To: Synoptic (WSW)
In Agreement with: Chuck Jones
In Response To: David Mealand
On: Stylometric Evidence for or against Lk 1-2
Let me add my note of thanks for David's stylometric report (on Lk 1-2) to
the thanks of Chuck, just posted.
David comments, "It is by no means a clear cut issue." I think that is
exactly right, and may I take a moment to say why?
From 40 years' experience with stylometric measures in various languages,
including awareness of other people's results, I find the following.
1. Stylometrics measures style. It does not measure authorship as such. It
is not designed to, and in principle it cannot be designed to. Authorship
lies one stage behind the directly measurable; for the investigator, it can
be no more than inferential.
2. The styles of two authors may differ, but may also be similar, either for
random reasons, or because of intentional imitation, or because of
unconscious imitation. An example of the latter is the one Madison letter
(in the Federalist Papers; I have in mind the Mosteller-Wallace classic
study) which ran way outside the usual ambitus of Madison results for the
particular algorithm Mosteller-Wallace used (not a bad one, by the way,
though it has proved capable of further refinement). This was most plausibly
because in that letter Madison had in mind a tract by an earlier British
political science writer, and was influenced by its style, almost to the
point of imitation. A second example: Analects 18:4-6 (a Confucian text), a
response to several passages in the Jwangdz (a hostile Dauist text) departs
from what we might call the stylistic norm for the Analects, and resembles,
not just the Jwangdz in general, but three passages which, upon examination,
prove to be plausible as the exact passages to which the Analects writer
(not Confucius, it is hardly necessary to say) was responding. This is at
the opposite end of the emotional spectrum from the Madison case - conscious
opposition rather than perhaps unconscious agreement - but the result in
stylometric terms is the same: the writer has been influenced by the style
of an outside text. Since most writing is done in a context of other
writing, some of it exemplary for that author and some of it impinging from
a hostile direction, these influences are in principle always present.
3. The style of one author may vary from its own known average, for what
amount to psychological reasons. It is possible, in scanning the stories of
James Thurber, to detect points where a stylistic disturbance occurs. These
points turn out to correspond very closely to the deep psychiatric issues
identified in the Bernstein biography of Thurber (which was written after
those particular stylometric results were obtained). Another example is the
anonymous letters written by Jonathan Swift for the Examiner, a
pro-government paper of the time. One of the twenty deviated violently from
the stylistic average of the other nineteen. It turned out to describe an
opium dream. It is a not unreasonable conclusion that Swift was reporting an
actual experience, and in the language of that experience; he may have been
under the effects when writing the piece, or he may have been recalling
those effects. In any case, the piece itself gives a not implausible
explanation for its own atypical style, and the conclusion that this letter
is not by Swift should be rejected. Again, it is possible, by scanning
certain long stories in the Dzwo Jwan, to detect, not so much the presence
of an interpolation, as the point at which the author's own interest in his
story peaks - the thematic climax of the story. Locating such points (and
merely literary analysis can differ about where they are, or whether they
exist) can be very helpful in detecting the author's purpose in writing the
story. They do not indicate the sudden presence of another author. These are
examples of stylometrics used in what I might call its home country - the
analysis of style.
4. Stylometric results are sensitive to the particular algorithm employed.
Different indicators give different results for the same two passages. In
calibrating an algorithm, to decide what measures to incorporate, one works
from known material, of the same author in the same mood, etc, The
prerequisite is that the material be in fact known, and though it is not as
much of a problem with recent authors, with ancient ones there is a constant
risk of authorially inauthentic material. The results of Walker on
interpolations in the Pauline letters may stand as a sufficient
methodological caution in this area. Are we (is our algorithm) measuring the
difference between Nice Paul and Angry Paul? Or is it registering the
difference between Real Paul and Deutero Paul? Are we finding a deviation in
the style of Luke, or are we detecting an early scribal interpolation into
Luke (one of the Western non-Interpolations)? These are unignorable
questions, and they much complicate all such investigations. As for the
Deutero Author of the Pastorals, who is trying as hard as possible to make
his work resemble the genuine letters of Paul, or the author of Ephesians,
who is doing the same (with more literary success) with Colossians in
particular, we can perhaps measure how far they succeed, but those familiar
with the literature of these problems will know that different measures
(vocabulary, theology, tone) give different results as to authorship.
5. Stylistic results are sensitive to genre. There is probably no such thing
as a neutral sample of authorial style; authors (and the rest of us) are
always *doing something.* Beethoven characteristically used E-flat for grand
statements; Mozart has a special momentous Innigkeit that is often
associated with his D minor movements. Not to notice this is to miss part of
the expressiveness of those pieces. D major, for fingering reasons, is a
convenient key for a violin concerto, by whomever. Writing a transition is
like writing an apotheosis, except that the rules are different and the
precedents are different, and the effects on anything measurable as "style"
will be different. This is one of the problems with regarding "seams"
(transitions and introductions and characteristic summaries) as authorially
neutral. They are not. Nothing is. Nothing can be. The only authorially
neutral sample (for vocabulary-based tests, and mutatis mutandis) is the
cumulative frequency curve of that language. We can compute an average for a
number of tested passages, but it is our average, not the author's, and it
is our selection of passages to test, not the author's. And if it IS the
author's, his selection bias has its own obvious problems. A recent
biography of Robert Frost has been faulted for including and discussing at
length only out-of-copyright poems. And so on, for a thousand variations.
One must calibrate, and not to do so is a lapse of technique, but there are
difficulties with the calibration process.
6. More generally, but by definition, stylometrics considers only part of
the evidence presented by two passages. But the evidence omitted from the
test may be relevant, perhaps in some cases more relevant, to the final
diagnosis. A basic blood test, though obviously useful, may not pick up a
condition that is detectable in other ways (or, as in the famous case of the
mammogram, may result in what the clinicians call a false positive). This is
something of which the basic GP physician needs occasionally to be reminded.
A similar caution applies to the text worker.
Careful work can greatly diminish the effect of some of these considerations
on a given stylometric test, and the results of that test become
proportionately more interesting and suggestive, whether as the
psychological profile of one author (the home ground of these tests), or as
an indication of the presence of another author (an extended use, and by
definition not as comfortable for the procedure itself). Even a less careful
result may have its interest. But as I remarked more or less, at the end of
a partly stylometric-based lecture in Leiden (2003; online), "in this area,
everything works a little bit, but nothing works perfectly. All stylometric
results must in the end be referred to the literary judgement of the
researcher, and of the larger scholarly community."
I think that still holds. Stylometrics is relevant, it is interesting, it
can sometimes be *very* interesting, but it is not DNA.
With repeated thanks, but also with repeated cautions,
E Bruce Brooks
Warring States Project
University of Massachusetts at Amherst