Introduction

In the traditional linguistic model of medieval England, the Norman Conquest of 1066 caused English, which had previously been an acceptable language for literary and cultural production, to be displaced by French and sidelined in aristocratic and courtly domains. In this traditional model, English only regained its status within these ‘high domains’ after about two hundred and fifty years. A growing body of research has pointed to the significant structural problems with this traditional linguistic model, and it is now generally accepted that French persisted as an important domestic and aristocratic language in England for much of the late medieval period.¹ In light of this increasingly important body of research, the status of French in this period, including the contexts and implications of its use, is being re-examined.²

To date, studies of the status of French in medieval England have been focused on isolated examples—either of individual cases of sociolinguistic interest, or of the interplay of languages within single manuscripts or texts.³ This example-based approach has allowed the field to productively challenge the traditional ‘grand narrative’ of England’s linguistic situation, but since examples are, by nature, highly specific, they provide only limited insight into patterns of language use across social groups, gender, and temporal periods.

The goal of investigating these language patterns on a broader scale lay behind this project: the creation of a digital database of manuscripts containing French literature that were copied in medieval England.⁴ Manuscripts—hand-written collections of texts—were chosen as the focus of this quantitative project because they provide unmatched insight into language use for a period in which no audio or spoken evidence is available. Since manuscripts are, by definition, handmade objects, they are distinct witnesses to the social contexts, patrons, and copyists that produced them.

Manuscripts functioned for their medieval users much like a binder does for modern ones: as a compilation of material to be consulted later. A medieval individual or group would select which text or texts should go in the manuscript and then either copy them out by hand or assign this task to one or more scribes. Texts in a manuscript, much like those in a binder, could be removed or added after the manuscript was originally compiled, either by the original compiler or by a later user. Manuscripts therefore represent the deliberate and conscious choices of one or more medieval users, and each manuscript, and each of its unique stages of compilation, can therefore serve as an information-dense data point about medieval language use. Taken together, these data points can be plotted to identify language patterns.

Manuscripts are a particularly good source of evidence for tracking language use in this context since they survive in far larger quantities than any other medieval textual witnesses. In particular, manuscripts can provide insight into the languages that were considered suitable for literary and documentary culture, and about who owned texts in certain languages and in which contexts. But there are also limitations to the kinds of sociolinguistic insight that manuscript data can provide. First, manuscripts, as written artifacts, cannot be taken as representatives of spoken language use in any straightforward way, so the evidence they provide, while valuable for understanding England’s written culture, must be approached with caution when exploring broader sociolinguistic questions. And the division between written culture and everyday life in the medieval period could be significant; producing a medieval manuscript was an expensive and time-consuming process, which means that manuscripts were usually owned by people with social status, such as those within the Church or those with financial means. Manuscripts and their ownership patterns, then, typically provide information about the literary tastes of an elite subsection of medieval society. Nevertheless, the texts in manuscripts did reach a more diverse audience beyond their original owners, and in the absence of audio recordings or other, more populist forms of language data, manuscripts are an unmatched source of information about medieval language use.

The approach taken here has been made possible through two important developments: digital tools that enrich and assist quantitative analysis and the increased availability of digitised manuscripts and their catalogues. Medieval manuscripts have traditionally been studied almost exclusively through qualitative methods—either in isolation or through smaller-scale comparative approaches. These methods, which remain central to manuscript studies, can yield valuable information about medieval reading communities, book production and textual exchange. Over the past few decades, though, scholars have increasingly used digital technology to look for large-scale patterns in medieval manuscript datasets. Among the promising developments in this area are studies aimed at exploring the circulation and production of manuscripts in the medieval world. So, for example, Michael Sargent has explored a dataset of medieval ‘bestsellers’ and identified a correspondence between the number of surviving manuscript copies of a given text and its circulation numbers.⁵

Quantitative approaches—not all of them digital—have also yielded exciting results in the field of medieval paleography–that is, the study of handwriting in medieval manuscripts. The new horizons in this area are suggested by the project undertaken by a team of researchers headed by Carla Bozzolo and Ezio Ornato, which analysed a large dataset of Psalm manuscripts to gain statistical information about abbreviation practices in the medieval world.⁶ Quantitative approaches have also offered new insights into changes in how manuscripts were copied; these include Erik Kwakkel’s analysis of approximately 350 manuscripts, which revealed several key changes in letter forms that took place in the twelfth century—changes that can now be used to support the dating of other manuscripts.⁷ Although new quantitative approaches often face significant barriers—some of which are discussed below—the developments in this domain of the past few decades point to the value of such approaches for illuminating patterns that would otherwise have gone unnoticed.

