Disputed Authorship: 30 Biographies and Six Reputed Authors A New Analysis by Full-Text Lemmatization of the 'Historia Augusta'

Introduction
This paper describes current research on authorship attribution that we are conducting by means of
a computer assisted analysis of the thirty biographies in the ‘Historia Augusta’, a set of lives of
Roman emperors and usurpers of the second and
third centuries of this era.
The programming system that we have developed
for research in textual analysis to provide these
solutions has been used specifically, in this case,
for an analysis of the original Latin, but has also
been used successfully in analysis of English text:
it is language independent. The system permits the
production of a word-by-word comparison of the
vocabularies of several texts, in which the actual
words are presented either in original form or as
lemmas, in order to permit immediate stylometric
analyses of the given works. In this paper we
demonstrate the results of some of the analyses
which we have undertaken using its assistance.
Textual Analysis
Over the past twenty years, the uses of computer
methods in the analysis of texts have grown both
in scope and in numbers. The extensive overview
of authorship attribution measures by Holmes
(1994) discusses in detail methods which have
been used in content analysis. It is clear from the
discussion, however, that it may be possible to
improve upon some of the techniques used in the
past. For example, in many cases, a small set of
function words, or a pre-selected small subset of
commonly-occurring words, has been used to describe the attributes of an entire text. Given the
power available in the standard office or home
computer, however, unless there is compelling
reason for the choice of a particular sample, it is
no longer necessary to restrict one’s efforts to
small samples of a full vocabulary. Various subsets of available vocabulary will undoubtedly be
used finally in any research project, but those
subsets can be selected after full consideration of
the entire vocabulary, and can be capable of immediate modification at any point in the research
cycle.
The Historia Augusta
The problem which is inherent in the Historia
Augusta lies in the necessity of attempting to
reconcile the conflicting claims of the manuscript
tradition, which names six authors (otherwise quite unknown) of the late third and early fourth
centuries of this era, with the results of very intense literary and historical analyses of the past 100
years, which point quite strongly to an origin in
the late fourth century, or even to the very early
fifth century, and to perhaps but one author. This
study will refer to ‘authors’, without being committed to any particular number.
The question of the authorship of this text must
surely be one of the ultimate challenges in such
studies, for the six authors occasionally mention
each other, claim for themselves lives which have
been attributed in the manuscripts to others of the
six, in a few lives make most unlikely dedications
to the emperors Diocletian and Constantine, and
claim to use sources which range from acceptably
solid Greek and Latin works, to apparently fake
histories concocted in their own imaginations.
This free-wheeling conduct on the part of the
authors has therefore led to a series of works that
range in quality from those of reasonably secure
sources of historical fact, to fantasies which have
their ‘historical’ sources in the fabrications of the
authors. (Momigliano, 1954) The quality also ranges from plodding, but reasonable accounts of
men such as Antoninus Pius or Marcus Aurelius
(author of the ‘Meditations’), to lengthy and scabrous descriptions of the activities of Heliogabalus, a reported pervert who exercised the absolute
power of imperial rule for a number of years after
218 of this era. It even contains one so-called
‘biography’ that describes some 33 different usurpers or attempted usurpers, all in insufficient detail
to permit an understanding of their careers.
The importance of the Historia Augusta is unde-
niable, since it is virtually the sole Latin text of any
substantial length of that age, and can be matched
only by sections of histories written in Greek. For
some of the period, it is the only source, and for
this reason, the question of authorship is indissolubly linked to the problem of disentangling any
solid historical information that is contained in
parts of the work, from the pure fantasy that infects
too much of it.
The Historia Augusta as a whole comprises some
111,000 words, or, without personal names, some
101,000. The individual lives vary in length from
the incomplete life of Valerian, with 1,030 words
of text, of which 400 are dictionary head-words
other than personal names, to the text of Alexander
Severus, with 11,000 words of text, including
1,900 dictionary head-words.
Design Decision
The varying quality of the works, and the risks of
making subjective judgements through statistical
analysis based upon a limited number of critical
words, brought us to act upon two decisions:
a) The complete text has been lemmatized and
disambiguated so as to be available for stylometric
analysis (by a system described at the ALLC/ACH
conference in Paris in 1994). This produces a file
of parallel lines of the original forms and of their
lemmas, with all matching words and punctuation
being vertically aligned.
b) This text is used to generate a file of word
frequencies that provides an array (or, matrix) of
approximately 5,600 disambiguated lemmas (dictionary head-words), together with their frequencies of occurrence in each of the thirty lives. At the
discretion of the researcher, comparable files of
the original forms can also be generated, and personal names can optionally be included in this
matrix. The design also allows more than 20 other
arrays to be generated; for example:
– frequencies of occurrence of specific lemmas, or classes of lemmas, per arbitrary number of input words;
– frequencies of only those lemmas that are
represented in all 30 lives;
– identification of the biographies that use each
specific lemma fewer than an arbitrary number of times (usually less than 30 occurrences
overall in the H.A.);
– 30×30 (symmetric) matrices giving the numbers of lemmas unique to, and also common
to (but not unique to), each pair of biographies;
– various frequencies of spacings between first
appearances of lemmas (i.e.: the frequency of
introduction of new lemmas);
– lengths of words and sentences; uses of the
various punctuation marks and of the letters
of the alphabet;
– frequencies of use of named function words,
and of coordinating and subordinating conjunctions (for research especially in certain
topics in psycho-linguistics);
– various word frequencies, suitable for the
differing requirements of Zipf’s Law and of
Yule’s Characteristic.
