Intertextuality in the Digital Age

Abstract

This paper describes a new digital approach to intertextual study involving
the creation of a free online tool for the automatic detection of parallel
phrases. A test comparison of Vergil’s Aeneid and Lucan’s Civil War shows that
the tool can identify a substantial number of meaningful intertexts, both previously
recorded and unrecorded. Analysis of these results demonstrates how automatic
detection can provide more comprehensive and accessible perspectives
on intertextuality as an aggregate phenomenon. Identification of the language
features necessary to detect intertexts also provides a path toward improved automatic
detection and more precise definitions of intertextuality.

Full text

Access Provided by SUNY @ Buffalo at 12/03/12 3:38PM GMT

Neil Coffee et al.
Intertextuality in the Digital Age*
neil coffee, jean-pierre koenig, shakthi poornima,
roelant ossewaarde, christopher forstall, and
sarah jacobson
University at Buffalo, SUNY
Transactions of the American Philological Association 142 (2012) 383–422
© 2012 by the American Philological Association
summary: This paper describes a new digital approach to intertextual study in-
volving the creation of a free online tool for the automatic detection of parallel
phrases. A test comparison of Vergil’s Aeneid and Lucan’s Civil War shows that
the tool can identify a substantial number of meaningful intertexts, both previ-
ously recorded and unrecorded. Analysis of these results demonstrates how au-
tomatic detection can provide more comprehensive and accessible perspectives
on intertextuality as an aggregate phenomenon. Identification of the language
features necessary to detect intertexts also provides a path toward improved au-
tomatic detection and more precise definitions of intertextuality.
when does vergil refer to the work of another author, and what
does it mean? Variations of this question have been asked about a range of
classical works from the height of Alexandrian scholarship in the 3rd century
b.c.e. to our own day. Formulations have changed, from ancient notions of
imitation, emulation, and theft, to the modern emphasis on artistic repur-
* We would like to thank the graduate student participants in the fall 2010 seminar
on intertextuality at the University at Buffalo, Kevin Roth, Sarah Skelley, and Valerie
Spiller, for their contributions to the testing process described below. We are indebted
to Neil Bernstein, Damien Nelis, Paul Roche, and the anonymous reviewers of TA PA for
suggestions that greatly improved this piece. As members of our Advisory Board, David
Bamman, Marco Büchler, Gregory Crane, and Stephen Hinds have offered invaluable
guidance on the project as a whole. We are likewise grateful to the generations of scholars
who laid the digital foundations for this work, including Professor Crane and his col-
leagues at the Perseus Project. This work was made possible by the support of the Digital
Humanities Initiative at Buffalo.

384 Neil Coffee et al.
posing and concerns with the ontological status of textual connections.1
In classical studies, Latin poetry has been the subject of the most intensive
investigation, but scholars of Greek poetry and of Greek and Latin prose
have become increasingly interested in how authors reuse and refer to their
predecessors and contemporaries.2
Digital approaches now in development can accelerate this research and
provide new large-scale perspectives on intertextuality.3 In recent decades,
computational investigation of intertextuality has generally consisted of in-
dividual word and phrase searches. Although such searches are now common
practice, they have rarely been the subject of methodological reflection. An ex-
ception is the 1997 essay by the theoretically adventurous Latinist Don Fowler
in which he critiqued the state of intertextual studies and addressed criticism
of computational methods. While others maintained that “correspondences
between texts [that] are not the product of reading but of computer searching
... tend to be too small, and have too much made of them,” Fowler disagreed
on both theoretical and practical grounds.4 How one located a correspondence
was immaterial, he argued, since the critic still needed to demonstrate that
it was an instance of marked language with some significance. More subtle
correspondences could create this significance as part of a larger thematic
program. And Fowler suspected in any case that “the computer is used more
often to check whether a correspondence is common than to find something
1 D’Ippolito 2000 reviews ancient conceptions of text reuse. Giangrande (e.g., 1967)
and his students have taken a lexical approach rooted in Alexandrian traditions that “as
a rule, [does] not treat allusion as would a literary critic” (Farrell 1991: 14). Pasquali
1968 provided an influential exposition of allusion as artistic practice and is taken as a
starting point by major works on Latin intertextuality such as Barchiesi 1984: 92, Conte
1986: 22–39, Thomas 1986: 171, Farrell 1991: 13, Wills 1996: 15n2, Pucci 1998: 13–14, and
Edmunds 2001: 12. Hinds 1998: 20n7 gives a précis of Pasquali’s Anglophone reception.
Conte 1986: 24–25 notes that the German scholars under whom Pasquali studied were
already engaged with the expressive possibility of intertextuality. Important intertextual
work was also done in the early modern period. So, e.g., De La Cerda 1617 became the
foundation for later work on Vergilian intertextuality, such as Knauer 1964.
2 Some examples are the collection Schepens and Bollansée, eds. 2005 on Polybius’s
use of his predecessors, Telo 2010 on Aristophanic intertextuality, Swift 2010 on the use
of lyric features in classical Greek tragedy, and Levene 2010 on Livy’s reuse of passages
from his predecessors.
3 We use the word “intertextuality” to denote the broadest range of relationships be-
tween texts. For an account of the differences in terminology (“allusion,” “intertextuality,”
“reference”) used by scholars of Latin literature, see Farrell 2005, esp. 98n2. We discuss
definitions further in our concluding section.
4 Fowler 2000 [1997]: 122.

385
Neil Coffee et al. Intertextuality in the Digital Age
new, and has actually acted as a brake on critical practice, rather than inspir-
ing a generation to waste its time on minutiae.”5
Fowler’s claim for the potential significance of finer correspondences is now
widely accepted. But his practical argument can be more strongly formulated
in light of technological and methodological advances in the years since his
essay was published. In 1997, it would have been very difficult to take full
and consistent account simultaneously of word rarity, diversity of diction,
alliteration, consonance, assonance, affective markers, and level of concrete-
ness in the analysis of one long poem, let alone hundreds. A recent study has
taken advantage of digital resources to conduct just this sort of assessment
of contemporary poetry. The study authors argue that decorated American
poets disregard Aristotle’s advice to employ rare words, but instead favor
features such as diverse diction, concrete imagery, and connotative rather
than denotative expressions of emotional states.6 While the study relied on
resources not yet available for classical languages (e.g., online semantic maps),
it nevertheless demonstrates that digital tools are quickly becoming part of
the interpretive process of literary studies, rather than a mere supplement.
Franco Moretti has referred to such digital use of quantitative models to
investigate literary corpora as a form of “distant reading” that contrasts with
and complements traditional close reading.7 Distant reading of intertextuality
has focused on identifying phrases or letter sequences in multiple locations
that are exactly identical, or nearly so. These matching techniques have been
used to demonstrate that classical Roman and late antique authors favored
different parts of Plato’s Timaeus, and that Diderot and d’Alembert borrowed
heavily from the very Jesuits they were contesting in composing their 18th-
century Encyclopédie.8 The study of intertextuality also encompasses a variety
of textual relationships considerably more elusive than quotation. To address
these with digital means requires identification of other computationally
tractable features, in addition to exact word form or letter-sequence iden-
5 Fowler 2000 [1997]: 123.
6 Kao and Jurafsky 2012, which compares 100 poems of contemporary professional
American poets with those of 100 amateurs. Aristotle’s discussion is at Poetics 1458a–59a.
7 Moretti 2005: 1, who conducts literary historical inquiries that, inter alia, trace the
development of novelistic sub-genres. Another form of textual analysis involves text
mining to identify recurring themes that might escape the reader’s notice. A convenient
example of the latter is the ongoing work of Cameron Blevins on the diary of Martha
Ballard, a midwife living in Maine in the late 18th and early 19th centuries (http://his-
torying.org/martha-ballards-diary/). Unless otherwise noted, all websites cited were last
checked on April 25, 2012.
8 Büchler et al. 2010; Horton, Olsen, and Roe 2010.

386 Neil Coffee et al.
tity, that constitute intertextual phenomena. Bamman and Crane 2008 have
demonstrated how detection of three features—word similarity (exact form
and dictionary headword), word order, and syntactic similarity—can identify
a range of potential intertexts, with the further suggestion of accounting for
meaning, meter, and sound.9
This article reports on a comparison of one digital approach to intertextu-
ality with traditional methods, and explores the consequences of automatic
intertextual detection for practical research and theoretical horizons. Our
comparison was performed using a free online intertextual comparison tool
we created, available at http://tesserae.caset.buffalo.edu/. As tested, the Tesserae
tool recovers approximately a third of the parallels captured by traditional
commentators, and adds a third not previously recorded. These results show
that the tool can find valuable new parallels in even the best-studied authors,
and suggest that it might be particularly useful in identifying intertexts in
works that have received less attention. In combination with information
from commentators, our method allows us to draw conclusions about the
overall intertextual artistry of ancient authors, as well as the intertextual
reading habits of scholars.
In what follows, we begin by describing our search and testing processes.
We then present a sample of our results showing how the Tesserae tool can
be used to identify individual parallels and groups of parallels. From there,
we explore the potential for distant reading opened up by automatic search,
and conclude with observations on the consequences of digital approaches
for intertextual study.
search method and testing
Our search focused on parallels containing at least two similar words in each
passage. This search was meant to approximate the traditional scholarly
identification of loci similes, which takes two-word pairs as the most basic and
common form of intertextuality. We left out of consideration the small- and
large-scale interactions that cannot be identified by lexical similarity alone.
Vergil’s corrupitque lacus, infecit pabula tabo (G. 3.481)10 resembles Lucretius’s
9 Other digital efforts are afoot to further expand the range of criteria that can be
captured. The eTRACES project, a collaborative effort funded by the German government
through 2014, has the goal of producing advanced tools for intertextual study (http://
etraces.e-humanities.net/home-etraces.html). The Italian Musisque Deoque project
(http://www.mqdq.it/mqdq/) is endeavoring to build on its digital textual collection to
produce a search engine for matching metrical patterns.
10 The plague “fouled the lakes and infected the pastures with disease.”

