Digitising Death: Benchmarking Genealogical Data and Recovering Women’s Histories in Early Modern Ireland

(1.1) Beyond bureaucratic records: the Funeral Entries

The Funeral Entries span seventeen volumes and more than three thousand manuscript pages and extend from the late sixteenth to the early eighteenth century (National Library of Ireland, n.d.; McAnlis, 2014). Compiled by the Ulster King of Arms, they were intended as bureaucratic instruments: tools for verifying genealogical claims and maintaining heraldic order within the colonial elite (Ohlmeyer, 2018). Yet because lineage and inheritance lay at their core, women appear throughout the collection as daughters, wives, mothers, and widows. Their presence is relational, but that relationality is not trivial; it is central to how these records construct social continuity and legitimacy. A gendered reading of the Funeral Entries highlights the active role women played in linking families, offices, and regions through patterns of marriage, descent, and kinship (Mac Curtain and O’Dowd, 1991; O’Dowd, 2005).

The entries are succinct yet densely layered. Many include a date and place of death, a burial site, and a concise account of family connections, husbands and former husbands, children, in-laws, and significant kin. At times, they evoke the cosmopolitan texture of early seventeenth-century Dublin, where marriages linked Old English officials to continental artisans. At other moments, they expose the social negotiations of confessional division, as unions bridged Catholic and Protestant households (Ohlmeyer, 2018, 2023; McShane, 2019). Occasionally, the bureaucratic façade softens and emotion surfaces: a woman is buried with most of her issue, or her three children have died in infancy. These brief asides remind us that these documents are not only instruments of lineage but also records of grief and continuity, capturing the affective dimensions of family life that underpinned genealogical logic (O’Dowd, 2005).

For historians of women, the scale of inclusion is remarkable. A preliminary survey of one volume (GO MS 66) shows that roughly 38 percent of entries concern women between 1600 and 1620, with the broader series likely approaching two in five. Such proportions are extraordinary by early modern standards, where women’s appearances in state or legal records are often sporadic and contingent upon litigation, deviance, or crisis (Mac Curtain and O’Dowd, 1991; Palmer, 2018; Tait, 2022). The Funeral Entries provide a more systematic, if socially selective, view of women’s lives, offering data that support both individual reconstruction and quantitative analysis (Mac Curtain et al., 1992; Nolan and McShane, 2022; McShane and Vanden Borre, 2025).

This combination of scale and consistency distinguishes the Funeral Entries from other core sources used in Irish women’s history. The 1641 Depositions, wills, and Chancery pleadings tend to capture women in moments of rupture, when property is disputed, when violence disrupts communities, or when death necessitates formal arrangements of goods (O’Dowd, 2005; Nolan, 2020). The Funeral Entries, by contrast, record death as a routine genealogical event. They allow us to see women not only in exceptional circumstances but as integral to the ordinary reproduction of family and status. The careful noting of successive marriages, who married whom, in what order, and with what issue, enables historians to trace how women’s marital alliances helped sustain civic, confessional, and regional elites over time (MacCurtain and O’Dowd, 1991; Ohlmeyer, 2018).

The source is also revealing in its treatment of ethnicity and confession. Early modern Ireland encompassed Gaelic Irish, Old English, New English, and smaller migrant groups such as Huguenots and Italian artisans (Ohlmeyer, 2018, 2023). The Funeral Entries reflect this diversity, recording women as daughters of Gaelic and Old English families, wives of New English officials, and sometimes as figures whose names point to continental origins. These entries offer a microcosm of cultural negotiation, showing how confessional and ethnic identities were redefined through marriage and memorialisation, and capturing the paradox of Catholic women memorialised within a heraldic record administered by a Protestant state (Mac Curtain and O’Dowd, 1991; O’Dowd, 2005; Palmer, 2018).

Because the entries note burial places in addition to kinship ties, they open further insight into women’s spatial and religious worlds. Patterns of burial in specific churches or vaults may reflect denominational preference, local attachment, or continued association with family estates (O’Dowd, 2005). When read across generations, these records trace persistence and change, familial clusters, shifts in confessional allegiance, or patterns of mobility as women marry across regions. For historians interested in women’s engagement with spatial and devotional life, the Funeral Entries offer unusually concrete coordinates (Tait, 2022).

The language of reproduction that recurs throughout the manuscripts adds yet another dimension. Phrases such as “without issue”, “leaving no issue”, or “having many children” illuminate how fertility and lineage structured early modern understandings of value and continuity. When linked to age, marital sequence, or notes on deceased children, these formulae permit cautious reconstruction of reproductive histories, rarely possible elsewhere in Irish material (O’Dowd, 2005; Palmer, 2018). Even in their brevity, they expose how women’s bodies and reproductive capacities were woven into the genealogical imagination.

