Bernard of Clairvaux and Nicholas of Montiéramey: Tracing the Secretarial Trail with Computational Stylistics

Although the past few decades of medieval studies have witnessed some renewed interest in the collaborative process by which medieval Latin prose was composed, such interest has nevertheless remained all too scant, and only few solutions have been offered to cope with the difficulties that rise in cases of dubious authorship. Whereas it has been rightly acknowledged that scribes, notaries, and secretaries should be regarded not merely as instrumental in the literary process, but as active participants in the composition process who have left a considerable impact on the image and style of the dictator and on the materialization and dissemination of the text, this acknowledgement has nevertheless been accompanied by difficulties

1 A more general introduction to twelfth-century notaries, especially in respect to epistolography, can be found in Giles Constable, "Dictators and Diplomats in the Eleventh and Twelfth Centuries: Medieval Epistolography and the Birth of Modern Bureaucracy," Dumbarton Oaks Papers 46 (1992): 37-46, at 38, where he stresses that "[notaries] took on a new importance in the eleventh and twelfth centuries, when they formed a distinct group of recognizable personalities whose activities extended beyond the scriptorium."Lynn Staley Johnson, albeit focusing mainly on late medieval female authors, has drawn attention to how "scribes not only left their marks upon the manuscripts they copied, they also functioned as interpreters, editing and consequently altering the meaning of texts.Writers, however, did not simply employ scribes as copyists; they elaborated upon the figurative language associated with the book as a symbol and incorporated scribes into their texts as tropes," in "The Trope of the Scribe and the Question of Literary Authority in the Works of Julian of Norwich and Margery Kempe," Speculum 66 (1991): 820.
2 See Bernard Cerquiglini, In Praise of the Variant: A Critical History of Philology, trans.Betsy Wing (Baltimore, 1999); Stephen G. Nichols, "Introduction: Philology in a Manuscript Culture," Speculum in defining the exact extent and sphere of influence of such secretarial mediation.This challenge is especially marked in the oeuvre of Bernard of Clairvaux (1090-1153).By 1145, the abbot's acclaim as the icon and figurehead of the Cistercian movement had brought along such a considerable administrative workload that the assistance of a group of secretaries-which, it could be argued, amounted to a kind of chancery3 -was indispensable.These secretaries acted as Bernard's standins and spared him the time and effort it would have cost if he had had to take up the quill himself at every single occasion. 4The reportatio, as it was called, entailed that the contents of Bernard's letters or sermons be engraved on wax tablets in a tachygraphic fashion.The cues, keywords, and biblical references that Bernard had spoken aloud provided a framework that captured the gist of his diction. 5Afterwards, the scribe reconstructed what he had heard as a text on parchment, which could pass for Bernard of Clairvaux's in its literary allure.Among these amanuenses, Nicholas of Montiéramey ( † 1176/78) was a focal figure and a highly skilled imitator of his master's writing style.The influence of Nicholas's mediation on several particular texts within Bernard's corpus, and more generally on his entire oeuvre, has been subject to much debate.This article revisits the authorship of a selection of texts from Bernard's corpus.A detailed listing of the texts under scrutiny can be consulted in the Appendix of Tables (Tables 3-11).Generally, the corpus comprises Nicholas of Montiéramey's letters and sermons 6 and Bernard of Clairvaux's letter corpus (Corpus epistolarum), Sermones de diversis (hereafter De diversis), and Sermones super Cantica canticorum. 7s a means of determining the extent of Nicholas's stylistic presence in the aforementioned works, this article advocates computational stylistics (or stylometry), which is a method that detects stable and recurring patterns of writing style in texts that have been reduced to a range of marked style features.These features are consequently quantified to numerical data in order to gain objective and measurable ground for distinguishing among works of varying authorship.Computational stylistics claims to offer a scientific and objective "distant reading" 8 of literature, as opposed to human expert-based methods, which are often liable to intersubjectivity. 9A stylometric approach to the case of Bernard of Clairvaux and Nicholas of Montiéramey will prove rewarding on two levels.On the first level, we will determine the authorship of the aforementioned texts, some of which have been subject to long-standing dispute.This will address an elementary concern within Bernardian studies, which Gillian Rosemary Evans has aptly put forward as such: "A turn of phrase can bear only so heavy a load of interpretation as it exactly reflects the author's thought." 10 On a second level, however, this study should also raise awareness of how the (mis)attributions of disputed medieval texts, as they have occurred in earlier studies, are often guided by personal intuitions or theoretical convictions of medieval authorship. 11This implies that, as we begin to challenge the established 7 These works are edited in the Sancti Bernardi opera (hereafter SBO), ed.Jean Leclercq et al., 8 vols.(Rome, 1957-77): the Corpus epistolarum (vols.7-8), Sermones de diversis (vol.6), and Sermones super Cantica canticorum (vols.1-2). 8The term "distant reading" was coined by Franco Moretti, "Conjectures on World Literature," New Left Review 1 (2000): 54-68, and was further developed in his monograph Graphs, Maps, Trees: Abstract Models for Literary History (Brooklyn, 2005). 9It was Frederick Mosteller and David L. Wallace's influential study on the disputed authorship of the eighteenth-century Federalist Papers in the early 1960s that would launch statistical approaches as tools by which to objectively determine authorship, currently known as nontraditional authorship attribution: see Frederick Mosteller and David L. Wallace, Applied Bayesian and Classical Inference: The Case of The Federalist Papers (New York, 1964).For four excellent state-of-the-art surveys on the history of nontraditional authorship attribution and the current debate within the field, see Patrick Juola, "Authorship Attribution," Foundations and Trends in Information Retrieval 1 (2008): 233-334; Efstathios Stamatatos, "A Survey of Modern Authorship Attribution Methods," Journal of the Association for Information Science and Technology 60 (2009): 538-56; Moshe Koppel, Jonathan Schler, and Shlomo Argamon, "Computational Methods in Authorship Attribution," Journal of the Association for Information Science and Technology 60 (2009): 9-26; and Walter Daelemans, "Explanation in Computational Stylometry," in Computational Linguistics and Intelligent Text Processing 7817, ed.Alexander Gelbukh (Berlin, 2013), 451-62. 10Gillian Rosemary Evans, Bernard of Clairvaux, Great Medieval Thinkers (New York, 2000), 20. 11 The bibliography of scholarship on medieval authorship is extensive and cannot be listed here in full.The following titles should nevertheless point any reader who is interested in the theory of medieval authorship in the right direction.A major work of reference is by Alastair J. Minnis, Medieval Theory of Authorship: Scholastic Literary Attitudes in the Later Middle Ages (Philadelphia, 1984), whose approach to medieval authorship is largely based on study of the prologues of commentaries and exegetical works especially from the later Middle Ages (the scholastic period).In this later period he describes how the schools defined for themselves a framework of literary theory from the newly translated Aristotelian logic, through which they could approach the biblical texts and patristic auctores more literally, and therefore more literarily, as "a new type of exegesis emerged, in which the focus had shifted from the divine auctor to the human auctor of Scripture," 5. Some other indispensable publications on medieval au-authorship of some of these texts, we also may offer metahistorical reflections on how earlier, more intuitive scholarly approaches have both enriched and yet predetermined our current understanding of medieval authorship.The case of Bernard and Nicholas is particularly suitable to demonstrate how computational stylistics provides new answers to old questions, and raises new questions about old problems.

