Better Solutions to Old Problems?

(1)

Better Solutions to Old Problems?

 

Using Computational Methods to Attribute the Authorship of 

Henry VI, Part Two and Part Three

Wessel Verhoef  s2201828

A thesis submitted for the degree of  Master of Arts in Digital Humanities

University of Groningen  April 2018  15,243 words Supervisor  Dr. John Flood Second reader  Dr. Leonie Bosveld-de Smet 

(2)

Introduction

The New Oxford Shakespeare is the most expansive edition of Shakespeare’s collected works to date. The printed version comes in three editions. The Modern Critical Edition presents Shakespeare’s works for students and teachers in modernized spelling and punctuation. The two-volume Critical Reference Edition presents the same works for literary and bibliographic scholars, with the original spelling, punctuation, capitalization, abbreviations, typographical contrasts, ambiguities, and inconsistencies. The Authorship Companion focuses on the canon and chronology of Shakespeare’s works. Additionally, there is an Online Edition which integrates all of the printed materials.

The chapters of the Authorship Companion are divided into methods and case studies. The first part gives an overview of the latest authorship attribution methods. The second part presents the edition’s case studies on the authorship of problematic plays and passages associated with Shakespeare. Gary Taylor, general editor of the New Oxford Shakespeare, claims that the Companion represents the state of the art in authorship attribution: “It combines the best of previous scholarship with twenty-first century tools to find better solutions to old problems.” (The New Oxford Shakespeare, 2018). These tools include online databases such as Early English Books Online (EEBO) and Literature Online (LION), which oﬀer digitized versions of many thousands of early modern texts. These databases facilitate computational research on a much larger scale than before. Taylor declares that the new techniques have improved attribution research:

Shakespeare has now fully entered the era of Big Data. Earlier attempts to answer all these questions about Shakespeare’s collaborative work have been limited to intuitions, or to hand counts of a small number of features. We can now situate Shakespeare within large digital databases of almost all extant early modern plays, and almost all early modern printed poetry and literary prose. This makes it possible to distinguish, more precisely and more confidently, Shakespeare’s stylistic identity from that of his fellow playwrights. (The New Oxford Shakespeare reveals the Bard had a collaborator, 2016)

But for all of Taylor’s confidence, the jury is still out as to whether these tools truly provide better solutions to these longstanding authorship problems.

The New Oxford makes some controversial editorial decisions. With the addition of the anonymous plays Arden of Faversham and Edward III, the Shakespeare canon now stands at forty plays, twelve of which are listed as collaborations. Its main source of publicity and controversy is the identification of Marlowe as the co-author of the Henry VI plays. The editors base this on the findings of two independent studies, one published by Shakespeare Quarterly in 2016, the other included in the Authorship Companion.

Within the field of authorship attribution, Shakespeare is a notoriously diﬃcult subject. The Henry VI plays in particular have been described as “perhaps the thorniest problem in

(4)

attribution in the Shakespeare canon” (Craig, 2009, 40). Hugh Craig observes that any new attribution or de-attribution “meets an interesting range of responses” (Craig, 2016). Some maintain that authorship is irrelevant in appreciating a play, and that as such there is no need for authorship attribution. Others say that genre and culture are much more decisive factors in creating a play. Still others say and that attribution is important but impossible, because of the many uncertainties that are involved. As for himself, Craig believes that that “while the other factors do matter, they don’t drown out the voice of the author” (Craig, 2016). In other words, he believes that it is possible to attribute Shakespeare’s plays in spite of these uncertainties.

The importance of ideology in Shakespeare studies cannot be overstated. E. K. Chambers famously proclaimed the “disintegration” of Shakespeare in 1924. He was referring to the eﬀorts of several nineteenth-century scholars to ascribe parts of the recognized Shakespeare canon to other authors. It now appears that the term reflected contemporary cultural anxieties: Chambers feared that the disintegration of Shakespeare, and of English culture in general, would lead to the disintegration of English society at large (Jowett, 2014, 172). Still, Chambers did accept John Fletcher’s hand in Henry VIII and Shakespeare’s contribution to Sir Thomas More, Two Noble Kingsmen, Edward III, and Pericles.

For Victorian scholars, the supposed presence of two authorial voices would be reason to suspect that Shakespeare revised the work of other dramatists. Modern scholarship is likely to explain the same phenomenon in terms of collaboration. Today, most scholars agree that Renaissance playwriting was collaborative at its core. Brian Vickers, a leading attribution scholar, writes: “Against the Romantic notion of individual inspiration, free of any financial considerations, we need to conceive of an artifact produced by a work-sharing process, in which certain elements of the composition are delegated to other hands under the supervision of a master craftsman” (Vickers, 2007, 312). According to Vickers, the playwriting process is best compared to “the practice of Renaissance painters and sculptors who worked together in a bottega” (Vickers, 2007, 314). Taylor stresses the similar point that early modern playwrights were more artisan than artist (Taylor, 2017).

Unfortunately, the term disintegrator continues to be used by attribution scholars to “castigate and condemn unwelcome propositions” (Jowett, 2014, 172), one of the main oﬀenders being Vickers. When Laurie Maguire and Emma Smith proposed that All’s Well that Ends Well contains traces of Middleton’s writing, Vickers dismissed their eﬀorts as the work of “the new disintegrators” (Vickers and Dahl, 2012, 14). Again, the label is reserved for attributions that are unwelcome. Vickers argues in his book Shakespeare, Co-Author (2002) that Shakespeare wrote Titus Andronicus with George Peele, Timon of Athens with Thomas Middleton, Pericles with George Wilkins, and Henry VIII and The Two Kingsmen with John Fletcher.

Between the last two Oxford editions, the collaborative portion of Shakespeare’s plays has almost doubled: the Oxford Shakespeare of 1986 identified 39 Shakespeare plays with eight of them being collaborations (Alberge, 2016), while the New Oxford Shakespeare contains 44

(5)

plays, of which 17 are identified as collaborations. Vickers is amazed that Taylor would present this as an advance: “To most other people it will seem like a grievous loss—always assuming that the attributions and deattributions are justified” (Vickers, 2017). Taylor’s lack of grief compels Vickers to question the purity of his intentions: “Taylor suﬀers from Shakespeare envy, an odd mixture of acknowledging his greatness while attempting to whittle away at the canon or introduce foreign bodies” (Vickers, 2017).

Attempts to whittle away at the canon are the mode du jour, writes John Jowlett: “For in the postmodern era the imperative to preserve organic unity holds little command. In our dispersed and variegated notion of the subject area of English, its primary instance, Shakespeare, is subject to energetic disintegration—to use the term in a neutral sense, without its historical baggage of the 1920s” (Jowlett, 2014, 182). The New Oxford Shakespeare comes at a time when the humanities are increasingly regarded as soft and impractical. The phrase “More hack, less yack” encapsulates digital humanities’ promise of a more practical approach. Franco Moretti pioneered digital humanities in literary studies by introducing the concept of “distant reading” as the antithesis to close reading, which he likened to “secularized theology” (Moretti, 2000b, 208). Moretti argued that modern technology enables us to look beyond the western canon to “the great unread.” Close reading, which necessarily depends on a small canon, is a part of the problem: “At bottom, it’s a theological exercise—very solemn treatment of very few texts taken very seriously” (Moretti, 2000a, 57). This reads like a caricature and may have been intended as such. Moretti was proposing a fundamentally diﬀerent manner of reading, focusing on “units that are much smaller or much larger than the text: themes, tropes—or genres and systems” (Moretti, 2000a, 57). In 2010, Moretti founded the Literary Lab at Stanford University, its very name suggesting scientific rigor. Now retired, Moretti concedes that the “revolution in knowledge” which he had hoped for has not taken place (Schuessler, 2017).