The Funeral Entries matter, then, not simply because they record women more often than other sources, but because they situate women at the intersection of family strategy, colonial governance, religious identity, and emotional life (Mac Curtain and O’Dowd, 1991; Ohlmeyer, 2018, 2023; Nolan and McShane, 2022). Precisely because of this interpretive richness, the methods used to digitise and analyse the Funeral Entries directly shape the histories that can be written from them. If women’s names are mistranscribed or misrecognised, or if Irish-language patronymics are ignored by NER models, the very women made visible by the manuscript risk slipping back into digital invisibility (Sadek et al., 2024; Tual et al., 2023; Provatorova et al., 2020; McShane and Vanden Borre, 2025). By placing benchmarking at the centre of our workflow, we aim not only to unlock the evidential potential of the Funeral Entries but also to ensure that digital tools expand rather than constrain the field of vision (Wilkinson et al., 2016; Nolan and McShane, 2022).

Repository location

https://doi.org/10.5281/zenodo.18174069

Repository name

Zenodo

Object name

VOICES-FuneralEntries-NER-0.0.3

Format names and versions

The repository (version 0.0.3) contains multiple transcription iterations of the first three folios of NLI GO MS 66 (.csv, .ods), Named Entity Recognition results (.json), and validated results (.csv) along with documentation and data curation guides (.md) and Notebooks used in the creation of the dataset (.ipynb).

Creation dates

From 2025-08-11 to 2025-10-07.

Dataset creators

The dataset creators are Bronagh Ann McShane (transcriptions), Diego Rincon-Yanez (NER curation), and Felix Vanden Borre (transcriptions, NER validation).

Languages

English (early modern).

License

Creative Commons Attribution 4.0 International (CC BY 4.0)

Publication date

2026-01-07.

(3.1) From manuscript sources to processable text

At the outset of this work, the Funeral Entries manuscripts existed as digital images on the National Library of Ireland’s website but were not text searchable. A first challenge in using computational methods on this historical source has therefore been to transform digital facsimiles into processable text, using the HTR software Transkribus (READ-COOP, 2025). Transkribus was designed for the transcription of historical documents and helps to expedite what would otherwise be a labour-intensive component of the research process. Users can mark up the specific regions of a page from which transcriptions should be generated.

The material intricacies of the Funeral Entries complicated this first step. A single page can contain anywhere between one and thirteen entries. Many entries are accompanied by additional information: a heading that contains the surname of the deceased, marginal notes, a heraldic shield, or some combination of these elements.

To manage this complexity, we designed a tagging protocol to keep each entry separate while correctly capturing all the information required. Transkribus enables users to label text regions with a set of tags that can be activated or customised at any time. We distinguished individual Funeral Entries by creating at least one heading region and one paragraph region per entry, with the paragraph region containing the main body of text (Figures 1 & 2). Where one or more marginalia existed, these were also captured and labelled accordingly. The tagged regions could then be parsed into individual entries after export from Transkribus in PAGE-XML format. Because drawings or symbols such as coats of arms can interfere with text recognition, we avoided capturing them when marking up text regions.

Figure 1

Figure 2

One fundamental material characteristic of the manuscripts that this system cannot fully address is the presence of several different hands within the corpus. The early modern period saw the development of multiple handwriting styles and typographical practices, which varied across countries, contexts, and individual scribes. Several hands sometimes coexist within a single volume, and the poor quality of some entries suggests that, amid the turmoil of seventeenth-century Ireland, speed sometimes took precedence over clarity.

Transkribus contains a range of open-access models that can be used by researchers and also allows users to train their own model based on a specific source or group of sources. Because of the brevity of the Funeral Entries and the constant variation in handwriting, we did not have enough homogeneous training data to build a stable custom model, nor was this feasible within the time constraints of the project. Instead, we opted to use one of the platform’s most powerful super models, Text Titan I (READ-COOP, 2023).

Even this state-of-the-art HTR model has struggled with aspects of the manuscripts. The presence of Irish names and place names throughout the corpus means that the transcription process would benefit from a model trained on historical Irish-language data. Together, these factors have resulted in some excellent, and also some disappointing transcriptions.

The funeral entry of Mary Lady Tyrell illustrates a relatively accurate output:

Aside from the Segrave family name not being capitalised, the main HTR errors are the misspelling of infaucie and an intrusive comma after St. Audoens. The overall transcription, however, preserves the historical spelling and typography sufficiently to allow identification of key entities and relations.

A second example, the entry for Alice Segrave, demonstrates the limits of the model:

The first sentence should read: “Ales dr. of Walter Segrave of Dublin Alderman Maior & wi: of John eldest sonne of Richard Fagan Ald: M. of the same, deceased the xxiiith of May MDCVIII and is bur: in St. Audoens Church.” It is easy to miss that Also is a misrendering of Ales (Alice) or that Pagan should read Fagan. Roman numerals are also particularly challenging for HTR.

Our transcriptions are therefore processable but uneven in quality. For the purposes of this article, the crucial question is how such variation, encompassing historical spellings, Irish-language forms, early modern typography, and HTR errors, affects the performance of off-the-shelf NER models, and how benchmarking can make these effects visible and addressable.