Nicholas of Montiéramey
The daily routines and workings of Clairvaux's chancery are rather poorly documented.We rarely know any of the scribes by name, and for those whom we doa select group of six-only three give us a faint clue of their specific tasks and responsibilities. 12Nicholas began serving Bernard as an emissary around 1138-41, carrying letters concerning Abelard's heresy to Rome.At this time he was still chaplain of Hato, bishop of Troyes, and Peter the Venerable's friend and secretary, but he must already have been collaborating with Bernard from 1140 onwards. 13He would officially become a monk at Clairvaux around the end of 1145.His literary qualities, likely to have been acquired through his education in the Benedictine abbey of Montiéramey, 14 enabled him to enter the scriptorium immediately and officially become Bernard's closest secretary.He appears to have been responsible for supervising the workings of the chancery, 15 and he may have been the monastery's librarian. 16But their friendship knew an abrupt and painful ending in the final years of Bernard's life, around 1151-52, when Nicholas must have severely breached his master's trust.In a letter to Pope Eugene III, we find Bernard disconcerted over the fact that letters had been sent out under his name and seal by "false brethren" without his permission. 17Later, Bernard would identify Nicholas as the culprit among these brethren, 18 although the exact reasons why the latter deserved this accusation are nowhere explicitly disclosed. 19In any case one can assume from his correspondence and his own words 20 that Nicholas's talent as a writer and his "versatility" ingratiated him with the greatest men of his time. 21Equally so, Nicholas appears to have had-perhaps through this flamboyance and self-confidence-a talent for making enemies as well. 22 1152 and assisted Bernard with his sermons as well as his letters, but he continued to visit Cluny and to serve Peter the Venerable, one of whose letters, we have seen, he presented to Bernard orally"; and Anne-Marie Turcan-Verkerk, "L'introduction de l'ars dictaminis en France: Nicholas de Montiéramey, un professionel du dictamen entre 1140 et 1158," in Le dictamen dans tous ses états: perspectives de recherche sur la théorie et la pratique de l'ars dictaminis (XIe-XVe siècles), ed.Benoît Grévin and Turcan-Verkerk (Turnhout, 2015), 70: "Ami de Pierre le Venerable, [Nicolas] avait déjà servi les intérêts de Bernard en portant au pape, en 1140-1141, des lettres concernant Abelard-à la rédaction desquelles il avait peut-être déjà participé, comme le suggère le manuscrit Phillipps 1732. . . .Trois billets de recommandation envoyés par Bernard à Innocent II entre 1138 et 1143 semblent le concerner [Epp.434-36], et montrent que s'il servait Hatton, il le faisait en obéissant à Bernard." 14 "Nicolas fit ses études à l'abbaye bénédictine de Montiéramey, près de Troyes en Champagne.On parle souvent de lui comme d'un Magister": see John Benton, "Nicolas de Clairvaux," in Dictionnaire de Spiritualité, 17 vols.(Paris, 1982), 11:256-59. 15Leclercq, "Lettres de S. Bernard: Histoire où littérature?,"Recueil d'études, 4:148. 16Giles Constable, "Nicholas of Montiéramey and Peter the Venerable," in Constable, The Letters of Peter the Venerable, 2 vols.(London, 1967), 2:321.
17 "Periclitati sumus in falsis fratribus," Bernard, Ep. 284, Ep. 298, likewise addressed to Eugene III (SBO 8:214).19 Constable, The Letters of Peter the Venerable, 2:327. 20Nicholas, Ep. 56 (PL 196:1652), "Ab ineunte aetate mea placui magnis et summis principibus hujus mundi."21 The word is Jean Mabillon's in PL 183:26: "Vir fuit ingenii facilis, versatilis, facile in aliorum affectus influens." 22 An example can be found in Nicholas's dispute with Peter of Celle.The two "were at odds over a substantive matter, a theological point about how to treat the attributes of God, and Abbot Peter took offense that Nicholas, who should have possessed the power to triumph with his own (verbal) arms, had used against him the authority and words of great philosophers:" see John Van Engen, "Letters, Schools, and Written Culture in the Eleventh and Twelfth Centuries," in Dialektik und Rhetorik im früheren und hohen Mittelalter: Rezeption, Überlieferung und gesellschaftliche Wirkung antiker Gelehrsamheit vornehmlich im 9. und 12. Jahrhundert, ed.Johannes Fried (Munich, 1997), 27:109; and Julian Haseldine, "Peter of Celle and Nicholas of Clairvaux's Debate on the Nature of the Body, the Soul, and God," in The Letters of Peter of Celle, ed.Haseldine (Oxford, 2001), 706-11.The "versatile" aspect of Nicholas's personality, as noted by Mabillon (above, n. 21), therefore also seems to display itself on the level of language.Nicholas was accused of "inverting words and their meaning," a