387
Neil Coffee et al. Intertextuality in the Digital Age
uastauitque uias, exhausit ciuibus urbem (6.1140)11 in rhythm, sound, syntax,
and general subject matter (plague), absent any shared words.12 These sorts of
similarity were not captured. We also did not address large-scale intertextual
correspondences formed by thematic material or homologous structures such
as book divisions.13 Furthermore, our test considered only intertextuality
between Latin texts, without accounting for the critical influence of Greek
literature on Latin.14 Rather than attempting to address the full complexity
of intertextuality at once, we chose instead to explore what automatic detec-
tion of a key subset of intertextual phenomena could achieve, and whether
it could constitute a step toward automatic detection and analysis of the full
range of intertextual phenomena.
To evaluate our search process, we performed two tests comparing book
1 of Lucan’s Civil War (BC) with all of Vergil’s Aeneid. These works were
chosen because they were long enough to give representative results, and well
studied enough to allow comparison with traditional approaches. Each of our
two tests was conducted with a different algorithm.15 Version 1 search found
minimum two-word phrases with exactly identical words in either order,
separated by no more than four other words. So Lucan’s horrida ... dumis
(1.28), describing Italian fields overgrown during civil war, matched Vergil’s
dumis ... / horrida (9.381–82), referring to obstacles to the flight of Nisus and
Euryalus. Version 2 matched minimum two-word phrases by dictionary head-
words (lemmata), in any order, with sentences as the unit of comparison.16
11 The plague “emptied the roads, and drained the city of its inhabitants.”
12 Thomas 1986: 180.
13 Fantuzzi and Hunter 2004: 467 discuss Catullus 64 as an example of structural rather
than lexical imitation.
14 The necessary techniques for cross-linguistic matching are still in development;
see the closing section of this piece. The Tesserae site currently offers Greek-to-Greek
matching of a limited number of texts, and Latin-to-Latin matching of a larger number.
Current plans foresee substantially increased availability of texts in both languages in
the near future.
15 To focus on results more likely to be meaningful, the tests employed a stop list of the
most common word forms compiled from a sample of classical authors. Matched pairs
containing any of these words were excluded. The algorithms are available as versions 1
and 2 on the Tesserae website, http://tesserae.caset.buffalo.edu, where users also have the
option of excluding high frequency words.
16 We used the Archimedes Project Morphology Service (http://archimedes.mpiwg-
berlin.mpg.de/arch/doc/xml-rpc.html, accessed August 21, 2011) to identify the lemmata.
The parser sometimes returned more than one lemma for a given word. So, e.g., bello
returned the adjective bellus, a, um, as well as the noun bellum, i. In the case of multiple lem-
mata for one word, all lemmata were checked for matches, leading to some false positives.

388 Neil Coffee et al.
In the Version 2 search, Lucan’s notae fulsere aquilae (1.244), referring to the
glimmer of Roman eagles that frightens the citizens of Ariminum, matched
Vergil’s notis fulserunt cingula bullis (12.942), describing the glimmer of the
baldric of Pallas that incites Aeneas to kill Turnus. Test results were compared
to parallels compiled from commentaries to explore the efficacy of the digital
intertextual matching and inform site users of the kinds of results they could
expect. The current third version available on the website should provide
faster and more accurate search, but awaits further testing.
results: individual parallels and sets of
parallels
The last example above was found by our lemma search, but does not appear
in any of the major commentaries on Lucan, including the ample new volume
on BC 1 by Roche.17 We start the discussion of our results with consideration
of this example to demonstrate how computational methods can reveal new
instances of the sort of local correspondences that have traditionally interested
scholars, even in well-studied texts.
If we look for sources for Lucan’s notae fulsere aquilae, we find no phrase
consisting of forms of the adjective notae and the verb fulgeo in extant Latin
before Vergil.18 When Lucan uses the phrase to describe Caesar’s invasion of
Italy, he starts his war narrative with a complex of emotions and perspectives
taken from the very end of the Aeneid, where Aeneas kills Turnus. In general,
Lucan’s connection of his opening with the end of the Aeneid suggests that fea-
tures of civil war found in the Aeneid were perpetuated down to the republic.
More specifically, Aeneas’s recognition of the bosses (notis ... bullis) suggests
that his motivations for killing Turnus are allegiance and sympathy toward
Pallas. In BC 1, the resplendence of arms highlights the tragically misdirected
excellence of Roman troops. Amid the institutional forces driving civil war,
the sorts of personal ties that animate Aeneas have little influence.19
17 Roche 2009. Other commentaries and articles consulted are: Heitland and Haskins
1887, Thompson and Bruère 1968, and Viansino 1995. As the length of commentaries
continues to expand (e.g., Horsfall 2000 at over 567 pages), Roche looks slim by compari-
son, but he offers a wealth of parallels, the greatest number of any Lucan commentator
by far (see further below).
18 Silius Italicus later employs the same phrase in the Punica to describe the arms of a
Roman mistakenly killed by his son the night before the battle of Cannae: notis fulsit lux
tristis ab armis (9.107 “the mournful light shone from the well-known arms”).
19 Lucan’s contradictions of Vergil in fact rework readings available in the Aeneid. The
futility of Aeneas’s allegiance with Latinus and Aeneas’s own inability to save Pallas also

389
Neil Coffee et al. Intertextuality in the Digital Age
Because automatic detection reveals numerous parallels at once, it can be
particularly useful in identifying constellations of intertexts. One such con-
figuration involves a single passage in a later text containing multiple parallels
with an earlier text. When Lucan’s Caesar is about to make his momentous
crossing of the Rubicon and invade Italy, he is suddenly brought up short
(BC 1.185–87):
ut uentum est parui Rubiconis ad undas,
ingens uisa duci Patriae trepidantis imago
clara per obscuram uoltu maestissima noctem.
As Caesar approached the waters of the slender Rubicon,
before him appeared the giant specter of his shaken fatherland,
shining brightly through the night’s gloom, but downcast.20
Three commentators noticed that words uisa ... maestissima recall the words
maestissimus Hector / uisus of Aen. 2.270–71, where the ghost of Hector ap-
pearing in a dream to Aeneas corresponds to the specter of Roma appearing
to Caesar at the Rubicon.21 Roche notes that, unlike Aeneas, Caesar is in the
midst of creating the crisis of which the vision warns, and disregards the warn-
ing. The intertext thus contrasts Caesar’s indifference to the destruction he
wreaks with Aeneas’s deep concern for Trojan losses.22 Searches by word-form
and lemma revealed two further Aeneid parallels not noted by commentators
that support this contrast. Lucan’s phrase Patriae trepidantis imago (1.186)
echoes the words patriae strinxit pietatis imago (Aen. 9.294, also with the
imago at line end), which describes the filial affection that inspires Ascanius
to promise to care for the mother of Euryalus as the young warrior sets forth
on a dangerous mission. Vergil uses the same phrase in the context of the pity
that Aeneas feels for Lausus, whom Aeneas killed as he rushed to avenge his
father: mentem patriae subiit pietatis imago (10.824). Lucan’s Patria refers to
Rome’s tutelary goddess, while Vergil’s term is in both instances an adjective,
patrius, -a, -um, “fatherly.” Whereas Caesar in the end simply disregards his
apparition, Ascanius responds by committing a generous act and Aeneas is
moved to sympathy.
suggest the feebleness of personal ties in the face of larger historical processes. Lucan’s
primary difference from Vergil on this theme is in his emphasis on impersonal social
forces, in contrast to Vergil’s invocation of divine causation and fate.
20 All translations are our own unless otherwise noted.
21 Thompson and Bruère 1968, Viansino 1995, and Roche 2009.
22 Roche 2009 on 1.186 observes in addition that the phrase uisa ... imago is commonly
used to introduce apparitions in epic, appearing several times in the Aeneid: 2.773, 4.557,
5.636–37, 5.722.

390 Neil Coffee et al.
Individually these parallels highlight Caesar’s indifference; collectively they
foreground his social isolation. The relations between Hector and Aeneas,
Aeneas and Ascanius, and even Aeneas and Lausus, are deeply personal.
Caesar’s encounter with the goddess Roma points rather to his lack of human
social ties anywhere in the poem: he is approached only by an abstract divinity,
not by any friend, son, or even pitiable enemy. The closest Caesar comes to a
personal relationship is his desire to possess Rome itself.
Constellations of intertexts can occur not just in one phrase, but across
various locations in the receiving text. Thus we find four separate instances
in BC 1 where Lucan recalls the Aeneid while referring to Italian city walls
(moenia). The second example was found by our automatic Tesserae search
and commentators, the other three only by automatic search.
1. Lucan’s mention of tottering house walls in Italian cities at the opening of
his epic (semirutis pendent quod moenia tectis / urbibus Italiae, 1.24–25)
reverses the notion of progress through the foundation of Rome’s walls at
the opening of the Aeneid (dum conderet urbem / / ... altae moenia Romae,
1.5–7).
2. As Lucan’s divine Roma appeals to Caesar to stop at the Rubicon, she invokes
Capitoline Jupiter’s guarding of Rome’s walls ([Iuppiter] magnae qui moenia
prospicis urbis / Tarpeia de rupe, 1.195), recalling the apparition of Hector
encouraging Aeneas to seek new walls for his people (his moenia quaere /
magna, 2.294–95). Lucan’s reference here is concurrent rather than adversa-
tive: Caesar threatens the great and storied walls of long foundation.
3. When the people of Rimini, alarmed by the prospect of invading Roman
soldiers, complain that their city is too close to the dangerous Gallic border
(o male uicinis haec moenia condita Gallis, 1.248), their lament pointedly
reverses Aeneas’s admiration for the rising walls of Carthage (o fortunate,
quorum iam moenia surgunt!, 1.437). The walls of Ariminum thus seem far
from providing the sort of secure home Aeneas longs for. Lucan’s moenia
condita at 1.248 also evoke the telos of founding Rome announced at the
opening of the Aeneid (dum conderet urbem, 1.5; altae moenia Romae, 1.7),
as do the other passages here.
4. Caesar riles up his soldiers by suggesting that they have been denied land
due for military service (quae noster ueteranus aret, quae moenia fessis?,
1.345). His phrasing recalls the words of Aeneas’s senior counselor Nautes,
who advises allowing Trojans exhausted from wandering to settle in Sicily
(et his habeant terris sine moenia fessi, 5.717). The concern of Nautes and
Aeneas for their people contrasts with Caesar’s use of one argument among

391
Neil Coffee et al. Intertextuality in the Digital Age
many to advance his personal interests. Roche notes the parallel between
this phrase of Lucan and Aen. 3.85, where Aeneas implores the god Apollo
to direct him to the new Trojan homeland with the words da moenia fessis.
The clear verbal similarities of the last parallel make it the most individually
compelling, but all four together create a small program of contrast between
BC 1 and the Aeneid. Vergil’s moenia suggest permanence, solidity, and safety,
the achievement of a home and a refuge. Lucan writes of the collapse of city
walls, their failure to keep citizens safe, and their utility only in rhetoric.23
Lucan’s repeated contrast undermines Vergil’s image of the city as a secure
refuge and suggests that Aeneas’s efforts to found one are pointless.24 The
complex of allusion prepares the way for Lucan’s extended development of
the paradox that, at Caesar’s approach, Romans seek safety by abandoning
their city walls rather using them for defense (1.484–520).
large-scale intertextuality
So far, we hope to have demonstrated that automatic search can reveal the
sort of individual intertexts and groups of intertexts typically found to be
meaningful. An evaluation of our results requires not just individual examples
of successful searches, however, but a full-scale comparison with parallels
found by commentators.
Amid the many studies addressing classical intertextuality, few have exam-
ined these relationships in the aggregate. So Hinds:
Modern scholarship of allusive relationships can be broadly divided ... into
studies of local contact (which tend to bracket out more systematic implica-
tions) and studies of systematic contact (which tend to bracket out details of
local contact). In line with the Latinist’s traditional preference for concrete and
isolable effects over intertextual open-endedness of any kind ..., purely local
approaches have usually been predominant.25
In the field of classics, the major exceptions are Knauer’s study of the Aeneid
and Homer and Nelis’s study of the Aeneid and Apollonius.26 Each of these
23 Walls are a theme of BC 6 as well, as discussed by Saylor 1978.
24 Again, Vergil anticipates Lucan. The fact that all the Aeneid passages in this constel-
lation are from the poem’s first half is instructive. By the time that Lavinium is burned
in Aeneid 11, the ideal of a secure city has receded.
25 Hinds 1998: 101.
26 Knauer 1964 and Nelis 2001. Casali 2011: 81n2 writes that “a study on the intertextual
relationship between Lucan and Virgil analogous to Nelis (2001) would be desirable.”
We hope that the work presented here offers a step in that direction, though along a
different path.

