Bibliographic metadata as relational data: A cross-disciplinary methodological reflection

Identifying Historical Actors: Disambiguation and Authority Alignment

Catalogs often record the names of both personal and corporate entities associated with a given holding, in line with the MARC 21 schema.³ These entries may include additional identifying information such as life dates or floruit, numeration, pseudonyms, religious affiliation or noble titles, and places of activity, as well as the role under which the individual appears in the publication. One of the most challenging steps in working with bibliographic metadata is the identification and disambiguation of historical actors based on these name strings. This difficulty arises from variant spellings, inconsistent orthography, and uneven role attribution across entries. These challenges may be more or less pronounced depending on the uniformity of cataloging conventions and the degree of authority control implemented, and this may apply to curated collections and union catalogs alike. To construct meaningful relational data, it is essential to reconcile these inconsistencies and unambiguously identify historical individuals. The disambiguation step ensures that relationships in the data are constructed around unified entities instead of fragmented or duplicated name strings. This process relies on accompanying biographical attributes to distinguish homonyms and resolve pseudonyms. In the case of homonyms, these additional attributes can help differentiate distinct individuals with similar names, while in the case of pseudonyms or variant forms referring to the same person, shared dates or roles may allow for accurate consolidation even when string similarity alone is insufficient. Ideally, at the end of this reconciliation process, one should be able to assign a persistent identifier to each person or corporate entity.

In practice, two main strategies are typically employed: a bottom-up approach, based on string similarity and clustering, or a top-down strategy that aligns name strings with external authority files. Both techniques can be employed at different stages of the workflow and often both might benefit from a semi-automated approach. For instance, fuzzy string matching and clustering algorithms can be used to group likely name variants, which are then manually reviewed for validation. Similarly, a name search can be launched against a selected external authority file to identify potential matches, but the final assignment of identifiers should remain under human supervision.

In my pilot study using the Collectio academica antiqua, I implemented a structured pipeline to extract, standardize, and cluster personal and corporate names from MARC 21 records. I began by identifying the relevant MARC 21 tags containing names of interest, then used pymarc to extract their content into a flat, tabular format while retaining the original tag associated with each entry. I grouped all name strings from subfield $a into a single column, regardless of the MARC field they originated from. I applied the same principle to other associated attributes–such as numeration, life dates, and roles–by creating standardized columns based on subfield content, across different MARC tags.

At this stage, I cleaned the strings to ensure consistency: I removed leading and trailing spaces, standardized the formatting of numeration and dates, and eliminated unnecessary punctuation. While reviewing the grouped values, I also checked for misplacements (e.g., numeration incorrectly stored in a name subfield) and reassigned them to the appropriate column when necessary.

I then processed the resulting dataframe in OpenRefine, which allows for semi-automated clustering with high user oversight.⁴ I applied different clustering techniques and manually evaluated the suggested matches. I recommend working on a duplicate column and proceeding in successive passes: first clustering on the combination of surname, given name, numeration, and parsed life dates (possibly formatted as YYYY–YYYY or as YYYY in case only one date is available); then reapplying clustering without the dates; and lastly filtering by common life dates to identify remaining variants that earlier steps may have missed. After this bottom-up procedure, I adopted a top-down approach by matching the clustered names against an external authority file. Given that my case study mainly involved early modern persons active in Europe, the Consortium of European Research Libraries (CERL) Thesaurus proved particularly useful.⁵ I relied on the dedicated Python library, cerl, which is a wrapper for the CERL Thesaurus API.⁶ However, other reconciliation workflows are equally viable. Many thesauri, including Virtual International Authority File (VIAF), Gemeinsame Normdatai (GND), and Wikidata, offer dedicated reconciliation services within OpenRefine or provide public APIs, which allow users to build custom pipelines for automated queries. Crucially, querying authority files often returns more than one potential match per name. These results must be carefully reviewed and manually approved, ideally by inspecting supporting metadata such as life dates, places of activity, or roles. This step requiring human validation is essential to ensure accurate reconciliation, especially when dealing with common names or minimal contextual information.