Bernard of Clairvaux and Nicholas of Montiéramey
Speculum 92/S1 (October 2017) The scandal at Clairvaux and the breach of Bernard's trust has for a long time upheld the portrayal of Nicholas as a disreputable Judas by Bernard's side, an analogy for which Bernard himself was responsible. 23Conversely and simultaneously, Bernard's status as a saint continued to grow during the intense process of canonization and idealization following his death. 24These respective caricatural depictions, in which Nicholas was deplored as the mistrusted secretary and Bernard praised as the saint who had become victim of textual theft, show through on an academic level as well.Dom Jean Leclercq, one of the most prominent Bernard scholars of the twentieth century, was as relentless as Bernard in accusing Nicholas of deceit, shamelessness, and plagiarism. 25Nicholas's most striking example of seeming textual theft presents itself in his letter to Henry the Liberal, count of Champagne, to whom he humbly offered his services as a secretary shortly after his expulsion from Clairvaux.Accompanying the letter we find nineteen sermons originally attributed to Peter Damian, 26 nine sermons attributed to Bernard of Clairvaux, 27 and seventy-four short commentaries to the Psalms that are ascribed to Hugh of St. Victor. 28In the letter, Nicholas asserts that these writings are "of my invention, of my style, aside from what I have taken from others in a few places." 29We know this assertion to be true of the nineteen sermons also found among those of Peter Damian, which have been identified by Leclercq as stemming from Nicholas.Bernard's and Hugh's writings, on the other hand, appear to have been copied almost literally, not merely rearranged or paraphrased "in a few places" (paucis in locis), as Nicholas seems to suggest.Most of the nine sermons can be found in Bernard's De diversis.It is striking that, months characteristic that Peter interpreted as equal to a falsification of language: "verba quoque et sensus verborum praesumis quandoque invertere": see Peter of Celle, Ep. 66 (PL 202:512). 23Bernard literally made the analogy with Judas, which he significantly did not make often in his letters: see Brian Patrick McGuire, "Loyalty and Betrayal in Bernard of Clairvaux," in Loyalty in the Middle Ages: Ideal and Practice of a Cross-Social Value, ed.Jörg Sonntag and Coralie Zermatten (Turnhout, 2015), 317-18. 24As it was first initiated by his biographer William of St. Thierry ( † 1148).He wrote the Vita prima Bernardi, a biography with a hagiographical, panegyrical slant.He shares the authorship of the entire Vita with Bernard's secretary Geoffrey of Auxerre and the Benedictine abbot Arnaud de Bonneval.Geoffrey was a strong advocate for Bernard's canonization, in which William's texts played a fundamental role.
Such incriminating evidence contributed to his reputation as a plagiarist, this reputation in its turn provoking prejudicial conclusions in other attribution issues.Henri M. Rochais, for instance, in a codicological approach to the question of determining the disputed authorship of three other lengthy sermons in Bernard's De diversis corpus (De diversis 40, 41, and 42),30 pointed out these sermons' close similarities to two of Nicholas's works and to chapter 100 of Hugh of St. Victor's sixth book of Miscellanea (the well-known writer somehow seems to be involved again);31 yet stated with confidence that Nicholas stole the texts from Bernard under false pretenses.This hypothesis Rochais sees corroborated in "the secretary's unscrupulous personality." 32At the same time Rochais casts aside Jean Mabillon's belief that the literary style of these sermons hardly seems that of Bernard as an all-too-subjective and unscientific argument. 33To our view, Rochais' own subjective mistake was thatdespite being fully aware of Bernard's collaboration with his secretaries-he treated codicological unity as identical to stylistic or authorial unity: "Cette tradition manuscrite ne donne donc aucun motif de doute sur l'authenticité bernardine des trois sermons étudiés, et, au contraire, elle constitue une telle probabilité en faveur de cette authenticité, qu'il faudrait des arguments incontestables pour dénier à Bernard leur composition." 34eclercq's and Rochais' attributions still stand in their editions, widely used today, although medievalists have seriously contested their highly subjective and speculative approach towards authorship attribution and their prejudiced view of Nicholas of Montiéramey's alleged deceitfulness and falsification, as we will show below.Moreover, the temptation for scholars to draw lines between imitation and plagiarism in order to categorize writings and collate them in attributed editions, valuable as it is, can also be rather anachronistic or even unbefitting in a medieval context.A fundamental rationale of the New Philology, in the wake of poststructuralist approaches to authorship and texts, is that in a medieval culture there is no place for the idea of an original author, a logic leading to the conclusion that Leclercq's and Rochais' quest for such an author only takes us further from the truth. 35Medieval lit-erature depended strongly on its oral component (dictare had supplanted scribere to designate the act of authorship); 36 it "circulated through networks" and formed part of a "shared culture [that was] characterized by a knowledge of the same erudite language as well as a common foundation of texts and memories." 37Giles Constable touched on the core of the issue by noting that in the Middle Ages "there are infinite shadings between correction, revision, imitation, and falsification and, in works of art, between repair, restoration, reproduction, and copying." 38In this light, Nicholas's appropriation of some of his former master's works in his letter to Henry the Liberal is rather the continuation of a dialogue, 39 not a spiteful act of revenge.Stephen Jaeger has similarly argued that Nicholas indulges in the kind of imitatio that would have made little distinction between "honest" and dishonest intentions. 40ike any distinguished writer of his time, Nicholas carefully applied for a new position by showcasing his complete immersion in a prevalent literary network. 41he juxtaposition of Leclercq's and Rochais' historical positivism with the more recent New Philology lays bare the dilemma that has arisen in medieval text studies.Although both practices have contributed immensely to the field, neither of the two stances is entirely satisfactory, leaving most scholars to agree to a compromise in cases of doubtful authorship.The first, rather positivist, approach acknowledges that the act of textual appropriation is suitable and possible.It presupposes that personal authorship is a retrievable aspect of the text, whose idealized state can be reconstructed from a hierarchical stemma.The disadvantage of this approach is that it of the Author," Aspen Magazine 5-6 (1967); and Michel Foucault, Bulletin de la Société française de philosophie (1969): 73-104; or in Dits et écrits, vol. 1, 1954-1988, ed.Daniel Defert, François Ewald,  and Jacques Lagrange (Paris, 2001), 817-49, is apparent in this New Philological approach towards medieval authorship.Also see Virginie Greene, "What Happened to Medievalists after the Death of the Author?," in The Medieval Author in Medieval French Literature, ed.Greene (New York, 2006): 205-27.
37 Pascale Bourgain, "The Circulation of Texts in Manuscript Culture," in The Medieval Manuscript Book: Cultural Approaches, ed.Michael Johnston and Michael Van Dussen (Cambridge, UK, 2015), 150.Also see Rebecca Moore Howard, Standing in the Shadow of Giants: Plagiarists, Authors, Collaborators (Stamford, CT, 1999), 65: "In the Middle Ages, mimesis was the means of establishing one's authority, as well as being an expression of humility.The notion of the individual author, autonomous, original, and proprietary, played only the smallest role in this economy of authorship.With those textual values so much in decline, plagiarism was hardly an issue." 38 Giles Constable, "Forgery and Plagiarism in the Middle Ages," in Constable, Culture and Spirituality in Medieval Europe (Aldershot, 1996), 1-41, at 3. 39 Moreover, the assertion that Bernard never heard of Nicholas again after he left Clairvaux is far from certain: see Constable, The Letters of Peter the Venerable, 2:330.
40 "Here then is a case in which a 'skilled student of the ars dictaminis' with alleged inclinations to forgery imitated a near-contemporary model, and we can assume that there would have been little difference between the 'honest' and dishonest imitation of Bernard's style," in Stephen Jaeger, "The Prologue to the Historia calamitatum and the 'Authenticity Question,'" Euphorion 74 (1980): 13.
41 After all, Nicholas was applying for a position as Henry's new secretary, and a familiarity with the greats of the twelfth century would have been one of the prerequisites.Leclercq's assertion that Henry the Liberal must not have noticed Nicholas's blatant plagiarism because he was a layperson unfamiliar with clerical texts seems unlikely: "Il était moins facile à un laïc comme Henri le Libéral qu'à un clerc de déceler le plagiat," in Leclercq, Recueil d'études, 1:57.Henry's recognition of the extent to which Nicholas's compositions were indebted to other authorities would have been the whole point and is likely not to have been conceived of as problematic.
reduces authorship debates to anachronistic, binary classification problems, whose conclusions often lack substantial evidence and set into motion more unsubstantiated debate.Therefore it both originates in bias and establishes a circulus in probando in which only additional factual evidence can provide closure.The second and more recent approach, on the other hand, comes to terms with the impossibility of closure through its recognition that-in a medieval context-knowing a text's authorship was subordinate either, on the one hand, to acknowledging its implied authority (for example, the authority of a writer's predecessors, such as the church fathers or God); or, on the other, to acknowledging the authority of its unique, materialized appearance: "There are as many texts . . .as there are scribal redactions; and there are as many authors of medieval texts as there were scribes composing new works in the act of writing them down in their [manuscripts]." 42By embracing the variance, this approach evades the impasse.Yet one is wary of where this might lead.Lena Wahlgren-Smith, who is preparing a critical edition of Nicholas of Montiéramey's letters, has quite rightfully expressed her concern regarding a "wholesale adoption" of the New Philological approach, which "assumes that all medieval literature, in all languages, all genres, and all periods, operates in the same way." 43Such an approach is counterintuitive to those medieval attestations where value is attached to titled authority, where there is an outspoken preference for unviolated text, or where personal literary style is cultivated. 44Bernard's denunciation of Nicholas for sending out texts without his consent serves as a firsthand example.Correspondingly, Nicholas's bold statement that Bernard's texts are in fact his own-"meo sensu inventos, meo stylo dictatos" 45 -also suggests that an explicit appropriation of texts was not unknown in the twelfth century.It is important not to underestimate the degree to which the Middle Ages was a "charismatic culture" in which texts were regarded in relation to "the body and the physical presence [that were] the mediators of cultural values." 46From this perspective, Leclercq and Rochais had justifiable reasons to care about the interdependence of text and physical author (or performer).Constable has referred to a "rising tide of concern" over textual theft in the late twelfth and thirteenth century, possibly instigated by rapidly changing approaches to "literary individuality." 47In the midst of such an impasse, historians and 46 Jaeger, "Charismatic Body, Charismatic Text," 122. 47Constable, "Forgery and Plagiarism," 18, 32.He also noted that Nicholas "may have inspired the Cistercian legislation of 1157 defining the punishments for the falsifiers of charters and seals," 19.However, the idea of literary individuality in the twelfth-century Renaissance that Constable here addresses is a subject of immense debate: see Caroline Walker Bynum, "Did the Twelfth Century Discover the Individual?," Journal of Ecclesiastical History 31 (1980): 1-17; Colin Morris, The Discovery of the Individual, 1050-1200 (New York, 1972); and Aron Iakovlevič Gurevič and Katharine Judelson, The Origins of European Individualism (Oxford, 1995).