392 Neil Coffee et al.
works contains a wealth of individual analyses and summative judgments, but
studies of this type are rare precisely because few scholars have the requisite
skills, industry, and endurance to undertake such vast projects. Approaching
these studies can also be sufficiently daunting that even scholars explore them
selectively, and thus still obtain only a partial picture of full intertextual rela-
tionships. To come to grips with multiple such volumes is an overwhelming
prospect for even the most capable critics.27
One way out of this aporia is to develop methods of describing intertex-
tual patterns that can be absorbed at once, and that can be played off against
readings of individual loci. That is, we can employ forms of distant reading
to complement traditional close reading. In Moretti’s conception of distant
reading:
distance is ... not an obstacle, but a specific form of knowledge: fewer elements,
hence a sharper sense of their overall interconnection. Shapes, relations, struc-
tures. Forms. Models.28
Engagement with distant reading may require an adjustment in perspective for
literary scholars accustomed to treating each individual instance with rigorous
scrutiny. Moretti’s methods assume a level of comfort with the quantitative
patterns of inference native to the natural and social sciences, which allow
for abstraction of trends from large sets of individual instances. But we can
hope that by working back and forth between distant and close reading, it
will be possible for scholars to arrive at satisfying synthetic interpretations
of intertextuality without being overwhelmed by its totality. No less impor-
tantly, simplified representations of intertextuality can help make intelligible
to students what would otherwise seem an arcane element of language and
literary artistry.
automatic detection: ranking system
Our approach to distant intertextual reading began from two basic questions.
First, can our method detect a significant number of meaningful parallels?
Second, if we were assured that we had a substantial body of meaningful
parallels, what observations could we make about them? To answer the first
question, we manually scored our automatically generated results along with
27 Farrell 2005: 107: “Of course the mind recoils from the thought of a library full of
books entitled, ‘The Aeneid and Homer,’ ‘The Aeneid and Apollonius,’ ‘The Aeneid and
Ennius,’ and so forth. At the same time, however, I at least have no doubt that many such
books could be written.”
28 Moretti 2005: 1, emphasis his.

393
Neil Coffee et al. Intertextuality in the Digital Age
those from a set of commentators. Scoring both sets together helped to ensure
that we applied a consistent standard to each. We devised for our ranking the
system depicted in Table 1. “Formal similarity” here refers to any linguistic
features of the words identified or the phrases they constitute, including those
searched for (word and stem similarity) and others not specifically searched
for (e.g., meter, sound, syntax).29
The main purpose of the scoring system was to distinguish between mean-
ingful (5–3) and not meaningful (2–1) parallels. This distinction follows that
of Thomas, who writes that intertexts are either “susceptible to interpretation
or meaningful” or not.30 The first group ranked not meaningful included those
parallels that did not meet our intended search criteria due to limitations of
our search process, which were given a score of 1. For example, false positive
matches appeared when a given form in the text had more than one possible
lemma. Thus BC 1.52 uelis ... regnum matched Aen. 4.587–91 uelis /// ... regnis.
Lucan’s uelis, referring to Nero’s preference of position among the stars, is a
form of the verb uolo, whereas Vergil’s noun uelis refers to the sails of Aeneas
departing from Carthage.
Another group, those ranked with a 2, were not returned erroneously, but
were nevertheless judged not meaningful because the two words of the phrase
were so far apart or so ordinary that we believe few critics would recognize
them as a parallel or choose to investigate them further. So, for example, the
following instance:
certe populi quos despicit Arctos
felices errore suo, quos ille timorum
maximus haut urguet leti metus. (BC 1.458–60)
To be sure the pole star looks down upon peoples
fortunate in their error, for the greatest of fears,
the fear of death, oppresses them not at all.
Pandarus, ut fuso germanum corpore cernit
et quo sit fortuna loco, qui casus agat res,
portam ui multa conuerso cardine torquet
obnixus latis umeris. (Aen. 9.722–25)
29 For his large-scale study of the relationship between the Aeneid and Homer, Knauer
1964 ranked parallels based upon various types of relationships translated into symbols
on pp. 363–64. For purposes of clarity, our scoring system distilled these and other rela-
tionships down to the two criteria of formal similarity and context analogy.
30 Thomas 1986: 117.

Table 1. Tesserae Intertextual Parallel Scoring System.
Meaningful Not Meaningful
Interpretable Not Interpretable
More significant Less significant
5 4 3 2 1
• High formal
similarity in
analogous
context.
• Moderate formal
similarity in analo-
gous context, or
• High formal simi-
larity in moder-
ately analogous
context.
• High / moderate
formal similarity with
very common phrase
or words, or
• High / moderate for-
mal similarity with no
analogous context, or
• Moderate formal sim-
ilarity with moderate
/ highly analogous
context.
• Very common words in
very common phrase,
or
• Words too distant to
form a phrase.
• Error in discovery al-
gorithm, words should
not have matched. E.g.,
phrase across punctua-
tion boundary judged
not meaningful.

395
Intertextuality in the Digital Age
Once Pandarus realized his brother was strewn on the ground
and recognized the adverse circumstances, what misfortune was upon them,
he turned the gate on its hinge with all his might, pushing against it
with his broad shoulders.
As critics have reminded us,31 whether a potential moment of intertextuality
appears interpretable depends upon the subjectivity of the reader: can he or
she see some similarity between the two loci that admits further significance?
That may indeed be possible in instances such as this.32 Researchers investigat-
ing linguistic patterns could potentially examine such parallels profitably, and
indeed we hope to enable such linguistic study. Critics investigating literary
style could evaluate the similarity of two instances of anaphora in the forms
of qui ... qui. But in the regular course of literary interpretation, positing any
link between the passages based on such common words, commonly expressed,
would, in the words of Hinds, put “strains ... upon a philological decorum of
interpretability.”33 Conversely, for a phrase to be declared meaningful and be
given a score of at least 3, it needed to represent a coherent, distinctive idiom.
This criterion was meant to exclude phrases that consisted of pairs that were
so common as to be colorless (e.g., aut ... aut) or so widely separated that they
did not constitute or occur in one phrase.
Among the meaningful parallels discovered (types 5–3), we distinguished
between passages that generated new significance through their association
(types 5–4) and instances where text was reused without creating such sig-
nificance (type 3). We have called the former “interpretable” and the latter
“not interpretable,” following a distinction made by critics. In the case of
Lucan, for example, Roche writes of the “incidental influence of the language
and expression of Senecan prose upon Lucan’s Latin.”34 Roche’s “incidental
influence” indicates an instance of mere language reuse, without the genera-
tion of meaning by the interaction of the two loci. We attempted to make this
distinction more precise by introducing surrounding context as a criterion.
To be deemed potentially generative of meaning, and so interpretable (5–4), a
given parallel needed to have analogous contexts in both works that were not
31 Notably Martindale 1993.
32 Hinds 1998: 26: “There is no discursive element in a Roman poem, no matter how
unremarkable itself, and no matter how frequently repeated in the tradition, that cannot
in some imaginable circumstance mobilize a specific allusion. This is a truth often sup-
pressed by professors of Latin for reasons of pedagogy and (perhaps) peace of mind; but
it is a truth none the less.” Wills 1996 explores this subject in depth.
33 Hinds 1998: 45.
34 Roche 2009: 29. Cf. his statement at p. 24 that “frequently [Lucan’s] diction is influ-
enced directly or indirectly by Horace’s own expression.”

396 Neil Coffee et al.
wholly typed. This criterion likewise originates in current critical practice. In
a study of intertextuality in Martial, Hinds asks:
When Martial’s elegiac couplets speak the cadences of Ovidian elegy, when and
how far are the echoes to be read as thematically grounded, when or how far
as utterly indifferent to content or context?35
Hinds’s “thematically grounded” parallels are equivalent to those that we
deem generative of meaning, and so interpretable, and those that are “indif-
ferent to content or context” are equivalent to simple language reuse, and so
not interpretable.
We can also describe the distinction between interpretable (5–4) and non-
interpretable (3) parallels with respect to genre. In an influential formulation,
Conte has distinguished between a source text as “exemplary model,” where
the referring author directs the reader’s attention to a particular moment
in another work, and as a “code model,” where the referring author draws
instead on a set of generic gestures laid out by another author or authors.36
We would suggest that the parallels we have classified as interpretable (types
5–4) correspond roughly to Lucan’s use of Vergil as an “exemplary model,” in
that they generate meaning by evoking specific loci in the Aeneid. Instances
of Lucan’s use of Vergil as a “code model” generally correspond to text reuse
we have designated not interpretable (type 3), in that they most often consist
of distinctive language repeated by Lucan, but in phrases that do not evoke
any particular resonance. When this language comes only from the Aeneid, or
when it is stamped by Vergilian influence, it can be considered an instance of
Lucan using epic language for its generic qualities without evoking a mean-
ingful analogy—in other words, a use of epic “code.” Some of the parallels
we have classified under text reuse may simply be common phrases, metri-
cally compliant but otherwise not particular to epic. But some preliminary
inquiries into this topic have suggested that most BC 1 phrases shared with
the Aeneid are found first there.
The analysis of the following parallel, scored as a 3, shows how these cri-
teria were applied:
patriae sedes remeamus in urbis,
inpiaque in medio peraguntur bella senatu. (BC 1.690–91)
But now I return to the site of the city,
where the insidious civil war is conducted in the senate itself.
35 Hinds 2007: 119.
36 Conte 1986: 31.

