This paper demonstrates the use of the ESTC as a representation of material history through the extraction and parsing of its data in a way which allows it to be used in social network analysis. In doing this it makes two contributions. The first is methodological, outlining how such a transformation of data is possible. The second is historical, by demonstrating how this data can be used to support historical claims.
The English Short Title Cata logue (ESTC) , according to the British Libra ry, is the comprehensive, international union catalogue listing early books, serials, newspapers and selected ephemera printed before 1801. Covering over 480,000 documents, wh ich are held by mo re than 2,000 libraries, it is an essential record of early English print culture, used by researchers to discover and locate early modern te xts. However, like other libra ry catalogues, it is “a greatly underestimated source of knowledge” (T olonen et al. 2018: 1).
Historians of the booktrade have noted the relevance of shifting relat ionships b
etween authors, printers, publishers, and other actors when studying the history of
print. The early modern develop ment of the English ma rket for printed bo oks has
been described as a process fro m an individual cantered medieva l c raft towards a
tightly organised wholesale business orchestrated through a London monopoly (Po
llard 1978). Publishing networks naturally, therefo re, p lay a central ro le when it co mes
to understanding the nature of these changes
The first contribution comes fro m the transformat ion of historical record into histo rical re lic; specifica lly, the process of taking a discrete set of catalogue records and turning them into a harmonized relational dataset rich in historical data. 2.1
As noted, the ESTC is the comprehensive, international union catalogue for Englis h titles. Its construction aimed to include a ll books (1475-1801), pa mphlets, broadsides, and other miscellaneous documents. Each document’s record can contain as many as 420 d iscrete data points, covering details fro m the document’s location in the colle ctions it is held, to its former owner. It is, there fore, an incred ibly rich source of histo rical materia l, a ll of which is recorded using the Machine -Readable Cataloging (MARC) 21 standard.1
While this means that the data is no mina lly machine readable, th e process of creating a linked dataset should not be trivialized. Originally designed to make the reproduction of catalogues simple r, the content of MARC records was not designed with e xtraction of linked data, and quantitative analysis of this data, in mind. There are a number of issues which ma ke using it for purposes other than locating te xts proble matic. Specifica lly, a more structured data model is required for catalogue data to be useful for systematic quantitative research.
To this end, the Helsinki Co mputational History Research Group (COMHIS) has an ongoing project wh ich a ims to e xt ract a ll data points fro m the ESTC.2 For the purposes of this paper, however, we focus on the ext raction of MA RC fie lds 100 and 700 (personal names); 110 and 710 (corporate names); and 260 (imp rint), and how we have created a relational data model in which each actor is an unique and identifiab le object connected to all t itles they were reported to have been involved with, wh ile maintaining the document-specific details of their particular role.3 1 The Library of Congress has an overview of M ARC 21 (https://www.loc.gov/marc/). 2 This paper is a part of this larger research project being conducted by COM HIS (https://www.helsinki.fi/en/researchgroups/computational-history). 3 The code used to parse the ESTC was written in R and Python.
MARC tags 100, 110, 700, and 710 contain informat ion with regard to actors involved with a specific docu ment. While there may be as many as 20 discrete pieces of data, the key points of interest include names; letters, initia ls, abbreviations, phrases, or numbers used in place of a name; tit les or other words associated with a name; dates associated with a na me (birth/death years, active years); the re lationship b etween a name and a work; attribution informat ion when the responsibility is unknown, uncertain, fictitious or pseudonymous; and other miscellaneous information.
With regard to d iscrete records, we have been able to e xtract 557,847 refe rences to actors from 397,061 documents (those for which there were actors named). Once e xtracted, these names were processed to correct nume rous unicode issues, data entry errors, and standardize accents. The ro le of each actor was either e xtracted (when available ), deduced fro m the MA RC tag, or marked as unknown. If a single entry had mu ltip le roles (e .g., “Publisher and printer”), new entries for the same actor we re created to reflect these multiple roles on the document.
Importantly, there is no recorded relationship between references to actors recorded in the ESTC. Thus, there is no way of knowing whether one entry for Willia m Shakespeare is those found elsewhere. Or, mo re proble mat ically, whether any or all of the 137 records for “A. B.” are the same person. Thus, before analysis was possible, extensive processing, harmonization, and validation were necessary.