In order to identify language use patterns in this quantitative way, this project gathered information about the contents and contexts of manuscripts that would be useful for understanding how French literature circulated in medieval England. For this reason, information was gathered about the number of pages (sides of a manuscript folio) dedicated to French in a given manuscript relative to Latin and English. To identify and track changes in language use and distribution, information was also gathered about a given manuscript’s date of composition, and, where possible, about its medieval owners and areas of circulation. This information was compiled in a digital format so that it could be analysed in a quantitative way—a traditional humanities approach that can be greatly enhanced by the predictability, efficiency, and accessibility offered by digital technology.

With the goal of providing a set of blueprints for those seeking to undertake or support similar projects, this article describes the methodology used for compiling and analysing linguistic data from manuscripts and describes the patterns of textual transmission that can be traced particularly effectively using this methodology. Since its aim is sharing a particular framework for this kind of quantitative manuscript-based analysis, this article is focused on the methodology and digital approach of the project rather than its specific sociolinguistic findings, although some discussion of these findings has been included where relevant.⁸ Quantitative approaches to medieval manuscript data are still in their infancy, so we discuss some of the current barriers to this kind of research and propose ways of structuring manuscript metadata that will facilitate future linguistic-based manuscript research.

Method

Preparing Manuscript Data for Quantitative Analysis

The dataset for this project was based on a list compiled by Ruth Dean and Maureen Boulton that was aimed at recording every known work of French literature from medieval England. The source list, produced in 1999, remains the most comprehensive list to date of French manuscripts from medieval England, and contains 958 items.⁹ For the goal of this study, this list had to be modified somewhat, and two categories of manuscripts on the list had to be omitted: those copied on the continent (indicated by a * in Dean and Boulton’s list) and those that contain no French but were included on the list for other reasons (marked with an ‘r’ in Dean and Boulton’s list). The former group had to be omitted since, as continental productions, they do not reflect textual production and linguistic exchange in the area under investigation. The latter category had to be omitted because the project was aimed at gathering data about manuscripts containing French literature, so those without it did not belong in the dataset.¹⁰ The remaining manuscripts were assigned an ID number.

In the first stage of the project, traditional catalogue data were gathered about the manuscripts, including their contents, date, linguistic profile, and owner marks—such as library tags, inscriptions and patron illustrations. This traditional information was structured in a relational model to ensure that it would be machine-readable and ready for digital analysis.¹¹ Manuscript identification and metadata information were therefore entered into a main ‘Manuscript’ table. This metadata information includes current holding information, period and place of production, explanatory notes and references for the information.

For each manuscript entry—either an individual manuscript or, where necessary, a part of a composite manuscript—data about texts and owners were entered into separate ‘Text’ and ‘Ownership’ tables. For each text in a manuscript, information recorded includes the title or incipit, the language, the start folio and the end folio. For each known owner, recorded information includes a name or reference of the owner, the period that they were in possession of the manuscript, source references and where available, whether they held clerical status at time of ownership and their gender.

The structuring of these files was aimed at upholding the internationally-recognized framework for data management known as the FAIR principles (Findable, Accessible, Interoperable and Reusable).¹² All data were stored in CSV files. CSV was chosen for several reasons; first, CSV, as a format that is compatible with existing and readily available tools like Microsoft Excel, allowed data entry to begin quickly and efficiently. Using CSV files as input and output also made archiving the results for reproducibility and reuse straightforward.

The columns in the table files and the relations between the files were described using JSON schema files that follow the CSV on the Web standard. With these schemas the CSV files can be checked for consistency and converted into RDF using available tools like COW and the RDF.rb suite.¹³ This use of schemas gives the project data greater machine actionability, which is one goals of the FAIR principles.