External authority files offer three main advantages. First, they often have already resolved duplications across records by grouping known name variants. Second, they enrich the dataset with additional information such as birthplaces, deathplaces, or institutional affiliations, which can be repurposed in later stages of our own analysis. Third, if the metadata needs to be merged with other datasets, records reconciled against the same thesaurus can be matched more reliably and seamlessly by relying on the same identifier. Of course, authority files come with some limitations, namely, they are inevitably incomplete, and many lesser-known historical figures remain unlisted. In such cases, these individuals will likely be unmatched in the thesaurus. Still, the majority of complex name variation tends to cluster around well-documented individuals, precisely those who are more likely to appear in authority files. In practice, researchers may need to consult multiple thesauri. To bring consistency to the resulting patchwork of matches, it becomes essential to generate a consistent internal identifier system, one that accounts for entities matched across authority files, those identified through clustering, and those that remain unmatched.

In the context of my pilot study on the Collectio academica antiqua, the disambiguation process began with 5,018 name strings extracted from MARC 21 records. I first applied a bottom-up procedure using OpenRefine’s clustering functions, combined with biographical information such as life dates and numeration. I then adopted a top-down approach by matching these entities to external authority files. I ended up identifying a total of 4,192 distinct individuals. Among these identified individuals, 47.5% (2,033) were matched to the CERL Thesaurus and a further 5% (213) to other authority sources such as Wikidata or Onderzoekssteunpunt en Databank Intermediaire Structuren (ODIS). The remaining 2,032 (47.5%) had no external match and were retained as internally disambiguated entities.

Disambiguation remains one of the most time-consuming steps in preparing bibliographic metadata, yet it is essential. Without it, individual identities would not be consistently established, and relationships would rest on unreliable associations. Crucially, it is what gives the data meaning for relational study. The process can also create synergies: matching entities to authority files often yields additional information that proves useful even beyond the immediate project.

Geolocating Place Names

Bibliographic metadata is typically rich in geographic information because it often records the place of publication, manufacture, and distribution for each library holding. As with personal and corporate names, place names are typically transcribed as they appear on the title page. This practice, once again, prioritizes fidelity to the original source over internal consistency of the catalog. As a result, the same city may appear under multiple variants. These may reflect different languages, historical spellings, political jurisdictions, or typographical inconsistencies. To make this information usable for relational analysis, place names must be standardized and geolocated. This task is not fundamentally different from disambiguating historical actors, but it is often somewhat easier due to the more stable nature of place identities. Nonetheless, caution is essential. Even with present-day place names, it is surprisingly easy to misidentify a location because many places can share the same name within or across countries. This issue may be less common for cities that are on average better known and thus more consistently recorded, like printing centers. Still, historical geolocation poses a challenge. The process often requires close attention to historical context and supporting metadata.

For harmonizing place names, best practice is to reconcile each string against a standardized entry in a historical urban gazetteer or geographic authority file. These resources contain a wide range of name variants, including multilingual and obsolete forms, which significantly facilitate disambiguation and geolocation. Among the most comprehensive options are Wikidata, GeoNames, the CERL Thesaurus, the RBMS/BSC Latin Place Names File, and the World Historical Gazetteer, all of which are particularly well suited to historical case studies.⁷ In my pilot study of Louvain’s Collectio academica antiqua, I processed 631 valid place-name strings, reconciled into 166 distinct locations. Table 2 reports, for the ten most frequent printing locations, the number of distinct place-name variants and the corresponding number of imprints.

After harmonization, locations can be geographically referenced. Although coordinates typically refer to a single point rather than the full spatial extent of a city or region, they still enable meaningful spatial representation. For example, in Schich et al. and de la Croix and Scebba, coordinates are used to visualize city attractiveness based on the mobility of, respectively, notable individuals and medieval and pre-modern scholars affiliated to European academic institutions.

