College of the Holy Cross
Overview
The Venetus A manuscript of the Iliad (Marciana 454 = 822) is uniquely important for Homerists. [1] The scholarly notes (or scholia) that often fill the margins of the manuscript’s pages cite scholars as early as the first Alexandrian editors of the Iliad, and in some cases quote them for readings of the Iliad not known elsewhere in the manuscript tradition. Between 2010 and 2018, the main focus of the Homer Multitext project (HMT) was to prepare digital editions of the texts in the Venetus A manuscript. In this paper, we first outline the design of our digital editions (Section 1), then show how this architecture supports new kinds of research. Specifically, we apply new methods to two classic problems.
- How should we interpret the physical organization of this uniquely rich manuscript? Can we identify the function of the carefully planned, distinct zones of the page layout? (Section 2)
- Can we find evidence for the work of antiquity’s most influential editor, Aristarchus of Samothrace? (Section 3)
These questions offer complementary perspectives on the transmission of ancient scholarship on the Iliad to us. The first perspective is synchronic. Using topic modeling, we show that the zones of the page layout are characterized by distinct patterns of language. The scribe of the Venetus A, in other words, planned the page layout to accommodate different kinds of material. This is most easily explained as material deriving from distinct sources, and so suggests that the tenth-century scribe is creating a unique manuscript drawing on multiple sources no longer extant.
The second perspective is diachronic. We construct a machine-learning model to identify verbatim language of Aristarchus and show how this model can find material preserving Aristarchus’ own words in any zone of the manuscript. This shows us that Aristarchus’ words reached the scribe of the Venetus A via different sources, and therefore appear, in different proportions, in multiple zones of the Venetus A.
Section 1
The Venetus A scholia
Since the discovery of the Venetus A manuscript by Villoison [2] in 1788 and the extended discussion of its scholia in Wolf’s seminal Prolegomena ad Homerum of 1795, the manuscript’s scholia have been central to scholarship on “the Homeric question.” It is all the more remarkable that they have never been fully published.
The Homer Multitext project is filling this gap. The work we describe here was completed in 2017 and analyzes eighteen books from the Homer Multitext edition of the Venetus A, comprising roughly 8,000 scholia. [3] Because each scholion is indexed to visual evidence and its location on a page, our analysis can combine the textual content of the scholia with information about their layout, and with other features such as the editorial signs associated with more than 10% of the Iliad lines. Because our edition is digital, we can systematically and comprehensively apply our analyses to the HMT editions. Prior to the HMT, the fullest edition was that of Dindorf, [4] but, as we discuss below, he does not distinguish scholia placed in the distinct intermarginal and interior zones of the page. The most recent edition of the scholia is by Erbse, [5] which is also the basis for the digital text included in the datasets of Northwestern University’s WordHoard software. [6] Erbse’s edition omits 20% of the content in the Venetus A, including only those scholia he classifies as “vetera.” [7] It is therefore impossible to draw a non-biased sample of data from this digital data set. By contrast, since the eighteen books we analyze here include every word, symbol or other mark seen by HMT editors, we were confident that the broad outlines of our conclusions would be valid for all of the Venetus A. Subsequent work has to date validated this assumption, [8] and our conclusions can be tested against the publicly available editions of the HMT.
When is a digital document a digital edition?
Within the Homer Multitext project, the experience of editing the Venetus A manuscript has prompted extended reflection and debate about what it means to edit a text. What changes when an edition is digital? All of us are excited about emerging possibilities of digital reading at multiple scales: not just systematic searching, but “reading” to identify latent patterns we don’t even know to search for. How do we create digital editions that can support a full range of digital reading, including digital methods we have not yet imagined?
We take as a point of departure the central idea of the Greek term ἔκδωσις, and its calques in various modern languages (editio, edition, Ausgabe…). An edition is first and foremost giving away a version of a text. Using analog technologies, giving away a physical artifact (whether printed book, vellum codex or papyrus scroll) makes available to the new owner every possibility of analog reading. To support every possibility of digital reading available when we give away a digital edition, we need to plan for three kinds of access. A digital edition must by legally, technically and semantically accessible.
Legal accessibility was a well understood problem by the time of our work 2017, and well-defined solutions are readily available. A digital edition must be licensed for reuse so that automated reading of the edition can proceed without manual intervention to seek special or exceptional permissions. The Homer Multitext project uses Creative Commons licenses permitting reuse for all of its digital content.