In many cases records within the ESTC themselves were robust enough for this (for e xa mp le, when an actor’s na me included specific years of activ ity). In many cases, however, s uch informat ion was not available. In these situations we turned to the Virtual International Authority File (VIA F: https://viaf.org/), a harmonized record of mu ltip le international library cata logues which has collected and unified individual records and identifiers for actors. The aim was to take these actor identifie rs and apply them to records in the ESTC. To do this, mu ltiple matching steps were made, for e xamp le: e xact string matches; string matches minus particular types of punctuation; bag-of-word matches for records which contained data considered unique such as years of birth and/or death. Once these matches were e xhausted, we made use of the VIAF API to conduct fuzzy searches matching both actor name and the title of the document they were linked to. Th is resulted in unification of 523,955 records into 68,230 actors, with 34,288 references to actors unmatched.
There were proble ms, however. First, VIAF has duplicate records for the same a ctor, and these duplicates made their way into our own data. Additionally, there are harmonized records in VIAF which represent mu ltip le individuals with a shared name. Due to these issues, a number of verification steps were imp le mented to validate and further harmonize actors. Specifically, there is an extensive semi-automated verification process which looks for false positives (names un ified under the sa me ID, but which are not recorded as identical in the ESTC) and false negatives (names which were identical in the ESTC, but do not share an ID following the ha rmonization process). In these cases, historical knowledge was used to make a judgement. For the actors which we were unable to un ify, a final match was made between those with an e xact name string which we re attached to records for multip le editions of the s ame text. Actors which d id not meet any of the previous criteria were treated as distinct, regardless of similarity. At this point 558,243 refe rences to actors have been harmonized into 92,044 unique actors. However, the process has been written to make additional unification a simple task, and thus improvement is an iterative process.
MARC tag 260 covers the imprint field, notes relating to the publication, printing, distribution, issue, release, or production of a work. While the fie ld itself contains fewer data points (nine) than the actor fields, it is recorded in a much mo re proble ma tic manner. Specifically, the na mes of entities rep resenting booktrade actors are co ntained within a te xt chunks replicat ing the fu ll publisher statement. Thes e can be very amb iguous for standard computational e xtract ion methods. 4 To parse these statements, first relevant data was identified, differentiated, and separated from the rest of the imprint using natural language processing. This data is then identified as entities representing persons, roles, names of organisations, and locations.5
Once e xtracted, a simila r unification process to that outlined above is followed. First, post processing of names is necessary. This includes correcting spelling e rrors and normalizing spelling variations when sensical (i.e., Iohn becomes John; VVoo dcocke becomes Woodcock), and location info rmation and years of act ivity are used to match in itia ls with full na mes (e.g., J. Newbery and John Newbery). Entities are then unified as fa r as possible by cross -referencing names with e xisting records e xtracted fro m the ESTC, the VIAF, and the Brit ish Book Trade Inde x (BBTI: http://bbti.bodleian.o x.ac.uk/ ). As with the VIAF, there are issues with the BBTI – in particular duplicate entries which require further processing. Currently we are able to harmonize 35,252 unique actors from 332,410 discrete entries.
To conclude, we have extracted extensive metadata from the catalogue, including titles, publication years and places, edition and imprint informat ion, and all na med actors involved or related to a publicat ion. Although this process is largely automated, it has required e xtensive verification and post-ext raction c leaning. Particu lar attention was given to careful e xt raction and unification of na med entities. Additional enrichment steps are a lso applied. For e xa mp le, genders were e ither e xt racted fro m e xternal sources (VIAF) or assigned using a historical gender-name d ictionary wh ich we constructed using historical parish records.6 Overall, however, the processing and cleaning of this data is an iterative process, and we co ntinue to refine the data. 4 Two offer two examples: "printed for Bernard Lintott at the Cross-Keys, between the two Temple-Gates, in Fleet-Street. The Double Gallant: Or, the Sick Lady's Cure. A Comedy. Written by M r. Cibber"; “printed by E: Coates. 1655. Sould by Thomas Heath in Covent garden, and Henry Herringman at the Ancker on the lowest side of the New-Exchange.” 5 The current method makes use of the Stanford Natural Language Parser. 6 See: https://github.com/COM HIS/names_and_genders
As the aim of this paper is to use the e xtracted data for social network analysis, we needed to further transform the bibliographic records. To this end, we created node and edge tables in which documents are edges and unified actors are nodes. The cu rrent outputs result in a network of 72,066 nodes connected by 328,996 edges.7 Ho wever, due to the temporal co mple xity of the booktrade, further re fine ment was nece ssary. In particular, in order to discard actors (in particula r authors) who continued to be published after their deaths, the dataset was refined to only create links during an actor’s lifet ime, or active years. Additionally, because we were interested in changes over the entire t imespan of the ESTC, as we ll as during an individual’s life, further subsets, covering ten year periods which overlap by five years, were created.