For this project, planning for the reusability of the data was particularly key. So, with the goal of ensuring that data, in keeping with the FAIR principle of reusability, would be ‘richly described with a plurality of accurate and relevant attributes’ (R1) and ‘meet domain-relevant community standards’ (R1.3), the project gathered some traditional manuscript data that, while not needed for the particular goals of this project, would benefit those working with the dataset for other purposes.¹⁴ This extra data included manuscript headnotes and old foliation system information. A complete list of columns for each table is presented in Figure [fig:class-diagram].

Issues in manuscript metadata collection

The manuscript description data for this project was gathered through traditional research methods since at present an automated gathering method is not practical. Historical and structural challenges faced by manuscript collections mean that many descriptions of medieval manuscripts have not been updated since they were first written during the wave of enthusiastic—but not necessarily purely academic—wave of interest in medieval sources that marked the late nineteenth and early twentieth centuries. While these centuries-old manuscript catalogues have the advantage of being free of copyright restrictions and therefore well suited to open scholarship projects such as this one, they have the disadvantage of being out of date, and relying on them too heavily could therefore lead to unnecessary mistakes in a dataset. For this project, the potential pitfalls of using out-of-copyright catalogues were mitigated through the use of more recent dating information, which was gathered wherever possible from Ruth Dean and Maureen Boulton’s Guide.¹⁵

Converting traditional manuscript description data to a machine-actionable, reusable, and interoperable format also presents several notable challenges. Manuscript cataloguing, as a practice that began in the medieval period itself, evolved differently to meet the needs of the various contexts and regions in which it was used, and has historically been a widely variable practice. Armando Petrucci, in his La descrizione del manoscritto, distinguishes between shorter summary catalogues and longer analytical catalogues, but within these categories there exist marked differences in the contents, layout, and languages of descriptions between regions, libraries, and even within individual collections.¹⁶ The significant variation between the ways in which catalogues described manuscript contents posed a particular challenge for this project. Many catalogues—and most notably those produced by M.R. James which describe the Cambridge manuscript collections—provide only the starting points of texts within manuscripts, omitting their ending points. While these starting points are sufficient for someone looking for a particular text in a manuscript, they are unfortunately insufficient for a quantitative study such as this one, since the end point of one text within a manuscript cannot necessarily be inferred from the starting point of the text that follows it.

Aside from causing gaps in information, the inconsistencies between manuscript descriptions pose a significant barrier to quantitative analysis more generally. Digitization of nineteenth and early twentieth-century catalogues has made these more accessible for analysis, and it is theoretically possible to create a program that could gather and process description information from diverse catalogues in a large-scale way, but at present, such an endeavor would be hindered by the significant differences between description structures and would undoubtedly prove less reliable than collecting such data manually.

There have been various attempts to establish a standard for manuscript description, including the widely used method given in Raymond Clemens and Timothy Graham’s Introduction to Manuscript Studies, but none have caught on.¹⁷ Within the digital realm, while there have been attempts to design online manuscript catalogues with interoperability in mind, there is currently no established metadata standard for manuscript description.¹⁸ Nor is such a standard likely to be established in the near future, given the piecemeal ways in which online manuscript catalogues have been developed, and the sheer number of institutions that provide online descriptions of their manuscript collections–two challenges which may be counted among the broader barriers to semantic web development. Moreover, libraries and archives are often prevented from adopting the metadata structures that have been developed for manuscripts due to funding or structural staffing issues.¹⁹

For this project, the most prevalent and significant omission in manuscript metadata was information about the languages of individual manuscript texts; this information is not included consistently in most existing manuscript catalogues and databases.²⁰ This is due, in part, to a lack of clear guidelines for recording this information. The influential Descriptive Cataloging of Ancient, Medieval, Renaissance, and Early Modern Manuscripts (AMREMM), a guide for adapting the MARC21 record structure of library catalogues for manuscript description, is largely open-ended in its guidelines for recording language use; it encourages cataloguers to note the ‘language or languages employed in an item’ and to ‘provide more detailed notes in the records for individually analyzed works’ but only ‘if desired.’²¹ It is perhaps not surprising, in light of the well-documented funding and staffing issues encountered by many special collections, that including such detailed information in manuscript metadata has often not been a priority. Moreover, not all libraries have opted to follow the AMREMM guidelines and library cataloging systems may not support the description of manuscripts in a way that makes their contents information suitably findable.²²

On a positive note, the new electronic Bibliotheca Neerlandica Manuscripta (eBNM+) is an example of a database with detailed contents information; it treats individual texts within a manuscript as separate entities and includes language information as a property of each separate entity, although the catalogue does not provide information about all the texts within each manuscript.²³ This database contains medieval manuscripts produced in the Low Countries and has a focus on Middle Dutch manuscripts, so it could not be used as a source in this project. Nevertheless, the structure of the catalogue shows that there is interest in, and initiatives in support of, this level of cataloguing.