Technical accessibility has also been extensively studied. To ensure that our editions remain technically accessible they must be in openly specified, plain-text formats. The ideal is the simplest possible format that is also expressively complete. Like most digital editorial projects today, the Homer Multitext uses XML to produce semantically encoded editions following the guidelines of the Text Encoding Initiative (TEI). [9]
If the scholarly community has well defined solutions to these problems, however, we lack any comparable consensus on how to make an edition semantically accessible. It may once have been a goal of the TEI to facilitate interchange of texts with common semantics, but the reality today is that using XML following TEI guidelines guarantees that you can not simply exchange texts with other projects. The guidelines are rich, and frequently recommend multiple possible solutions to an encoding problem; even when the editor intends to follow a single recommendation, the generic schemas for validating TEI following the guidelines allow multiple possible embeddings and sequences of elements. [10] Alternative readings are a further complication: every form of TEI choice element presents an ambiguity that a digital reader has to resolve.
More generally, XML is limited to validating a given syntax; real semantic accessibility would require that we validate contents as well. Like many projects, therefore, the HMT has defined not only its own project-specific guidelines for using a subset of the TEI guidelines, but also editorial standards for encoding the contents of the documents.
In 2014 and 2015, we began to go beyond simply giving editors a document with prose guidelines. We identified abstract features of our editions (for example, “editions must distinguish numeric tokens from alphabetic tokens,” or “personal names are disambiguated with unique identifiers”). We specified an XML syntax for representing each feature (e.g., “numeric tokens are tagged with TEI num,” or “the n attribute of the TEI persName should have a valid URN value from the HMT collection of personal name identifiers”). At the same time, we implemented each feature in an object model that could parse XML and evaluate whether it complies with our specifications. At the time of our work in 2017, we worked with an implementation of the full model of HMT editions written in Scala. [11] In 2020 and 2021, we migrated to an interactive system using Pluto notebooks, and to support this have implemented the model in Julia, the language used to write Pluto notebooks. [12]
Our initial motivation was to create an editors’ manual specifying syntax and restrictions on contents (such as allowed character encodings, and controlled vocabularies for identifiers), and a parallel implementation in code so that all contributors could validate their work in progress against a single, machine-enforced standard. [13] In 2016–2017, the co-authors began to leverage the same code library to generate multiple, semantically unambiguous editions from a single, rich XML source. TEI XML documents validated against our project standards serve as archival editions, from which we create plain-text editions without any XML markup. We can, for example, generate pure diplomatic plain-text editions, or plain-text editions reflecting scribal corrections, or editions normalized against modern orthographic practices. Because all of our plain-text editions are derived from a single, canonically citable source, they can be automatically aligned. We can identify exactly which token in one edition corresponds to a given token in another edition.
At this point we are no longer limited to producing editions for immediate reading by humans: reading is not just for people anymore. We can create editions for specific purposes, including specific kinds of automated reading.
Section 2
Analyzing layout and content in the Venetus A
Figure 1 illustrates folio twelve recto of the Venetus A manuscript, the first page to contain text from the Iliad. The text of the Iliad is in the middle of the page written in a larger hand and darker ink. The colored overlays highlight distinct zones of scholia found on a typical page: main scholia (shown in yellow), intermarginal scholia (red), interior scholia (blue), and exterior scholia (green). A fifth zone is made up of the interlinear scholia (not highlighted with a color here, but visible as the notes in a smaller script in between lines of the Iliad text). While not every folio of the Venetus A contains all five types of scholia, every folio maintains the spatial integrity of the five zones: if there are no exterior scholia on a particular folio, for example, the scribe will respect the space allowed for exterior scholia, and will not expand other scholia into this zone.
While the difference in placement on the page for these five distinct zones of Venetus A scholia is obvious to any casual reader of the manuscript, the difference in function among these zones is much less evident. Why should a scribe place a scholion in any particular zone as opposed to another?
Nineteenth-century scholars debated this question extensively. Karl Dindorf, in the preface to his 1875 edition of the Venetus A scholia, argues that the interlinear zone presents no more than glosses on the words which lie below them (Dindorf 1875, ix), and indeed this does seem to be their primary function. As for the exterior scholia, T. W. Allen [14] argued that these scholia are the result of an editorial pass through the manuscript (181). For Allen, the main, intermarginal, and interior scholia were all written by the same scribe, but the exterior scholia were undoubtedly the work of a different scribe checking the work of the previous scribe against the original sources for the manuscript, and making corrections when necessary in the outer margin (172).