397
Neil Coffee et al. Intertextuality in the Digital Age
haud procul hinc saxo incolitur fundata uetusto
urbis Agyllinae sedes, ubi Lydia quondam
gens, bello praeclara, iugis insedit Etruscis. (Aen. 8.478–80)
Not far from here lies the site of the city of Agyllina,
founded upon ancient stone, where once the Lydian people,
renowned warriors, settled in the Etruscan hills.
This parallel avoided being assigned a score of 2 because its two words were
close enough in each text and infrequent enough to constitute an instance
of text reuse. On the other hand, the words are rather ordinary, and their ar-
rangement in the line differs substantially between the two passages. Finally, it
is difficult to find analogous elements in the topic and context of the passages
that would make this a meaningful recollection by Lucan. In Lucan’s passage,
the bacchant who predicts civil war for Rome returns in her prophetic and
visionary journey to foretell dissent in the Roman senate itself. In Vergil’s
passage, Evander describes the city of Agyllina, where Aeneas may find allies
willing to fight against Mezentius and the Latins. Beyond a general context of
war shared by both epic poems, there seemed little obvious analogy, concur-
rent or adversative, between the passages.
The final and most subjective distinction we employed was between
stronger (5) and weaker (4) forms of meaning-generating parallels, which
we called respectively “more significant” and “less significant.” This was an
effort to indicate, among all meaning-generating parallels, those that seemed
most meaningful according to traditional critical standards. The example
given above of the parallel between Lucan’s notae fulsere aquilae (1.244) and
Vergil’s notis fulserunt cingula bullis (12.942) represents a strong parallel that
merited a score of 5. The following parallel earned a rank of 4:
tum rura Nemetis
qui tenet et ripas Atyri, qua litore curuo
molliter admissum claudit Tarbellicus aequor,
signa mouet. (BC 1.419–22)
Then the soldiers who guarded the Nemes changed
camp, and those holding the banks of the Atyrus, where
the Tarbellian people surround the sea that flows gently
to their curving shore.
feror huc et litore curuo
moenia prima loco fatis ingressus iniquis
Aeneadasque meo nomen de nomine fingo. (Aen. 3.16–18)

398 Neil Coffee et al.
I am borne here [to Thrace], destined for failure,
and place my first walls on the curving shore,
and for my people fashion the name of Aeneadae from my name.
Apart from the exception noted in our discussion of the moenia theme, the
passages presented so far were found only by the Tesserae search. In this case,
we give an example noted by Roche as well, to illustrate why it falls short of
the highest ranking. Roche observes only that we find the words litore curuo
ending lines at Aen. 3.16, as well as Aen. 3.223, Ov. Fast. 3.469, and Valerius
Flaccus 1.275.37 His lack of fuller exposition suggests that he finds little sub-
stantial to explicate in the parallel between the two passages, which might, in
the terms used here, earn the echo the ranking of a 3 for pure language reuse.
In our reading, it deserved a higher score both because of the exact repetition
of the phrase, and because of a potentially analogous context. The analogy is
between Aeneas founding a fortification he will have to abandon, and Roman
legionaries in BC 1 abandoning their fortifications for civil war, both on or
near curved shores. The parallel fell short of a ranking of 5, however, because
the connection between the words litore curuo and this analogous context
seemed slight.
In light of continued critical reminders of the subjectivity of meaning-
making by the reader, the classification of parallels by the elementary criteria
we have outlined may seem crude. Some classification system is required,
however, to make a basic distinction between more and less useful results;
to evaluate results from automatic discovery against those from traditional
methods; and to present the outcome in intelligible form. More subtle distinc-
tions—including those we posit between pure language reuse and different
types of meaning-generating parallels—will be open to debate, but we hope
our criteria will help bring greater transparency to the major types of inter-
textual phenomena. Our results will be posted online, and we hope to develop
a system to allow users to comment on marked-up texts, so that future clas-
sification can benefit from the insights of a broader interpretive community.
results of automatic detection
Applying our ranking system to Tesserae parallels and those of the com-
mentators produced the results depicted in Table 2. The table indicates that
automatic discovery with manual examination of results can reveal substantial
numbers of the parallels most significant to literary interpretation. Tesserae
search added 279 meaningful parallels to the commentators’ 364, an increase
37 Roche 2009: 288.

Neil Coffee et al.
Table 2. Intertextual Parallels Between Lucan BC 1 and Aeneid Found by Tesserae and Commentators, by Type.
TYPE ALL Tesserae Tesserae Tesserae All Heitland & Thompson Viansino Roche
Word Word Lemma Commentators Haskins & Bruère
Lemma
5 103 36 18 26 93 4 14 30 85
4 115 57 25 43 79 3 12 18 66
3 425 280 41 262 192 6 13 33 168
2 2289 2241 70 2182 55 1 1 8 50
1 486 486 5 481 0 0 0 0 0
TOTAL 3418 3100 159 2994 419 14 40 89 369
5–3 643 373 84 331 364 13 39 81 319
5–4 218 93 43 69 172 7 26 48 151
Note: Testing conducted Fall 2009–Fall 2010. In totals numbers, individual parallels found by more than one source are counted only once, so totals may be smaller
than the sums of the individual values. See key to table overleaf.

400 Neil Coffee et al.
key
ALL Total of all Tesserae and commentators
Tesserae Word Tesserae identical word form match
Tesserae Lemma Tesserae lemma match
All Commentators Total of all following commentators
Heitland and Haskings Heitland and Haskins 1887
Thompson and Bruère Thompson and Bruère 1968
Viansino Viansino 1995
Roche Roche 2009
Table 2. cont.
38 A point made by Bamman and Crane 2008.
of 43%, for a total of 643 BC 1–Aeneid parallels detected (as seen in the type
5–3 totals). For interpretable (types 5–4) parallels, Tesserae word (43) and
lemma (69) searches separately returned results comparable to those of
Viansino (48), but fewer than half those of Roche (151), though the combined
Tesserae total excluding duplicates (93) comes closer. Tesserae returns fewer
meaningful results because it currently lacks sensitivity to features such as
semantics and syntax.
For the moment, the automatic search available on the Tesserae website
can serve as a check on and complement to traditional methods. More signifi-
cantly, it can identify parallels unrecognized by commentators in substantial
numbers, as illustrated in Figure 1. The circle on the left represents the total
of all type 5 and 4 parallels found by all commentators combined (172).The
circle on the right represents all such parallels (93) found by both Tesserae
search methods (exact word and lemma). The numbers within show, from left
to right, parallels found only by commentators (125), by both Tesserae and
commentators (47), and only by Tesserae (46). Although the Tesserae search
captured only a quarter of the high interest type 5–4 parallels found by the
commentators, this recovery rate nevertheless shows the capacity of auto-
matic search to replicate traditional methods. Equally compelling is the fact
that Tesserae search returned a high proportion of previously undiscovered
parallels. This result suggests that the systematic application of fixed criteria
can detect parallels traditional criticism may tend to overlook.38
We can explore this last question in greater depth by considering which
parts of a source text are referred to by the target text, or, in the case of our
sample, how often intertexts from each book of the Aeneid appear in BC 1.
Figure 2 gives the number of interpretable intertexts between BC 1 and in-

Neil Coffee et al.
Figure 2. Distribution of Type 4 and 5 BC 1–Aeneid Parallels by Aeneid Book.
Figure 1. Unique Type 5 and 4 BC 1–Aeneid Parallels Found by Tesserae vs. All
Commentators.

402 Neil Coffee et al.
39 By the ordinary measure of standard deviation, Roche’s parallels fluctuate three
times as much as those of Tesserae or Viansino. Roche has a mean number of high value
parallels per Aeneid book of 12.6, with a standard deviation of 6.7. For Tesserae the values
are respectively 7.8 and 2.6, and Viansino 4 and 2.4.
dividual books of the Aeneid, as found by each of three sources: combined
Tesserae word and lemma search, Viansino, and Roche. The y-axis gives the
number of parallels found, while the x-axis shows in which book of the Aeneid
the parallel occurs. Each source finds more parallels to the first six books of
the Aeneid than to the last six. Roche finds more parallels overall. The dis-
tribution of Roche’s parallels in the Aeneid varies more than those found by
Tesserae and Viansino, as indicated by the large gap between the highest and
lowest numbers of parallels Roche finds (28 to Aeneid book 2 vs. 5 each to
Aeneid books 5 and 10).39
Where all sources find consistent variation in intertextual connections
with different books of the Aeneid, this likely represents a real difference in
Lucan’s practice. Just as Servius devotes a much greater portion of his Aeneid
commentary to the first half of Vergil’s epic than the second, Lucan may have
had greater interest in, or use for, the first half of the Aeneid in composing BC
1. Both poet and commentator may have responded in a similar way to the
work itself or been influenced by an inherited interpretive emphasis on Aeneid
1–6. We explore these possibilities further below when we combine results
from all sources to obtain an overall picture of Lucan’s intertextual patterns.
Where the sources diverge, comparison can illuminate their different
critical practices. The principal difference is the greater variability in the
results of Roche, who finds a high proportion of parallels between BC 1 and
the perennially popular and well-studied Aeneid books 2, 4, and 6. Roche’s
higher variability may be a matter of outlook and interests—perhaps he had
certain books of the Aeneid foremost in his mind, while Viansino had others.
Alternatively, the difference in variability may be a function of different meth-
ods. Roche conducts a deep analysis of BC 1, while Viansino and Tesserae find
fewer results: Viansino because he treats all of BC, and Tesserae because it did
not capture all commentator intertexts. So a deeper look at BC 1 may simply
reveal greater variation in Lucan’s intertextual practice. The development of
more comprehensive automatic search systems may help distinguish more
clearly between observer preferences and textual differences.
If we consider not just interpretable parallels (5–4), but all meaningful
parallels (adding type 3), we find another significant trend. More recent
investigations, both by traditional commentators and our automatic search,
have returned a significantly higher proportion of pure language reuse/code

403
Neil Coffee et al. Intertextuality in the Digital Age
model language (type 3) intertexts, as well as ordinary language (type 2).
These types typically represent, respectively, literary craftsmanship at work
and prevailing linguistic patterns. The increase in type 3 and type 2 parallels
would seem to be a natural result of subsequent commentators examining
BC 1 more minutely. It would also seem to result from the increasing use of
digital search. In the latter respect, Roche represents something of a mid-point
between earlier commentators and Tesserae. The promotional materials for
Roche’s commentary point out that his was the first commentary on BC 1
to make full use of electronic search.40 The consequences are evident from
comparison with Viansino. As Figure 3 shows, more than half of Viansino’s
parallels are interpretable (types 5 and 4 at the top of the second chart),
whereas types 2 and 3 make up more than half of Roche’s instances (at the
bottom of the first chart). Note the expansion of the segments for types 3 and
2 as we go from Viansino to Roche, indicating the recovery of more common
phrases and ordinary language, possibly as a result of electronic search.
The trend toward more parallels of types 3 and 2 is still more evident in
the Tesserae word and lemma searches. Tesserae lemma matching by whole
sentence naturally returned far more total candidates (2,994) than exact word
search within six-word windows (159). Their distribution by type, along with
those of Viansino and Roche, are shown in Figure 4. Beyond illustrating the
increase in type 3 and 2 parallels with forms of electronic search, these re-
sults allow us to draw a further conclusion. The upward slanting lines in the
two forms of Tesserae search indicate that automatic detection of two-word
similarity returns a high proportion of ordinary language. The commentators
naturally filter most of this out, so that, unlike Tesserae search, they return a
much lower proportion of type 3 and type 2 parallels. Even if the commenta-
tors are willing to admit more common phrases and ordinary language when
using electronic search, and so expand the definition of intertextuality in this
way, better replication of their results will nonetheless require the ability to
sort parallels automatically to allow for selective focus by type.
If rather than considering the differences between the sources, we combine
them, we arrive at the best overall picture of Aeneid intertexts by type in BC 1
(see Figure 5). Some of Lucan’s Aeneid intertexts originate with Vergil, others
Vergil himself adapted from the common stock of Latin epic, while others
are simply elements of the Latin language selected for metrical tractability.
Distinguishing precisely between these three groups would require further
40 Roche per litteras writes that the main electronic tool he used in researching the
commentary “was the Packard Humanities Institute disk of Latin texts ... using Musaios,
typically for paired, rare or juxtaposed words, or for particular clausulae.”