Once coordinates are available, they can be plotted on vectorized shapefiles or digital atlases, which represent geopolitical or cultural regions as spatial polygons. In historical analyses, it is particularly important to work with maps that reflect time-consistent boundaries. Using modern administrative boundaries for early periods can misrepresent political geography and bias spatial patterns. A wide array of historical vectorized digital atlases or shapefiles are available to support this kind of analysis. Both computational humanists and applied economic historians are increasingly contributing projects in this direction. However, the limited integration between disciplines means that humanists are often unaware of the spatial datasets used by economists, and vice versa. These resources may vary in geographic coverage and temporal resolution, where some provide century-scale snapshots, while others offer yearly updates with shifting polity boundaries. Below, I present a selection of the most commonly encountered and widely used resources across computational humanities and quantitative economic history.

• The Centennia Historical Atlas: Academic Research Edition offers high-resolution, time-sensitive boundaries of European polities between the late Middle Ages and the nineteenth century (Reed).

• The Cliopatria database records overlapping and successor polities from 3400 BCE to the present (Bennett et al.).

• EurAtlas provides digital cartographic reconstructions of European boundaries spanning two millennia.⁸

• Developed by economist Victor Gay and hosted by Harvard Dataverse, the Third Republic France Geographic Information System (TRF-GIS) maps French administrative constituencies spanning the period 1870–1940.⁹

• The IPUMS Mosaic project, as the name suggests, provides a patchwork of harmonized historical census microdata and corresponding geographic boundary files for selected countries in the European area at different points in time.¹⁰

• The Historical GIS Collection at ETH Zurich, compiled by computational sociologist Ramona Roller, supplies static polygon boundaries for territories within the sixteenth-century Holy Roman Empire, along with attributes such as foundation dates and confessional status.¹¹

• The project (Re)counting the Uncounted, led by historian and digital humanist Rombert Stapel, produced the Historical Atlas of the Low Countries, a detailed GIS dataset designed to anchor premodern population censuses to historically accurate administrative units.¹²

A comparative overview of these resources is provided in Table 3, which summarizes their temporal and geographic coverage, resolution, and accessibility. While the Centennia, Cliopatria, and EurAtlas shapefiles offer broad historical coverage—either pan-European or global—and span from antiquity to the modern era, they were primarily designed as general-purpose tools and are widely used in applied economic history. By contrast, the other atlases are more narrowly focused, both in geographic scope and in time frame. Figure 1 plots publication centers of the Collectio academica antiqua’s holdings over time, employing time-consistent and evolving boundaries.

Figure 1

The maps show printing centers in the Low Countries and surroundings at fifty-year intervals. Bubble area is proportional to the number of printed holdings from the Collectio academica antiqua associated with each city. Historical boundaries are time-varying and drawn from the Centennia Historical Atlas.

Classifying Content: Assigning Topical Terms

Library catalogs typically include some indication of what each book is about. In the MARC 21 schema, this information tends to be stored in tags 650 and 655, which refer respectively to topical words and genre categories. However, this metadata might be incomplete, inconsistently applied, or skewed by cataloging practices. While a close reading approach might allow for a direct understanding of the subject of each book, this method is not feasible in large-scale, data-driven studies. In these cases, a systematic and ideally automated workflow for classifying the content of library holdings becomes essential. Most of these approaches draw on a key element, that is, textual input, most commonly the book title, which in early modern imprints is often lengthy and content-rich. Although full-text content could, in theory, offer even richer input, the focus here is on completing and standardizing existing catalog metadata to enable structured comparisons across records. Assigning accurate subject metadata is an essential preprocessing step for higher-level analyses tracking thematic trends over time, identifying disciplinary clusters, or studying the diffusion of ideas across communities.