Here we focus on the much less clear relationship among the other three zones of scholia: the main, intermarginal, and interior scholia. Dindorf observed that many of the intermarginal and interior scholia present the same or nearly the same content as the main scholia, and sometimes even use exactly the same wording. To him, these shorter intermarginal and interior scholia appeared to be nothing more than redacted versions of the main scholia (Dindorf 1875, vii). From these observations, Dindorf concluded that the intermarginal, interior, and main scholia must have come from similar, though still subtly different, sources (vii). However, it is worth noting that when Dindorf uses the phrase “intermarginalia scholia” in his preface, he is actually referring to what the Homer Multitext would consider the distinct intermarginal and interior zones (Dindorf, vii). While the precise wording of Dindorf’s definition of the “intermarginalia scholia” in his preface is potentially confusing, he explains that his edition marks all the scholia which he refers to as the intermarginalia with an asterisk so as to avoid confusing them with the main scholia (Dindorf, ix). It is clear from his edition that scholia which the HMT would call interior are marked with an asterisk alongside those which the HMT would also call intermarginal. Dindorf was not alone in this treatment of these two zones of the scholia: T. W. Allen similarly uses the single term “intermarginal” to refer to the two physically distinct zones (181). Thus, for Dindorf and Allen, there was no meaningful difference between what we call “interior” and “intermarginal” scholia; they were considered functionally equivalent.
We used topic modeling to explore the content of scholia in these three zones, and to highlight ways the content of the various zones might differ quantitatively. Specifically, we used the R-language implementation of the Latent Dirichlet Allocation (LDA) algorithm in Thomas Köntges’ ToPan. [15] Each scholion was treated as a distinct document.
Topic modeling requires a carefully prepared text. We derived a series of plain-text editions from our archival TEI source until we had a citable edition that was optimized for topic modeling.
We began from a reading of the archival text that expanded abbreviations to their full form. From this, we derived a second edition that normalized Byzantine orthography to modern orthographic standards, and resolved named entities to unique identifiers. (Each of the many scholars named “Ptolemy” could then be distinguished, for example.) This edition formed the basis of a lemmatized edition. We used the Morpheus parsing system from the Perseus project to “edit” our text, so that all tokens were presented as identifiers for lemmatized lexical entities. Because our citation scheme was aligned across all these editions, we could readily compare surface forms in modern or Byzantine orthography with the lemmatized tokens.
Finally, we created a version suitable for topic modeling by curating a list of stop words, and removing those tokens from the orthographically normalized edition.
Because topic modeling requires users to define the number of topics the program should produce, we manually analyzed the results of experiments with differing number of topics. Generating too many topics risks not observing meaningful patterns since they may be arbitrarily split up into more topics than necessary. On the contrary, producing too few topics can render topics meaningless by lumping together topics that we would like to distinguish. Ultimately, we found that fifteen was a good compromise for our analysis.
One of the advantages of ToPan is that it produces a clear visualization of the topic modeling results, as illustrated in Figure 2.
In this visualization, each of the circles on the left represents a different topic; when a topic is highlighted, the words which are most strongly associated with the topic are listed on the right. (For clarity, we have supplied English translations of each of the most strongly associated words in Figure 2.) Though the topic modeling results actually exist in a multidimensional space, they are reduced in this visualization into a two-dimensional plane using principal components analysis. The distance between circles reflects how closely related the topics are.
For the purposes of this paper, we will illustrate one important result with the example of a single topic. [16] Recall that a “topic” in a topic model is defined as a repeated pattern of co-occurring words. In the topic illustrated in Figure 2, topic 9, we see that the names of Aristarchus of Samothrace, Zenodotus of Ephesus and Aristophanes of Byzantium are prominent. All were head of the library of Alexandria at some point in the third and second centuries B.C.E., and all three produced editions of the Iliad known only through references such as the scholia to the Venetus A. Looking at the most significant terms in this topic and reading specific scholia classified as predominantly composed of topic 9, we interpreted this recurring pattern as the “Aristarchus topic,” where the editorial work of Aristarchus is discussed and often compared to his predecessors.