404 Neil Coffee et al.
investigation, but the distribution of our types 5–3 offers a useful approxima-
tion. Lucan has 218 interpretable (type 5 and 4) parallels to the Aeneid in BC
1, as compared to 425 instances of type 3 parallels. In the terms offered by
Conte, this means that Lucan creates “code model” intertexts roughly twice as
often as “exemplary model” intertexts. He uses ordinary (metrically compat-
ible) language some 100 times as often. This distribution of intertext types
generally accords with Zipf’s Law for natural language, which holds that the
most common words in a language occur vastly more frequently than oth-
ers, making for a roughly asymptotic distribution curve.41 Although it makes
intuitive sense that there will be far fewer meaningful intertexts than instances
of ordinary (metrically compatible) language, it will require further study,
including advances in automatic detection and classification, to determine
whether this pattern in fact holds true for intertextual relationships generally.
Such study can also illuminate the further question, more compelling for
literary scholars, of how types of intertexts are distributed within an author’s
works and among different authors.
analysis by location in source text
The discussion of occurrence of intertexts by type has brought us to consid-
eration of what combined information from commentators and automatic
detection can say about the relationship between Lucan BC 1 and the Aeneid.
Here we set out to answer three further questions: 1) Which books of the
Aeneid does Lucan draw from most in BC 1?; 2) How are his references
41 Zipf 1949. Recent discussion in Sampson 2003.
Figure 3. Detection of BC 1–Aeneid Parallel by Type in Commentators.

Neil Coffee et al.
Figure 4. BC 1–Aeneid Parallels by Commentator and Type.

406 Neil Coffee et al.
Figure 5. All BC 1–Aeneid Parallels from Tesserae and Commentators by Type.
distributed within BC 1?; and 3) How do these conclusions affect existing
interpretations of the relationship between BC and the Aeneid?
The first of these questions was addressed above by looking at the BC 1
intertexts assigned by various sources to books of the Aeneid. Combining these
sources and indicating distribution by type gives us the results illustrated in
Figure 6. This figure shows the number of parallels by type to each book of
the Aeneid. If we begin by looking at the total height of each column, we get
an image of Lucan’s total use of references types 5–3. On the whole, he is less
interested in the games of Aeneid 5, which gets only 5% of his total references
(35 of 643), and most interested in books 1 (10%; 64), 2 (10%; 62), and 4
(11%; 68). If we mentally exclude the darkest, bottom portion of each column
and focus on the top two segments, we see greater differences in Lucan’s use
of interpretable references (types 5–4). The larger size of these combined seg-
ments for in the first half of the chart, representing books 1–6 of the Aeneid,
show Lucan devoting 60% of his interpretable references to the first half of
Vergil’s epic. If we proceed to consider only the topmost segment of each bar,
representing the most meaningful intertexts, we find a still stronger preference
for phrases from Aeneid book 2, as well as to a lesser degree for those from
books 3, 4, 6, 7, 11, and 12.
To a certain extent, Lucan’s preferences in BC 1 mirror those of ancient
and modern audiences. We noted above the greater attention given by Servius
to the first half of the Aeneid. Augustine was enthralled by the story of Dido.
Books 2, 4, 6, and 12 of the Aeneid are standard parts of modern school cur-

407
Neil Coffee et al. Intertextuality in the Digital Age
ricula because the fall of Troy, the Dido story, the underworld, and Aeneas’s
final confrontation with Turnus remain among the most exciting and engaging
parts of the poem. It could be that Lucan too found these the most compelling
parts of the poem and thus had them closest to mind.
Closer scrutiny of context suggests more specific reasons for Lucan’s choic-
es. We might have conjectured that the Neronian poet would draw the greatest
number of his parallels from Aeneid 2. In BC 1 he sets out his major theme of
the Roman civil war as the destruction of Rome, for which he naturally draws
upon Vergil’s descriptions of the fall of Troy. Thus we find parallels such as BC
1.24–6 semirutis ... tectis / ingentia ... / saxa iacent ≈ Aen. 2.489 pauidae tectis
matres ingentibus errant, where the tottering walls of Italian cities desolated
by civil war recall the terror in Priam’s palace at the invasion of the Greeks.
Lucan’s reliance on Aeneid 4 seems a less obvious choice. Lucan adapts
language from Aeneid 4 for several purposes, but a significant subset of his
type 5–4 parallels (8 of 27) borrow notions of fear and mania from the Dido
story to show the same conditions gripping Rome before the outbreak of
civil war. In one of these instances, Lucan refers to Vergil’s combination of
these two motifs: BC 1.676 attonitam rapitur matrona per urbem ≈ Aen. 4.666
concussam bacchatur Fama per urbem.42 Lucan implicitly compares the frenzy
Figure 6. BC 1 Intertexts by Book of Aeneid.
42 “The [prophetic] matron is swept through the city” ≈ “Rumor runs riot through
the stunned city.” The other instances are: BC 1.469 ≈ Aen. 4.173; 1.472 ≈ 4.183; 1.472 ≈
4.189; 1.495-96 ≈ 4.300-1; 1.674-76 ≈ 4.300-1; 1.676 ≈ 4.68; 1.678 ≈ 4.110.

408 Neil Coffee et al.
of a matron predicting civil war to the dark rumors and madness that grip
Carthage in the wake of Dido’s death. Added to this are other allusions that
evoke inauspicious similarities between Carthage and Rome, such as when
Caesar casts off Roman law (procul hinc iam foedera sunto, BC 1.226) in a way
that recalls Dido’s preemptive rejection of future treaties between Carthage
and a future Rome (nec foedera sunto, Aen. 4.624). This complex of allusions
to Aeneid 4 in BC 1 suggests that Romans of the civil war period are as mad,
rumor-driven, and (in the case of Caesar) lawless as passionate Dido and her
Carthaginians ever were.
Lucan makes equally specific use of other Aeneid books. Aeneid 1 and 7
respectively initiate the poem as a whole and its major war narrative. Lucan
does both at once in BC 1, borrowing from these books to create resonances
of poetic and martial beginnings.43 Lucan uses Aeneid 6 to show how the
crime and corruption Vergil sequesters in the underworld is rampant in the
real world of Roman affairs.44 Aeneid 11 and 12 provide Lucan with military
language, showing through premonitions that Rome’s civil war will be as dire
as the full-blown battle of the Latins and Trojans.45
The discovery of a number of significant parallels to the relatively un-
derstudied book 3 by all sources suggests the sensitivity of each method to
Lucan’s artistry apart from his reception. A significant strand in Lucan’s use of
Aeneid 3 involves reversing that book’s optimistic prophecies of a new Trojan
homeland to suggest the woeful future in store for the Roman people.46 One
parallel discovered only by Tesserae search exemplifies this motif. At Aeneid
3.417–18, the seer Helenus advises Aeneas to avoid Scylla and Charybdis, and
43 Aen. 1 parallels begin already with BC 1.2 canimus ≈ Aen. 1.1 cano. For Aen. 7, note,
e.g., BC 1.68 immensum aperitur opus ≈ Aen. 7.45 maius opus moueo; 1.311 longa ... pace
≈ 7.47 longa ... pace; 1.8 licentia ferri ≈ 7.461 amor ... ferri; 1.62 limina Iani ≈ 7.610 limine
Ianus.
44 E.g., BC 1.13 heu quantum terrae potuit ≈ Aen. 6.828 heu quantum inter se bellum;
1.178 rapti fasces pretio ≈ 6.622 fixit leges pretio atque refixit; 1.681 tela manusque ≈ 6.57
tela manusque, with the reversal 1.256 hac iter est bellis ≈ 6.542 hac iter Elysium nobis.
45 E.g., from book 11: BC 1.237 clangorque tubarum ≈ Aen. 11.192 clangorque tubarum;
1.388 it ... clamor ≈ 11.192 it clamor; 1.4–5 rupto foedere regni / certatum totis concussi
uiribus orbis ≈ 11.313 toto certatum est corpore regni. From book 12: 1.157 dat stragem ≈
12.454 dabit ... / stragemque; 1.366 usque adeo miserum est ciuili uincere bello ≈ 12.646
usque adeone mori miserum est; 1.388 it ... clamor ≈ 12.409 it clamor.
46 E.g., BC 1.196 Phrygiique penates ≈ Aen. 3.148 Phrygiique penates; 1.224 Hesperiae
... aruis ≈ 3.171 Dictea negat tibi Iuppiter arua; 1.524 manifesta fides (of ills to come) ≈
3.375 manifesta fides (of divine guidance). O’Hara 1990 shows that such prophecies are
in fact already proved false in the Aeneid itself.