Automated classification techniques broadly fall into two categories: supervised and unsupervised methods. Supervised methods rely on a labeled dataset, that is, a subset of records for which the correct classification (e.g., field or topic) is already known. Machine-learning models are trained to learn these associations and predict labels for unclassified data. An example of supervised topic modeling in the case of bibliographic metadata is introduced in Koschnick and it involves fine-tuning a pretrained transformer model such as DistilBERT (Sanh et al.) on a dataset of titles or abstracts paired with subject headings. The model learns to classify each input into one of the predefined topics, based on patterns in the text. Unsupervised methods, by contrast, do not require labeled input. Topic modeling techniques, such as Latent Dirichlet Allocation (LDA) (Blei et al.) attempt to infer hidden thematic structures by analyzing word co-occurrence patterns in a corpus. Similarly, algorithms such as k-means clustering can group texts based on similarity in their vector representations without any prior knowledge of categories. Applications of k-means clustering to bibliographic metadata include the analysis of genre-indicating subtitles in German literature (Gittel) and the definition of academic fields from publication titles (Curtis and de la Croix).

While unsupervised methods can help identify thematic patterns, they are often less appropriate when the goal is to complete or standardize existing subject metadata. In such cases, including the present paper, supervised approaches are favored because they directly learn from cataloged examples and preserve alignment with established classification schemes.

In my case study of the Collectio academica antiqua, I used a supervised approach. I identified a topical classification term for 38.4% of the holdings (1,450 out of 3,777 records). This subset served as the labeled data for training. I then proceeded to derive a higher-level classification by grouping subject headings in more general categories. To assign topics to the books in the Collectio academica antiqua that lacked subject metadata, I fine-tuned a DistilBERT model to classify titles based on the labeled records. Each title was mapped to one of three broad disciplinary categories—theology, humanities, or sciences—based on the existing catalog labels. Because the model was pretrained on English text, I first translated titles to ensure compatibility. The model was then trained over four epochs, that is, it went through the entire training dataset four times to refine its predictions.

To evaluate the model, I set aside 20% of the labeled data (290 titles) as a validation set. Model performance differs markedly across categories. As displayed in Table 4, precision, recall, and F1 scores are high for theology and humanities, the two dominant classes in the labeled data, indicating that the classifier learns stable and interpretable patterns for these categories. By contrast, the sciences category is severely underrepresented in the training set, and the model fails to predict this class in the validation data, resulting in zero recall for this category. As a consequence, overall accuracy is driven by the two majority classes, while macro-averaged performance metrics provide a more informative summary under class imbalance.

Figure 2 illustrates this imbalance, reporting both raw counts and row-normalized percentages in order to make performance comparable across unevenly sized classes. Each row of the matrix represents the actual category of the entries from the validation set, while each column shows the predicted category assigned by the model. If the model worked perfectly, each class count would fall in the diagonal, where predicted labels match the actual ones, and the off-diagonal entries, which indicate misclassification, would be empty. Most errors occurred between theology and humanities, which reflect their historical overlap. The sciences category is severely underrepresented in the training data, and the model fails to predict this class in the validation set, reflecting the limits imposed by class imbalance. Once the model was trained, I used it to assign topics to the remaining set of titles that lacked subject metadata. These predicted labels were then combined with the original metadata to generate a complete, standardized topical classification across the entire collection.

Figure 2

Confusion matrix for the validation set of the labeled holdings in the Collectio academica antiqua. Each row shows the true topic based on catalog metadata, and each column shows the model’s predicted topic. Values along the diagonal (off-diagonal) indicate correct (incorrect) classifications.

Adopting a Network Perspective

Catalog records are not inherently networked systems. However, they do encode structured associations among entities—authors linked to books, books linked to subjects, people linked to other people via shared roles in a publication, and so on. A network representation emerges only when we interpret these associations under a specific set of assumptions. For example, if two agents appear together on a title page, one might interpret that co-occurrence as a connection between them. Many studies in the computational humanities adopt this strategy. For instance, Gavin and Hill et al., “Reconstructing Intellectual Networks” on early English literary criticism and book trade, respectively; Greteman on the English print network from its origins through the eighteenth century; Ladd on early modern dedications; Valleriani et al. on networks of printers and publishers and their influence on the evolution of scientific knowledge; Ryan and Tolonen on Scottish Enlightenment publishing; and Heßbrüggen-Walter on interdisciplinarity in seventeenth-century German dissertations. All of these works use the co-occurrence of individuals in print artifacts as an interpretative lens to reconstruct intellectual, textual, or social communities. This raises two broader methodological questions: Under what interpretive assumptions can bibliographic metadata be treated as network data, and what are the limits of such representations? The Collectio academica antiqua provides a particularly fitting test for these questions because it documents a well-bounded institutional world (i.e., an academic print production orbiting around the university center of Louvain) where the logic of collaboration can be observed directly in the paratext and imprint data. This discussion directly addresses the second research question mentioned early on in the paper, by making explicit the interpretive assumptions under which bibliographic metadata can be treated as network data and by delineating the limits of co-occurrence-based representations.