It would be interesting and telling if this topic were to appear with a greater frequency in any of the three zones of interest (main, intermarginal, and interior). While the visualization in ToPan shows relations of different word patterns, the visualization alone does not give information about what scholia are strongly or weakly associated with each of the topics. However, ToPan does offer robust data tables providing exactly this information. Using these tables, we were able to analyze how strongly the Aristarchus topic is associated with each scholion since this relationship is scored quantitatively by the theta-score, scaled from 0 to 1. Were a scholion to have a theta-score of 0.0 with respect to a particular topic, this would indicate that this particular scholion contains no content from that particular topic. Conversely, a theta-score of 1.0 would indicate that the content of this scholion would likely come from this topic alone.
Our goal was to move beyond analyzing individual scholia to analyzing scholia by zone. Using the theta-tables, we identified the proportion of scholia within each zone (main, intermarginal, and interior) which were strongly associated with the Aristarchus topic. We defined a “strong association” as a scholion having a theta-score equal to or greater than 0.9. The following table summarizes the results.
Table 1: Distribution of “Aristarchus” topic by zone compared to all scholia
| Zone | Total distribution in Venetus A | Distribution within “Aristarchus topic” |
| Main scholia (yellow) | 3597 (45.3%) | 45 (32.37%) |
| Intermarginal scholia (red) | 1219 (15.4%) | 65 (46.76%) |
| Interior scholia (blue) | 819 (10.3%) | 28 (20.14%) |
The second column presents the percentage of our corpus of eighteen Venetus A books represented by each zone of scholia. The third column presents the percentage of the scholia that are strongly associated with the Aristarchus topic (the percentages do not add up to 100% because this table does not include the interlinear or exterior scholia that are not the focus of this analysis). We highlight the middle row: the intermarginal scholia make up only about 15% of all the scholia in the eighteen books of the Venetus A we analyzed, but about 47% of all scholia strongly associated with the Aristarchus topic are from the intermarginal zone. Not only does the intermarginal zone preserve a greater proportion of the Aristarchus scholia than the main zone (47% to 32%): the proportion is more than twice the proportion of intermarginal scholia within the manuscript overall (only 15% vs. 45%). The Aristarchus topic is disproportionately concentrated within the intermarginal scholia.
Since a “topic” is defined as words which co-occur frequently, this means that the particular pattern of co-occurring words identified by the topic model occurs more in one zone than another. There are other ways to talk about Aristarchus (his name does appear as a significant term in other topics), but this pattern in which Zenodotus, Aristophanes and discussion of editorial decisions all figure prominently along with Aristarchus is heavily concentrated in the intermarginal zone. At the same time, we can see that the main and interior scholia still constitute significant proportions of the scholia strongly associated with this topic (32.37% and 20.14%, respectively).
That points to a larger conclusion. The distribution of topics across zones of the scholia can vary greatly: the zones demonstrably differ from each other in the balance of topics they present. Some topics are more evenly distributed across zones, but the concentration of topics such as the Aristarchus topic shows that the layout of the manuscript’s pages predefines zones that are used for distinct patterns of language.
For most of these patterns, we can readily interpret the subject, just as we interpreted topic 9 to be about discussion of Aristarchus’ editorial activity. Viewed systematically, the topics that predominate in the interior and intermarginal zones cannot be simply reduced to summaries or alternate versions of material appearing in the main scholia.
To account for the concentration of some topics in particular zones while at the same time recognizing that no topic is entirely exclusive to a single zone, we believe that the simplest explanation is that the scribe intended the five zones for material from five distinct sources. There is some overlap in topics across zones because the sources used by the Venetus A scribe incorporate some common material that made its way from antiquity to the tenth century via different paths in different forms.
Section 3
Finding the language of Aristarchus
How influential were the Alexandrian editors in shaping the written Iliad that we have today? What role did the editions of Zenodotus, Aristophanes and Aristarchus play in narrowing down the multiformity of the earliest written versions of the Iliad to the relatively homogeneous text we find in the medieval manuscripts?
The scholia of the Venetus A offer a window on the editorial work of scholars from the Alexandrians through late antiquity. The manuscript itself provides very little information about the sources of its scholia, however. At the end of several books of the Iliad, a subscription cites four scholars of the first and second centuries C.E. However, the scholia themselves frequently refer to scholars from later periods, so the subscription does not provide us with a complete view of the sources lying behind the scholia.