It is important to reflect on the mean ing of these connections. The pre mise is that a
shared effort on a part icular document represents a real world h istorical conne ction.
While it is difficult to know whether a printer, for e xa mple , could really be said to
have had a close relat ionship with an author, the records nonetheless represent
relationships between actors taken fro m historical objects
The history of the English booktrade is t ied to London.9 By a large marg in the majority of ea rly modern publications in Eng lish ca me fro m London (317,679 to runnerup Edinburgh’s 31, 887). Th is is impo rtant because, although it was a large c ity (Lo ndon’s population grew to roughly one million people by 1800), the booktrade functioned on a human scale in which personal networks mattered, and geographical location reflected personal and intellectual connection . As Raven (2007: 155) reports, booksellers located themselves in pro ximity to their c lientele, and thus clustered amongst competition. Additionally, these locations could be fu rther linked through networks of publishers and printers specializing in g iven genres or types of public ations. A printer, for e xa mp le, could be attached to a specific bookselle r or publisher, even sharing a pre mise. Th is is important because it means that the networks which emerge in the ESTC can reflect very real spatial and personal relationships.10
The second aspect of the booktrade which suits it for network analysis is one of
technological limitation - specifically that of the hand-operated printing press. For the
entirety of the period covered by the ESTC, the key factor li miting increases in print
production was labour. As the numbers of publications increased, the industry (and
the humans who made it ) had to grow in tande m. The upshot of this reality (as we ll as
the centrality of London within the industry) is that the booktrade was made up of,
and dependent on, human networks
The results presented in this paper are meant as proofs of co ncept that bibliographic metadata can be transformed into historical network data (not to form a cohesive historical argu ment with regard to the booktrade), and to this end, the next section of this paper is divided into four parts wh ich apply network analysis to the ESTC: an overview of the network data e xtracted in re lation to authors; individual case studies; a uthorial centrality over the early modern period; and historical community detection. 3.1
As a starting point, an overview of data, as recorded and thus reported by the ESTC is offered (Fig. 1).
While the data for much of the beginning o f the period covered is sparse, this in itself
allo ws for co mparison. 11 As noted above, the growth of the booktrade was tied to
11 The jump in the amount of records covering the era of the civil war may be related to an
increase in political writings from the time, but the records of these writings can be tied to
increased labour - and this relationship is visible immediate ly: as the number of do
cuments printed went up, the size the net work g rows in tandem. However, the
transformation is more co mple x than this, as visible in the increase of cliques and drop in
edge density. These are both signs that the network was becoming more d iverse,
rather than a singular densely interconnected network, indicat ing an emerging structure
in which part icular nodes played central roles. 12 That is, as new individuals were
brought into the booktrade to increase production, aspects of the trade become
increasingly specialized and vert ically integrated with specific actors e merging as key
in managing and developing the industry
It should be noted, however, that there is much historical nuance and debate with regard to the details of the production end of the booktrade. Trade publishing, distr ibution networks, proofreading, imp recision in both the titles assigned to professions within the trade, and what was actually recorded on imprints, all ma ke the history of publishing an incredibly difficult fie ld to document, and this comple xity is c learly visible in the figures above.13 It is therefore important to, again, note that the data e xtracted fro m these networks do not represent claims as to a ne w understanding of the historical reality of the booktrade. Instead, this data should augment e xisting historical knowledge in a way which supports it and provides further insights.
With this in mind : when we investigate the relationships between roles specifica lly, we can see that this comple xity is not universal. Publishers, needing to develop more robust networks to advance their co mmerc ia l ends, gre w larger networks over t ime, while printers became, if not less entangled in these networks, less visible in the i mprints reporting them. 14 Authors, however, ma intained a simila r nu mber of conne ctions (Fig. 2); it was who these connections were with that shifted.
the existence of the Thomason Collection of Civil War Tracts, and thus claims with regard
to this increase should be tempered. On the other hand, the Great Fire is clearly responsible
for the loss of historical works, the records of works, and the potential for production.