First and foremost, library catalog data were never intended to capture explicit relationships between historical actors. Their purpose is providing bibliographic description and information management. Nonetheless, converting a collection’s metadata into a co-occurrence network means establishing links based on the structured associations present in the records (people, texts, places, and topics connected via shared bibliographic entries). This reinterpretation is not without pitfalls. A key risk is assuming that co-occurrence automatically implies a meaningful social relationship. Simply sharing a title page does not guarantee contemporaneity, collaboration, or even mutual awareness, especially in cases like posthumous editions or honorary dedications.

Yet, despite this ambiguity, the connections we derive from bibliographic metadata are grounded in historical evidence, that is, the sources themselves: title pages, imprint statements, colophons, and dedications. They are not arbitrary, ad hoc constructions imposed by the researcher. While the meaning of each connection may be open to interpretation, the very co-occurrence in the production of a shared work constitutes material evidence derived directly from the historical record (Hill et al., “Reconstructing Intellectual Networks”). At the same time, bibliographic metadata can sometimes support a stronger definition of network ties than simple co-occurrence. For instance, using bibliographic metadata from the Sphaera corpus, a curated collection of 359 astronomy and cosmology textbooks published between 1472 and 1650, Valleriani et al. reconstruct a network of book producers by defining “awareness relationships” between early modern printers and publishers. Such relationships arise when two similar editions are attributed to different producers, typically indicating imitation or participation in the same print run. Here, bibliographic fingerprint of editions and additional metadata on printers’ and publishers’ biographical timelines are used to establish, respectively, chains of editions and contemporaneity, yielding ties that can be interpreted more robustly than co-occurrence alone.

The Multilayer Approach

With these caveats in mind, bibliographic metadata have been used to construct various kinds of networks depending on the entities and relationships of interest. One of the most common projections is a person-to-person network where an edge represents two individuals’ joint participation in the same publication. This co-occurrence network is frequently employed to study scholarly or intellectual communities. Even this straightforward representation involves important modeling choices. A first decision concerns an edge’s directionality. Defining edges as directed or undirected depends on the interpretative aim. For example, cases reflecting unilateral acknowledgment like links from authors to dedicatees might call for directed edges, whereas undirected edges more accurately capture the symmetric nature of shared participation in other contexts. A second choice concerns encoding some data as node attributes (e.g., relevant life dates, biographical information, and religious or institutional affiliation), or as edge attributes (e.g., the publication year and the number of shared works).

However, there is one critical dimension that resists both types of attribute encoding: the role of each person within the book. This is because roles are not fixed properties of individuals. Rather, a node may be an author in a publication and a dedicatee in another. Nor do they characterize the tie itself, because each tie reflects individual involvement in a shared work, not the work as a whole. For example, the connection between a printer and an author cannot be reduced to either an attribute of the person or an attribute of their tie.¹³ Basically, individuals’ roles in each publication are asymmetric and two-way, which disqualifies both node and edge attributes as adequate carriers of this information. Capturing these asymmetries is essential for understanding early modern print economies, where the same person could occupy different positions across publications.

Hence, I argue that the most coherent solution is to model the co-occurrence network as a multilayer graph. Following the definition offered by Kivelä et al., a multilayer network extends beyond nodes and edges to include layers as core components, allowing each node to appear in multiple layers and enabling edges to connect any pair of node-layer instances. Additionally, multiple aspects, or features—that is, different sets of layers—can be modeled on top of each other (209). In practice, an aspect defines a distinct dimension of variation in the data, such as the type of activity, the temporal slice, or the institutional context, so that multiple aspects can coexist within one unified, multilayer structure.