For this project, missing catalogue information could on occasion be inferred from the titles of the constituent works included in the catalogues, but these titles are not always reliable or helpful indicators of language— especially for medieval French texts, which often appear under Latin titles. This, and other limitations of existing catalogue descriptions, led to gaps in the project dataset; where possible, these have been remedied through work with digital manuscript facsimiles or, in some cases, archival work, but of course neither approach is practical for large-scale data collection.

Aside from the limitations posed by existing catalogue records, any attempt to render manuscript data machine readable is met with additional challenges posed by the complex, multivalent and often enigmatic nature of the manuscripts themselves. To give a simple example, the so-called Black Book of Christ Church College Dublin was copied in a series of stages, which adds complexity to the process of assigning a single date to the manuscript in its current form, or to analyse it as a single language use data point. To represent such multi-stage manuscripts accurately, these manuscripts were entered into the catalogue as a series of individual stages, with each stage assigned a unique manuscript ID.

Adding to the challenges of recording manuscript data in a machine-readable format is the difficulty involved in arranging medieval heritage data into the kind of clean categories that facilitate machine-assisted data analysis. Among the difficulties involved in this project was the process of classifying manuscript owners as man or woman and as lay or clerical, which was done in order to identify linguistic patterns among these owners.²⁴ The process of compiling and sorting this data revealed significant limitations with the established classification structure; first, determining the gender of medieval people generally required guesswork based on their first names. And some medieval lives, such as those of anchorites or students, resisted the lay/clerical binary required for this classification. These challenges can be counted among the broader complications that arise when working with humanities data, which is often found deeply embedded in social and historical contexts. But rather than signaling inherent problems with digital humanities approaches, these challenges point to the importance of approaching the results of quantitative humanities projects with a sensitivity to, and awareness of, its multivalent social and historical contexts.

Calculating Language Distribution

With the aim of calculating the distribution of languages in each manuscript, the number of folia occupied by each text was recorded in each manuscript’s ‘Text’ table. This table noted the specific language of each text if it was in one of the languages under investigation (French, Latin or English), and described the language as ‘other’ in the rare cases of other languages—including Hebrew and Greek. For the sake of consistency, texts were recorded as such only when they occupied four or more ruled manuscript lines; some interesting or noteworthy writing under four lines was also recorded, but this information was placed in the ‘Notes’ field. An exception to this approach had been made for glosses, which were recorded in the ‘Text’ tables; although these often consist of fewer than four ‘ruled manuscript lines’ they operate as part of a larger textual framework and therefore were considered noteworthy linguistic data.

In this system, pages containing multiple texts in different languages presented a challenge to data representation. A page (one folio side) that contained more than one text was assigned partially to each of its constituent texts; so a page containing both a French and an English text, for example, was counted as .5 of a page of French and .5 of English. Greater accuracy for page ranges may have been achieved by counting the number of lines or words of a given text on each page but these methods, which prioritize either length of words on one hand or wordiness of a text on the other, misrepresent data to some extent in their own way and any extra accuracy they would have offered would not have been worth the enormous time commitment involved in counting individual lines of text for each manuscript.

Calculating the number of pages occupied by a given text was done using a Python script and the Pandas code library, which is excellent for working with tabular data.²⁵ The code works by calculating the number of pages covered by each individual text and assigning them to their associated language category. To allow a mathematical calculation of page ranges, the collected folio numbers for the start and end points of each text (e.g. 1r, 3v) were converted to their ordinal variants (1, 6, respectively). This method worked well for most folia, but not for those assigned roman numerals, which are generally used by manuscript cataloguers to indicate the flyleaves or endleaves. Roman numerals were therefore initially converted in the same way, but after conversion, the code increased them by 100000 in order to avoid reusing the same ordinal page numbers in calculations. So, for example, IIIr was converted to 100005 for its page range calculation. Since no manuscript has page numbers even close to that number, this method prevented the unintended reuse of page numbers.