12 These changes also reflect the reality of imprint records not always acknowledging all
involved in production. Additionally, the increase in the size of network cliques, particularly
in the 18th century, is representative of the historical practise of shared copyrights – a
situation which is clearly visible in Fig. 4.
13 To note how these historical difficulties may impact our data: as the records being used are
largely extracted from imprints, and as printers became less important in terms of the
distribution of works to the audiences purchasing them, their details may have been less likely to
be included. Additionally, as printing houses became larger, or the work more frequently
distributed between houses, the pragmatics of including all printers may have been
questioned. There is also ambiguity in imprint st atements which has previously been noted as
problematic. As Raven writes: “Recovering these trading relationships is not easy. M any
distribution agents are unspecified.” Having said this, the ESTC remains the best record of
these relationships we have, and as Raven importantly concludes: “Despite these
reservations, we can at least be certain that in most cases inclusion of a name in an imprint line
indicates some sort of financial involvement. This is particularly the case where the author is
mentioned”
Feather (1994: 79) has written that:
By the middle of the eighteenth century, it was possible to ma ke a decent liv ing fro m writing … Pat ronage was no longer essential even to literary a uthors, and the idea of the author as a partic ipant in co mmerc ial activ ity was beginning to be accepted. At the same time as authors were becoming more overtly me rcenary, there was also developing a more e levated idea of the author as artist or creator. The very concept of ‘originality’, in the sense of an ‘origina l work’ of art or literature, is an eighteenth-century idea.
What we may be witnessing in the figures above, then, is the shift fro m the “selfpublishing” author (if such a conceptual category could be said to have e xisted) to the established “professional author” wh ich, by the e ighteenth century, turns to e xisting publisher-headed networks, rather than printers, to manufacture and distribute their works (Raven 2004: 4). For such a claim to be considered, however, we must acknowledge another co mple xity in the data: the recorded drop in authoria l connections to printers is not uniform.
Fro m the middle of the 18th century the re lationship begins to reverse as pu blisher connections drop and printers once again increase in author ego networks. There are, at least, three historical reasons for this. First, the g rowth of the colonia l booktrade in the 18th century, where publishing had yet to take the form it had in London, meant printers were more frequently identified on imprints than publishers (Fig. 3). The impact of this distinction towards the end of the 18th century was ampl ified as, by 1792, non-London-based locations outnumber London for the first time.
The second cause for the recorded increase in printers was the 1710 Copyright Act. While the legality of the statute is complicated, in theory it ended perpetual co pyright after 28 years. Th is resulted in reprints of wo rks which entered the public do ma in first in the 1740s by printers outside London, and then in the 1770s by those in Lo ndon (both events visible in Fig. 4) (Feather 1994: 81). The third reason was piracy; while a work legit imately tied to an established publisher would ma ke this c lear on the imp rint, pirated works were, unsurprisingly, more vague in this regard (Feather 1994: 68).
Overall, it is clear that there is extensive historical data availab le in these network
overviews which o ffer longue durée insights - which is an important contribution in
itself fro m the historians perspective
The rea lit ies of the booktrade for individual authors are better seen when e xa mined independently. To demonstrate this we provide two case studies: Willia m Shakespeare and Bernard Mandeville . To begin, we o ffer a co mparison of the ego network of Bernard Mandeville as e xtracted automatically fro m the ESTC, and as constructed by a scholar of Mandeville (Fig. 4).
The two networks are similar in construction: relationships and names reoccur, and a cluster of publishers who shared the copyright of Mandeville ’s Fable of the Bees is visible in both. When discounting unknown actors and booksellers fro m the ESTC, there is only one actor in the hand-crafted network wh ich was not found in the ESTC network: the publisher John Peele.
There are, however, three addit ional nodes found in the ESTC ne twork: two authors and one publisher. Additionally, the ESTC network contains nine additional public ation records (five of wh ich record Mandeville as “Unknown,” and three as an attributed author) and 49 further unique connections between actors, a detail wh ich would be e xponentially mo re difficult record when constructing a network by hand. It is, therefore, both historically accurate and robust, and when turning to the makeup of the relationships further historical detail can be noted.