In the present case, I only focus on a single aspect, the co-occurrence dimension in the catalog, where each layer corresponds to a distinct role and where nodes may participate in one or more role-layers depending on their involvement in each publication. This framework allows for both inter-layer edges (i.e., connections stemming from shared works in different roles) and intra-layer edges (i.e., ties between individuals sharing the same publication and role). This representation makes it possible to measure how role multiplicity structures collaboration and to distinguish between within-role cohesion and cross-role integration.

Figure 3 offers a visual example of what a multilayer co-occurrence network can look like.¹⁴ I built a person-to-person co-occurrence network from the bibliographic metadata of the Collectio academica antiqua. The layers are defined from the roles recorded in the metadata, which I grouped into broader categories as detailed in Table 6 in Appendix A.1.¹⁵ The figure illustrates the methodological shift from descriptive catalog lists to an explicitly relational view, where each role layer isolates one channel of participation. Although the full network comprises several thousand individuals and ten role layers, the figure focuses on the fifty most active nodes and four main layers to improve legibility. In this representation, layers are the loci where nodes appear when a given role emerges in at least one shared publication. Thus, if an individual has different roles in different works, they appear in multiple layers. The layer classification preserves the co-occurrence principle: Edges are still defined by co-occurrence in publications. Dashed lines denote inter-layer edges and solid lines denote intra-layer edges.

Figure 3

Multilayer person-to-person co-occurrence network constructed from the bibliographic metadata of the Collectio academica antiqua. Each layer isolates one type of role-based connection. Nodes represent individuals (only those with highest degree are shown for legibility). Intra-layer edges (blue lines) and inter-layer edges (red lines) connect individuals who co-occur in at least one book, respectively in the same or in different roles. The plots were generated with pymnet (Nurmi et al.).

Thinking of co-occurrence in multilayer terms allows for several modes of analysis. One can examine each layer separately to study the network of collaborations within a single role. This is particularly useful when intra-layer ties dominate or when the aim is to identify role-specific structural patterns. In practice, within-role co-occurrence in early modern data tends to be sparse. In the Collectio academica antiqua network, fewer than one-quarter of the ties link individuals who share the same role in a publication; more than 75% of connections are between people in different roles (inter-layer links; see Table 8 in the Appendix). This is consistent with the low density observed in each individual layer (Table 7), which in turn reflects the historical reality that collaborative roles, like for instance co-authorship, were relatively rare in early modern print culture. If we were to apply this method to modern scientific publications, we would expect to see a much denser authorship co-occurrence layer, whereas the early modern dataset under analysis shows that authors typically connect to others through complementary roles (e.g., author to printer, author to dedicatee, etc.) rather than directly to other authors. This dominance of inter-layer ties underscores how production and intellectual labor were intertwined: Most connections bridge material and scholarly functions rather than occur within a single occupational sphere.

Second, the layers can also be collapsed into a single projection when applying classical metrics such as centrality or modularity, though doing so sacrifices the interpretive nuance that multilayer distinctions offer.

Third, and most distinctively, one can analyze patterns that involve the interplay between layers, something not possible in a single-layer projection. In this case, the layers linked to material production (printers, publishers, and booksellers) provide a controlled setting to study how the combination of roles across publications relates to the university ecosystem from which much of the corpus originated.¹⁶ Focusing on the production side is particularly revealing because the Collectio academica antiqua primarily contains academic works—dissertations, textbooks, and reprints of classical authors—commissioned for or produced within the orbit of the Old University of Louvain. In such a context, holding multiple production roles can be read as an indicator of alignment with the university’s printing economy, where a small number of trusted workshops frequently combined several functions under one roof. Analyzing these overlaps therefore allows one to test whether functional integration within print workshops corresponded to closer institutional or intellectual integration with the university world itself.