Bernard of Clairvaux and Nicholas of Montiéramey
Speculum 92/S1 (October 2017) philologists should hope to find more solid ground on which to establish the authorship of dubious texts.Authorship attribution-a seemingly trivial question concerning who wrote which text-forms a vital stepping-stone towards a more scientifically responsible understanding of how the medieval author perceived the proportional relationship between one's personal authority over a text and one's personal contribution to its composition.

Computational Stylistics
Instead of treating the medieval text as collective and impersonal, the methodology of computational stylistics traces the linguistic traits of a text that originate from a highly individual stylome, a set of features that betray personal writing style (note here that mapping out individual stylistics should not eliminate the possibility of collective authorship). 48Moreover, its ambitions-whether realistic or not-may extend to extracting information concerning the author's sex or age 49 or to measuring a text's genre or degree of "literariness." 50Although experiments and debates as to which textual features best capture stylistic difference are still ongoing, many state-of-the-art studies employ function words, which still prove to be the most robust discriminators for writing styles.Function words are usually short and insignificant words that pass unnoticed-such as pronouns, auxiliary verbs, articles, conjunctions, and particles-whose main advantages are their frequent occurrence, their less conscious use by authors, and their content-or genre-independent character.Their benefit and success for the study of stylometry in Latin prose have been convincingly demonstrated before, 51  questions, which keep stylometrists on the lookout for alternatives. 52Before returning to our case study of Nicholas of Montiéramey and his alleged falsification, we will expound on the availability of the data, on the preprocessing steps involved, and on the statistical technicalities of computational stylistics.
For our subsequent analysis, we relied upon the digitized texts of Bernard of Clairvaux's Corpus epistolarum, Sermones de diversis, and Sermones super Cantica canticorum as they appear in the state-of-the-art scholarly edition of the Sancti Bernardi Opera by Leclercq et al. included in the online Brepols Library of Latin Texts. 53For Nicholas of Montiéramey's letters we are provisionally still reliant on the digitally available Patrologia Latina. 54All text data are available in an online GitHub repository for experimental replication, yet in a camouflaged form so that the copyright protection on the original text editions is respected.Only the texts' function words were retained in their original form, whereas all content-loaded words were filtered out and replaced by dummy words. 55Since Leclercq's editions and the Patrologia Latina make use of different orthographical conventions, and since Latin is a synthetic language with a high degree of inflection, Bernard and Nicholas's texts required some preprocessing for the sake of data alignment and feature culling.The result of the latter is that texts are more easily mined for information: thus, the lexemes are lemmatized (which means that a specific instance of the word is referred to its headword) and a text's words (tokens) are classified according to grammatical categories (parts of speech).For this purpose we applied the Pandora lemmatizer tagger on the texts, a piece of software developed to achieve specifically this. 56ken Lemma PoS-tag (simplex) harum hic PRO imo immo ADV An advantage of Pandora's design is that it normalizes orthographical variants to a classicized headword if necessary.Both imo and immo were categorized under immo and identified accordingly as one and the same word. 57The part-of-speechtag (PoS) displayed in the third column in the above diagram allowed us to restrict the culling of the most frequent words to those word categories that make up the collection of function words: conjunctions (CON), prepositions (AP), pronouns (PRO), and adverbs (ADV).This likewise filtered out some noise caused by ambiguities or homonyms like secundum, which can be either a preposition or the accusative case of the adjective secundus.Afterwards, some lemmata in the list that did not qualify as style markers were culled and filtered out, such as tu, tuus, vos, and vester.Vos and vester betray a vernacular influence in Bernard's unpublished letters as formal, polite forms similar to the French vous.Bernard was known to adapt his sermons to his audience, 58 and in a literary or classicizing text he would maintain tu and tuus when addressing his correspondent.Therefore these pronouns were not regarded as suitable features for stylistic difference but as content-dependent, conscious authorial choices linked to register. 59Aside from lemmatization, smaller interventions were undertaken, such as separating the enclitics -que and -ve from the token in order to be recognized as a feature.Once procedures of this sort were carried out in full, we arrived at a list of the 150 most frequent function words (MFFW) of the corpus examined in our experiment. 60Tables 1 and 2 of the most frequent function words correspond to the two experiments (and their two respective corpora) described in this article and are listed in the appendix.
The corpora under scrutiny were subsequently segmented into parts with a fixed size, in other words, text samples.Sampling yields the advantage of "effectively [assessing] the internal stylistic coherence of works," 61 as it also allows for a more finegrained comparative analysis with segments from external works.The sample sizes, however, can differ depending on the requirements of the experiment.As will become apparent in the appendices and the figures, Bernard's letters were segmented into 3,000-word samples, whereas his sermons were segmented into 1,500-word samples.The decrease of sample size in the second experiment was necessary due to the fact that it treats shorter texts.It should be noted that whereas 3,000-word 57 Other pairs include tanquam and tamquam, quoties and quotiens, nunquid and numquid, quanquam and quamquam, nunquam and numquam, etc. 58 Constable, "The Language of Preaching in the Twelfth Century," 131-52. 59See Leclercq, "Notes sur la tradition des épitres de S. Bernard," Recueil d'études, 3:317.It should be noted that our decision to disregard these features, or in other words to succumb to manual feature selection, implies a degree of supervision and subjectivity.We do not really see this as a problem, firstly since we have supplied evidence that including these features could only distort the results for obvious historical reasons, and secondly since this is in line with our approach that was already strongly determined by how we set limitations to the culling of function words (only prepositions, conjunctions, adverbs, and pronouns were taken into consideration).The omission of too-characteristic corpus features considerably improves precision and historical validity.David L. Hoover has demonstrated this for personal pronouns as well: see Hoover, "Testing Burrows's Delta," Literary and Linguistic Computing