With regard to author popularity, death would perhaps be unsurprising as a limit ing factor. While this may not be entire ly intuitive in te rms of the nu mber of publications attached to an author, it is should certainly be the case with regard to social networks. 15 Data for hand-crafted ego network provided by M ikko Tolonen. Our data shows, however, that death need not entirely hinder one's social network; posthumous publications - and those who worked on them - need to be taken into consideration when using this data (Fig. 5).
This is an important point. Methodologically, it highlights the issue of simply e
xtracting all reported relationships. If one is interested in examining potentially real
historical relationships, data must take into consideration life and/or active years.
Historically, however, there are reasons one may be interested in exact ly these types
of posthumous records. While, as an author, Shakespeare remained in print after his
death, there are noticeable peaks and troughs in his connection counts which are
indicative of known editions issued by different publishers.
When looking at ego maps from d ifferent periods, we can clearly identify these
publishers (Fig. 6). What is more, the networks themselves offer addit ional contextual
detail - specifically, co mpetition and shifting changes in importance. For examp le, in
1730-1739 we see distinct connections between feuding publishers Robert Walker
and the Tonsons, the latter o f which would co me to do minate the network in the
middle of the century. By 1770-1779, however, the market had changed. John Bell,
whose more cheaply printed works, “satisfied public demand for inexpensive
editions,” emerged as a key actor, and remained vis ible until the end of the century
although further competition in the discount-edition market can be seen in Edward
Harding’s arrival
While these examp les are, purposefully, built upon previous research, they offer demonstrations as to the novel ways in which the ESTC may be used to track authorial relationships at different historical moments. 3.3
As Wasserman and Faust (1994: 169) have noted, “[o]ne of the primary uses of graph theory in social network analysis is the identification of the ‘most i mportant’ actors” through centrality measurements – a quantitative method which, by taking into account the relative relat ionships between actors, can identify key nodes. What makes a node key depends on the measurement chosen, with each emphasizing particular types of relationships. Thus, careful assessment of one’s data and the usefulness of each method is necessary when calculating centrality. With this in mind, this section of the paper applies four centrality measurements - degree, betweenness, closeness, and eigenvector - to the ESTC with the aim of both assessing their particular uses, as well as validating the historical richness of the data.
We extracted the ten most central authors per ten year time slices (to ensure contemporarily meaningful results) and compared them: first, against the most published authors for the same t ime period; second, for instances of direct overlap (i.e., were the same people identified by different measurements); and third, mo re qualitatively, in terms of how representative those authors were with regard to the historical period they were extracted from. Below is an overview of these results.
Fro m 1500-1535, an era which one may identify with renaissance humanism, d egree, betweenness, and closeness centrality measurements most frequently return to humanist authors (e.g., Erasmus, Tho mas More, Sebastian Brant, Baptista Mantuanus) while eigenvector and the most-published pointed to religious authors and grammarians. During the reformat ion (1535 to the mid-16th century) all measurements identified reformation and counter-reformation figures (e.g., Martin Luther, William Tyndale, Tho mas Beccon, John Kno x, Jean Calvin, Theodore Beza). Ho wever, those found in the most-published category are almost exclusively English authors, some of who do not show up in any of the centrality measurements (John Hooper and Ed mund Bonner). The do minance of the reformation continued for the first half o f the Elizabethan era, although by the 1570s familiar dramat ists, poets, and other literary figures emerge (although Puritan authors remain present throughout). Authors include: George Gascoigne, Tho mas Newton, Thomas Nash, Robert Greene, Shakespeare, Thomas Heywood, and Michael Drayton. This trend generally continues during the Jacobean era (1605-1640) with the exception of eigenvector centrality which, interes tingly, takes an explicitly Anglican and royalist turn, returning (as authors) James I and Charles I, Lancelot Andrewes, John White, William Laud, and John Williams. Unsurprisingly, during the civil war (1640-1655), literary authors disappear and are replaced by overtly political and religious authors (especially nonconformists). On the political end of the spectrum we find, amongst others, Charles I, Oliver Cro mwell, Thomas Fairfax, Ed ward Montagu, John Pym, Robert Devereu x, and Christopher Love. Religious figures include George Fo x, James Naylor, W illiam Prynne, Richard Baxter, and the first colonial-based author identified, John Cotton. Fig. 7. Great Fire of London’s impact on publications. The redline marks the year of the fire (1666).