To examine this mechanism empirically, I traced the involvement of academic actors across the corpus. The Studium.AI research infrastructure is developing a comprehensive list of the students enrolled at the Old University of Louvain, but academic scholars can already be identified through the Repertorium Eruditorum Totius Europae (RETE) project, a pan-European prosopographical database documenting the university careers of medieval and early modern professors.¹⁷

Through the disambiguation process described in Section 3’s Identifying Historical Actors, I linked each person in the Collectio academica antiqua to a unique CERL and Wikidata identifier. This, combined with a manual search, allows the matching of individuals appearing in the collection to scholars from the RETE database. Given RETE’s pan-European scope, the linkage allowed me to identify which books involve professors in general and which specifically involve professors affiliated with Louvain. The integration with RETE data adds a socio-institutional dimension to the bibliographic co-occurrence network: Some nodes now carry information on whether the person held a university position, thus allowing the relationship between print production and academia to be tested empirically. I find that around three-fourths of the books of the Collectio academica antiqua see the participation of a university professor, specifically Louvain professors in most of the cases but also professors active at other European universities (around 7%). These figures confirm that although the collection is institutionally rooted in Louvain, its intellectual network extended beyond the city. After all, the transregional dimension of the book trade in the Low Countries is a well-established fact in the literature (De Ridder et al.).

I then tested whether production agents who held multiple roles, for instance those appearing as both printer and bookseller in different publications, are more frequently associated with books linked to professors. The test is based on a simple comparison of proportions. For each book, I classified the involved production agents as single-role or multi-role and recorded whether the book involves at least one professor. I computed the share of professor-linked books among multi-role and single-role agents and assessed whether the difference is larger than expected by chance using a difference-in-proportions (z) test.¹⁸

Results in Table 5 show that books involving multi-role production agents are significantly more likely to involve professors. The share of professor-linked books is about 83% for multi-role agents, compared with 73% for single-role agents, a difference of 9 percentage points (z = 6.96, p = 3.35×10^–12). The association is even stronger for books involving professors affiliated with Louvain, where the corresponding shares are 77% and 63%, yielding a difference of 14 percentage points (z = 9.53, p = 1.61×10^–21). These patterns illustrate how preserving role differentiation in a multilayer co-occurrence representation makes visible systematic associations between production structure and academic involvement that would be obscured in a flattened network view. In this sense, the case study demonstrates how modeling choices are not merely technical, but substantively shape the kinds of institutional and historical patterns that can be recovered from catalog data.

The same reasoning applies when considering how many distinct layers each person appears in, sometimes called the actor’s “multiplexity,” or the number of role dimensions in their activity. In the Collectio academica antiqua network, only about 12% of individuals participate in more than one role layer (see Table 9 in the Appendix). Those who do are often individuals who also served in another capacity (e.g., editors of other work, authors who also wrote dedications, etc.). Table 10 in the Appendix shows the share of multi-role individuals in each layer. Unsurprisingly, roles like authorship and paratextual attributions account for a large fraction of the multiplex individuals, whereas very few censors, translators, or illustrators in this dataset also took on additional roles. The prevalence of paratextual connections (e.g., authors who share a printer also often share an editor or a contributor) is partly a reflection of how thoroughly this particular catalog records those secondary roles. By contrast, the censorship role appears underrepresented in the network. Historically, from the 1520s onward, works printed in places like Louvain often included an approbation or censor’s note (Cammaerts 241), but in the data derived from the Collectio academica antiqua, such connections are few. This likely indicates that many approbations were not captured in the catalog metadata, not that censorship was absent. A comparison with a dataset where censorship information is fully recorded could yield a denser censorship layer and alter the overall network structure. This example highlights how the depth and focus of cataloging practices (which information was recorded or omitted) directly influence the networks we can extract. In this perspective, a comparative analysis using this analytical lens could substantiate how differences in cataloging practices shape resulting network structures.