The New Oxford may not be quite so revolutionary, there are similarities to Moretti’s manifesto. While close reading has not been abandoned altogether, the computational methods are given center stage. But the New Oxford has its own agenda. Throughout his career, Taylor has been working at demystifying Shakespeare. as a singular genius and draw attention to some of Shakespeare’s peers. The New Oxford makes a point of presenting Shakespeare not as a singular genius but as a writer among other writers, who had considerable talent of their own.

Taylor has long been an enfant terrible of Shakespeare studies. He served as a general editor for the Oxford Shakespeare of 1986, which took a decidedly historical approach to Shakespeare’s collected works. Reasoning that Shakespeare’s plays were written for the stage, the editors focused their eﬀorts on “recovering and presenting texts of Shakespeare’s plays as they were acted in the London playhouses which stood at the centre of his professional life” (Wells and Taylor, 2005, xxxix). The edition came as a shock to Frank Kermode, who noted “a certain archaeological rigor in the editors’ procedures.” The editorial decisions include the changing of Falstaﬀ’s name to Sir John Oldcastle, which was the character’s name in early performances, though not in any printed form of the plays (Kermode, 1988). For the plays

(6)

commonly known as Henry VI, Part Two and Part Three, the editors chose the more historical titles The First Part of the Contention and Richard Duke of York. For Kermode, the emphasis on what the plays originally were went at the expense of what they came to be for modern readers:

To an extent one can’t really suggest in a review of this kind the Oxford editors, tearing like a hurricane across the terrain of the Complete Works, have altered many familiar features. If they sometimes seem overbold, the laborious and civilized arguments of the Companion volume should exempt them from the charge of having been simply épatants, though the suspicion lingers that they have been capable, all else being equal, of choosing the more spectacular alternative. (Kermode, 1988)

In the case of the Henry VI authorship issue, the editors of the New Oxford have undoubtedly chosen the more spectacular alternative. But their endorsement of Marlowe as co-author does not automatically mean that the attributions will stand. Donald Foster (1989) attributed the obscure poem A Funeral Elegy to Shakespeare, using a form of computational stylistics. The attribution received front-page coverage in The New York Times and the poem was subsequently included in three American editions of Shakespeare’s works. Some years later, Gilles Monsarrat (2002) compared the same poem with the work of John Ford and concluded that the writing was in fact his. Many scholars who had accepted the attribution were relieved to discover that Shakespeare was not the author after all. Foster himself recanted: “No one who cannot rejoice in the discovery of his own mistakes deserves to be called a good scholar” (Niederkorn, 2002). The Elegy, which was included in not one but three editions of Shakespeare’s works, is rarely associated with Shakespeare today. It is not inconceivable for the Marlowe attributions to suﬀer the same fate.

The aim of this thesis is to establish whether nontraditional methods are more reliable and objective in attributing the authorship of Shakespeare’s works, specifically the second and third part of Henry VI. Chapter 1 outlines traditional scholarship on the authorship of the Henry VI plays. It aims to paint a picture of the debate surrounding the plays since the eighteenth century. The second half of the chapter discusses the origins of nontraditional attribution methods.

Chapter 2 discusses the two studies (Segarra et al., 2016; Burrows and Craig, 2017) which convinced the Oxford editors of Marlowe’s co-authorship of the Henry VI plays. The chapter concludes with a discussion of the two studies and of Joseph Rudman’s critique of Shakespeare attribution studies as a whole. Chapter 3 discusses some recent articles by Darren Freebury-Jones, whose research focuses on Shakespeare’s fellow playwright Thomas Kyd, and who holds entirely diﬀerent views on the authorship of the Henry VI plays. His criticism is directed not so much at the nontraditional methods that are applied by the Oxford studies, but at the conclusions that are drawn from them. 

(7)

1 Methods

The attribution scholar Joseph Rudman makes a useful distinction between traditional and nontraditional attribution studies. The latter refers to “those employing the computer, statistics, and stylistics” (Rudman, 1998, 351) while traditional studies rely on close examination of internal and external evidence. Both traditional and nontraditional studies on the Shakespeare canon are characterized by heated disagreement.

1.1 Traditional scholarship on the Henry VI plays

While it is commonly accepted that Marlowe influenced Shakespeare’s early work, merely suggesting that Marlowe cowrote the Henry VI plays is likely to cause an uproar in Shakespearean circles. Vickers categorically rejected the attributions before having had an opportunity to look at the Authorship Companion:

As anyone who has studied Marlowe knows, his dramaturgy is wholly unlike Shakespeare’s, his vocabulary is far more Latinate, he seldom wrote prose, and his prosody is completely diﬀerent, as quantitative studies have known. If he had in fact written hundreds of lines of verse for these plays, his presence would be detectable by traditional observation. (Vickers, 2017)

The fact of the matter is that traditional scholars have been detecting Marlowe’s presence in these plays since the eighteenth century. In recent times, 2 Henry VI has been described as being “virtually saturated in Tamburlainean statements of proud self-assertion.” (Cartelli, 1991, 86). Margaret Owens writes: “So insistent is the verbal and visual imagery of decapitation in 2 Henry VI that one might be tempted to speculate that the dramatist(s) took a deliberate risk, indulging in an exercise in theatrical audacity of a kind we more typically associate with Marlowe” (Owens, 1996, 379). Marlowe has been repeatedly associated Marlowe with Shakespeare’s early plays. Nicholas Brooke wrote of “a number of eruptions in Shakespeare’s work of passages which are unmistakably Marlovian in tone and attitude, to a degree which would almost justify a disintegrator in identifying them as Marlowe’s work” (Brooke, 1961, 34). The association of Marlowe with the Henry VI plays is by no means new, but it may be true that the plays have never been attributed to Marlowe as confidently as they are now by the New Oxford.

Attribution scholars often distinguish between internal and external evidence. This distinction is fairly straightforward: internal evidence refers to any evidence gathered from the text itself, whereas external evidence is derived from sources outside of the text. The presence or absence of an author’s name on an edition of a play may serve as external evidence—but never as proof—of his authorship. There are editions of The London Prodigal and A Yorkshire

(8)

Tragedy with Shakespeare’s name on them, even though most scholars agree that he wrote neither one of these plays. Conversely, Pericles and The Two Noble Kingsmen are not included in the First Folio, the 1623 published collection of Shakespeare’s plays, while Shakespeare is commonly believed to be the author of both. During the 1590s, at the start of Shakespeare’s playwriting career, editions of plays did not routinely print the author’s name. This explains why the earliest known versions of 2 and 3 Henry VI, printed in 1594 and 1595 respectively, are anonymous.