With the digital edition of the HMT, can we pinpoint specific source material for the scholia with more precision than has been achieved previously? In particular, can we identify material stemming directly from the editions of Aristarchus? To address these questions, we constructed a machine-learning model for identifying scholia that can be directly associated with Aristarchus.
Defining a feature set
In contrast to the unsupervised classification of topic modeling, we build our classifier from a manually analyzed set of training data. A second set of manually analyzed data is then used to evaluate the performance of the model. The resulting model can be applied to arbitrarily large sets of data.
Our first step, then, was to identify features that mark scholia as “Aristarchan.” The Venetus A uniquely preserves roughly 2800 critical signs marking noteworthy lines of the Iliad, and indicating that a comment discusses some aspect of that line. In the Venetus A, these marks therefore call our attention to lines of main text discussed in scholia, but the system was invented by Aristarchus, whose edition of the Iliad and commentary on the Iliad were two separate documents. The critical signs were especially helpful for readers trying to coordinate two papyrus scrolls: the appearance of a sign in the edition would notify them to check the commentary. The Venetus A alone has preserved these Aristarchan signs, and while their presence on a line does not guarantee that the corresponding scholion is Aristarchan, it unambiguously means that Aristarchus’ commentaries originally discussed this line. [17]
Scholia on lines marked by Aristarchan critical signs are often introduced with the word ὅτι, “because” (transliterated “hoti”). This elliptical expression is short for “The [critical] sign is there because…” and indicates that the comment is meant to explain the presence of the critical sign. Because “hoti” occurs so frequently in scholia which are associated with a critical sign, it is likely that the comment introduced by “hoti” was transmitted alongside the critical signs and is therefore Aristarchus’ own way of explaining why he placed a critical sign next to a line in his edition. We therefore considered the presence of an initial “hoti” in a scholion an Aristarchan feature.
We looked at n-gram frequencies of the text of scholia and stumbled across an unexpectedly clear formula that is characteristic of scholia marked by a critical sign and introduced by “hoti.” Aristarchus’ edition responded directly to the edition of his predecessor, Zenodotus. While Zenodotus’ name appears frequently in the Venetus A, and not just in “hoti” scholia, we found that his edition is cited in two distinct ways in the Venetus A scholia. Outside of the “hoti” scholia, he is regularly cited with the preposition παρά (παρὰ Ζηνοδότῳ, “in [the work of] Zenodotus”), as are other editors and scholars. In the “hoti” scholia, however, he often appears as the subject of the verb γράφω (literally, “to write”) when the scholion cites his edition. (In English, we say that editors “read” a particular variant; in Greek, they “write” one version or another.) Remarkably, no other individual appears as the subject of the verb γράφω in the eighteen books of scholia we examined. This makes perfect sense if it was Aristarchus’ practice to cite editors with γράφω, and the practice of later scholars to cite editors with παρά. We therefore considered the expression “Ζηνόδοτος γράφει” (“Zenodotus writes…”) to be an indicator of directly Aristarchan language in a scholion.
For an automated classification, it is equally valuable to identify features which indicate that a scholion is likely not Aristarchan. Citing Zenodotus with the phrase “παρὰ Ζηνοδότῳ” is negatively correlated with the Aristarchan features we have defined: it occurs 27 times in the books we examined, but only five of those scholia were on lines marked with a critical sign. Reference to Zenodotus using this formula suggests a later source or later reworking of Aristarchus’ commentary, so we used the phrase “παρὰ Ζηνοδότῳ” to identify post-Aristarchan language.
The presence of Aristarchus’ name in a comment is also an indication that the language of the scholion is post-Aristarchan. This is based on the assumption that Aristarchus did not write about himself in the third person. The assumption is strengthened by the fact that Aristarchus’ name rarely appears in scholia on lines with critical signs. On the other hand, the names of Aristarchus’ predecessors appear much more frequently in the scholia which are connected with the critical signs, since Aristarchus often disagreed with them by name in these scholia. Thus, the presence of Aristarchus’ name in a scholion indicates some scholar at a later time was writing about him and his work on the Iliad.
Similarly, the presence of the name of a post-Aristarchan scholar obviously indicates that the language of the scholion is later than Aristarchus.