Insights from Quantitative Economic History

Some studies in history and literary scholarship that use bibliographic metadata often combine meticulous preprocessing pipelines with network visualizations and descriptive commentary, typically focused on centrality measures. These exercises undoubtedly elevate our understanding of the collections and catalogues used as material, yet, as Lemercier and Zalc observe, such work frequently remains at the stage of mapping and visual speculation, without fully mobilizing networks as formal analytical tools. The patterns observed in Section 4.1 (e.g., the sparsity of within-role links or the rarity of multi-role connections) highlight both the potential and the limitations of co-occurrence networks derived solely from catalog data. On one hand, such networks can reveal hidden structures in the catalog or collection studied; on the other, relying on networked bibliographic data exclusively, without complementary data or further analytical techniques, may limit the types of historical questions we can answer. To overcome this limitation, bibliographic network studies can draw on frameworks that have demonstrated how relational data can be embedded in explicitly articulated analytical designs, most notably, quantitative economic history.

In recent years, bibliographic metadata have begun to feature in applied economic history, often alongside other sources. These studies go beyond descriptive and exploratory scopes and use a range of empirical tools beyond network analysis. Importantly, bibliographic metadata are not treated as econometric data but instead as surviving historical evidence, interpreted and combined with other sources within composite datasets tailored to specific historical questions, for example, building on library catalogues, biographical sources, and manuscript data. Chaney, “Religion and the Rise and Fall of Islamic Science”, and “Modern Library Holdings Historic City Growth” constructs a georeferenced dataset of authors from which he derives metrics such as author counts, the share of authors engaged in scientific topics, and number of authors’ deaths. He uses this resource to address two distinct questions, respectively, that of the decline of scientific output in the medieval Islamic world and that of the enhancement of preindustrial city growth estimates.

Chiopris analyzes how spatial connections, namely the introduction of the railroad network in nineteenth-century German-speaking regions, shaped the creation and diffusion of ideas. Using comprehensive library catalogues from the German Collective Library Consortium, the author measures the emergence and spread of new ideas, broadly defined as novel publication topics, new words, or new combinations of existing concepts derived from enriched catalogue metadata.

Koschnick links author-level publication data from the English Short Title Catalogue (ESTC) to collegiate records from Oxford and Cambridge Universities to examine the diffusion of ideas between teachers and students from 1600 to 1800. In another study, de Pleijt and Koschnick combine ESTC data with information on scholars’ socioeconomic backgrounds, and using sentiment analysis on the titles of printed works, they find that limited career prospects for less advantaged academics coincided with a rise in dissenting religious publications.

Finally, Cervellati et al. use metadata from the epistolary union catalogue Early Modern Letters Online (EMLO) to study the role of the Republic of Letters in explaining why Britain experienced a sustained economic take-off at the time of the Industrial Revolution. These contributions do not leverage networks as an analytical lens, but they do use bibliographic metadata as part of broader quantitative frameworks to address questions about scholarly output, institutional change, and idea diffusion. What they also show is that bibliographic metadata can sustain formal empirical strategies once linked with complementary datasets, to the end of the pursuit of a specific research question.

The disciplinary silo between quantitative economic history and digital or computational history has left analytical strategies developed in the former largely unexplored in the latter. From this perspective, large-scale patterns revealed by bibliographic metadata can be connected to deeper historical questions when analyzed with the appropriate combination of domain knowledge and formal methods. A crucial step in this direction involves enriching catalog metadata with external sources, based on the research question at hand. The multilayer model proposed in Section 4 might operationalize this integration as follows. The bibliographic co-occurrence network can be conceived as an initial aspect or feature, that is, a first set of layers, onto which additional layers sets could be stacked to encode relations for the same entities drawn from complementary sources. This approach facilitates the incorporation of external knowledge into digital history analyses and lay the groundwork for applying more sophisticated network techniques, like, for example, peer-effects models.

Mapping of Roles

This subsection presents the grouping strategy used to translate detailed role descriptors into broader layer categories for multilayer network construction. The table clarifies how fine-grained roles, such as “writer of preface,” “censor,” “engraver,” and others, are aggregated into interpretable layers like Paratextual Attributions, Censorship, and Engraving/Illustration. The grouping reflects a deliberate analytical choice designed to balance representational fidelity with visual and conceptual clarity. Other researchers may opt for alternate classifications according to their research aims.

Per-layer Network Analysis

Inter-layer Network Analysis