S201
Speculum 92/S1 (October 2017) samples correspond to a state-of-the-art norm, 1,500-word samples run the risk of increased imprecision, a consideration that should nuance any interpretation of the results. 62Once we divided our corpus, each of the text samples needed to be translated to a format that is also readable to computers, namely document vectors.A text sample is represented as an array of tallies for each of the one hundred and fifty function words on the checklist, as follows: These raw counts were TF-IDF normalized, a procedure that divides the function word frequencies by the number of text samples that respective function word appears in.As a consequence, less common function words received a higher weight, which prevents them from sinking away (and losing statistical significance) in between very common function words. 63Once the data was preprocessed and regulated, two statistical techniques were applied to visualize its dynamics.
The first is k Nearest Neighbors (hereafter k-NN); the second is principal component analysis (hereafter PCA).Their respective results will prove to be similar in a general sense, yet crucially different in the details.We argue that such an additional statistical validation provides for a more accurate, nuanced interpretation and a better intuition of the data.In Figs. 1 and 3, the k-NN networks, we first calculated the five closest text samples to each text sample by applying k-NN on the frequency vectors. 64Accordingly, for each text the five most similar, or closest, texts were 62 There is debate over the adequate sample length to capture a stylistic signal.The risk has been addressed for twelfth-century Latin in Kestemont, Moens, and Deploige, "Collaborative Authorship in the Twelfth Century," 210-11.Also see Maciej Eder, "Does Size Matter?Authorship Attribution, Small Samples, Big Problem," Literary and Linguistic Computing 30 (2015): 167-82; and Kim Luyckx and Walter Daelemans, "The Effect of Author Set Size and Data Size in Authorship Attribution," Literary and Linguistic Computing 26 (2011): 35-55. 63The TF-IDF vectorizer is therefore a normalization procedure that penalizes more frequent function words in favor of rare function words: see Christopher D. Manning et al., Introduction to Information Retrieval (New York, 2008), 117-33.The TF-IDF vectorizer is not an obvious choice for feature extraction in authorship-attribution studies, since it readjusts Burrows's Delta presupposition (see above) that the high-ranked most frequent words generate better distinctions between authors.We argue that observing less common function words can yield interesting authorial preferences, which might go unnoticed in the standard Delta approach."In many ways, this model can be contrasted with the assumption that low-frequency items are bad predictors of authorial style.Nevertheless, a few studies suggest that it might be useful.Arguably, this model captures the intuition that if a highly rare feature is present in two documents, this increases the likelihood that the two documents were authored by the same individual.While the method might therefore be sensitive to overfitting on low-frequency properties, this might be an attractive characteristic in certain (e.g., single-domain) authorship problems," Kestemont et al., "Authenticating the Writings of Julius Caesar," Expert Systems with Applications 63 (2016): 89. 64The similarity metric applied for the pairwise distances is the Minkowski metric, a Euclidean metric that is "a very general metric that can be used in a k-NN classifier for any data that is represented as a feature vector": see Pádraig Cunningham and Sarah Jane Delany, "k-Nearest Neighbour Classifiers," Multiple Classifier Systems 34 (2007): 1-17, at 4.

Bernard of Clairvaux and Nicholas of Montiéramey
Speculum 92/S1 (October 2017) calculated, weighted in rank of smallest pairwise distance 65 and consequently mapped in space through force-directed graph drawing. 66What a k-NN network ultimately captures is which texts are most akin-or have the closest connectionwhen it comes to writing style (as defined by the distribution of function words).It should be noted that k-NN nearly always finds relationships, as it is very much a closed game.It is designed to link candidates to one another in terms of distance (every text sample needs to find its five neighbors) and can presuppose ties that are rather coincidental or nonexistent (for example, in the case of outliers).The network visualization can therefore be biased by a misleading directionality.
Secondly, PCA is a technique that allows us to reduce a multivariate or multidimensional data set of many features, such as our function word frequencies, to merely two or three principal components, which disregard inconsequential information, or noise, in the data set and reveal its important dynamics (Figs. 2 and 4).
The assumption is that the main principal components, our axes in the plot, point in the direction of the most significant change in our data, so that clustering and outliers become clearly visible.Each word in our feature vector is assigned a weighting, or loading, which reflects whether or not a word correlates highly with a PC and therefore gains importance as a discriminator in writing style.In a plot, the loadings or function words that overlap with the clustered texts of a particular author are the preferred function words of that author (see Figs. 5-6). 67PCA is built to find the most meaningful variance of observations along the axes of its principal components.In this sense it is not always interested in finding links between candidates, as k-NN is, but rather in finding links between variables.Disadvantages are that 65 The weights were derived directly from the calculated distances (see n. 64 for specifications on the metric).The intuition is then that the distances should be normalized to a (1,0) range.Note that this is not a (0,1) range, since smaller distances correspond to greater similarities and therefore require greater weighting: distances 5 distances2minðdistancesÞ minðdistancesÞ2maxðdistancesÞ . 66See, for instance, Maciej Eder, "Visualization in Stylometry: Cluster Analysis Using Networks," Digital Scholarship in the Humanities 32 (2015): 50-64.The algorithm used for the graphs in this article was Force Atlas 2, embedded in GEPHI, an open-source tool for network manipulation and visualization: see Mathieu Bastian et al., "Gephi: An Open Source Software for Exploring and Manipulating Networks," Proceedings of the Third International Conference on Weblogs and Social Media (ICWSM), ed.Eytan Adar et al. (San Jose, CA, 2009): 361-62.It must be noted that Eder's network iteratively runs over an increasing number of features (100-1000 MFW) to establish the consensus between different text samples (likewise by means of nearest neighbors).I have somewhat adjusted the method for the purpose of this article, since running up to 1000 MFW would surely overfit on the strong content-dependent connections that exist between Nicholas and Bernard's texts.Matthew Lee Jockers likewise applied GEPHI networks to detect stylistic differences in his seminal work Macroanalysis: Digital Methods and Literary History (Champaign, 2013).The difference with his approach and the one maintained in this article is that Jockers not only generated his networks through stylistic variables, but combined these with thematic linkages to discover trends of literary evolution on a far larger and diachronic scale than in this article (one could say that Jockers, as a quantitative formalist, demonstrated the popular formalist concept of the "defamiliarization" of literary language).As argued earlier, thematic variables should be disregarded in this case study.Nevertheless, I concur with Jockers's view that networks are a powerful tool to "demonstrate literary imitation, intertextuality, and influence" (156).
67 For an elaborate explanation of PCA and its applicability to stylometry, see José Nilo G. Binongo and M. Wilfrid A. Smith, "The Application of Principal Components Analysis to Stylometry," Literary and Linguistic Computing 14 (1999): 446-66.The PCA plots were drawn with the Matplotlib package available for Python: see John D. Hunter, "Matplotlib: A 2D Graphics Environment," Computing in Science and Engineering 9 (2007): 90-95.
PCA can never explain all the variance of the data, since it purposefully disregards many features and dimensions that it finds insignificant.It also has the tendency to produce somewhat nebulous scatter plots when texts are stylistically entangled (as is the case for Bernard and Nicholas).