The plays were printed again in Mr. William Shakespeare's Comedies, Histories, & Tragedies, better known as the First Folio, which is now considered the most reliable source for the plays. Scholars began to debate the authorship of Shakespeare’s plays during the eighteenth century. Lewis Theobald, in his 1733 edition of Shakespeare’s works, questioned Shakespeare’s sole authorship of the plays. In 1765, Samuel Johnson cited the First Folio’s ascription of the plays to Shakespeare, the reference to the plays in Shakespeare’s Henry V, and the close narrative connection between 3 Henry VI and Richard III as evidence of Shakespeare’s authorship. However, Johnson considered the early versions to be “so apparently imperfect and mutilated, that there is no reason for supposing them the first draughts of Shakespeare” (in Cox and Rasmussen, 2001, 159).

George Steevens’ 1778 edition of Shakespeare’s works included a piece of external evidence, discovered by Thomas Tyrwhitt, which was to define scholarship on the plays for centuries to come. A passage from Robert Greene’s deathbed pamphlet Greenes Groatsworth of Wit, bought with a Million of Repentance, printed posthumously in September 1592, alludes to Shakespeare and to 3 Henry VI:

those Puppets (I meane) that spake from our mouths, those Anticks garnisht in our colours. . . . Yes trust them not: for there is an upstart Crow, beautified with our feathers, that with his Tygers hart wrapt in a Players hyde, supposes he is as well able to bombast out a blank verse as the best of you: and beeing an absolute Iohannes fac totum, is in his owne conceit the onely Shake-scene in a countrey. . . . whilest you may, seeke you better Maisters; for it is pitty men of such rare wits, should be subiect to the pleasure of such rude groomes. (In Alexander, 1929, 41)

It is commonly believed that Greene was addressing his “fellow schollers” Christopher Marlowe, George Peele, and Thomas Nashe. The clear reference to the 3 Henry VI line “O tiger’s heart wrapped in a woman’s hide!” (1.4.137) and the pun on Shakespeare’s name (“Shake-scene”) suggest that this “upstart Crow” is Shakespeare. Greene speaks resentfully of actors as “Puppets . . . that spake from our mouths” and of Shakespeare as a glorified actor rather than a bona fide playwright like Greene. Tyrwhitt himself took the passage as evidence for Shakespeare’s authorship of the three Henry VI plays (Alexander, 1929, 39).

Scholars have argued over the “upstart Crow” passage since its discovery two centuries ago. Opinions have varied on whether “Crow” refers to Shakespeare as an actor, aspiring playwright,

(9)

Henry VI. Arguing that the quartos “are by no means mutilated and imperfect” (in Cox and Rasmussen, 2001, 160), Malone felt that “the inferior parts are not merely unequal to the rest . . . but of quite a diﬀerent complexion from the inferior parts of our author’s undoubted performances” (Wilson, 1952, xiv). Malone interpreted Greene’s allusion to 3 Henry VI as an accusation of plagiarism, using it as the “chief hinge” of his argument that the quarto texts were originally written by Greene and Peele and subsequently revised by Shakespeare. He later changed his opinion and attributed the quartos to Marlowe (Alexander, 1929, 94).

Eighteenth-century scholars such as Pope and Johnson believed that the quartos were sketches for the superior folio versions. During the first half of the twentieth century, Madeleine Doran and Peter Alexander suggested, working independently from each other, that The Contention and The True Tragedy were memorial reconstructions, written down from memory by actors who had played in them. Alexander proposed that these were the two actors playing Warwick and Suﬀolk, while Doran tentatively suggested a group eﬀort by several of the actors (Alexander, 1929, 115). Alexander also noted numerous interpolations from some of Marlowe’s plays into 2 and 3 Henry VI, but he accredited these to Marlowe’s popularity at the time, rather than taking them as evidence for Marlowe’s authorship (Alexander, 93–106). In recent years, scholars have begun to question the memorial reconstruction theory. Laurie Maguire allows that a handful of Shakespeare plays may have been memorial reconstructions, but maintains that the overwhelming majority, 2 and 3 Henry VI included, are not (Maguire, 1996, 324–325).

In constructing his argument, Alexander found additional evidence against Malone’s plagiarism theory in a passage from Greene’s Francescos Fortunes, the second part of Neuer too Late (1590). This passage describes a dispute between Cicero (“Tully”) and the actor Roscius:

It chanced that Roscius & he met at a dinner . . . where the prowd Comedian dared to make comparison with Tully: which insolencie made the learned Orator to growe into these termes; why Roscius, art thou proud with Esops Crow, being prankt with the glorie of others feathers? of thy selfe thou canst say nothing, and if the Cobler hath taught thee to say Ave Caesar, disdain not thy tutor, because thou pratest in a Kings chamber: what sentence thou utterest on the stage, flowes from the censure of our wittes, and what sentence or conceipte of the invention the people applaud for excellent, that comes from the secrets of our knowledge. (In Vickers, 2016c, 254)

The crow and feathers are references to Aesop’s fable of the jackdaw and his borrowed feathers. An Elizabethan reprint of Caxton’s translation clearly states the moral of this tale: “Of the Jay and the Peacockes, how none ought to be proud of that which is not theirs” (in Vickers, 2016c, 249). This anecdote serves much the same function as the Groatsworth passage: “The image of a crow decked out with other birds’ feathers means simply that actors are unfairly given the credit that ought to go to the author of the words they speak” (Potter, 2012, 99). There is nothing in either passage to suggest that Greene is accusing Shakespeare of plagiarism, or that “Crow” signifies anything other than actor. This did not stop Malone from arguing that “upstart Crow”

(10)

been written by Greene and left unaltered by Shakespeare in his revision. Neither argument is convincing.

John Dover Wilson revived Malone’s argument in 1952. Wilson approvingly cited Alexander’s memorial reconstruction theory but rejected his evidence against Malone, which he maintained “does not touch [Malone’s] main position at all” (Wilson, 1952, xvi). By ignoring Alexander’s evidence, Wilson was left “isolated in an echo chamber of Shakespeare, Greene, Peele, Marlowe and Nashe” (Knowles, 2001, 127). Wilson eventually subscribed to Malone’s initial theory that Shakespeare had “considerably enlarged and improved” two plays by Greene. Wilson argues that all contemporary records on the matter interpret the “upstart Crow” reference as a charge of plagiarism, but only oﬀers one example. He goes on to say that all Renaissance writers identified Aesop’s crow with Horace’s Third Epistle, in which the crow warns a friend not to steal from other poets. Greene’s “beautified with our feathers” echoes Horace’s “stolen colors,” while “bombast out a blank verse” alludes to Horace asking whether Titius “is still storming and swelling in the bombastic style of tragedy.” Wilson then declares that the crow fable “always suggests dishonest appropriation in Greene,” even though Alexander had presented evidence to the contrary thirty years earlier. Wilson cites another passage in which Greene refers to “Ezops Crowe” and states that here Greene was actually thinking of Horace’s crow. He surmises that Greene made his accusation out of desperation and anger, and that Shakespeare’s work on Henry VI was revisionary (Wilson, 1952, xvii–xix). Finally he states that Greene wrote the original Henry VI plays, with some help from Nashe and Peele (Wilson, 1952, xxviii). All of these claims are made with much more confidence than the material allows for.

