (Fig. 5); in the latter, the
two texts by the same author are separate from that of the
third on component one, but not from each other. In two
of the three cases tested, they also mingle on component
two (Fig. 6). T-tests involving two texts by one author
also typically produce far fewer discriminating words
(275) than those involving two authors (677). Finally,
in all tests involving three authors, the three texts remain
quite distinct, whichever authors are assumed to be different.
Once the three scenarios have been explored, we can
compare these results with tests on our authors of interest.
Assuming the two Wards are different and testing
Mrs Ferris as an unknown (narrative only, approximately
27,000 words for each of the Wards and 18,000 for Ferris)
yields Fig. 7. (Tests on the three whole texts produce
similar results.) T-tests involving one author, three authors,
or two authors (when the wrong assumption of difference
is made) can produce similar graphs, though the
small number of discriminating words supports the last
possibility. But Fig. 8. and Fig. 9, with one of the Wards
tested against Ferris and the other Ward as unknown,
produce patterns like that in Fig. 6, patterns consistent
with the results from Delta and the various cluster analyses
above: Mrs Ferris is unlikely to have written either
Female Life or Male Life, and Maria and Austin Ward
seem very likely to be two pseudonyms for the same person
(whose identity remains a mystery). Conclusion
Considering the level of difficulty of this authorship
problem, the results seems quite persuasive. They suggest
that combining several different approaches across
many analyses can help to compensate for the lack of
training texts. Finally, comparing results based on false
and true assumptions in simulations with known authors
to the results of similar tests involving the texts in question
can provide worthwhile results even under conditions
that are far from ideal.
References
Arrington, Leonard J., and Jon Haupt. (1968). Intolerable
Zion: The Image of Mormonism in Nineteenth Century
American Literature, Western Humanities Review,
22: 243-260.
Burrows, John F. (1992). Not Unless You Ask Nicely:
The Interpretative Nexus Between Analysis and Information,
LLC 7: 91-109.
Burrows, John F. (2002a). ‘Delta’: A Measure of Stylistic
Difference and a Guide to Likely Authorship, LLC
17: 267-287.
Ferris, Benjamin G. (1854). Utah and the Mormons.
New York: Harper & Brothers.
Ferris, Mrs Benjamin G. (1856). The Mormons at Home.
New York, Dix & Edwards.
Hoover, David L. (2004). Testing Burrows’s Delta, LLC
19: 453-475.
Review of Female Life Among the Mormons. (1855).
The New York Times. July 14: 3.
Rudman, Joseph. (1997). The State of Authorship Attribution
Studies: Some Problems and Solutions, Computers
and the Humanities 31: 351-65.
Ward, Austin N. (1857). The Husband in Utah. Maria
Ward (ed.). New York: Derby & Jackson; republished
as Male Life Among the Mormons, Philadelphia: John E.
Potter and Company, 1863.
Ward, Maria. (1855). Female Life Among the Mormons.
New York: J. C. Derby.
Worldcat. (2001-2008). Dublin, Ohio: OCLC Online
Computer Library Center, Inc. Online.