All arrays are produced also in ‘comma-delimited’
form, and can be transmitted directly to a database, spreadsheet, and to statistical packages such
as SPSS or SAS.
Basic to the design of the research was the need to
be able to take groups of segments of biographies
for statistical analysis. Hence the entire system is
under the control of a ‘batch’ file, so that once a
researcher has chosen the segments of text to be
analyzed (by use of a text editor), a single command results in the creation of the matrix or matrices of frequencies. In the case of the entire text
of the 30 biographies, this requires 93 minutes on
a 33 MHz DOS machine to produce the fundamental matrix of frequencies, or 5 hours for the complete set of files of matrices.
There was a further purpose behind this decision
to create the matrices. The question of acceptable
disambiguation is never closed, since many decisions are subjective, and are not accepted by all
scholars. For example, if an adverb has been created from an adjective that has itself been created
from a verb form, is the adverb to be lemmatized
to its form as an adverb, or to the dictionary
head-word of the original verb? Our design decision was made so that a scholar who disagrees with
the lemmatization and disambiguation, or even
with the text itself, and possibly with its punctuation, or with the lists of subordinating conjunctions or the function words, can modify the
original files in any manner desired (using a text
editor), then run the ‘batch’ file again, and develop
a new set of matrices in just over five hours.
Statistical Analysis
Many different analytic measures, such as correspondence analysis, factor analysis, cluster analysis and principal component analysis, have been
employed in this study. Principal component analysis, for example, has been used to examine the
occurrences of various frequently used lemmas
found in all 30 texts. The occurrences of words
found in only some texts has also been examined,
since this, too, is an element of style which may
be used to examine authorship (Burrows, 1992).
Vocabulary richness has been examined in many
ways. For example, the hapax/token ratio has been
examined for segments of text of approximately
124
equal lengths, and Honoré’s (1979) formula has
also been used in the examination of texts in
reference to hapax legomena and dislegomena.
The matrices provide an ease of use in these studies, in that the relative numbers of hapax legomena and dislegomena are available at a glance.
Furthermore, utility programs (or simple counting, in shorter texts), can find lemmas and forms
common to two or more texts (where forms, of
course, may be ambiguous) as well as sets of
words unique to any given text or to subsets of
texts.
The use of the original or the lemmatized text, in
close conjunction with the matrices, permits another valuable means of analysis of the richness of
vocabulary. For the text itself permits analysis of
the rate of introduction of new words, and of the
degrees of separation between new words over the
entire span of each of the texts being considered;
but since each specific lemma or form has its own
row in the matrix of frequencies, different rows
can be combined for analysis of the use of synonyms of words discovered in the earlier analysis,
and hence can provide a further measure of the
vocabulary of an author. The lemmatized text can
also, of course, be transmitted directly to other
systems of analysis, such as TACT.
Further Considerations
It should be noted that modern texts, which may
not require the intensive lemmatization and disambiguation that is necessary in highly inflected
languages such as Latin, can be transmitted directly to the programming system that generates the
various matrices of words and frequencies of occurrence.
Conclusion
The preponderance of evidence at present in our
research points to a degree of multiple authorship
(or of authors plus editor). Yet the question of
authorship of the ‘Historia Augusta’ may never be
solved to the satisfaction of all, and it probably
cannot be solved by stylometric methods alone;
but a judicious balance between stylometric and
literary and historical techniques will improve our
understanding of this interesting, controversial,
and infuriating work. It will also assist in solving
the fundamental problem of the Historia Augusta:
the question whether its analysis of the lives of the
emperors represents the frame-of-reference (and
hence the interpretation) of a single author, or is
the quite significantly less valuable result of perhaps a number of persons, writing over an unknown span of time.
Bibliography
Burrows, J.F. (1992). “Not Unless You Ask Nicely: The Interpretive Nexus Between Analysis
and Information”, Literary and Linguistic
Computing, 7: 91-109.
Chastagnol, A. (1994). (Traducteur du latin de
l’Histoire Auguste). Les Empereurs Romains
des IIe et IIIe Siècles. Éditions Robert Laffont,
S.A. Paris.
1991. Historia Augusta. Packard Humanities Institute CD-ROM 5.3.
Holmes, D. (1992). Authorship Attribution. A monograph produced for the Faculty of Computer
Studies and Mathematics, University of the
West of England, Bristol.
Holmes, D. (1994). “Authorship Attribution”,
Computers and the Humanities, 28: 87-106.
Honoré, A. (1979). “Some Simple Measures of
Richness of Vocabulary”, Association for Literary and Linguistic Computing Bulletin, 7:
172-177.
Momigliano, A. (1954). Journal of the Warburg
and Courtauld Institutes, 17: 22-46. Re-printed
in: “An unsolved problem of historical forgery: the ‘Scriptores Historiae Augustae’”, Studies in Historiography .(1969): 143-180.
Sichel, H. S. (1975). “On a Distribution Law for
Word Frequencies”, Journal of the American
Statistical Association, 70:542-547.