The Letters
Bernard's epistolary corpus is very complex, but a coherent structure has been recognized thanks to Jean Leclercq's editorial achievements.Leclercq, whose terminology will be adopted here, has divided the corpus into a literary (intra corpus) and a nonliterary section (extra corpus). 68Bernard intended the letters in the intra corpus to circulate as a literary collection and kept refining them intensely throughout his life, whereas the second group of letters is scattered across time and manuscript traditions. 69Then, within the first section, the literary or intra corpus, we can make another division.Manuscript transmission allows us to distinguish between letters written before Nicholas's arrival in the scriptorium and letters inserted later. 70The letters that date from before 1145 in an earlier first appearance are found in the brevis manuscripts, whereas those added to Bernard's literary corpus afterwards can be found in the perfectum manuscripts.The perfectum corpus was assembled after Bernard's death in 1153, possibly by Geoffrey of Auxerre, and contains, aside from the brevis letters, many new additions (Tables 3 and 4 give a detailed overview of which letters are included in either the brevis or perfectum samples). 71It is impor- 68 Leclercq separated the intra corpus (Epp.1-310) from the extra corpus (Epp.311-547) in the SBO. 69Leclercq, SBO 8:233-38. 70Leclercq, "Lettres de S. Bernard: Histoire où littérature?,"Recueil d'études, 4:158. 71The main difference between the brevis and perfectum manuscripts is that the latter contain versions of these letters that were clearly amended and lengthened.In the introduction to the edition of the intra corpus, Leclercq gives a full account of the arrangement of the different transmissions, in which he distinguishes three more or less homogeneous collections, two of which are the brevis and perfectum cycles, whose names have served as inspiration to how we labeled our chronologically ordered data.It should be noted that Leclercq mentions a third intermediary publication that we have decided to exclude from our main argument, namely the longior corpus, which was compiled by Geoffrey of Auxerre and was presumably published in 1145.The longior corpus already contains quite a few of the perfectum additions.However, this corpus is hard to date or reconstruct, making it less interesting for us to include in this study: see Leclercq, SBO 7:xv.We decided to make a distinction between the early brevis publication, when Nicholas of Montiéramey was certainly absent from Clairvaux, and the later publication, when both of them, or an even more developed chancery, could have exerted influence on Bernard's style.More specifically, the letters falling under the heading of brevis-written before 1145 and not to be confused with Leclercq's brevis manuscript collection-can be consulted under the following indices: Epp. 1,  2, 4, 8, 11, 12, 24, 25, 42, 65, 67-70, 72, 73, 78, 79, 82, 83, 85, 87-89, 91, 95, 96, 98, 102, 104, 106,  107, 111, 113, 114, 117-19, 124-27, 129-33, 136, 139, 141, 143, 150, 152, 156, 158, 159, 168, 169,  178, 212, and 254.The letters that fall under the heading of perfectum-not to be confused with the perfectum cycle as it is described by Leclercq, but corresponding to all letters that were written after 1145 and remained unmentioned in the brevis data-are to be found under these indices: Epp. 3, 5-7,  9, 10, 13-23, 26-41, 43-64, 66, 71, 74-77, 80, 81, 84, 86, 90, 92-94, 97, 99-101, 103, 105, 108-10, 112, 115, 116, 120-23, 128, 134, 135, 137, 138, 140, 142, 144-49, 151, 153-55, 157, 160-67,  170-77, 179-211, 213-53, and 255-310.However, these letters were categorized according to their date of publication (brevis and perfectum) and split up into samples.A full inventory of which letters can be found under which sample in the figures is provided in the Appendix of Tables (see Tables 3-7).tant to note that those letters which were already found in the brevis manuscripts and reoccur in the perfectum corpus have sometimes considerably changed in the eight years between the two appearances.After all, Bernard's aim was to compose a unified piece of literature.He corrected, rearranged, and selected throughout his life.Importantly, Leclercq's edition of Bernard's literary letter corpus, which we use in these experiments, is almost entirely based on the perfectum transmission, which enjoyed the most popular circulation. 72We therefore do not work strictly with the brevis corpus in its original pre-1145 form, but with a group of letters that was collectively reworked and jointly disseminated.This condition of the texts is a flaw in the experiment that should be kept in mind during the analysis.Moreover, as mentioned earlier, Bernard's extra letters have not known a homogeneous transmission but have been handed down to us under divergent circumstances.The corpus therefore required some reorganization.Since Leclercq's edition allows us to assign individual letters to discrete periods, we decided to divide the extra corpus into three time-bound parts to see if Nicholas's arrival came with a stylistic impact: the first part dates from before 1140, the second between 1140 and 1145, and the third from 1145 onwards.Those extra letters that are of questionable dating and addressee have been left out of our experiments, for they cannot contribute to a study of Bernard's stylistic evolution through the influence of his secretaries (Table 5 in the appendix gives a full overview of which extra letters were included or excluded).
In Fig. 1 (k-NN) and Fig. 2 (PCA) we have applied the statistical methods described above to calculate and visualize the stylistic differences between Bernard's letter corpus and Nicholas's authentic sermons and letters. 73For each of these techniques, we have provided three additional subplots, which highlight how the different corpora are positioned within the clusters.There appear to be two general, observable dynamics, confirmed both by k-NN and PCA.Firstly, the writing style in Bernard of Clairvaux's letters is fairly coherent and forms a distinguishable cluster separated from Nicholas's works.Nevertheless-and this is the second, more hidden dynamic-our chronological and codicological rearrangements in the corpus have laid bare a gradual, subtle disturbance in Bernard's stylistic signal from 1140 onwards, corresponding to the approximate time of Nicholas's arrival in Clairvaux and seemingly moving towards the latter's cluster. 74Yet, two major remarks are in order.Firstly, although the perfectum additions were indeed inserted into the literary corpus from 1140 onwards, some of them must have been first composed at a time before Nicholas's arrival.For example, sample in_10 of the perfectum additions, which draws closest to Nicholas's cluster of all literary samples (only in the PCA, not in the k-NN network), contains letters that revolve around the schism between Antipope Anacletus and Pope Innocentius II, a series of events that occurred between 1130 and 1138. 75Although Nicholas was not yet part of Bernard's entourage during these events and was therefore likely not involved in their first redac- 72 Leclercq, SBO 7:xvi. 73See n. 6 for a listing of the sermons under consideration.For Nicholas 's letters, see PL 196:1593's letters, see PL 196: -1654. 74 . 74 Turcan-Verkerk, "L'introduction de l'ars dictaminis en France," 70. 75The sample (in_10) contains Epp.128, 134, 135, 137, 138, 140, 142, 144, 145, 146, 147 (SBO  7:321-51).

S207
Speculum 92/S1 (October 2017) tion, he was nevertheless present when they were first sent out collectively with the other perfectum letters.However, if any refinement was imposed on these letters after 1140, it seems likely that Bernard, as the author, would have been the one to do so, not Nicholas, although technically the latter's interference is possible.
The second remark ties in to the problem we have just raised.Although both figures show a diachronic stylistic shift, PCA slightly adjusts k-NN's inference that this shift has a determined direction towards Nicholas.The ex_ samples rather "float" around Nicholas's vicinity but never fully coincide.This suggests that the disturbances in Bernard's stylistic signal should not necessarily be as "monocausal" or "directional" as the k-NN network suggests, a nuance that reciprocates the historical skepticism raised in our first remark.Countless other variables aside from Nicholas's interference could have contributed to the subtle stylistic change in Bernard's letter corpus.One factor could be the lapse of time and Bernard's personal development. 76 Another is the respective corpus's divergent transmission history.But perhaps the most crucial reason for PCA's less outspoken directionality is that Bernard did not have just one secretary.Although Nicholas was the scriptorium's headman, this experiment undoubtedly simplifies or fragmentizes its diversity of styles and personalities.We might even be surprised that Bernard's letters-considering the circumstances under which they were conceived-still display this amount of stylistic coherence (although there might have been a more outspoken divergence if we had been able to oppose the very original brevis corpus to the published versions).
This does not alter the fact that the plots' gravitation towards Nicholas's Latin style, which was of a very schooled nature, 77 might hold some historical ground.As Bernard more and more became a public figure, he increasingly began requiring the 76 There is a considerable amount of literature that argues that such an evolution of personal style through time can be captured computationally by applying so-called stylochronometrical methods.For a concise overview of this subfield in computational stylistics, see Constantina Stamou, "Stylochronometry: Stylistic Development, Sequence of Composition, and Relative Dating," Literary and Linguistic Computing 23 (2008): 181-99. 77Nicholas's style has often been deemed schooled and unoriginal: see Constable, The Letters of Peter the Venerable, 2:328; and Leclercq, "Les collections de sermons de Nicolas de Clairvaux," 1:55: "Ses exposés superficiels se développent selon un plan scolaire, en un style artificiel."Likewise, Dorette Sabersky argued that "the syntactical structure of his sentences is similar to Bernard's, but often clumsier, less clear, less elegant, and rhythmically less balanced.His frequent use of word plays is at times rather superficial and, in opposition to Bernard's use, of little importance to the development of the contents.Repetitions of certain phrases and topics occur every so often.He favors rather unusual words and likes to quote classical authors.His literary exertions are only too obvious.All these aspects evidence Nicholas' lack of Bernard's creative spontaneity and mastery of language," in "The Style of Nicholas of Clairvaux's Letters," in Erudition at God's Service, Studies in Medieval Cistercian History 11, Cistercian Studies Series 98, ed.John R. Sommerfeldt (Kalamazoo, 1987), 196.However, it is all the more peculiar and contradictory to these former statements that-even until very recently-attempts were made to attribute texts to Nicholas on the grounds of phrasing tics and certain lexical preferences (e.g., neologisms): see Patricia Stirnemann and Dominique Poirel, "Nicolas de Montiéramey, Jean de Salisbury et deux florilèges d'auteurs antiques," Revue d'histoire des textes 1 (2006): 173-88.Either such attribution methods should be challenged (perhaps rightly so; their word and phrase concordances form particularly dangerous grounds for attributing authorship in a twelfth-century context that boasts such a high degree of "plagiarism") or the statement that Nicholas has no style of his own should be withdrawn.I am convinced of the latter.Our computational experiments show that Nicholas, despite being an imitator, has a very controlled and rather clean authorial signal.

