Enrichment of the C-CLAMP Corpus

Metadata creation and cleaning

C-CLAMP includes two metadata sets: text-level and author-level. The former records year of publication, author name(s), and the title as listed in The Digital Library of Dutch Literature (DBNL). The author-level metadata, now considerably expanded, originally included name(s), dates and places of birth and death, and a link to the authors’ DBNL landing pages (Piersoul et al., 2021).

Recent cleaning and enrichment efforts substantially revised the author metadata’s first version. Through staged (semi-)automatic processing, with manual correction where necessary, the metadata were supplemented and corrected by linking it to three DBNL-related resources: (i) the KB’s (Koninklijke Bibliotheek) RDF-encoded linked data portal (data.bibliotheken.nl); (ii) the KB’s internal person database, to which access was granted; and (iii) the DBNL landing pages’ HTML source code. All additional sources were linked to the original TXT-based author metadata via the authors’ unique IDs.

Although these sources overlap, they each contribute distinct information. (see Table 2) The linked data portal, containing a thesaurus on the DBNL authors, provides structured information on occupations and gender; the person database adds publication period; and the HTML code specifies countries of birth and death through backend connections to the DBNL atlas. In addition to introducing new fields from the sources, existing fields were standardized (e.g., uniform date formatting) and incomplete records were supplemented where possible.

The resulting improvements concern both scope and coverage (Table 1).

Notably, 3,989 authors (45.5%) are associated with at least one occupation other than writer, spanning 67 distinct professions, from agriculture to diplomacy, and academia. This socioeconomic information distinguishes C-CLAMP from comparable literary corpora. To enable straightforward analyses, abstraction beyond raw occupational labels was achieved by mapping the occupations onto the Historical and International Social Class Scheme (HISCLASS) (van Leeuwen & Maas, 2011). HISCLASS classifies professions into twelve hierarchical classes based on four criteria: (non-)manual labour; skill level; supervision; and sector (van Leeuwen & Maas, 2005). Higher managers occupy class 1, while unskilled farmers are categorized into class 12. A residual category covers unclassifiable occupations.

C-CLAMP’s occupational labels were automatically linked to a Dutch translation of HISCLASS (Zijdeman, 2016), with minimal manual corrections (e.g., resolving the ambiguity of Dutch “bestuurder”, accurately translated as ‘director’ rather than ‘driver’). For authors with multiple occupations, each occupation was assigned its respective class number. An additional field was created to store the lowest class number per author, representing the highest social position attained, thereby capturing upward social mobility.

Figure 1 displays the distribution of authors across classes. The classes are ordinal categories ranging from 1 (highest) to 12 (lowest), with class 13 representing a residual category. Since the ability to publish in intellectual periodicals presupposes prestige, the concentration in the highest classes is expected.

Figure 1

Overall, the enrichments substantially increased the number of metadata fields while improving internal consistency, standardisation, and precision, thus enhancing analytical reliability.

Network analysis

Authors within C-CLAMP regularly engage with one another in their writings, through (i) offering critique and citing each other, and (ii) collaborating as co-authors. Together, these forms of interaction constitute the basis of the network: each author is modelled as a node, and the relationships between them are represented as edges (Figure 2), distinguished as either directional or reciprocal. A reciprocal (bi-directional) relationship arises when two authors co-author, whereas a directional (unidirectional) relationship occurs when Author X (source) references Author Y (target). To represent these dynamics, we constructed a directed network: a graph in which edges have an explicit orientation indicating the direction of interaction, preserving asymmetry in the edges. This distinction is essential: Author X frequently citing Author Y does not entail Author Y reciprocating with the same frequency.

Figure 2

The network edges are weighted to reflect interaction intensity. A given relationship’s weight equals to the sum of (i) the number of co-authored publications, and (ii) the number of times the source author cites the target author. Consider the following example. Authors X and Y co-authored five publications, and X cited Y in three separate texts. This results in an edge weight X→Y = 8 (5 co-authorships + 3 citations). However, Y cited X 7 times, yielding Y→X = 12 (5 co-authorships + 7 citations).

Periodicals occasionally include works by earlier writers, e.g., translations of Horace, and the C-CLAMP authors sometimes refer to them in their own texts. Consequently, these historical authors can appear as unintended targets within the network. To prevent distortion, targets born before 1700 were excluded from the dataset, as these authors largely represent classical or early modern reference figures rather than participants in the literary field captured by C-CLAMP. Self-referential ties were likewise removed. After filtering, the resulting network (Figure 3) contains 7,451 nodes (authors) and 156,123 edges (inter-author relationships).

Figure 3

To assess each node’s (author) significance in a directed network, various centrality measures can be employed. These provide insight into which nodes are particularly influential or structurally important according to different criteria. We provide In-Degree, Out-Degree, Degree, Betweenness, Closeness, Harmonic and Eigenvector centrality (Freeman, 1978).¹ Each metric captures a distinct dimension of importance.

In-Degree centrality highlights frequently cited authors, while Out-Degree centrality identifies authors who cite others; both measures are also affected by co-authorship patterns. Degree centrality combines these two perspectives, reflecting a node’s total number of connections. Betweenness centrality detects nodes that act as critical bridges connecting different regions of the network. Closeness centrality, in turn, identifies authors who can reach others efficiently, reflecting overall accessibility within the network. Harmonic centrality provides a similar measure of reachability, but remains robust in networks with disconnected components. Finally, Eigenvector centrality identifies authors connected to other highly influential authors, capturing influence that extends beyond direct ties.