Fro m 1660-1675 the data is, unfortunately, tarnished by the Great Fire o f London which, while taking place in 1666, resulted in the loss of texts going back to at least 1660, and impacted production for years to follow (Fig. 7). There is little cohesion in terms of authors identified as central during th is period, although non-conformists dominate the publication records .16
The restoration period and beyond - the “Augustan Age,” lasting roughly 1660 until 1785 - can be conveniently divided into three authorship -defined sub-eras: the Age of Dryden, the Age of Pope, and the Age of Johnson. The centrality of the na mesakes for these literary e ras is quic kly confirmed. During the first, Dryden is more frequen tly identified than any other author until 1695 (a lthough other important litera ry a uthors returned include Roger de Coverley, Elkanah Settle, Thomas D’Urfey, Abel Boywer, and Gilbert Burnet). This is followed by the Age of Pope (1700-1745) which, again, includes Pope himself, Danie l Defoe, Jon athan Swift, Richard Steele, Eliza Haywood, Co lley Cibber, A mbrose Ph ilips, and many others. The Age of Johnson, from 1750 until his death in 1784, is again dominated by the literary g iants of the era, such as Johnson himself (although he first appears in 1765), Oliver Go ldsmith, Tobias Smollett, John Lockman, George Colman, Hannah Cowley, etc.. This e ra of litera ry centrality ends abruptly, however, as politica l turmo il aga in e merged – this time round the American and French revolutions – and key figures shift to the likes of Thomas Paine, George Washington, Benja min Franklin, Joel Barlow, Willia m Cobbett, and Edmund Burke.
For the most part, these results are rema rkab ly in line with historical e xpectations. To better compare the measurements, however, they were corre lated into a series of tables from which the follo wing visualizat ion was constructed (Fig. 8). Each bar represents a specific measurement, and segments ten historical eras. The opacity of each segment represents how well each measure ment reported authors deemed to be relevant to a given historical era. The aim is to offer a simp lified representation from which the usefulness of the measurements can be assed. 16 There are two potential reasons for this: one is the prominence of the Dutch printing nonconformist texts in Europe during this time; another being nonconformist social groups in London being located in the East End, away from the fire.
The first finding to dra w one's attention to is that the centrality measure ments seem to successfully identify key figures, and often better than simple publication records (although Eigenvector is an outlie r). Additionally: there is a la rge a mount of diversity between who measurements report as central - especially towards the end of the period covered. Eigenvector was, by far, the most unique, with nearly 50% of a ll results not found in other measure ments. This was followed by closeness (36%), betweenness (27% ), and degree (11%). Additionally, 27% o f the authors - largely re ligious returned as most frequently published were not found in any of the centrality mea surements. Interestingly, authors often appear in the publication count category after they are first noted by a centrality measure ment with one e xception: the first half of the eighteenth century, a time coinc iding with the development of the publisher. While beyond the remit of this paper, this points to a comple x structural re lationship between publishing as an industry, and the popularity of an author.17
The results generated from the eigenvector centrality measurements are also worth noting. Figure 8 reports that, as a measurement, eigenvector was quite poor at identifying authors representative of a given period. There we re e xceptions to this, however - in particu lar, in the lead up to, and during the, civil war. To understand the reasons for this one must understand that eigenvector identifies centrality by weighing each node by the centrality of a ll other nodes. That is, any given node is as central (or as important) as the nodes it is connected to. The u pshot of this is immediate ly visib le in the data: nodes identified as central by eigenvector measurements often represent 17 There are, of course, even larger questions which one should ask: for example, do the results represent a central author because of their reception while alive, or do the connections which we are now identifying play a role in our perception? Additionally, the records for those who we have come to expect to be important are likely to be the bibliographic records which are more robust. more established institutional figures. That is, in the lead up to the civil wa r we find royalist and Anglican authors, but at the start of the civil war they are rep laced by parlia mentarian authors. This is perhaps further h ighlighted by Henry Scobell b eing identified in the 1650s. While ce rtainly an imp ortant figure in h is own right, what perhaps contributes to him being identified as central in the entire network was his role as editor of parlia mentary proceedings and licensor of newspaper and political pamphlets. This institutional centrality, however, means that the measurement is perhaps less appropriate in situations - or with data - in which a clear hiera rchy is not visible. For e xa mple , at the end of the 18th century, while most measurements ident ified leading polit ical authors, eigenvector returned various literary figures - with one exception: the very much institutional figure Edmund Burke. 3.4