(3.2) Generating multiple transcription variants for benchmarking

To investigate how transcription quality affects NER performance, we created multiple versions of a small, carefully controlled subset of the Funeral Entries. We selected the first twelve entries in GO MS 66 and generated a series of transcriptions that held content constant while varying levels of normalisation, correction, and interpretive intervention.

Five transcription variants were produced:

• Manual transcription (MAN): A historian-generated version retaining early modern spellings while correcting obvious errors.

• EyeCR transcription (EYE): Produced using an internal transcription tool that can connect to different LLMs, developed at the ADAPT Centre. For our case study, the LLM that was connected to EyeCR was GPT-4.1-mini model.

• Transkribus transcription (TRA): Raw HTR output exported directly from Transkribus.

• Curated Transkribus transcription (CUR): A diplomatically curated version of TRA, correcting clear HTR errors while preserving original orthography.

• Transkribus + VARD transcription (TKV): The TRA version processed with VARD2 to normalise historical spelling variants. VARD2 is an interactive tool designed to deal with spelling variation in historical corpora (VARD, 2025).

An additional manual transcription was created and annotated with named entities. This manually annotated set acted as an approximate ground truth against the transcriptions used for benchmarking, recognising that no transcription can be fully authoritative given the interpretive decisions required.

Curating this benchmarking dataset, including the ground truth, highlighted how historical and manuscript realities resist straightforward computational encoding. The funeral entry of James Barnewall, for example, mentions the parish of Stabannan in County Louth. In the manuscript, the scribe splits the place name across two lines and inserts a colon to indicate continuation, writing it as “Staba: nan” (Figure 3). The same device appears in “Ro: bart”. In both our manual transcriptions and the NER ground truth, these were joined into “Stabanan” and “Robart” (Robert). This decision makes sense from the standpoint of data regularisation but means that the ground truth is not, in a strict palaeographic sense, identical to the manuscript. Creating a benchmarking dataset thus made visible the interpretive steps required to turn manuscripts into machine-readable data and forced us to reflect on how such steps affect downstream model performance.

Figure 3

(3.3) NER models and configurations

Within VOICES, reproducible and scalable NLP workflows are essential, both because of the heterogeneity of the manuscript collection and because the project ultimately aims to construct a knowledge graph capable of linking persons and places across archival corpora (Debruyne et al., 2022). To support this, we implemented an experimentation pipeline to evaluate off-the-shelf NER models on the transcription variants described above.

In the benchmark experiment, each paragraph of a Funeral Entry was processed using two standard NER approaches: a transformer-based BERT model (Devlin et al., 2019) and the CRF-based Stanford CoreNLP (Qi et al., 2020) tagger (Figure 4). Their outputs were then manually annotated to quantify model performance using the synthetic metric outlined in Section 2.4.

Figure 4

Because one of our key aims was to assess how transcription quality shapes NER results, we ran each model across all five curated transcription variants (MAN, EYE, TRA, CUR, and TKV). We also tested a combined configuration that merges predictions from BERT and Stanford, retaining only those with a confidence score of 70 percent or higher and resolving overlapping spans consistently. The full structure of the experiment, including inputs, outputs, and expected behaviour, is shown in Figure 5.

Figure 5

The experiment evaluates three NER configurations: a BERT-based model (see Table 1), the Stanford CoreNLP tagger (see Table 2), and a combined scenario in which we take the union of predictions from both systems but retain only those with a confidence of 70 percent or higher (see Table 3). In this combined scenario, overlapping spans are resolved so that entities detected by BERT do not contradict those produced by Stanford. A detailed version of the experiment scenarios, inputs and expected results can be seen in Figure 5.

(3.4) Evaluation protocol and synthetic metrics

For each entry, and for each model and transcription combination, the output entities were manually annotated using a simple three-level scheme: 0 for incorrect, 1 for correct, and 2 for partial match. This annotation process produced a small but carefully controlled benchmark suited to our particular corpus and research questions.

Standard NER evaluation relies on counts of true positives, false positives, true negatives, and false negatives, together with derived measures such as precision, recall, and F1 score. In newly digitised humanities collections, however, a robust gold standard rarely exists, and the number and type of candidate entities varies across transcriptions. For this reason, we adopted a synthetic evaluation protocol.

Let N₁ denote the number of entities annotated as correct, N₂ the number annotated as partial matches, and N₀ the number annotated as incorrect. The total number of evaluated entities is then:

Under this annotation pattern, a strict and relaxed accuracy can be computed as the number of correct entities divided by the total number of evaluated entities, multiplied by 100 to give a percentage:

This simple metrics does not capture all nuances of entity recognition, but it provides a transparent and reproducible way to compare model performance across different transcriptions and configurations. Importantly, the annotation-based approach allows us to work with multiple plausible transcriptions rather than assuming the existence of a single, fully authoritative ground truth.