Once this method of conversion is in place, the number of pages per language for a whole manuscript can be calculated by grouping the texts per language and summing the number of pages in each group. The effect is that, once the individual tracts are described in tables, the script can identify the percentage of each language in each manuscript.

Presenting the Results in HTML

While the main output for this project was the archivable CSV files and their analysis, we also wanted to present the project data in a format that was more user friendly, so the project data were also presented as a website.²⁶ With the goal of sustainability in mind, the data and results of the language analysis were presented as a static, rather than dynamic, website. A static website was chosen for this project since maintaining the code for a dynamic website after the end of the project was deemed unfeasible and, as others have observed, static websites have the advantage over dynamic ones of offering greater flexibility in terms of preserving and moving websites that consist of text files alone.²⁷

Results: Preliminary Findings

The statistical analysis of language use in these manuscripts has yielded valuable results about the linguistic situation in England during the centuries following the Norman Conquest. Most notably, this analysis has revealed that French literature circulated on its own relatively infrequently.

As shown in Figure 2, from the twelfth to fifteenth centuries, monolingual manuscripts containing French literature remained in the minority. Only 11 of the 35 manuscripts dated to the twelfth century are monolingual (31%) and the proportion of monolingual French manuscripts in the dataset remains relatively consistent throughout the period under investigation; it increases only slightly in the thirteenth and fourteenth centuries and then decreases in the fifteenth. The extent and nature of the multilingual contexts of England’s French writing has not been commented on before and represents a significant result of this project.

Given that manuscripts were often designed with a purpose, and can often provide insight into the literary tastes of their patrons, these findings suggest that French literature was most often read in multilingual contexts and by multilingual patrons.²⁸ The multilingual contexts of French seem to have remained relatively consistent throughout the period under investigation, a finding that challenges the traditional narrative in which French was increasingly sidelined by English in the later period. This is significant because it suggests that the use of French in the centuries following the Norman Conquest was not usually dictated by a patron’s linguistic limitations and speaks to a high level of multilingualism of the patrons of French literature in medieval England.

The dating data provided by the manuscript catalogue has also provided valuable information about the use of French in written contexts in medieval England. In particular, it appears that there was an increase in the copying of French literature in the thirteenth and fourteenth centuries, as seen in Figure 3.

The increase in the thirteenth and fourteenth centuries is remarkable; it suggests that the period in which, in the traditional ‘grand narrative’, French was supposedly on the decline in England was, in fact, the most significant period for the copying of French literature.

But of course, these data on their own tell us very little, since the apparent increase in copying in the thirteenth and fourteenth centuries could, in theory, be skewed by the contingencies of manuscript survival or by broader patterns of manuscript copying. It is therefore necessary to contextualize this data within broader patterns of manuscript production and survival in England.

Ideally, this would be done using a large-scale dataset for the temporal distribution of manuscripts produced in England as a whole, but at present there is no such dataset available, and given the challenges that current manuscript catalogues present with respect to interoperability and findability already described, it is not currently feasible to produce one. However, some insight can be gleaned by comparing the temporal distribution of French manuscripts to one available catalogue of medieval manuscripts: Neil Ker’s Medieval Libraries of Great Britain (see Figure 4).²⁹ This catalogue, which was recently digitized and therefore available for quantitative analysis, contains information about all known manuscripts produced in England in any language that can be traced back to specific medieval libraries. It therefore provides some insight into manuscript production in England as a whole, although since institutional libraries were most often tied to monasteries and other clerical organizations in the medieval period, the dataset is skewed toward clerically-owned manuscripts and so may not accurately reflect manuscript production in England as a whole.³⁰ Nevertheless, in the absence of other evidence, this dataset provides a valuable point of comparison.