Vickers wrote a paper on the issue last year in which he repeats: “Nowhere in Greene’s repetitive writing is there any association of the bird’s borrowed plumage with the charge of plagiarism. That association was first made two hundred years later” (Vickers, 2016c, 258). Vickers is not so much concerned with the authorship of 3 Henry VI as with the accusation of plagiarism, “a pervasive smear that is usually applied by those wanting to diminish Shakespeare’s achievement and seldom feeling the need to cite evidence” (Vickers, 2016c, 245). Vickers asserts that Wilson was “an acute and independent scholar on many issues” but a disintegrator when it came to matters of authorship, causing him to disregard both Alexander’s evidence and the lack of evidence presented by Malone (Vickers, 2016c, 263).

An additional point is that plagiarism did not exist in Shakespeare’s time. The concept of plagiarism as something immoral emerged during the Romantic movement, together with the concept of originality as an ideal. The “upstart Crow” debate shows that disintegrators are not necessarily more objective and less susceptible to confirmation bias than those who would like the better parts of the Shakespeare canon to be the work of a singular genius. But even if traditional scholars have reached very diﬀerent conclusions after examining the same evidence, there at least seems to be some agreement on method.

(11)

1.2 Nontraditional methods

Nontraditional attribution studies are aimed at producing more objective and unambiguous results. At the same time, there seems to be little consensus among practitioners on either attribution or method. “What are non-experts expected to believe?” Rudman asks (2016, 309). One immediately obvious reason for this lack of agreement is the multidisciplinarity of the field. Authorship attribution is traditionally the domain of literary scholars, whereas nontraditional methods are by and large created by statisticians and information scientists. Attribution scholars seem to specialize in either literature or information science, but very rarely in both. Another problem is that some nontraditional methods are based on assumptions that are not universally accepted. Nontraditional methods can be divided into two main approaches: stylometry and corpus linguistics.

1.2.1 Stylometry

Stylometry is the statistical analysis of variations in literary style between writers. It is founded on the premise that every author has a unique style which can be measured, provided there is enough data available. Stylometry often focuses on frequent function words, which are considered better predictors of authorship than lexical words. The linguist Charles Fries defined function words as separate items signaling particular structural meanings (Fries 1952, 108). He listed the following sentences by way of illustration:

The boy was given the money  The boy had given the money  The boy vab given the money

In the first two sentences, was and had signal diﬀerent structural meanings, whereas the third sentence shows that with a nonsense word in the same position the entire sentence becomes nonsensical: it no longer clear whether the boy is the giver or recipient of the money. Function words include prepositions, pronouns, auxiliary verbs, and conjunctions. What does set function words apart from other words is that “it is usually diﬃcult if not impossible to indicate a lexical meaning apart from the structural meaning which these words signal” (Fries, 1952, 106). Fries stressed that this is not the same as saying that function words have no meaning, since they determine the meaning of the sentence (Fries, 1952, 88). It would be more accurate to say that their meaning lies in their function. Many stylometrists believe that function words are easily quantifiable and “relatively immune from conscious control” (Nerbonne, 2006, xvii), whereas lexical words appear to be largely determined by genre.

Frederick Mosteller and David Wallace (1964) popularized the use of stylometry in their authorship study on The Federalist. This collection of essays, first published in 1787–88, was written under the pseudonym “Publius” by Alexander Hamilton, John Jay, James Madison to promote the ratification of the United States Constitution. Mosteller and Wallace focused their

(12)

study on the authorship of twelve disputed papers, which were written by either Hamilton or Madison.

Their method was based on Bayesian analysis, named after the English mathematician Thomas Bayes (1702–61). Bayes’ theorem holds that the conditional probability of each of a set of possible causes for a given observed outcome can be computed from knowledge of the probability of each cause and the conditional probability of the outcome of each cause. (Modern computational linguistics makes frequent use of Naive Bayes, a family of probabilistic classifiers.)

In order to diﬀerentiate between the two authors, Mosteller and Wallace needed “variables that depend on authors and nothing else” (Mosteller and Wallace, 1964, 265). Previous research suggested that content words were mostly selected on the basis of subject matter. Additionally, most educated Americans in the late eighteenth century used the same language, employing the same stylistic devices and standard phrases and using similar sentence structures (Mosteller and Wallace, 1964, 9). Between the two candidates, they found that Madison used upon eighteen times more, and at a much higher rate, than Hamilton. Further, Madison used whilst whereas Hamilton preferred while. The word enough emerged as a Hamilton marker. Vickers observes that Mosteller and Wallace never discussed why the authors used these words so frequently or infrequently (Vickers, 2016b, 4). Any answer would be speculation, of course, but perhaps this is Vickers’ point.

Mosteller and Wallace made it clear that they were approaching the issue as statisticians, not as literary scholars or historians. Their ultimate goal was to further discrimination studies by testing a statistical method that had received some criticism in the past. Explaining their involvement with what is primarily a historical and literary problem, they write: “From the point of view of statistical methods, authorship problems fall into a general area called discrimination or classification problems. In these problems the task is to assign a category to an object or individual whose true category is uncertain” (Mosteller and Wallace, 1951, 1). According to Nerbonne (2006), the fact that they included insights from nonstatistical studies shows that they did recognize the need for involving literary scholars and historians.

The twelve disputed papers are increasingly thought to be collaborations of some sort between Hamilton and Madison (Rudman, 2012, 263). Another study’s findings suggested that the writing process for the disputed papers had been more collaborative than previously supposed (Collins et al., 2004). This may explain why both Hamilton and Madison claimed authorship of these papers, and why other investigations did not reach the same conclusions as Mosteller and Wallace. Collins et al. (2004) conclude that Mosteller and Wallace failed to correctly define the problem.

MacDonald P. Jackson pioneered the use of stylometric methods in Shakespeare attribution studies. Jackson counted the frequency of function words to identify Middleton’s contributions to Shakespeare’s plays (Jackson, 1979). He found that Middleton likely contributed to

(13)

Shakespeare’s Timon of Athens, as David Lake had done before him (Lake, 1975). Jackson is also one of the contributors for the Authorship Companion.

Hugh Craig and John Burrows have been involved with the majority of stylometric studies on the Henry VI plays. In 2009, Craig did a statistical analysis of the distribution of function words and lexical words in the plays. The results pointed to Marlowe as the possible author of “many of the Joan la Pucelle and Jack Cade scenes” in 1 and 2 Henry VI (Craig, 2009, 77). Craig reasoned that both characters have Marlovian qualities, being “of humble parentage but aspiring to supreme political power” (Craig, 2009, 73). Craig considers the evidence for Marlowe’s hand especially persuasive in the Cade scenes of 2 Henry VI: “If Shakespeare wrote these scenes, he was imitating Marlowe’s diction and syntactic habits as well as responding to his transgressions against decorum and social stability, his rampant individualism in attitudes and dramatic characterization” (Craig, 2009, 76). The results for 3 Henry VI, on the other hand, did not show clear signs of collaboration.