Bernard of Clairvaux and Nicholas of Montiéramey
Speculum 92/S1 (October 2017) support of scribes to take on administrative tasks, be it in Clairvaux or on exceptional occasions elsewhere. 78These scribes would have received a similar training or education.We can assume that most were under Nicholas's supervision, which meant that they departed from a common framework or set of rules from which they set out to imitate Bernard.This had become the nature of the epistolary writing art, or ars dictaminis. 79Letters were constructed on the basis of similar formulas, abounded in clever wordplay, and the rhythms of their prose pulsated under comparable cadences. 80Diplomats, ambassadors, and secretaries would inspire one another in a network of correspondence or share these rhetorical devices within their scriptoria. 81These practices might have considerably reshaped the stylistic homogeneity that is evident in the writings of Bernard from his earlier days, when he relied on a far smaller number of secretaries and had more time at his disposal so that he 78 Bernard would not necessarily have found help only in Clairvaux.It is conceivable that when he was occupied with the turbulent matters of the schism and was traveling through Italy he called for the assistance of papal scribes to whom he could dictate his messages.The papal notaries, educated in the ars dictaminis, would in fact have been schooled in a similar tradition as Nicholas, who had visited Rome and moreover corresponded with at least three popes during his lifetime: see Constable, "Dictators and Diplomats," 43.This could also explain why brevis samples in_12 and in_13, which likewise have the schism as their subject, somewhat pair with letters that were added to the corpus later and not with the other brevis letters, which cling more closely together.The samples in_12 and 13 contain Epp.126, 127, 129, 130, 131, 132, 133, 136, 139, 141, and 143 (SBO 7:309-43).For a concise overview of Bernard's interference in the papal schism and its importance for his public career, see the subchapter The Second Edwards Lecture Delivered within the University of Glasgow on 7th December (Glasgow,  1966), 20-21."Instructions about the framing of papal letters may be found in chancery ordinances and in guide-books for chancery clerks; these help to elucidate the legal principles which underly the phraseology."The writing style of the papacy's chancery must have served as an important model to all clerks and diplomats both in ecclesiastical and worldly contexts. 79See Ronald Witt, "Medieval 'ars dictaminis' and the Beginnings of Humanism: A New Construction of the Problem," Renaissance Quarterly 35 (1982): 1-35. 80With "comparable cadences" I am here referring to rhetorical devices such as the cursus: see Tore Janson, Prose Rhythm in Medieval Latin from the 9th to the 13th Century (Stockholm, 1975).The cursus has also been tested as a feature for authorship attribution: see Linda Spinazzè, "'Cursus in clausula,' an Online Analysis Tool of Latin Prose," Proceedings of the Third AIUCD Annual Conference on Humanities and Their Methods in the Digital Ecosystem, ed.Francesca Tomasi, Roberto Rosselli Del Turco, and Anna Maria Tammaro, Association for Computing Machinery (ACM) International Conference Proceedings Series (ICPS), (New York, 2014), 10:1-6. 81On the subject of the ars dictaminis, see Giles Constable, Letters and Letter-Collections (Turnhout, 1976), 34-35: "This tendency towards a personalization of style and contents in eleventhand twelfth-century epistolography was paralleled by a tendency, which was in some respects contradictory, towards formalization, which was represented by the emergence of the discipline known as the dictamen or ars dictandi, with teachers (dictatores), text-books (artes or summae dictaminis), and collections of model letters (formularies).Although dictamen now emerged for the first time as a discipline with clearly formulated rules, it had roots deep in the past and was connected in ways which are still not fully understood with the epistolographical rules and traditions which went back to Antiquity. . . .In the course of the twelfth century the number both of teachers and of text-books of dictamen spread rapidly, first in Italy and later, in the second half of the century, north of the Alps.Various schools developed with different styles, as at Bologna and Orleans; and although in the earlier twelfth century a certain number of writers, like St. Bernard and Peter the Venerable, who knew about dictamen, did not observe its rules, its influence was all but universal by the end of the century."

S209
Speculum 92/S1 (October 2017) could be present during the various phases of composition.We know of Bernard's increasing discomfort concerning the fact that he felt obliged to delegate the writing of his letters and sermons to assistants, and his dissatisfaction with some of them when it came to grasping the sensus of his message. 82Perhaps to his own frustration, Bernard was increasingly forced to have faith in the reliability of such scribes as Nicholas to reformulate his initial dictation in a letter that conformed to the style and content Bernard had intended.The extra letters would have received far less revision, resulting in the kind of hybrids that float towards middle ground in the figure.