The lack of French manuscripts prior to eleventh century is of course unremarkable given the relatively few French speakers in England prior to 1066. But the difference in the temporal distribution of manuscripts after the eleventh century is striking. A relatively high number of manuscripts on Ker’s list were produced in the twelfth century, providing quantitative support for Ker’s qualitative observation that the century after the Norman Conquest was the most significant for book production in England in general.³¹ But the number of surviving manuscripts containing French literature produced in this period was comparatively quite limited. This stands in powerful contradiction with the traditionally held view that the efflorescence of French literature emerged as a direct response to the Norman Conquest of 1066.³²

The comparison reveals that the first wave of manuscripts containing French literature, which dates to the thirteenth century, appears to have been part of a broader increase in written production in medieval England. Generally speaking, increased book production was undoubtedly both a result of, and contributed to, the rise of commercial centres, the growth of universities, the proliferation of monasteries, and the development new technologies for fitting text on the page.

Why did the first wave of French literary production fall during this period and not, as might be expected, in the century following the Norman Conquest? Following the hypothesis put forth by Michael Clanchy, this particularity may be at least partially explained by the growing role of French in the thirteenth century as a language for international affairs, mercantile exchange and business transactions.³³ Within England, French was also gaining a foothold within a legal context in this period; in the last quarter of the thirteenth century, the dominant language for written statutes changed from Latin to French.³⁴ We should not be surprised to find that an increase in legal and administrative writing in French would come accompanied by a new interest in literary production.

These findings, which will be explored in greater depth within a sociolinguistic framework in the future, provide the first quantitative evidence for the persistence of French writing in England in the centuries following the Norman Conquest and suggest that the role of French in medieval England was not, as was once thought, an immediate result of the Conquest itself, but rather of a network of complex economic, social and international developments that took place in the thirteenth century. More broadly, the data suggest that England’s literary culture remained multilingual throughout much of the medieval period and support an ongoing challenge to the traditional ‘grand narrative’ of England’s linguistic history.

Discussion: Facilitating Future Manuscript-Based Research

The script that was used to calculate language distribution in this project is available online and comes accompanied by documentation intended to facilitate reuse.³⁵ Its function in this project—calculating the language distribution in a set of manuscripts, could also be deployed for exploring other manuscript datasets in a quantitative way. For example, comparing the linguistic distribution of surviving manuscripts between different medieval libraries would undoubtedly reveal valuable quantitative information about language use in these communities. In the future, if more consistently structured manuscript catalogue metadata becomes available—at the level of individual texts in manuscripts—it would also be possible to explore the linguistic distribution within manuscripts produced in a given region. For example, one could identify whether English was used more commonly in manuscripts produced in the West Midlands—typically considered a conservative linguistic area—than in those produced in the East.

Future work could also incorporate the growing knowledge of manuscripts’ provenance that various projects and databases are publishing. The Schoenberg Database of Manuscripts, for example, provides a large user-edited knowledge base of manuscript transactions.³⁶ This database will, over time, likely offer more complete and more rigidly structured data on owners that those that were collected for this project.

On a broader scale, the code produced for this project, by calculating the number of pages occupied by a manuscript’s constituent parts, could be beneficial to those wanting to compare the manuscript contexts of various texts. For example, the code could help provide quantitative data about changes in the makeup of manuscripts containing the Canterbury Tales over time. With more consistent manuscript metadata, the code could also be adapted to enable largescale comparisons of how various types of texts were copied. For example, data could be gathered into what percentage of thirteenth-century literary production was dedicated to chronicles or to poetic texts. While scholars make claims about the importance of various types of writing in a qualitative way, preparing quantitative data could yield new insights into medieval tastes, interests, and desires.

Given the significant and documented divergences among current approaches to manuscript description, however, quantitative studies such as this project require a great deal of preparation in order to render data machine-actionable, and therefore face barriers. So while it would theoretically be possible to very quickly analyze patterns in, for example, the dating of all surviving manuscripts produced in England, at present the lack of established and adopted metadata standards makes such a task prohibitively labor intensive. This project therefore highlights the need for more consistently structured manuscript data. Greater structural and financial support for this type of structured cataloguing would enable more efficient investigation into the textual culture of the medieval period and, in so doing, shed new light on other emerging cultural history questions.

Code and Data Availability

The code for analysing and converting the data to HTML is hosted at GitHub (https://github.com/LeidenUniversityLibrary/manuscript-stats) and archived in Zenodo: https://doi.org/10.5281/zenodo.1472267.

The normalised input files and results have been archived in EASY: https://doi.org/10.17026%2Fdans-zxr-juar.