409
Neil Coffee et al. Intertextuality in the Digital Age
notes that their domain along the sea is where the sea once separated Sicily
from Italy, pontus ... / Hesperium Siculo latus abscidit. Vergil’s image is one of
opening up the straits of Messina for Aeneas’s voyage. At BC 1.547, Lucan uses
the language of Vergil to figuratively close the same sea passage. As a portent of
the coming civil war, Mt. Aetna erupts and with its lava rejoins Sicily and Italy:
ignis in Hesperium cecidit latus. Whereas the phrase of Helenus suggested the
unlocking of possibilities in the new Italian world, Lucan’s reversal describes
a calamity signaling the end of the Roman power Aeneas had established.47
The attention given to these books contrasts with Lucan’s relative neglect
of others. The poet draws least across all types from Aeneid book 5, and little
from book 8. We may surmise that the comparatively tranquil subject matter
of these books, including the Trojan games and Aeneas’s sojourn with Evander,
bore too little relation to Lucan’s story of Roman catastrophe. Exceptions show
the rule, as when Lucan borrows Vergil’s phrase for the cold blood of the boxer
Entellus to describe the hot blood of Caesar’s troops (Aen. 5.395 gelidus ... /
sanguis ≈ BC 1.363 calidus ... sanguis). Where he can thus intensify language
from the Aeneid in the direction of conflict, Lucan seizes the opportunity, but
otherwise passes over Vergil’s tenuous discourses of normalcy and harmony.
The poetic effects of Lucan’s intertexts with individual books of the Aeneid
can all be (and have been) investigated further. We hope this survey has suf-
ficed to shed some light on this aspect of Lucan’s artistry and to demonstrate
that large-scale analysis of intertextual connections by source location can
be productive.
analysis by location in target text
We offer a last perspective on large-scale intertextuality by considering the
distribution of Aeneid intertexts within BC 1. Figure 7 shows the locations of
parallels to the Aeneid in BC 1 by type. The top of the figure represents the
beginning of the book, the bottom the end, with episodes labeled on the left
and line numbers on the right. Each circle within represents one parallel to
the Aeneid, separated by type as indicated at the bottom. When one locus in
BC echoes more than one in the Aeneid, more than one circle appears at the
relevant BC locus. The nearly continuous line of type 3 circles on the left side
47 Lucan’s intertextual reversal seems almost argumentative, since he might instead
have chosen to echo Aen. 3.370–77, where Vergil describes the eruption of Aetna. As it is,
Lucan’s lines share only two significant words with this longer passage, widely separated
and in different cases and metrical positions (Aetna ... / / / ... flammarum; Aetnae / ... flam-
mas, BC 1.545–46). Lucan is reversing Ovid as well, though without the verbal parallels.
Ov. Met. 15.290–92 refers to accounts that Sicily was separated from the mainland by a
surge of the sea.

Figure 7. Distribution of Aeneid Intertexts in BC 1 by Type.

411
Intertextuality in the Digital Age
of the figure shows Lucan drawing on the Aeneid for epic (“code model”)
diction and metrically compatible phrases evenly throughout BC 1. The con-
sistency of Lucan’s use of Vergil at this level suggests the near-inevitability for
the Neronian poet of using the epic materials Vergil provided.
The distribution of interpretable parallels in BC 1 will be of greater interest.
We have added dark bars indicating the thickest clusters of these, with light
bars in areas relatively or completely lacking in parallels. Among types 4 and
5, the middle and rightmost lines of circles, the most conspicuous pattern is
the thick cluster at the opening of the poem: in the proem at lines 1–7 and
the apostrophe to Rome at lines 8–32. Scholars have demonstrated how in
these lines allusions to the Aeneid and Iliad identify Lucan’s poem as a new
sort of epic focusing on repetitive destruction rather than linear progress.48
Another cluster in the opening descriptions of Caesar and Pompey in lines
120–57 links each with figures in the Aeneid and epic tradition. A final group,
in the prophecy of the matrona, closes the book. Lucan thus shows a pattern
of employing interpretable references in opening and closing sections, as well
as in a highly significant passage establishing his principal characters.
If we look more closely at the group of intertexts that open the book, we
find a notable absence. Within a thicket of Vergilian language, references to the
Aeneid suddenly cease in the middle of Lucan’s praise of Nero (39–59). Roche
has noted that “the invocation [to Nero] is in a radically different mode from
the epic announced at 1–32. It raises questions of propriety of genre, made
more urgent by the cataclysmic subject matter which Nero is said by Lucan to
inspire.”49 The change can be explained in part by Lucan’s shift to the Georgics
as a model for his panegyric.50 But the cessation of Aeneid intertexts at the
height of his praise of Nero also means that Lucan foregoes an opportunity
to associate Nero with the great mythical heroes of epic in favor of references
to real-world Rome, creating a more prosaic tone that tempers his encomium.
In the other major lacuna in BC 1, at lines 639–72, Lucan ceases significant
references to the Aeneid during the astrological forecast of civil war by Nigidius
Figulus.51 Again the poet turns to the Georgics as well as Lucretius and other
48 Overview at Roche 2009: 22–23.
49 Roche 2009: 9.
50 Particularly G. 1.24–62. Roche 2009: 9.
51 Casali 2011: 92–95, which was published after our testing, suggests several parallels
between this episode and the Aeneid. Only one of these, cited on p. 95, meets the criteria
for an interpretable parallel in our classification scheme. This is BC 1.649–50 quod cladis
genus, o superi, qua peste paratis / saeuitiam? ≈ Aen. 2.361–2 quis cladem illius noctis, quis
funera fando explicet ...? We would rank the parallel as a type 5, so in fact we have one
intertext recalling the destruction of Troy as a precedent for the calamity at Rome to give
political point to the portents of nature Figulus describes.

412 Neil Coffee et al.
didactic writers.52 The silencing of high epic resonances seems consonant with
scientific prophecy in a way that it is not with praise of a princeps.
In addition to sections of concentrated references and lacunae, we also find
a tendency to cluster references at the beginning and ending of sections (indi-
cated, again, by dark bars in figure 7). As Lucan moves from the praises of Nero
to the causes of war (61–68), from the causes of Rome’s decline to Caesar’s halt
at the Rubicon (178–205), from the Rubicon to panic at Ariminum (224–44),
from Curio’s speech to Caesar’s (290–311), and from Caesar’s speech to that
of Laelius (345–76), he bridges his sections with some of his most notable and
significant references to the Aeneid. By clustering Vergilian references around
section openings and closings, Lucan brings density of meaning and Vergilian
authority to his transitions. Within these transitions, Lucan employs a specific
pattern that highlights his own virtuosity. Just prior to the close of a section,
he presents his audience with a Vergilian intertext, then finishes with his own
novel sententia, before beginning a new section freshly anchored and autho-
rized by references to the Aeneid. Thus toward the end of the enumeration of
the causes of Rome’s decline, we find the bribery of the Vergilian underworld
translated into the real world of Lucan’s Rome (BC 1.178 rapti fasces pretio
≈ Aen. 6.622 fixit leges pretio atque refixit). Lucan closes the accounting with
his own sententious vision of avidity leading to war (concussa fides et multis
utile bellum, 1.182). He then opens the section on Caesar at the Rubicon with
several references to the Aeneid. In one, noted above, the specter of the city of
Rome (1.186 ingens ... Patriae trepidantis imago), revealing Caesar’s betrayal
of pietas, recalls Aeneas’s thoughts of his relationship with his father after
unwillingly killing the boy-warrior Lausus (10.824). In another, Roma calls
on Caesar’s troops to refrain from fighting on their native soil in words used
to encourage disheartened Trojans to join the fighting (BC 1.190 quo tenditis
ultra? ≈ Aen. 9.781 quo deinde fugam, quo tenditis?). Lucan’s transition from
the speech of Caesar to that of Laelius follows a similar pattern.53 Here and
elsewhere, Lucan uses Vergilian authority and color to hold and excite his
audience during scene transitions, but reserves section endings for his own
master strokes.
The grouping of intertexts around transitions can also create the sort of
register changes seen in the praise of Nero and the prophecy of Nigidius
Figulus. We find Aeneid intertexts in the transitions to and from Caesar’s first
52 E.g., Roche 2009: 370–71 cites BC 1.657 ≈ Lucr. 2.413, Manilius 5.744–45, as well as
[Sen.] Her. O. 1385–87; 1.658–59 ≈ Ve rg. G. 1.34–35.
53 BC 1.345 ≈ Aen. 3.85. Sententious ending: detrahimus dominos urbi seruire paratae,
BC 1.351. Further significant Aeneid references: BC 1.354, 355, 363, 366.

413
Neil Coffee et al. Intertextuality in the Digital Age
speech to his troops. Caesar begins in the grand tones of Aeneas consoling his
men (BC 1.299 bellorum o socii, qui mille pericula Martis ... ≈ Aen. 1.198–99
o socii (neque enim ignari sumus ante malorum)), while denigrating the life
of peace which in the Aeneid seems a lost ideal (BC 1.311 ueniat longa dux
pace solutus ≈ Aen. 7.45–46 urbes / ... longa placidas in pace regebat). He ends
with the tendentious reference to city walls owed to his men considered above
(1.345 ≈ 5.717). Vergilian intertexts are not entirely absent from the middle
of the speech: Caesar borrows Vergilian language to compare Pompey to a
bloodthirsty tiger (1.328 matrum ... lustra ≈ 3.646–47 ferarum / lustra). But
this parallel is of a different order from his invocation of grander Vergilian
subjects. At the heart of his appeal, Caesar descends from epic grandeur into
petty deprecation of Pompey (314–40) and complains about his lost gains
(340–42). Caesar’s intertextual slander of Pompey as no more than a savage
animal contributes to this shift in register and the impression that Caesar’s
nobler opening and closing sentiments are just a veneer.
bc 1 and the aeneid
Having explored two large-scale views of the intertextual relationship between
BC 1 and the Aeneid, we can now ask what these perspectives add to our
understanding of Lucan’s epic. In terms of the artistic relationship between
the two poets, Lucan’s consistent use of Vergilian language pays tacit tribute
to his predecessor, and his careful deployment of Aeneid intertexts in section
transitions attests to the power he saw in the use of Vergilian precedent. In
thematic terms, Lucan’s references to the Aeneid have generally been taken as
oppositional: he subverts passages from the Aeneid in order to suggest that
the construction of empire, whose costs and benefits Vergil carefully balances,
was inevitably corrupt. An uncontrollable thirst for power destroyed pristine
Roman values of loyalty to family, community, and country.54
A large-scale view allows us to see how this oppositional tenor is articulated
in BC 1 through local strategies. Beyond amplifying and trumping the open-
ing of the Aeneid in the first book of his own epic, Lucan creates a series of
transformative identifications. Civil war recapitulates the destruction of Troy
(Aen. 2). Prophecies of progress (Aen. 3) turn to forebodings of catastrophe.
The madness and panic of Carthage (Aen. 4) spreads to all of Rome. Crimes
worthy of the underworld (Aen. 6) appear on earth. Fate as a guiding force
54 Syndikus 1958: 89–90 characterizes BC as a “Gegenbild” of the Aeneid. Roche 2009
gives an overview of this line of thought, writing of the “ideological opposition to the
surface narrative of the Aeneid” found in BC. The most recent treatment is Casali 2011.