The Sermons
In this second visualization we put Nicholas's word to the test.Firstly, assuming that the secretary speaks the "truth" 83 in his letter to Henry the Liberal, which Leclercq cited as the most striking example of his plagiarism, we expect that a small number of sermons that occur in Bernard's De diversis, namely 6, 7, 21, 62, 83, 100, and 104, could be attributed to him instead.On the side, we test if his claims to Hugh's commentaries on the Psalms, which were also mentioned in the letter, hold any ground. 84In a second phase, we follow up on Henri Rochais' conclusions that Bernard-not Nicholas-wrote De diversis 40, 41, and 42. 85In fact, the De diversis collection in its entirety is worth testing here, as it suffers from some considerable issues of authenticity, provenance, and dating and might contain other traces of Nicholas's presence.The corpus comprises an assembly of unpolished and rudimentary sermons found in various, heterogeneous manuscripts, conceivably written down by secretaries and granted little revision by Bernard (unless if they were reused elsewhere). 86Bernard never disseminated the De diversis sermons himself.They were gathered after his death and passed on for several centuries until Jean Mabillon enumerated and published them in the seventeenth century.Leclercq and Rochais maintained Mabillon's structure in their edition. 87Secondly, we have included the Sermones super Cantica canticorum, Bernard's literary masterpiece, as the cleanest possible specimen of Bernard's literary style to benchmark against these texts. 8882 "Multitudo negotiorum in culpa est, quia dum scriptores nostri non bene retinent sensum nostrum, ultra modum acuunt stilum suum, nec videre possum quae scribi praecepi," Bernard, Ep. 387  (SBO 8:355-56).
83 "The medieval idea of truth . . .was subjective and personal rather than, as today, objective and impersonal": see Constable, "Forgery and Plagiarism," 23. 84 Bernard,De diversis 6,7,21,62,83,100,295,324,367,.Nicholas used the phrase "aliosque sermones" in the prefatory letter in MS Harley 3073, Recueil d'études, 1:50, referring to the aforementioned sermons, a few other texts by Bernard, and, finally, Hugh of St. Victor's chapters on the Psalms gathered in the second book of his Miscellanea (PL 177:589).85 Rochais, "Saint Bernard est-il l'auteur des sermons 40, 41 et 42?," 324-45.86 Leclercq, SBO 6/1:59-71.87 Françoise Callerot, "Introduction," in Bernard of Clairvaux, Sermons divers, ed.Jean Leclercq,  Henri Rochais, and Charles H. Talbot, 3 vols.(Paris, 2006), 1:21.88 Bernard must have started composing its beginnings around the end of 1135, but never commentated the entire Song of Songs.They are, nevertheless, regarded as his life's work and greatest literary achievement: see Leclercq  Fig. 3 (k-NN) and Fig. 4 (PCA) feature the results of matching up these texts.Firstly, when examining the visualizations, it is striking how the diversity of Bernard's De diversis is indeed captured.PCA, especially, demonstrates a discernible stylistic incoherence, as the samples burst open all over the plot (especially along the vertical axis of the second principal component), at times suggesting the interference of writers other than Nicholas or Bernard in their composition.Other samples gravitate in between Nicholas and Bernard, and in some cases Nicholas's influence on the style is undeniable.Before discussing some contingent subjects of interest, let us focus on the primary questions at hand.De diversis 6, 7, 21, 62, 83, 100, and  104, which Nicholas included in the letter to Count Henry the Liberal (they are split up in two red samples labeled with le_ of Leclercq), do not betray an obvious affinity to Nicholas's style (although le_1 is not far off).Neither are they unambiguously Bernard's.Both samples diverge strongly from Bernard's cluster and seem too hybrid in nature to be restrained to either of the authors' clusters.The case rather demonstrates how difficult it is to defend such concepts as "single authorship" and "textual theft" in a medieval context: the le_ samples are clearly not of a "singular" style (neither Nicholas's style nor Bernard's) but defy classification.In fact, if we compare both k-NN and PCA, Nicholas's influence in sample le_1 seems considerably larger than Bernard's.It has by now become an untenable simplification to argue that Nicholas has stolen these sermons, especially if we review the results of our second case, that of De diversis 40, 41, and 42 (four red samples labeled with ro_ of Rochais): although the sermons emanate from Bernardian thought, k-NN and PCA unambiguously cluster all three sermons together with those written by Nicholas, not Bernard.
There are some less straightforward developments on the side.Hugh of St. Victor's presence in both attribution problems remains somewhat unclear.Nicholas included Hugh's commentaries on the Psalms in his collection, yet Figs. 3 and 4 show that he was unlikely to have been the (only) author of this incohesive text (see the purple hu_ samples, of which hu_9 comes closest to Nicholas). 89Vice versa, De diversis 40 (first part of the dubious ro_ samples) is collected in Hugh of St. Victor's Miscellanea.Would Nicholas have known Hugh well, and would they have collaborated before the latter's death in 1141?There is no proof of a direct acquaintance.Nicholas's musical sequences seem largely based on those of Adam of St. Victor, Hugh's choirmaster, but these texts enjoyed a popular circulation, so the sim-ilarity does not necessarily presuppose a personal tie. 90For Bernard and Hugh, however, the connections are less far-fetched.We know they corresponded. 91Hugh incorporated an entire letter he received from Bernard in his acclaimed masterpiece, the De sacramentis. 92Likewise, Figs. 3 and 4 show that samples di_46 and di_47 of Bernard's De diversis bear some affinity with Hugh's commentaries.These samples comprise the very last additions to Bernard's corpus, De diversis 112-25.They are shorter texts, which have not always been accompanied by the preceding sermons but must have circulated as a separate unit in manuscript transmission.Mabillon has argued that their provenance differs from that of the other De diversis sermons in a footnote, 93 thereby perhaps showing some wariness as to the authenticity of the works. 94Although they might be Hugh's, we find that the textual style of both Bernard's De diversis and Hugh's commentaries is too unreliable to provide closure.The case for the triangular writing relationship between these authors is compelling, but there is insufficient historical proof to corroborate speculations of a collaboration between Nicholas and Hugh.

Conclusion
Jean Leclercq, in aspiring to discern the psychological personality of the author behind any given historical text, conceded the difficulty of infiltrating the "screen of rhetoric" so characteristic to twelfth-century literature, referring to its predilections of imitation and formal rigidness. 95The surface of the medieval text can strike one as impenetrable.In a similar vein, Giles Constable has argued for medieval epis-90 "Since Nicolas is known for his plagiarism and incorporated the work of Hugh of St. Victor in the collection of his own opera dedicated to Count Henry, the suspicion arises that Nicolas modeled his work directly on that of Hugh's colleague, Adam of St. Victor": John F. Benton, "Nicolas of Clairvaux and the Twelfth-Century Sequence," 92 "Adding to the complications of De sacramentis as a text is Hugh's incorporation of passages not only from his own prior works but also from other theologians, patristic and contemporary, sometimes named but often without any attribution at all.In this respect, Hugh nicely represents the overall concern of twelfth-century authors to synthesize their sources": Paul Rorem, Hugh of St. Victor, Great Medieval Thinkers (New York, 2009), 60. 93 The earliest editions of Bernard's Opera omnia did not yet include these sermons.It was not until its publication by printer Johann Herwagen of Basel in 1566 that De diversis 112-25 found its place among the Sermones de diversis: see Gerhard B. Winkler, trans., Bernhard von Clairvaux: Sämtliche Werke, 10 vols.(Innsbruck, 1990-99), 9:882 Lh , 884 To .Jean Mabillon relied on Jacobus Pamelius's edition of these sermons; see note to PL 183:739. 94Jean Mabillon was aware of the fact that Herwagen's and Pamelius's editions were to be approached with great caution when it comes to attribution.Herwagen's and Pamelius's collections of Bede's works are examples of how these editors "ignored and altered rubrics, expurgated passages, disregarded section breaks, and lied outright about the Bedan origins of their material:" Nathan J. Ristuccia, "The Herwagen Preacher and His Homiliary," Sacris Erudiri 52 (2013): 188.

Fig. 2 .
Fig. 2. Principal Components Analysis of Bernard of Clairvaux' epistolary corpus compared to Nicholas of Montiéramey's letters and sermons.
"A Leading Figure in the Papal Schism 1130-38," in Brian Patrick McGuire, "Bernard's Life and Works," in McGuire, A Companion to Bernard of Clairvaux, 40-47.For information on the papal chancery, see Christopher Robert Cheney, The Study of the Medieval Papal Chancery: Fig. 3. Network visualization of Bernard of Clairvaux' Sermones super Cantica Canticorum and Sermones de diversis compared to Nicholas of Montiéramey's letters and sermons and Hugh of St Victor's commentaries on the Psalms.

Fig. 4 .
Fig. 4. Principal Components Analysis of Bernard of Clairvaux' Sermones super Cantica Canticorum and Sermones de diversis compared to Nicholas of Montiéramey's letters and sermons and Hugh of St Victor's commentaries on the Psalms.

S215Fig. 5 .
Fig.5.Principal Components Analysis loadings for Fig.2.Function words that swarm around author's cluster are most distinctive for that author.

Fig. 6 .
Fig.6.Principal Components Analysis loadings for Fig.2.Function words that swarm around author's cluster are most distinctive for that author.

Table 8
Description of Sample Contents (3,000 words) for Nicholas's Sermons and Letters in Figs.1-2