Vickers strongly opposes the application of stylometry to early modern plays. Having become involved with nontraditional methods late in his career, he is by his own admission “still basically a Shakespearean and text reader” (Vickers, 2016a). For Vickers, “an attentive reading and rereading of a text [is] the sine qua non of authorship study” (Vickers, 2007, 318). Stylometry, he is convinced, is unequipped to deal with the dynamic and collaborative nature of early modern drama. According to Vickers, it is no coincidence that the Federalist study and Alvar Ellegård’s study on the Letters of Junius, two foundational stylometric attribution studies, both focus on supposed single-author texts (Vickers, 2016b, 10).

Vickers’ main argument is that dramatists do not write in propria persona, but from the perspectives of their characters (Vickers, 2016b, 12). Shakespeare became increasingly adept at using language to diﬀerentiate between his characters: “Clowns speak prose, varied with rhymes and songs; servants or socially subordinate characters speak prose; their superiors speak in pentameter verse to each other, but can speak prose to their servants” (Vickers, 2016b, 15). From this perspective, King Lear consists not of Shakespeare’s words but of the words of its characters. Assuming that Shakespeare wrote all of Lear, the words are of course his words, but Vickers has a point. Just as research shows that the “style” of plays varies more by genre than by author, there may in some cases be a greater stylistic diﬀerence between characters than between writers.

Hugh Craig believes that there is a single Shakespearean idiom underlying all of his writing —“a grandiose claim” according to Vickers, if only because Shakespeare’s language changed over the course of his career (Vickers, 2016b, 17). Nor does Vickers subscribe to the idea that the selection of function words happens unconsciously, arguing instead that the choice of function words “is determined by a larger set of conscious choices” (Vickers, 2016b, 6).

(14)

1.2.2 Corpus linguistics

Vickers is a proponent of corpus linguistics, which he believes is “responsive to the fundamental feature of natural languages, that they ‘weave’ together words of all kinds in order to create meaning” (Vickers, 2011, 135). Vickers’ own attribution methods involve the use of plagiarism software called WCopyfind to search for n-grams, followed by the close reading of passages containing verbal parallels. An n-gram is a contiguous sequence of n items from a sequence of text or speech. These items can be phonemes, syllables, letters, words, or base pairs, but in authorship attribution they tend to be words. Sequences of one, two, and three items are usually referred to as unigrams, bigrams, and trigrams, respectively.

In his critique of stylometry, Vickers refers to Bengt Altenberg’s phraseological study of the London-Lund Corpus of Spoken English (LLC) (Altenberg, 1998). Compiled between 1959 and 1990, this corpus derives from two projects: the Survey of English Usage (SEU) at University College London, and its sister project, the Survey of Spoken English (SSE) at Lund University. It consists of nearly half a million running words, which constitute over 201,000 recurrent word combinations of varying length and frequency—over 80 percent of the total number of words (Altenberg, 1998, 101–102). Vickers does not mention that these word combinations include unintentional repetitions, stuttering, and combinations of little phraseological interest like the the, and the, in a, and out of the. After narrowing it down to combinations of at least three words occurring at least ten times, Altenberg ended up with some 6,000 examples of 470 diﬀerent types, only 3 percent of the total number of recurrent word combinations in the corpus (Altenberg, 1998, 103).

Vickers has himself received his fair share of criticism, most recently in two chapters of the Authorship Companion. MacDonald P. Jackson criticizes the attribution of Arden of Faversham, The True Chronicle History of King Leir, Fair Em, The Miller’s Daughter, and parts of 1 Henry VI and Edward III to Thomas Kyd (Jackson, 2017, 49). Jackson criticizes Vickers for using Kyd as the sole entrant in a one-horse race, without performing the necessary negative checks:

It is not enough to collect n-grams—whether short or long, exact or discontinuous—or any other textual constituents of dramatic scripts, that are shared by a disputed play and the canon of one favored playwright, and to then sift these for rarity or uniqueness, and suppose that a case has been made, however many such unique matches survive the sifting process. It must be shown that similar results are obtained by no other candidate for the disputed play’s authorship when the same processes are carried out with his canon as the target. (Jackson, 2017, 59)

In another chapter, Gabriel Egan expresses similar criticism of Vickers’ n-gram methods (Egan, 2017). For the attribution of Arden of Faversham, Vickers did not perform negative checks by looking for matching n-grams in the works of playwrights other than Kyd. Egan points out that Vickers demonstrates a misunderstanding of the diﬀerence between n-grams and collocations, the former being a group of words in a fixed order, whereas collocations can be

(15)

variations of a group of words: the three trigrams her green dress, green her dress, and dress her green make for one collocation. Egan attempted to reproduce the searches as reported by Vickers and found many things that Vickers said were not there, leading him to three possible explanations: that Vickers’ database was incomplete, his searching ineﬀective, or both.

Where stylometric methods are frequency-based, Vickers’ method is based on rarity. Craig notes the problems involved with methods that rely on rare verbal parallels:

Certainly it would seem perilous to argue from a set of such parallels alone for authorship. Then if the text may have multiple authors there is another set of complications. If a given section has no such parallels, does that argue for a diﬀerent author? How long should a section go without a significant parallel before we suspect a second author? Such questions of segmentation bedevil any method, of course, but weigh especially heavily on a method that relies on rarities. (Craig, 2009, 61)

This is another reason why some practitioners rely so heavily on high-frequency words: they are by definition a much more constant feature in a text than rare verbal matches. Craig is not suggesting that practitioners should refrain from using rarity-based methods, but he warns against using them to the exclusion of other methods. If there is one thing that many practitioners agree on, it is that attribution studies should strive to be as inclusive as possible. The more diﬀerent results point to the same candidate, the stronger the case for that candidate becomes.

The New Oxford editors attribute 2 and 3 Henry VI to Shakespeare and Marlowe (Taylor and Loughnane, 2017, 493–499) based on three attribution studies. As discussed before, Craig (2009) pointed to Marlowe as a potential co-author for 2 Henry VI. A new, more thorough study by Burrows and Craig (2017), which is included in the Authorship Companion, does the same for 3 Henry VI. Finally, a study by Segarra et al. (2016) identifies Marlowe as co-author of all three Henry VI plays. These last two studies will be discussed in the next chapter. 

(16)

2 The Oxford Studies

The Oxford editors refer to two recent studies to support the attribution of 2 and 3 Henry VI to Shakespeare and Marlowe. The first study (Segarra et al., 2016) looks at the authorship of the three Henry VI plays and introduces a new stylometric method. The second study (Burrows and Craig, 2017) focuses on 3 Henry VI and is included in the Authorship Companion.