Each of these features can be treated as a binary value encoded as 0 or 1: either a scholion begins with “hoti” or it doesn’t, either it cites Zenodotus with γράφω or it doesn’t, etc. In addition, we realized that we could use results from the previously described topic modeling as features for our classifier. Two of the topics from our fifteen-topic model were relevant for picking out Aristarchan material. One was the topic 9 we have examined, where the editorial work of Aristarchus is the subject of discussion. Contrasting with this, topic 6 was defined by terms associated with Aristarchus’s own language, such as “Zenodotus,” “writes” and “athetize.” We interpreted topic 6 as “Aristarchus’ own comments.”
We used the theta scores for topics 6 and 9 as features in our data for each scholion. To reiterate, these scores give an indication of how closely the words in a scholion relate to the words of a given topic, and are scaled from 0.0–1.0, with a score closer to 0 meaning the scholion is not strongly correlated with the topic in question, and a score closer to 1 meaning that the scholion is very strongly correlated with the topic in question.
Topic 6 consists of words which are strongly associated with Aristarchus’ own language, so a high theta score for topic 6 suggests that the language of the scholion is strongly Aristarchan.
Since topic 9 consists of words used to discuss Aristarchus, his editions, and other editors, high theta scores for topic 9 indicate language from a post-Aristarchan source or reworking of his editions.
The final list of eight features included four that we expected to characterize directly Aristarchan material, and four that we expected to characterize later material:
Aristarchan features:
- Critical sign
- γράφει (active)
- Initial ὅτι
- Higher topic 6 score
Post-Aristarchan features:
- παρὰ Ζηνοδότῳ
- Aristarchus’ name
- Post-Aristarchan name
- Higher topic 9 score
Training and evaluating a model
Computing the eight features for each scholion once again creates a new, citable edition of the text. In this edition, the text of each scholion is simply eight successive numeric values from 0 to 1. To train and evaluate an automated classifier, we set aside a randomly selected sample of 100 scholia from the main, intermarginal, and interior zones.
Because our data set was automatically aligned with all our other editions of the scholia, we could then read the text of each of the 100 scholia in the test and assign it a class. It is important to recognize that this manual classification is based on our own, traditional close reading. Our arguments for deciding that a scholion is Aristarchan or not are the same kind of arguments that have been used in Quellenkritik of the scholia for more than two centuries. The question an automated classifier can address is how closely the eight features we selected can reproduce those interpretations.
We assigned each of the randomly selected scholia to one of four classes. In addition to our initial goal of classifying scholia as “Aristarchan,” we split non-Aristarchan scholia into “paraphrase of Aristarchus” and “post-Aristarchan” classes. We applied the “paraphrase class to scholia mentioning Aristarchus or his editions, but no other later editor, when we interpreted these scholia as reworkings in third-person language of originally Aristarchan material. We also recognized that many scholia are difficult or impossible to classify confidently. We classified these as “indeterminate.”
In our random sample of 100 scholia, the “indeterminate” class was the most frequent (38), followed closely by “Aristarchan” (36), with the “paraphrase” and “post-Aristarchan” classes about evenly matched.
We next split the manually classified scholia into a training set and an evaluation set. While many machine-learning algorithms could be applied to learn a model from a training set of data, we chose to use a decision tree. Other algorithms are likely to produce better results, but the actual working of models built using methods such as neural networks are opaque. We wanted to build a model we could interpret directly, in order to understand what role different features played. Of course, our data sets could subsequently be analyzed with other classification methods, in hopes of getting better results, but our initial decision tree model produced usable results.
We repeatedly built a decision tree by training it on a randomly selected 70% of the 100 manually classified scholia (the training set), then scored the resulting model against the remaining 30% (the evaluation set). The resulting decision trees correctly classified the scholia in the evaluation set up to 90% of the time. This is significant, since, if you were to randomly assign one of four classes to the scholia, you would expect about a 25% success rate.
Automated classification
Having established that our classifier could identify Aristarchan scholia with a high degree of success, we applied it to all 8,000 scholia in our full data set. In the resulting classification, both Aristarchan scholia and post-Aristarchan scholia appear most frequently in the main zone. By contrast, scholia in the “Aristarchan paraphrase” category are most often found in the intermarginal zone. The distribution of the four categories in the interior scene resembled the distribution within the intermarginal zone, but with a narrower spread between most frequent and least frequent classes.
Given our earlier topic modeling work, we were not surprised to find that the four classes of scholia had different concentrations in each zone. We decided to add the zone of the scholia as a further feature (for a total of nine features per scholion), and train a new classifier on this expanded data set. The resulting classifier performed worse than the classifier trained on the initial eight features.