414 Neil Coffee et al.
in war (Aen. 7) is shown to be hollow.55 The purpose of military honor, even
as hesitantly posed by Vergil (Aen. 11), is questionable.56 Caesar is assimilated
to Turnus while Pompey appears weaker than Aeneas (Aen. 12).57 These are
just some of the ways in which Lucan refers to individual books of the Aeneid
to demonstrate that civil war is the latest and most devastating repetition of
previous mythical struggles, and that Vergil’s honorable mythology of Rome’s
foundation is but a shadow of the truth.
Digital means are not, of course, the only path to such large-scale under-
standings of intertextuality. We have given but a miniature version of the vast
comparisons carried out manually by Knauer and Nelis. The point is rather
that digital detection can accelerate large-scale analysis considerably, and that
digital manipulation of results allows for multiple ways to view intertextual
relationships and develop meaningful generalizations about them.
theoretical consequences and the future of
intertextuality
Among the most significant consequences of the development of digital
search is the potential for more precise definitions of intertextuality. Consider
Table 3, which lists numbers of type 5–4 parallels identified in our combined
tests (both Tesserae searches and all commentators) grouped by features.
Tesserae looked only for parallels containing at least two similar words, those
in the first category above. But every parallel identified by Tesserae or com-
mentators could have been found by a search sensitive to the seven feature
categories listed. Two-word similarity was sufficient to identify 67% of the
parallels identified by all sources. A search for single identical word in a highly
related context would have found 12%, a search for a single identical word
and synonym, 7%. A search for similar thematic material, without words in
common, would have found 6%, a search for paired synonyms another 6%,
and so forth. Of these features, one is already subject to automatic detection
55 BC 1.34 inuenere uiam ≈ Aen. 7.297 inuenere uiam; 1.635 di uisa secundent ≈ 7.259
di nostra ... secundent.
56 The cowardice of Drances becomes the greater concern of fields left vacant by war
(BC 1.29 desuntque manus poscentibus aruis ≈ Aen. 11.379 bella manus poscunt); the
narrator asks why Rome rushed to war, in contrast with the Latin leader Tarchon who
asks why his troops will not fight (1.8 quis furor, o ciues, quae tanta licentia ferri? ≈ Aen.
11.732–32 quis metus, o numquam dolituri, o semper inertes / Tyrrheni, quae tanta animis
ignauia uenit?).
57 Caesar-Turnus: BC 1.205–12 leo ... ///... erexitque iubam ... /... infremuit ... lancea ≈
12.4–9 leo ... comantis / excutiens ceruice toros ... / ... telum et fremit. Pompey weaker than
Aeneas: 1.134 nouas uires ≈ 12.424 nouae ... uires.

415
Neil Coffee et al. Intertextuality in the Digital Age
with the Tesserae tool (word identity). Linguistic techniques for identifying
semantically related words and word clusters could be employed to identify
synonyms and antonyms and to identify thematic contexts.58 Methods for
syntax matching have been demonstrated in intertextual analysis,59 and sound
matching techniques are in development.60
Full sensitivity to these seven features and feature combinations thus holds
the potential to capture all of the interpretable intertexts found by commenta-
tors in our sample and others in addition. To build this search configuration
into a full algorithm for intertextual analysis would require the addition of
at least two other capacities, however. One is the ability to sort parallels into
58 Latent semantic analysis, a technique developed by linguists to find words with associ -
ated meanings, is increasingly being used by humanists to search for semantically related
words and repeated themes. See Underwood 2011. Topic modeling is a related semantic
detection technique that identifies themes in the form of clusters of co-occurring words.
Blevins 2010, e.g., uses topic modeling to identify themes such as midwifery, church,
death, gardening, shopping, and illness in the diary of Martha Ballard, a midwife living
in late 18th- to early 19th-century Maine.
59 Bamman and Crane 2008.
60 Forstall and Scheirer 2010.
Table 3. Minimal Features of Type 5-4 BC 1–Aeneid Parallels.
Features Count % of total
1. Two-word (exact word or lemma) identity 146 67%
2. One identical word + semantic context 27 12%
3. One identical word + synonym 16 7%
4. Semantic context only 13 6%
5. Synonyms only 12 6%
6. One identical word + syntax 2 1%
7. One identical word + sound 2 1%
TOTAL 218 100%

416 Neil Coffee et al.
types. The table above accounts only for our type 5–4 parallels, which we
selected manually from the total Tesserae and commentator results. We are
currently exploring whether word frequency, phrase frequency, and other
feature matching may be helpful in creating an automatic ranking system.
Another necessary expansion will be the development of cross-linguistic
matching. For the study of Latin intertextuality, a Greek-Latin matching al-
gorithm should at a minimum recognize parallelism such as Homer Od. 1.1
ἄνδρα (“man”) ≈ Aen. 1.1 uirum (“man”). Such similarity would seem to be
detectable: cross-linguistic semantic and morphological (for case) matching
along with sensitivity to section boundaries would rate these words highly as
a potential parallel.61 As such methods are developed, they can be tested on
an expanding range of primary texts and their results compared with those
from the relevant secondary sources.
Once an algorithm has reached a satisfactory level of performance in
sample tests in terms of recall (capturing known parallels) and specificity
(sorting high and low interest parallels), it can be set to work on comparisons
of many individual texts with some confidence in the results it will return.
The full set of comparisons would then provide a baseline against which to
measure the intertextual activity of genres, authors, works, and work sections
synchronically or over time. We might thus think of the table above as a first
version of an intertextual “fingerprint” that in the future can be compared
with fingerprints for other authors and texts to arrive at a better idea of how
each author and text are distinctive and of overall intertextual practice.
This program of intertextual study presents us with a theoretical question:
if an algorithm identifies substantially all intertexts found by critics, does it
constitute a definition of intertextuality? We would answer yes, provided that
we distinguish between a definition and a description. The sort of detection
algorithm we have outlined cannot describe the whole phenomenon of in-
tertextuality any more than a recipe for apple pie describes the experience of
baking and eating one. As discussions of Latin intertextuality from Pasquali
onward have shown, intertextual phenomena show a range of subtlety in the
connections they establish and meanings they form. Even if search algorithms
proceed further into the detection of Alexandrian footnotes, window refer-
ences, structural similarities, and other complex phenomena, the definitional
truth they represent will always need to be complemented by the experiential
truth of the reader to form a fully satisfying description.
61 Work on cross-linguistic semantic matching has included alignment of translations
with target texts to find translation equivalents. See Bamman, Babeu, and Crane 2010.

417
Neil Coffee et al. Intertextuality in the Digital Age
Nevertheless, a definition can be a useful complement to such description.
A definition identifies only the minimal features necessary to distinguish one
thing from another. The most minimal definition of intertextuality might be
“similarity between any texts.” To make such a definition more useful, however,
we must make it more specific. Here we can turn for guidance to the intuitions
of scholars well versed in intertextual artistry. In a discussion of the problem
of authorial intentionality, Joseph Farrell appeals to the statistical unlikelihood
of the random production of two very similar lines,
Ἄλκανδρόν θ᾽ Ἅλιόν τε Νοήμονά τε Πρύτανίν τε (Il. 5.678), and
Alcandrumque Haliumque Noëmonaque Prytanimque (Aen. 9.767),
and makes this unlikelihood the ultimate justification for declaring the lines
related:
One does have to be a qualified reader to grasp the fact that Vergil’s line is inter-
textual with Homer. But once this relationship is grasped, one may ask whether
the similarity exists only in the mind of the reader and not at all in the text or
in the mind of the poet. I am not a statistician, but I think we can assume a
very low mathematical probability that any line in any Latin poem would, just
by chance, reproduce so accurately any line from any Greek poem. That is to
say, the exactitude of this similarity, and the unlikelihood that such a similarity
would occur by chance—neither of which factors is merely impressionistic—are
both so great as to make the line itself an unmistakable intertextual marker.62
Farrell is arguing for the improbability of Vergil unintentionally producing
such a close echo of Homer. To substantiate this argument, he needs to make
the case that Vergil’s line obviously is intertextual with the line of Homer, and
would have seemed as obviously so to Vergil as it does to us. When he looks
for the strongest argument that this is in fact an intertext, Farrell turns to
formal and quantitative features. That is, to produce the soundest definition
of intertextuality, Farrell turns to an algorithm. In his case, it is a hypothetical
one that accounts for the probability of a semantic and sound convergence
between lines of Homer and Vergil.
If we pursue Farrell’s reasoning further, we can make this definition yet
more precise. His argument suggests a notional spectrum, where parallels
with rare configurations of similarity are placed at one end and common
co-occurrences at the other. We could then place a line on the spectrum
separating intertexts from non-intertexts. In its fullest and most satisfying
form, the “intertext” side of the spectrum would consist of a composite of all
62 Farrell 2005: 100–1.

418 Neil Coffee et al.
combinations of features—the semantics and sound that Farrell addresses,
along with others—that qualified a proposed parallel as an intertext. Better
still, rather than dividing the spectrum into two parts, we could segment it
into types, where Farrell’s Homer-Vergil parallel would appear among inter-
pretable intertexts at one end. The full, segmented spectrum, or an algorithm
that produced it, could then be understood as an empirical definition of
intertextuality, much as animal species are defined by a set of morphological
and behavioral features.
As described, such a definition would largely replicate the criteria critics
have traditionally used to describe intertexts. The process of digital discovery
can also expand our definition, however. To a degree, this is already happen-
ing. Our comparison of commentators showed that those who use digital
detection methods find more instances of pure language reuse (type 3). The
collective scholarly understanding of intertextuality is thus shifting to include
phenomena considerably less salient than the most celebrated examples.
Digital detection may also bring to prominence other features that have passed
beneath the notice of critics, but which have nevertheless played a role in
the poet’s production and reader’s experience. When we listen to music, we
respond not only to melody and rhythm, but also to more obscure elements
such as structure, overtones, and established expectations. We have argued
for a similar effect when Lucan silences the otherwise steady drone of fore-
ground (type 5–4) and background (type 3) intertexts with the Aeneid. Such
arguments could be made for sound patterns and other features.63 To be sure,
digital detection will contribute to forming not one definition of intertextual-
ity, but many. Substituting other features for the seven we identified might
create another definition that might be equally comprehensive. Our results
were dependent upon a particular set of commentators and sample texts.
Other combinations might produce definitions with other features. It is not
unreasonable to expect, however, that from such investigations an algorithm
accounting for substantially all levels of intertextuality could emerge.64
63 T.S. Eliot writes “I cannot help suspecting that to the cultivated audience of the
age of Virgil, part of the pleasure of poetry arose from the presence in it of two metrical
schemes in a kind of counterpoint: even though the audience may not have been able to
analyse the experience” (Eliot 1942: 12, cited by Wilkinson 1970: 120).
64 Lausberg 1985: 1616–17 argues that Homer and Vergil had such weight in the epic
tradition that no lexical similarity is necessary to invoke comparison, an observation ap-
proved by Henderson 2010 [1987]: 479. Whether the precise levels of similarity Lausberg
refers to are automatically detectable, or would be defined by most critics as intertextuality,
requires further exploration.

