Mining Classical Greek Gender

This paper examines gendered language in classical Greek
drama. Recent years have seen the emergence of data mining in
the Humanities, and questions of gender have been asked from
the start. Work by Argamon and others has studied gender in
English, both for the gender of authors of texts (Koppel et al.
2002, Argamon et al. 2003) and for that of characters in texts
(Hota et al. 2006, 2007 on Shakespeare); Argamon et al. (in
prep. a) study gender as a variable in Alexander Street Press’s
Black Drama corpus (gender of authors and characters) and
(in prep. b) in French literature (gender of authors).
Surprisingly, some of the main fi ndings of these studies show
signifi cant overlap: Female authors and characters use more
personal pronouns and negations than males; male authors
and characters use more determiners and quantifi ers. Not
only, then, do dramatic characters in Shakespeare and modern
drama, like male and female authors, prove susceptible to
automated analysis, but the feature sets that separate male
from female characters show remarkable continuity with
those that separate male from female authors.
For a scholar of Greek, this overlap between actual people
and dramatic characters holds great promise. Since there
are barely any extant Greek texts written by women, these
results for English give us some hope that the corpus of Greek
drama may serve as evidence for women’s language in classical
Greek.
After all, if the results for English-language literature showed
signifi cant differences between male and female characters,
but no parallels with the differences found between male
and female authors, we would be left with a study of gender
characterization by dramatists with no known relation to the
language of actual women. This is certainly of interest from a
literary and stylistic point of view, but from a linguistic point
of view, the English data hold out the promise that what we
learn from Greek tragedy will tell us about gendered use of
Greek language more generally, which is arguably a question
of larger import, and one we have practically no other means
of learning about.
There is some contemporary evidence to suggest that Greek
males considered the portrayal of women on stage to be true
to life. Plato’s Socrates advises in the Republic against actors
portraying women. Imitation must lead to some aspects of
these inferior beings rubbing off on the (exclusively male) actors
(Republic 3.394-395). Aristophanes, the comic playwright, has
Euripides boast (Frogs 949f.) that he made tragedy democratic,
allowing a voice to women and slaves alongside men.
Gender, of course, also continues to fascinate modern readers
of these plays. For instance, Griffi th (1999: 51) writes, on
Antigone: “Gender lies at the root of the problems of Antigone.
(...) Sophocles has created one of the most impressive female
fi gures ever to walk the stage.” Yet there are no full-scale
studies of the linguistic characteristics of female speech on
the Greek stage (pace McClure 1999).
In this paper, we report our results on data mining for gender
in Greek drama. We started with the speakers in four plays of
Sophocles (Ajax, Antigone, Electra, and Trachiniae), for a total of
thirty characters, in order to test, fi rst of all, whether a small,
non-lemmatized Greek drama corpus would yield any results
at all. We amalgamated the text of all characters by hand into
individual fi les per speaker and analyzed the resulting corpus
with PhiloMine, the data mining extension to PhiloLogic (http://
philologic.uchicago.edu/philomine).
In this initial experiment, words were not lemmatized, and
only occurrences of individual words, not bigrams or trigrams,
were used. In spite of the modest size, results have been
positive. The small corpus typically resulted in results of
“100% correct” classifi cation on different tasks, which is to
be expected as a result of overfi tting to the small amount of
data. More signifi cantly, results on cross-validation were in the
80% range, whereas results on random falsifi cation stayed near
50%. We were aware of other work on small corpora (Yu 2007
on Dickinson), but were heartened by these positive results
with PhiloMine, which had so far been used on much larger
collections.
In our presentation, we will examine two questions in more
depth, and on the basis of a larger corpus.