This may seem counterintuitive, but it illustrates a subtle and important point about the relation between the automatically inferred topics in our topic modeling and the classes in our supervised classification using decision trees. We used topic modeling to classify scholia into fifteen groups based solely on their vocabulary (in the edition we created that had been optimized for topic modeling). We saw that we could plausibly interpret many of these classes as “topics” in the normal English sense of the word. The distribution of these topics shows that some concentrate more in particular zones of the page, and that the contents of the interior and intermarginal scholia cannot be summarized as condensed versions of the main scholia.
Two of those topics were particularly relevant for the problem of identifying Aristarchan material, topic 6 (“verbatim Aristarchan”) and topic 9 (“discussion about Aristarchus’ edition”). For our decision tree classifier, we incorporated the theta scores for these topics in a deliberately selected set of features that included the non-textual feature of the critical signs and emphasized formulaic phrases that we could associate with one of our four manually defined classes. These four classes were also distributed unevenly across the different zones of the page, but each zone included some material from all four classes. Equating one of our classes with a specific zone would therefore be simplistic: because the full set of other features in our data set more accurately predict which scholia are “Aristarchan,” etc., adding the zone as a feature actually worsened the performance of the classifier.
The topic modeling results are strong evidence that the scribe of the Venetus A planned the layout of the manuscript to accommodate material from multiple distinct sources. The Aristarchan classifier shows us that “Aristarchan” material, “Aristarchan paraphrase” and “post-Aristarchan” material are concentrated differently in the page zones, and therefore, we infer, were transmitted in sources that included these categories in different concentrations.
Conclusion
As our title indicates, these are preliminary results, but they are significant for Homerists. We have found strong evidence that the scribe of the Venetus A planned the layout of the manuscript so that he could incorporate material from a variety of no longer extant sources. Our work suggests that it is still possible in the twenty-first century to recover some of the work of Aristarchus in the second century B.C.E from a manuscript of the tenth century C.E. We can begin to see how the sources of the tenth century scribe preserved this verbatim language to differing degrees.
We were able to analyze the scholia of the Venetus A manuscript in part because of the architecture of the HMT editions. The abstract semantic model of the HMT defines everything encoded in the project’s archival XML edition. The archival edition is complex, and necessarily includes multiple possibilities for reading the text (abbreviated or expanded readings, original or corrected version, etc.). A code library implementing the semantic model makes it straightforward to generate univocal, semantically unambiguous editions from the XML source. We developed one edition optimized for topic modeling, and another consisting solely of eight features that we trained a classifier to assign to one of four classes of our own definition.
“Reading” these editions is no longer just for people. Automated processes can read an entire corpus and can interpret editions that are essentially unintelligible to human readers. The editions we created for topic modeling and for automated classification are algorithmically derived from our hand-crafted XML text, but nearly meaningless to the human eye. At the same time, since all the editions are automatically aligned through our use of canonical citation, we can read a version of the same text that is intended for a human audience.
One goal of the HMT is to make all of our work replicable and therefore falsifiable. Canonical citation plays an important role in this, too. We expect these preliminary conclusions to be refined and modified as later scholars work with a complete edition of the Venetus A and its scholia. As they do so, it will be possible to determine exactly where other editions of the Venetus might read the text differently, or where a topic modeling system might score the classification of a particular scholion differently.
Even when the HMT includes every word of the uniquely rich Venetus A manuscript, there will be no limit on the ways we can choose to read—and hence edit—these texts.
Footnotes
[ back ] 1. The work presented here was developed in independent but coordinated undergraduate theses by Wauke and Schufreider completed in the spring semester of 2017, for which Smith served as an advisor. The discussion of how distinct zones of scholia function is analyzed more extensively in Schuhfreider’s thesis; the identification of Aristarchan material is further developed in Wauke’s thesis. The digital edition of the Venetus A we worked with was created by more than 150 contributors: we owe all of them a debt of gratitude. On behalf of the whole project as well as for their individual help and encouragement, we would like to thank the Homer Multitext’s co-editors, Casey Dué and Mary Ebbott, and co-architect, Christopher Blackwell. At Holy Cross, we benefited from the opportunity to collaborate with members of the Manuscripts, Inscriptions and Documents Club. We are also grateful to Thomas Köntges, then at the University of Leipzig, who generously shared his time and invaluable advice on topic modeling and developing a classifier using topic modeling scores as features, as we describe in this paper.