2.1 Segarra et al.: Word-adjacency networks

Segarra et al. (2016) set out to attribute the Henry VI plays, using what they refer to as word adjacency networks (WANs). While networks of this kind are usually applied to sequences of characters, Segarra et al. base their method around frequent function words. Like most stylometrists, they believe that function words are better discriminators than lexical words, which appear to be primarily determined by genre. What sets their stylometric method apart from others is that it focuses not on frequent function words per se but on the relational structure between them:

We can . . . capture for diﬀerent authors just how far from one another they tend to place each of a large set of preselected words (“target words”) in which we are interested. It turns out that the habits of placement—that is, the choice of words and how far from one another they are placed—vary enough from author to author for this to be a reliable test for identifying authorship when it is unknown or disputed. (Segarra et al., 2016, 235)

In a WAN, the nodes are made up by individual function words. The edges pointing into and out of them represent the likelihood of the words appearing in a particular order. The values of these edges, which are relative rather than absolute, are meant to show that, for instance, Shakespeare is more likely than other playwrights to follow function word x with function word y. The WAN of a target text is compared with various authorial WANs to see which of them shows the most likeness to the target text. Each target word is ranked according to the ranking of the words that are connected to it. In this way, the highly frequent word and is ranked higher than the less frequent beneath. The authors compare this system to Google’s PageRank, which rates the importance of a web page according to the number and quality of links to that page.

Shannon’s entropy is used to calculate the diﬀerence between WANs. In 1950, Claude Shannon presented his method of estimating the entropy and redundancy of a language. Entropy is a measure of the average amount of information that is produced for each letter of a text, while redundancy measures the amount of constraint that a language imposes on a text. Examples of this constraint in the English language are the high frequency of the letter e and the high likelihood of q being followed by u:

(17)

The new method of estimating entropy exploits the fact that anyone speaking a language possesses, implicitly, an enormous knowledge of the statistics of the language. Familiarity with the words, idioms, clichés and grammar enables him to fill in missing or incorrect letters in proof-reading, or to complete an unfinished phrase in conversation. (Shannon, 1951, 54)

Shannon concluded that ordinary literary English is about 75 percent redundant. Segarra et al. use Shannon’s entropy to quantify writers’ preferences for placing words in a particular order and proximity to one another: “Where the symbols are whole words rather than letters, the collocation habits vary considerably from writer to writer, but the way to measure them is the same” (Segarra et al., 2016, 239).

Figure 1 Two WANs for extracts from Hamlet and Satiromastix (Segarra et al., 2016, 236).

Eight candidates were considered for the authorship of the Henry VI plays: Shakespeare, Marlowe, Fletcher, Jonson, Chapman, Middleton, Greene, and Peele. Segarra et al. constructed profiles for each candidate using what they consider to be their “noncontentious, well-attributed, sole-authored plays” (Segarra et al., 2016, 239). Fletcher and Jonson are included, even though they make for “most unlikely candidates for authorship of the play [sic]: Fletcher was in his early teens in the early 1590s and Jonson's earliest known plays did not appear until the end of the 1590s” (Segarra et al., 2016, 244). Thomas Kyd and Thomas Nashe are excluded from the study, because both have only one reliably attributed play to their name.

Segarra et al. tested their method by attributing 94 known plays, treating them as if unknown. They achieved attribution accuracies of 93.6 percent for whole plays and 93.4 percent for acts. The accuracy drops sharply for individual scenes, except when the choice is between two

(18)

authors, in which case the accuracy is 91.5 percent. Hence, they “trust this scene-by scene approach only where existing scholarship gives good reason to consider just two candidates” (Segarra et al., 2016, 244).

The comparison of WANs can point to potential authors by showing which authorial WAN diﬀers least from the target text WAN, or to unlikely authors: “Where the aim is to exclude candidate authors, large diﬀerences in the relative entropies give large degrees of confidence in the exclusion” (Segarra et al., 2016, 240). The most discriminating words—those that most improve the method’s accuracy—vary from comparison to comparison. Segarra et al. included the full list of function words in their appendix, together with the lists of words used for plays, acts, and scenes. These lists understandably become smaller as the analysis moves from full plays down to acts and finally to scenes.

Segarra et al. give the first act of 2 Henry VI to Marlowe and the remainder to Shakespeare. Ten scenes go to Marlowe and fourteen to Shakespeare. They concede that “the run from 4.3 to 4.8 is really too close to call either way” (Segarra et al., 2016, 246). This run includes most of the Jack Cade scenes which Craig ascribes to Marlowe (2009). For 3 Henry VI, thirteen scenes go to Shakespeare and nine to Marlowe. The six remaining scenes are deemed “too close to call” (Segarra et al., 2016, 246). The first act goes to Marlowe and the remaining acts to Shakespeare. The attribution of Act 1 to Marlowe is very strange when three of its four scenes are given to Shakespeare. It would be more understandable if Marlowe only scored slightly lower than Shakespeare for those scenes, but this information is not provided in the article.

In an earlier study, Segarra, Eisen, and Ribeiro observed that combining WANs with frequency-based methods improves the attribution accuracy (Segarra et al., 2015). In the case of 3 Henry VI, Craig and Burrows (2012) had seen to the frequency-based part. Segarra et al. (2016) report that the results of that study are “consistent” with their own, but perhaps this is being too generous. Rather, it seems that the results are for the most part not explicitly at variance: they “agree” for nine scenes, “agree or cannot tell” for twelve, “cannot tell” for two, and disagree for five. This means that they only agree with conviction for one third of the scenes.

Table 1 The level of agreement between Segarra et al. (2016) and Craig and Burrows (2012) in their attribution of 3

Henry VI. See Table 2 (Appendix) for the attributions of individual scenes.

Agreement Number of scenes

Agree 9

Agree or cannot tell 12

Cannot tell 2

Disagree 5

(19)

2.1.1 Strengths and weaknesses

Segarra et al. do not seem to have a problem with the fact that their method is essentially a black box: “No one knows why methods that count frequencies of common words are able to distinguish authorship, and we oﬀer no explanation for why our method of measuring their proximate adjacency is equally successful” (2016, 251). This not-knowing can get in the way of assessing the results, particularly when there are no other results to compare them to. Fortunately, they conducted another study in which they applied the same WAN method to a larger corpus of early modern English plays. In order to find out whether their method could discriminate between genres, they divided the plays into comedies, tragedies, and histories. The results showed that comedies and tragedies were more stylistically similar than histories and tragedies:

This is interesting because it is common to consider history and tragedy more thematically singular than comedy and tragedy. Our results, by contrast, suggest that the writing styles of comedy and tragedy are more closely linked than the writing styles of either comedy and history plays or tragedy and history. (Segarra et al., 2015, 16)

This finding, which is based on function words, challenges the notion that function words vary more by author than by genre. It also points to other possible nuances in understanding an author’s style. There may, for example, be a stylistic diﬀerence between a young Shakespeare imitating Marlowe’s histories and a mature Shakespeare with a more personal style.

Segarra et al. observe that Marlowe’s profile is very dissimilar from Shakespeare’s profile, but ranks “uncharacteristically high” for the Henry VI plays, in addition to scoring well for Henry V, King John, Richard II, and Richard III:

These seven plays have in common that they are history plays, a genre in which Marlowe wrote Edward II and Massacre at Paris, comprising a third of his profile. Thus, there is a genre bias of history plays towards Marlowe. Focusing on the Henry VI saga, . . . we see a particularly strong signature of Marlowe in the three plays compared to Shakespeare’s other plays. (Segarra et al., 2015, 8)

Genre bias is one way of explaining these results. Another possibility is that in the course of writing these seven histories Shakespeare went from being heavily influenced by Marlowe to gradually developing his own style. This would be consistent with most chronologies of Shakespeare’s works, which place Henry V, King John, Richard II, and Richard III (c.1595– 1599) after the Henry VI plays (c.1591–92).