First, there is the overlap found between work on English and
French. Argamon et al. (2007) laid down the gauntlet:
“The strong agreement between the analyses is all the more
remarkable for the very different texts involved in these two
studies. Argamon et al. (2003) analyzed 604 documents from
the BNC spanning an array of fi ction and non-fi ction categories
from a variety of types of works, all in Modern British English
(post-1960), whereas the current study looks at longer,
predominantly fi ctional French works from the 12th - 20th
centuries. This cross-linguistic similarity could be supported
with further research in additional languages.”
So do we fi nd the same tendencies in Greek, and if so, are
we dealing with human, or ‘Western’ cultural, universals?
Our initial results were mixed. When we ran a multinomial
Bayes (MNB) analysis on a balanced sample, we did indeed see
some negations show up as markers for female characters (3
negations in a top 50 of ‘female’ features; none in the male top 50), but pronouns and determiners show up in feature sets
for both the female and male corpus. An emphatic form of the
pronoun ‘you’ turned up as the most strongly male feature in
this same analysis, and two more personal pronouns showed
up in the male top ten, as against only one in the female top
ten. Lexical items, on the other hand, were more intuitively
distributed. Words such as ‘army’, ‘man’, ‘corpse’ and ‘weapons’
show up prominently on the male list; two past tense forms of
‘die’ show up in the female top ten. A larger corpus will allow
us to report more fully on the distribution of function words
and content words, and on how selections for frequency
infl uence classifi er results.
Secondly, after expanding our corpus, regardless of whether we
fi nd similar general results for Greek as for English and French,
we will also be able to report on variation among the three
tragedians, and give more fi ne-grained analysis. For instance,
in our initial sample, we categorized gods and choruses as
male or female along with the other characters (there are
usually indications in the text as to the gender of the chorus
in a given play, say ‘sailors’, ‘male elders’, ‘women of Trachis’).
Given the formal requirements of the genre, we expect that it
will be trivial to classify characters as ‘chorus’ vs. ‘non-chorus’,
but it will be interesting to see whether gender distinctions
hold up within the group of choruses, and to what extent
divinities conform to gender roles. The goddess Athena was
the character most often mis-classifi ed in our initial sample;
perhaps this phenomenon will be more widespread in the full
corpus. Such a fi nding would suggest (if not for the fi rst time,
of course) that authority and gender intersect in important
ways, even as early as the ancient Greeks’ conceptions of their
gods.
In conclusion, we hope to demonstrate that data mining
Greek drama brings new insights, despite the small size of the
corpus and the intense scrutiny that it has already seen over
the centuries. A quantitative study of this sort has value in its
own right, but can also be a springboard for close readings
of individual passages and form the foundation for a fuller
linguistic and literary analysis.
References
Argamon, S., M. Koppel, J. Fine, A. Shimoni 2003. “Gender,
Genre, and Writing Style in Formal Written Texts”, Text 23(3).
Argamon, S., C. Cooney, R. Horton, M. Olsen, S. Stein (in prep.
a). “Gender, Race and Nationality in Black Drama.”
Argamon, S., J.-B. Goulain, R. Horton, M. Olsen (in prep. b).
“Vive la Différence! Text Mining Gender Difference in French
Literature.”
Griffi th, M. (ed.), 1999. Sophocles: Antigone.
Hota, S., S. Argamon, M. Koppel, and I. Zigdon, 2006.
“Performing Gender: Automatic Stylistic Analysis of
Shakespeare’s Characters.” Digital Humanities Abstracts
2006.
Hota, S., S. Argamon, R. Chung 2007. “Understanding the
Linguistic Construction of Gender in Shakespeare via Text
Mining.” Digital Humanities Abstract 2007.
Koppel, M., S. Argamon, A. Shimoni 2002. “Automatically
Categorizing Written Texts by Author Gender”, Literary and
Linguistic Computing 17:4 (2002): 401-12.
McClure, L., 1999. Spoken Like a Woman: Speech and Gender in
Athenian Drama.
Yu, B. 2007. An Evaluation of Text-Classifi cation Methods for
Literary Study. Diss. UIUC.