[ back ] 2. Homeri Ilias ad veteris codicis veneti fidem recensita. Scholia in eam antiquissima ex eodem codice aliisque, nunc primum edidit cum asteriscis, obeliscis aliisque signis criticis Joh. Baptista Caspar d’Ansse de Villoison (Venice, 1788: Typis et Sumptibus Fratrum Coleti).
[ back ] 3. Since the completion of Wauke and Schufreider’s work, the number of undergraduate contributors to the HMT project has surpassed 200, and an initial edition of all the scholia in the Venetus A manuscript, numbering nearly 10,000, was completed in 2018.
[ back ] 4. Scholia Graeca in Homeri Iliadem. Edited by Karl Wilhelm Dindorf. (Oxford UP, 1875). Available online from the HMT project at http://www.homermultitext.org/pd-pdfs/Dindorfius1875a.pdf.
[ back ] 5. Scholia Graeca in Homeri Iliadem (Scholia vetera) (De Gruyter, 1969–1988) in five volumes with two additional volumes of indices.
[ back ] 7. We would like to acknowledge the work of Dr. Melissa Toomey and Frank Hartel who in 2010 edited books 3 and 4 of the Venetus A scholia, and cross-indexed their editions with the digital text of Erbse in WordHoard. Comparing their complete text with the digital Erbse revealed that, measured in numbers of words, approximately 20% of the text content of the Venetus A is absent from Erbse’s edition.
[ back ] 8. In the course of summer research at Holy Cross in 2020, Sophia Sarro ’21, Luke Giuntoli ’21, Anne-Catherine Schaaf ’22, Natalie Dimattia ’22 and Augusta Holyfield ’22 applied topic modeling to the entire corpus of Venetus A scholia. Sophia Sarro expanded on this in an undergraduate thesis completed in the spring of 2021. In the summer of 2021, Dimattia, Holyfield, Schaaf and Rebecca Rose Kaczmarek ’23 were able to build on this foundation to develop network models of the transmission of material in the scholia. Although it was not their primary goal to review Wauke and Schufreider’s work, their research in 2020 and 2021 collectively has validated the principal conclusions presented here.
[ back ] 10. Consider a simple example that can enormously complicate digital reading of a text. We can use the TEI persName element to identify personal names, and TEI abbr to identify abbreviations. But what is the proper encoding of an abbreviated personal name? The TEI-all schema allows an example like <persName><abbr>ΑΡΙΣΤ</abbr></persName>; it also allows <abbr><persName>ΑΡΙΣΤ</persName></abbr>.
[ back ] 11. Like all our work, the implementation is available from publicly visible GitHub repositories. We would especially like to thank HMT Project Manager Stephanie Lindeborg of Boston Latin Academy for her invaluable contributions to this part of the project.
[ back ] 12. Anne-Catherine Schaaf, Rebecca Kaczmarek, and Neel Smith presented this interactive system at the conference, European Association for Digital Humanities 2021: Interdisciplinary Perspectives on Data (September 2021): “Scalable, machine-tested, collaborative editing using Pluto notebooks.” Published abstract forthcoming.
[ back ] 13. For a discussion of some of the ways this has made it possible for the HMT to accept contributions from more than 150 team members, see Neel Smith, “Citizen Scholarship in the Homer Multitext Project” in Pushing the Boundaries of Historia (edited by Lee Fratantuono and Mary English) (Routledge: 2019).
[ back ] 14. Allen, T. W. “On the Composition of Some Greek Manuscripts: The Venetian Homer,” Journal of Philology 26, 1899, 161–181. Available online from the HMT project at http://www.homermultitext.org/pd-pdfs/Allen-JP-1899.pdf.
[ back ] 16. Schufreider’s 2017 thesis presents a more detailed discussion of the results of this topic modeling.
[ back ] 17. On the Aristarchan critical signs, see Graeme Bird, “Critical Signs—Drawing Attention to ‘Special’ Lines of Homer’s Iliad in the Manuscript Venetus A” in Recapturing a Homeric Legacy (= Hellenic Studies 35), edited by Casey Dué, pp. 89–115. Available from http://www.homermultitext.org/Pubs/Due_Recapturing_a_Homeric_Legacy.pdf.