Segarra et al. state that their method is, “as far as we can tell, impervious to mere literary impersonation” (Segarra et al., 2016, 249). This may or may not be true, but there are signs that it is not impervious to genre bias or literary influence. Another possible explanation for

(20)

Marlowe’s ostensible presence is that Shakespeare was imitating Marlowe in the early stages of his career. While Segarra et al. conclude that Marlowe’s presence in the Henry VI plays “is now undeniable” (Segarra et al., 2016, 249), the results seem to be too ambiguous for such a claim. 

(21)

2.2 Burrows and Craig: A multivariate approach

John Burrows and Hugh Craig concluded their previous study on 3 Henry VI with the division of the play into sixteen Shakespeare and twelve non-Shakespeare scenes (Craig and Burrows, 2012). Greene, Kyd, Marlowe, and Peele were identified as being the most likely candidates for the non-Shakespeare scenes. Of the plays that were considered, Kyd’s Spanish Tragedy showed the greatest likeness to the Folio 3 Henry VI as a whole, followed by Shakespeare’s Richard II and Marlowe’s Edward II. The latest study focuses on attributing the non-Shakespeare scenes 1.1–2, 2.3, 3.3, 4.2–9, and 5.2.

Burrows and Craig recognize the diﬃculties associated with 3 Henry VI and other plays written around the same time:

It may be that at this stage of the early modern English drama, the phenomena of imitation, shared materials and dramatic approaches, and mutual influence were unusually dominant in relation to the impulse to establish a distinctive authorial style on which attribution relies. In these circumstances, it is more than usually obvious that attempts at attribution have to rely on presuppositions which simplify the problem and restrict the possibilities to a manageable number. (Burrows and Craig, 2017, 197)

To simplify the problem, they presuppose that any given section of the play was written by one writer at a time, that any revision or corruption did not fundamentally alter the underlying authorial style, and that their division of Shakespeare and non-Shakespeare scenes is accurate. Since most of the authorial canons “vary in size from minimal to modest” (197), Burrows and Craig start with Delta tests on individual plays in order to work around the statistical inequality between the authors. They test 3 Henry VI against 63 non-Shakespeare plays dated between 1580 and 1599.

As determined for this study, Shakespeare’s canon consists of the seventeen plays he is thought to have written before 1600, following the dates of first production from the Database of Early English Playbooks. Marlowe’s canon consists of seven plays, Greene’s of four, and Peele’s of five. In addition to The Spanish Tragedy, the Kyd canon includes Soliman and Perseda and Cornelia, even though the latter is a translation of the French play Cornélie and may carry a competing authorial signal from the original. Arden of Faversham, Fair Em, The Miller’s Daughter, and King Leir, which Vickers attributes to Kyd, are excluded on the grounds that “these attributions are not as yet supported by a detailed study” (201).

Burrows and Craig state their objective as trying to capture both the paradigmatic and syntagmatic aspects of language. Vickers, whose conception of language leans toward the syntagmatic, has openly disapproved of “atomistic” approaches which ignore the fact that “words are not independent but interdependent: one word looks for another.” Dividing a text into single words, Vickers believes, “reduces language to a severely limited lexicon” (Vickers, 2011, 117). The paradigmatic aspect of language is that grammatical structures provide “slots” which

(22)

can be filled in diﬀerent ways. Burrows and Craig try to be as inclusive as possible by acknowledging both aspects.

An earlier study tested which n-grams work best for which methods (Antonia et al., 2014). The results suggest that unigrams perform best for frequency-based methods, strict trigrams for methods using rare markers, and skip bigrams for methods using a defined list of function words. (Skip n-grams owe their name to the fact that they skip over any word that is not in the defined list.) Burrows and Craig incorporate most of these findings in their latest study, with the exception that is it not clear whether the trigrams are indeed strict trigrams.

In their previous study, the Delta test showed that The Spanish Tragedy was most similar to 3 Henry VI proper (Craig and Burrows, 2012). The same Delta test is applied here to the Shakespeare and non-Shakespeare parts of 3 Henry VI. Marlowe’s Edward II is closest for both, with The Spanish Tragedy and Soliman and Perseda respectively coming in second and third for the non-Shakespeare scenes and third and sixth for the Shakespeare scenes. This suggests that there are overall similarities between 3 Henry VI and the two Kyd plays. Unfortunately, it expresses little similarity or diﬀerence between Shakespeare and Marlowe and Kyd, the three plays being more or less equally close to the Shakespeare and non-Shakespeare scenes.

In order to calculate how much of this is attributable to chance, Burrows and Craig divide the diﬀerence between the expected and the observed number of plays for each author by the expected number. Kyd scores highest for both the Shakespeare and non-Shakespeare segments, once again pointing to a general likeness between the two Kyd plays and 3 Henry VI. Marlowe, on the other hand, scores unexpectedly high for the non-Shakespeare scenes and unexpectedly low for the Shakespeare scenes. It becomes diﬃcult at this point to compare Kyd and Marlowe, given their unequal canon sizes. Similarities and dissimilarities can be expected to become more pronounced as the sample size increases.

Following this, Burrows and Craig divide the plays and the two divisions of 3 Henry VI into “rolling” segments of 2,000 words, with the remaining words going into the last segment for every text. The first Zeta test is applied to single non-function words. They compare two sets at a time with the aim of identifying marker words for the various authors. An ideal marker word is one that appears in every segment of the base set and in none of the segments in the counter set. Burrows and Craig chart the percentage of words in a target segment, which can be Shakespeare or non-Shakespeare, to find out to what extent that segment’s vocabulary matches a particular author. Of the four candidates, Marlowe comes closest to showing any kind of overlap with the 3 Henry VI segments.

The next Zeta test is applied to pairings of function words. In this way, “I wonder how the King escaped our hands?” yields the skip bigrams I + how, how + the, and the + our. This allows for more potential markers than single words whilst considering the syntagmatic aspect of language. For every author, they collect the 500 pairings that are commonest in the base set and rarest in the counter set. The results once again point to Marlowe as the likeliest candidate.

(23)

The four non-Shakespeare segments are on the Marlowe side of the perpendicular bisector, as are two of the Shakespeare segments.

The third Zeta test is a direct comparison between Kyd and Marlowe, this time using lexical-word trigrams. Whether these are strict or skip trigrams is not entirely clear: “We excluded function words and proper names and found all the trigrams in the remaining, lexical words, but this time passing over function words to find sequences of three lexical words” (Burrows and Craig, 2017, 210). Without exception, the 3 Henry VI segments are closest to Marlowe.

The fourth Zeta test is a direct comparison between Shakespeare and Marlowe, the only candidate to show “persistent and marked likeness” to the non-Shakespeare scenes of 3 Henry VI. Marlowe’s extended canon of seven plays includes the suspected collaboration Doctor Faustus, as well as Dido and Aeneas, which is ascribed to Marlowe and Thomas Nashe on the title page. Burrows and Craig removed the comedies from Shakespeare’s pre-1600 corpus “to make a better generic fit with the Marlowe canon,” leaving six tragedies and two histories (Burrows and Craig, 2017, 210).