419
Neil Coffee et al. Intertextuality in the Digital Age
By offering a definition of intertextuality, we have perforce entered into the
theoretical discussions of recent years over the status of intertexts.65 In these
discussions, three main alternatives have been on offer: intertexts are created
by an author; they are features of texts; or they are constituted by an observer
connecting two texts while reading or listening. Digital detection and analysis
do not change the terms of this debate. But they can bring an increased level
of precision to the positions.
A search algorithm might be construed as simply an extension of the
reader, in the form of a systematic projection of a set of reader preferences and
expectations. From this vantage point, our efforts can be seen as an attempt
to reverse-engineer an intertextual detection algorithm latent in the mind
of the critic. However we might conceive of this operation, making readerly
criteria clear for all to see allows them to be compared, contested, and refined.
Alternatively, digital search may be the ultimate means of focusing on the
text alone. The fact that our algorithm identifies unobserved instances and
minimally-recognized features of intertextuality suggests that it removes at
least one layer of subjectivity by presenting to the critic a set of examples that
all fit consistent criteria. Finally, although we cannot recover the author’s inten-
tion, developing a stylistic fingerprint for each author can fill in the picture of
an artist’s preferences and working methods. Critics have long observed that
Ovid employs dactyls in the Metamorphoses more often than Vergil does in
the Aeneid. Reading this difference together with the content of the poems,
we conclude that Ovid’s swifter meter contributes to a lighter narrative style
and tone.66 If we are willing to accept a certain amount of distant reading,
digital analysis can provide an analogous and complementary account of
each poet’s intertextual practice. When intertextuality, and potentially other
complex poetic features, can be thus modeled, we can build a more satisfying
profile of the artistry of each poet and richer interpretations of their works.
works cited
Bamman, D., Babeu, A., and Crane, G. 2010. “Transferring Structural Markup Across
Translations Using Multilingual Alignment and Projection.” Proceedings of the 10th
ACM/IEEE-CS Joint Conference on Digital Libraries (JCDL 2010) 2010–6: 11–20.
Available at: http://www.perseus.tufts.edu/hopper/about/publications.
65 Farrell 2005 provides a useful entry point.
66 54.6% of Ovid’s first four hexametrical feet are dactyls, to 44.1% in the Aeneid
(Wilkinson 1970: 131, with further references). Wilkinson writes that Ovid’s “fantastical
stories could only bear the lightest touch.”

420 Neil Coffee et al.
—— and Crane, G. 2008. “The Logic and Discovery of Textual Allusion.” Proceedings
of the Second Workshop on Language Technology for Cultural Heritage Data (LaTeCH
2008) Marrakesh. Available at: http://www.perseus.tufts.edu/~ababeu/latech2008.pdf.
Barchiesi, A. 1984. La traccia del modello: effetti omerici nella narrazione virgiliana. Pisa:
Giardini.
Blevins, C. 2010. “Topic Modeling Martha Ballard’s Diary.” historying, April 1, 2010.
Available at: http://historying.org/2010/04/01/topic-modeling-martha-ballards-diary/.
Date last accessed: February 26, 2012.
Büchler, M. et al. 2010. “Unsupervised Detection and Visualisation of Textual Reuse on
Ancient Greek Texts.” Journal of the Chicago Colloquium on Digital Humanities and
Computer Science, North America, June 1, 2010. Available at: https://letterpress.uchicago.
edu/index.php/jdhcs/article/view/60/71. Date last accessed: May 9, 2012.
Casali, S. 2011. “The Bellum Ciuile as Anti-Aeneid.” In Asso, P. ed. Brill’s Companion to
Lucan. Leiden: Brill. 81–109.
Conte, G. B. 1986. The Rhetoric of Imitation: Genre and Poetic Memory in Virgil and Other
Latin Poets. Ithaca: Cornell University Press.
D’Ippolito, G. 2000. “Il concetto di intertestualita nel pensiero degli antichi.” Bécares Botas,
V. et al. eds. Intertextualidad en las literaturas griega y latina. Classica Salmanticensia
2. Madrid: Ed. Clásicas. 13–32.
De La Cerda, J. L. 1617. P. Virgilii Maronis libri Aeneidos, argumentis, explicationibus, notis
illustrati. Lyon: Sumptibus Horatii Cardon.
Edmunds, L. 2001. Intertextuality and the Reading of Roman Poetry. Baltimore: Johns
Hopkins University Press.
Eliot, T. S. 1942. The Music of Poetry. Glasgow: Jackson, Son & Company.
Fantuzzi, M. and Hunter, R. 2004. Tradition and Innovation in Hellenistic Poetry.
Cambridge: Cambridge University Press.
Farrell, J. 1991. Vergil’s Georgics and the Traditions of Ancient Epic: The Art of Allusion in
Literary History. New York: Oxford University Press.
—— 2005. “Intention and Intertext.” Phoenix 59: 98–111.
Forstall, C. and Scheirer, W. 2010. “Features From Frequency: Authorship and Stylistic
Analysis Using Repetitive Sound.” Journal of the Chicago Colloquium on Digital
Humanities and Computer Science, North America, June 1, 2010. Available at: https://let-
terpress.uchicago.edu/index.php/jdhcs/article/view/56. Date last accessed: May 9, 2012.
Fowler, D. 2000 [1997]. “On the Shoulders of Giants: Intertextuality and Classical Studies.”
In id. Roman Constructions: Readings in Postmodern Latin. Oxford, Oxford University
Press. 115–37. Originally published in MD 39 (1997): 13–34.
Giangrande, G. 1967. “‘Arte Allusiva’ and Alexandrian Poetry.” CQ 17: 85–97.
Heitland, W. E. and Haskins, C. E. 1887. M. Annaei Lucani Pharsalia. London: Bell.
Henderson, J. 2010 [1987]. “Lucan/The Word at War.” In Tesoriero, C., Muecke, F., and
Neal, T. eds. Lucan. Oxford: Oxford University Press. 433–92. Originally published
in Ramus 16 (1987): 122–64.
Hinds, S. 1998. Allusion and Intertext: Dynamics of Appropriation in Roman Poetry.
Cambridge: Cambridge University Press.
—— 2007. “Martial’s Ovid/Ovid’s Martial.” JRS 97: 113–54.
Horsfall, N. 2000. Vergil, Aeneid 7: A Commentary. Leiden: Brill.

421
Neil Coffee et al. Intertextuality in the Digital Age
Horton, R., Olsen, M., and Roe, G. 2010. “Something Borrowed: Sequence Alignment
and the Identification of Similar Passages in Large Text Collections.” Digital Studies/
Le champ numérique 2.1. Available at: http://www.digitalstudies.org/ojs/index.php/
digital_studies/article/view/190/235. Date last accessed: May 7, 2012.
Kao, J. and Jurafsky, D. 2012. “A Computational Analysis of Style, Affect, and Imagery in
Contemporary Poetry.” NAACL Workshop on Computational Linguistics for Literature.
Available at: http://www.stanford.edu/~jurafsky/kaojurafsky12.pdf. Date last accessed:
May 26, 2012.
Knauer, G. N. 1964. Die Aeneis und Homer: Studien zur poetischen Technik Vergils, mit
Listen der Homerzitate in der Aeneis. Göttingen: Vandenhoeck & Ruprecht.
Lausberg, M. 1985. “Lucan and Homer.” ANRW 2.32.3: 1565–622.
Levene, D. S. 2010. Livy on the Hannibalic War. Oxford: Oxford University Press.
Martindale, C. 1993. Redeeming the Text: Latin Poetry and the Hermeneutics of Reception.
Cambridge: Cambridge University Press.
Moretti, F. 2005. Graphs, Maps, Trees: Abstract Models for a Literary History. London: Verso.
Nelis, D. 2001. Vergil’s Aeneid and the Argonautica of Apollonius Rhodius. Leeds: Francis
Cairns.
O’Hara, J. J. 1990. Death and the Optimistic Prophecy in Vergil’s Aeneid. Princeton: Princeton
University Press.
Pasquali, G. 1968. “Arte allusiva.” In id. Pagine Straviganti. Florence: Biblioteca Sansoni.
275–83.
Pucci, J. 1998. The Full-Knowing Reader: Allusion and the Power of the Reader in the Western
Literary Tradition. New Haven: Yale University Press.
Roche, P. 2009. Lucan: De bello civili Book 1. Oxford: Oxford University Press.
Sampson, G. 2003. “Statistical Linguistics.” In Frawley, W. J. ed. International Encyclopedia
of Linguistics. 4 vols. 2nd ed. Oxford: Oxford University Press. 4.162–65.
Saylor, C. 1978. “Belli Spes Improba: The Theme of Walls in Lucan’s Pharsalia VI.” TAPA
108: 243–57.
Schepens, G. and Bollansée, J. eds. 2005. The Shadow of Polybius: Intertextuality as a
Research Tool in Greek Historiography. Proceedings of the International Colloquium,
Leuven, 21–22 September 2001. Studia Hellenistica 42. Louvain: Peeters.
Swift, L. 2010. The Hidden Chorus: Echoes of Genre in Tragic Lyric. Oxford: Oxford
University Press.
Syndikus, H. P. 1958. “Lucans Gedicht vom Bürgerkrieg: Untersuchungen zur epischen
Technik und zu den Grundlagen des Werkes.” Diss. Munich.
Telo, M. 2010. “Embodying the Tragic Father(s): Autobiography and Intertextuality in
Aristophanes.” ClAnt 29: 278–326.
Thomas, R. F. 1986. “Virgil’s Georgics and the Art of Reference.” HSCP 90: 171–98 (repr. in
id. 1999. Reading Virgil and His Texts: Studies in Intertextuality. Ann Arbor: University
of Michigan Press. 114–41).
Thompson, L. and Bruère, R. T. 1968. “Lucan’s Use of Vergilian Reminiscence.” CP 63: 1–21.
Underwood, T. 2011. “LSA is a Marvellous Tool, but Literary Historians May Want to
Customize it for their own Discipline.” The Stone and the Shell, Oct. 16, 2010. Available
at: http://tedunderwood.wordpress.com/2011/10/16/lsa-is-a-marvellous-tool-but-
humanists-may-no-use-it-the-way-computer-scientists-do/. Date last accessed:
February 26, 2012.

422 Neil Coffee et al.
Viansino, G. 1995. Marco Anneo Lucano: La Guerra Civile (Farsaglia). 2 Vols. Milan:
Mondadori.
Wilkinson, L. P. 1970. Golden Latin Artistry. Cambridge: Cambridge University Press.
Wills, J. 1996. Repetition in Latin Poetry: Figures of Allusion. Oxford: Clarendon Press.
Zipf, G. K. 1949. Human Behavior and the Principle of Least Effort: An Introduction to
Human Ecology. Cambridge, MA: Addison-Wesley Press.