Centrality is not the only way in wh ich institutions or establishments can be ident ified in networks. With a lgorithmic co mmunity detection the ESTC can be used to identify diffe rent and changing intellectual groups. To do this, we e xtracted (up to) the ten largest commun ities for each overlapping ten year slice using the fastgreedy algorith m.18 This provided 577 co mmun ities varying fro m as fe w as two actors to over 1,200. To test whether these represented actual historical co mmunit ies we developed a semi-auto mated classificat ion method. First, we e xtracted a ll actor IDs attached to a given community, purging non-authors (e.g., publishers). 19 These IDs were then lin ked back to ESTC IDs, which a llowed us to ext ract the titles and subtitles for the entirety of each commun ity. The tit les representing the 577 co mmun ities we re still too numerous to investigate manually (the total collection of ra w te xt tit les was over 350 megabytes). We therefore token ized tit les, counted the occurrences of each type per community, and converted them into proportions. We then measured each individual community against its peer-communities as a single entity. From this, we e xtracted the tokens for each community wh ich had the highest distance measurement, and a Z score for two population proportions which was statistically significant (p < 0.05). The results were less than a megabyte, and much easier for a human to parse.
When exa min ing the time slices covering the English civil war, the contextual
mean ing of the tokens for two h istorical groups looked immediate ly pro mising. In one
community we identified words such as: army, lieutenant, parlia ment, forces, co
mmons, protector, souldiers, victory, wa rre, and cro mwell. In the other, we found: king,
ma jesties, procla mation, parlia ment, declarat ion, gracious, subjects, charles, royall,
kingdom, and dominion. That is, the two co mmunities appeared to represent the pa
rliamentary and royalist sides of the civil war (plotted below in Fig. 9).
18 Fastgreedy was chosen due to the amount of data that needed to be processed. Better results
are likely to be achieved using a different method
To turn to another e xa mp le, one which does rely on as clea r h istorical divisions : In the 1645-1654 time slice one can find a co mmunity of religious authors, including Quaker founders George Fo x, Ed ward Bu rrough, and Francis Howgill, as we ll as various other nonconformists and Puritans. Interestingly, when one follows these actors through time slices, a Qua ker-specific co mmun ity develops, made up of many more pro minent Quakers (inc luding Margaret Fe ll, George Whit ehead, Ja mes Nay ler, and nearly half of the ‘Valiant Sixty’).
While e xp loratory, these methods, again, offer a number o f pro mising avenues for researchers interested in identifying, exploring, and tracking historical authorship.
The goal of this paper has been to offe r an overview of the processes used to turn the
ESTC into a historical re lic , rather than produce a historical a rgument with regard to
authorship as a whole. This re mains impo rtant future research. Howe ver, we believe
that there are many pro mising contributions in this work with regard to both its
methodological and historical cla ims. With regard to the forme r contribution, we envision
a wide variety of scholars who could ma ke use of this work - inc luding historians,
linguists, and literary scholars. Fo r e xa mple , h istorical sociolinguists aimin g to
reconstruct and analyse authorial social networks may find the processing and parsing
procedures particularly info rmative (Be rgs 2005; Sairio 2009). Additio nally, there is the
possibility of co mbining this type of metadata with massive full-te xt databases like
Ea rly English Books Online (EEBO) and Eighteenth Century Collections Online
(ECCO). The latter - historica l - contribution made by this paper is of particular
relevance to book history. By introducing the statistical measurements unique to social
network analysis, this paper augments and improves upon the existing, albe it sparse,
quantitative analysis of the early English booktrade.20 Additionally, we a re part
icularly interested in its potential contribution to intellectual h istory, and the possibility of
using these methods to reconstruct intellectual conte xts , such as the relationships
between specific actors involved in producing and disseminating philosophical and
politica l te xts. Bib liographic records have a lready shown themselves to be an i
mportant foundation for research in this fie ld