Burrows and Craig identify 1,000 marker words for both authors and run a validation test by dividing the Marlowe and Shakespeare plays into 2,000-word rolling segments and anonymizing them each in turn. 83.1 percent of the Marlowe segments are correctly attributed. The Jew of Malta was a “spectacular failure,” with only 29 of its 81 segments attributed correctly. The authors comment:

The test has some power to recognize a Marlowe segment, presented as part of an anonymous play, as Marlowe not Shakespeare, though it is far from infallible. The test can be deceived by a play like The Jew of Malta, which evidently departs from the Marlowe style as represented by the other six plays, and which intuitively might be regarded as more confessional, more farcical, and more comic than they are, though still Marlovian. (Burrows and Craig, 2017, 212)

The Zeta test’s failure to correctly attribute The Jew of Malta seems to be an anomaly: the attribution accuracy for Marlowe’s other six plays is 93 percent. For seven random pre-1600 Shakespeare plays, all but one of its 667 segments were correctly ascribed. Burrows and Craig gather from this that Shakespeare’s corpus is more homogeneous than Marlowe’s.

Burrows and Craig run the same test on the 1623 Folio and 1595 octavo versions of 3 Henry VI. It is not clear why they chose to include the octavo version at this particular stage of the experiment, but the results show a broad similarity between the two texts. The Zeta test assigns 54 percent of the Folio segments and 55 percent of the octavo segments to Shakespeare. These assignations hold up fairly well in comparison to the earlier results, which identified 21 of the 27 Shakespearean segments and 16 of the 27 Marlovian segments as being respectively Shakespeare and non-Shakespeare.

Finally, Burrows and Craig run some comparisons between Shakespeare’s pre-1600 tragedies and histories and the extended Marlowe and Kyd canons. The Delta test shows the Shakespeare

(24)

plays to be most diﬀerent from the Shakespeare and non-Shakespeare scenes alike for both the Folio and octavo versions of 3 Henry VI. Marlowe is closest to the non-Shakespeare scenes of the octavo 3 Henry VI, while Kyd has the lowest scores for the remaining three blocks of scenes. In one-on-one Zeta comparisons, all four blocks are closer to Marlowe than to Kyd and closer to Shakespeare than to Kyd. Between Shakespeare and Marlowe, the Shakespeare scenes conveniently go to Shakespeare and the non-Shakespeare scenes to Marlowe.

2.2.1 Strengths and weaknesses

Burrows and Craig conclude that Kyd is an unlikely candidate for the authorship of 3 Henry VI, “though certainty is diﬃcult to attain, given his small surviving canon.” They are more confident in excluding Greene and Peele. They restate that the scenes assigned to Shakespeare “include the most memorable in the play.” Of the non-Shakespeare scenes, they only mention one instance that “does come close to obviously Marlovian flamboyance” in the lines:

How sweet a thing it is to wear a crown,  Within whose circuit is Elysium  And all that poets feign of bliss and joy  (1.2.29–31)

Wilson noted that these lines echo and reference 1 Tamburlaine (Wilson, 1952, 134), but Burrows and Craig infer that it must have been Marlowe himself who wrote these lines. Still, these three lines are scant evidence when half of the play is attributed to Marlowe.

Throughout the study, the Delta method repeatedly points to Kyd as the most likely author, but the Zeta method repeatedly does not. Burrows and Craig attribute this to the fact that the Delta test does not make a preselection of authorial markers. In an earlier study, John Burrows applied the Zeta method to seventeenth-century poetry: “Most poets usually write within their customary repertoires. Whenever they do so, the Zeta test shows high levels of accuracy. When they do not, a reader should know it” (Burrows, 2007, 44). Sadly, this does not translate to early modern drama.

The main issue with Kyd’s canon is that it is much too small to derive a customary repertoire from it, or for its lexicon to diﬀerentiate itself from those of the other authors, as seems to be the purpose of the Zeta method. Burrows and Craig themselves note that Marlowe and Shakespeare are the only candidates whose “canons are of suﬃcient size to allow some cross-validation” (Burrows and Craig, 2017, 197). Even Shakespeare’s pre-1600 corpus is larger than that of any of the other canons. Marlowe’s canon includes one suspected and one apparent collaboration, while Kyd’s canon for this study consists of one translation and two original plays, neither of which are directly attributed to him. The question remains how Kyd would fare in these tests if the plays attributed to him by Vickers were included.

(25)

2.3 Discussion

The two studies conclude with varying degrees of confidence. Segarra et al. state that their study has made Marlowe’s presence in the plays “undeniable.” This would be acceptable if one stretches the meaning of “presence” to include influence and imitation. But assuming that it refers to authorship, it seems overly confident. One obvious diﬀerence between the two studies is the fact that Burrows and Craig spent many years fine-tuning their methods, whereas Segarra et al. published their first paper on the WAN method in 2015. It should also be noted that Egan is the only Shakespearean involved, the other three researchers being engineers. As a consequence, proving the validity of the WAN method may have taken precedence over the authorship question itself. This may help to explain the finality of their conclusion. Burrows and Craig conclude in a more prudent manner: “We submit our necks to posterity, the great executioner, with such grace as we can muster” (217).

The direct comparison between Shakespeare and Marlowe demonstrates two things. First, the results can be confounding and inconvenient. In Burrows and Craig’s study, we cannot be sure that Zeta’s “failure” to properly attribute The Jew of Malta is in fact a failure. To say that the Zeta method has been “deceived” by an atypical Marlowe play hinges on the assumption that Marlowe is in fact the sole or at least principal author of that play. Burrows and Craig understandably go with this interpretation, Marlowe’s canon being small enough as it is. To say that the results point toward a whole new authorship problem would complicate matters even more. However, this does not eliminate the possibility that people other than Marlowe were involved with the writing of The Jew of Malta. The fact that Zeta achieved an attribution accuracy of 93 percent for the remainder of the Marlowe canon—and 99.9 percent for the Shakespeare canon—certainly speaks for its validity, in addition to reaﬃrming that the size of Shakespeare’s canon improves the attribution accuracy.

The Kyd canon is even smaller and more problematic than the Marlowe canon. It may not be a stretch to say that Kyd did not stand a chance from the start. But again, the fact that the accepted Kyd canon did not survive this study’s elimination process does not thereby exclude Kyd as a candidate. Again, the outcome might be diﬀerent if Vickers’ attributions were included.

2.3.1 Rudman’s objections

A more general critique of Shakespeare attribution studies worth mentioning in this context is provided by Joseph Rudman. Having written several surveys of attribution studies, Rudman is not entirely optimistic about the attribution of the Shakespeare canon. Among the reasons given are lack of data, lack of scientific rigor, and most importantly, the nature of the texts themselves:

If this were a court of law, the evidence (First Folio) would be inadmissible as the chain of custody was broken (shattered!) and the potential for contamination too great. Continuing in

Better Solutions to Old Problems?