In the digital humanities, datasets inherit and perpetuate biases through multiple channels: individual and institutional biases, discriminatory language in archives, unequal representation in collection practices, and algorithmic biases in AI-assisted processing. These biases are compounded throughout the research process, yet the term “bias" itself lacks a clear definition, often causing “bias paralysis." This paper proposes treating “bias" as a productive category of analysis for digital humanities research through the development of a “Bias-Aware Framework" for dataset creation and contextualisation. It has three components: a Bias Thesaurus creating shared vocabulary across disciplines to address the conceptual instability of “bias" by breaking down this nebulous concept into interrelated issues like representation, gaps, positionality, CARE, etc; a Bias-Aware Dataset Lifecycle Model showing where biases enter the research process; and Guidelines for documenting, describing, and mitigating bias. We approach bias not simply as an error, but as a revealing analytical lens that shapes knowledge production. By explicitly describing these conditions of production, researchers can improve transparency, improve dataset documentation, and enable more informed reuse of their data.
compounding efect means that seemingly minor biases at early stages can result in significantly skewed outcomes during analysis and interpretation.
The Dutch East India Company (Verenigde Oost-Indische Compagnie, henceforth VOC) testament
archives provide a revealing case study of how digitisation can inadvertently preserve and amplify
historical bias.1 When the Dutch National Archives digitised this collection in 2017, they made the
archive accessible online by scanning its pages and digitising a 19th-century index. However, this digital
transformation preserved a significant historical bias: while the index includes approximately 10,000
European male testators, it omits female co-testators, individuals from diverse ethnic backgrounds, and
enslaved persons who appear as beneficiaries, partners, debtors, properties, or witnesses [
To address these silences, Luthra et. al. [8] worked with transcribed versions of the testaments and used information retrieval methods to develop more inclusive finding aids. They used named entity recognition and classification (NERC)—a common natural language processing method for identifying and categorising entities such as people, organisations, and locations [9]. However, standard NERC schemas like CoNLL [10] and ACE [11] were insuficient for the complexity of colonial records, which often reference unnamed or marginalised individuals and include vital context like roles, gender, and legal status. By designing a custom typology tailored to colonial archives, the project was able to surface individuals who were previously obscured both by the original records and the digital tools built on them.
This example illustrates a critical point: while tech- Figure 1: Silences in the historic index; a: niques like NERC are powerful for information extrac- 19th century index; b: Testator: tion, they can also embed existing power dynamics, only his name is indexed; c: “Free reinforcing or challenging historical biases depending Christian woman Magdalena van on how they are designed and applied. Similarly, in Boegis” is present in the document semantic web technologies, choices in ontology de- but not findable in the digitised insign directly influence whose histories are made vis- dex (NA, VOC, 6847 folio number ible and whose remain obscured. When datasets and 40, page 119) knowledge structures fail to account for marginalised perspectives, the colonial and patriarchal biases embedded in original archives are not only encoded into digital systems—they can become further entrenched. Awareness of these multiple types of bias and their compounding efects is essential for developing better digital humanities methodologies.
Despite growing attention to bias and its mitigation [12, 13, 14, 15], there is still no coherent framework
for understanding what bias actually is. The term ‘bias’ carries diferent meanings across contexts:
1“1.04.02 Inventaris van Het Archief van de Verenigde Oost-Indische Compagnie (VOC), 1602–1795 (1811) | Nationaal Archief,”
https://www.nationaalarchief.nl/onderzoeken/archief/1.04.02
2https://www.nationaalarchief.nl/onderzoeken/archief/1.04.02/invnr/6847/file/NL-HaNA_1.04.02_6847_0119
archivists emphasise issues of inventorisation and deceptive categorisation; historians examine historical
power structures [
Several valuable interventions have emerged, from documentation templates to tools for identifying harmful language and replacing ofensive terminology [ 25, 14, 26, 27, 28, 29, 23]. However, these approaches typically address specific manifestations of bias. Digital humanities researchers regularly encounter an entangled spectrum: historical, technical, descriptive, and representational biases—many of which are reinforced through infrastructures like metadata schemas, machine learning pipelines, and semantic web technologies. While these semantic systems can encode dominant world-views, recent work also shows their potential for mitigating bias—by enabling more nuanced representations, identifying disparities across groups, and supporting fairer information retrieval and classification [ 8, 18].
What remains missing is a cohesive approach that makes visible how these forms of bias interact and transform across the data lifecycle. Without such a framework, mitigation eforts risk remaining fragmented—treating symptoms rather than confronting the underlying systems through which bias is produced, sustained, and reproduced in digital knowledge production.
To efectively identify, articulate, and mitigate bias in digital humanities research, three fundamental questions need to be answered:
2. Where does bias occur in the dataset creation process? 3. How can researchers efectively address bias within resource constraints? In response to these questions, we are developing a “Bias-Aware Framework" for dataset creation: 1. A Bias Thesaurus: A comprehensive list of the concepts connected to bias (such as representation, ofensive language, FAIR, CARE, silences, etc.) that creates a shared vocabulary for discussing bias across disciplines. 2. A Bias-Aware Data Lifecycle Model: A visual and conceptual model mapping where and how diferent types of bias arise across the research process, enabling targeted reflection and intervention. 3. Practical Guidelines: A set of reflective questions, examples, and “good–better–best” recommendations tailored to each stage of the data lifecycle, supporting practices of bias identification, description, and redress.
This framework addresses a recognised gap in digital humanities: “a set of guidelines is missing, a serious lack when one might want to think through ethical concerns" [30]. It is a framework that demystifies ‘bias’ and transforms it into a productive tool for improving knowledge production. 3For instance the Words Matter [23], a publication on sensitive words in the museum sector, doesn’t use the term ‘bias’, but projects such as DE-BIAS [24], based at the Dutch Institute of Sound and Vision, use the term in context of developing an automated tool to identify harmful language in archives.
Focus on Datasets as Critical Intersection Points. Our framework centres datasets as the primary unit of analysis within the digital humanities landscape. This focus is strategic for several reasons. Datasets function as critical nexus points where four key elements converge: the data itself (from archives, born-digital sources, or interviews); the researchers who structure this data; the users who access and build upon these resources; and the computational methods that process this information. For Semantic Web Technologies in particular, datasets form the foundation upon which ontologies and knowledge graphs are constructed, making them crucial sites for bias intervention before problematic representations become encoded in semantic structures.
Most importantly, datasets should not be viewed merely as areas of ‘risk’ requiring intervention, but as sites of tremendous opportunity when created with critical awareness of biases. Our project highlights how dataset creation can function as a form of “deconstructing archival sources", enabling researchers to view historical materials through new analytical lenses [31]. For example, it was through dataset creation and subsequent analysis that researchers uncovered the pivotal role of a 17-year-old woman named Flora in orchestrating the escape of nineteen enslaved people—efectively re-inscribing her into historical narratives despite her name appearing only fleetingly in primary sources [32].
The Bias-Aware framework development follows a three-phase approach combining theoretical analysis, practitioner insights, and practical validation:
To gain a better overview of current theories about and strategies against bias, we systematically reviewed literature in the fields of archival studies [ 23, 26, 33, 34, 32], epistemology [35, 36] and computer sciences [37, 38, 39, 40, 18]. These are fields in which bias has received much attention. From here, we expanded out to include sources that were not academic and/or text-based, such as videos, art installations, and fiction [ 41, 42, 43]. The importance of including these resources was to critically confront our own biases for the written and academic. For each resource, we focused on what forms of bias or strategies to mitigate bias were present, and made a note of that under the column ‘concepts used’. We compiled these in an open-to-access list of resources.4
Our framework development draws on semi-structured interviews with partners from four major digital infrastructure projects focusing on colonial and slavery archives: Slave Voyages5,GLOBALISE6, Exploring Slave Trade in Asia7, and the Historical Database of Suriname and Caribbean8. These partnerships provide crucial insights into practical implementation challenges. We also engaged advisors with diverse expertise across cultural heritage, critical archival studies, community memory work, ethnomusicology, natural language processing, and FAIR data principles. This plurivocal approach [44] ensures our framework ofers adaptable strategies and examples suitable for diverse project contexts and resource levels.
We are validating and refining the framework through two parallel tracks: expert consultations and interactive workshops with digital humanities and social science projects. The workshops serve as practical testing grounds where participants apply the framework and its methodology to analyse bias in 4Combatting Bias Resources List. We are now working on the bias thesaurus to establish better categories for organising the readings before making the resource list fully collaborative. 5https://www.slavevoyages.org/ 6https://globalise.huygens.knaw.nl/ 7https://esta.iisg.nl/ 8https://www.ru.nl/onderzoek/onderzoeksprojecten/historische-database-van-suriname-en-de-cariben their own datasets. This implementation phase aims to reveal the framework’s strengths and limitations across diferent domains and identify potential blind spots. Participant feedback and documented use cases will drive iterative improvements, ensuring its broader applicability and efectiveness.
To conceptualise bias meaningfully, we first examined its etymology. “Bias" entered English in the 1570s from the game of boules, describing balls weighted to curve obliquely. This technical meaning evolved into the figurative sense of “a one-sided tendency of the mind" and later “undue propensity or prejudice." The French origin biais means “sideways, askance, against the grain"—suggesting movement contrary to an expected direction. This etymology raises a crucial question: when labelling something “biased," what is our assumed “true" path? What constitutes an unbiased space, description, or archive—and is such a thing even possible? Rather than pursuing an impossible “bias-free" ideal, we draw inspiration from the textile meaning of bias: the diagonal stretch between warp and weft where fabric shows greatest flexibility. Garments cut “on the bias" follow this diagonal orientation, creating fluidity and adaptability.
We employ this sense of bias in our framework: just as fabric’s bias exposes structural tensions and possibilities, biases in datasets highlight gaps, conditions of production, overlooked questions, and unconsidered perspectives. This shifts our focus from attempting to “solve" bias to using it as a critical tool for systematic analysis.
The Bias-Aware Framework consists of three integrated components designed to transform how researchers understand, identify, and address bias throughout the dataset creation process. Each component builds upon the others to create a comprehensive approach to bias as a category of analysis. 3.2.1. Bias Thesaurus
bias functions as a heuristic addressing interconnected concerns about power, inequality, positionality, silences, and representation. Drawing from Scott’s [36, 45] concept of gender as an analytical category and Foucault’s [46] understanding of power as relational, we view bias as dynamic—actively shaping and being shaped by social and historical contexts.
The bias Thesaurus maps the various expressions of bias—concrete forms bias takes in research practices, such as harmful language, uneven descriptive depth, Figure 2: Bias-Aware Dataset Lifecycle or limiting categorisation schemes. The thesaurus creates a shared vocabulary across disciplines, visualises interconnections between diferent expressions of bias, and provides researchers with a conceptual map for navigating bias-related concerns. 3.2.2. Bias-Aware Dataset Lifecycle Model The dataset creation lifecycle forms the structural backbone of this framework, grounding abstract bias considerations in familiar research workflows while addressing a gap in digital humanities methodology. Our model builds upon the Research Data Alliance’s harmonised Research Data Lifecycle (RDL) model9 9https://www.rd-alliance.org/wp-content/uploads/2024/09/D1_The-creation-of-a-harmonised-research-data-lifecycle-RDL-model-and-crosswalk pdf which identifies five key stages (Set Up, Collect, Process, Analyse, Preserve & Share). We extend this framework by mapping how diferent expressions of bias defined in the Thesaurus manifest at each stage.
A key insight from our research is the “stickiness" of certain bias expressions across multiple stages, though they manifest diferently depending on the stage’s focus. For example, representation concerns appear throughout the lifecycle: in Set Up, they relate to whose scholarship informs the project; in Collect, they concern whose perspectives are captured in the data; in Process, they involve how categories represent complex realities; and in Analysis, they address whose stories are highlighted in ifgure 2. 3.2.3. Practical Guidelines
standing into practical action through structured guidelines for each stage of the dataset lifecycle.
These guidelines provide reflective questions, curated resources, documentation templates, “good-betterbest" recommendations that accommodate varying resource constraints, and example strategies drawn from successful digital humanities projects.
Figure 3 illustrates our guideline approach for considering CARE as a principle at the funding step, while ifgure 4 illustrates our guideline approach for ad- Figure 3: Bias-Aware Dataset Lifecycle with dressing archival silences, ofering tiered intervention reflective questions strategies from basic documentation to participatory community engagement.
The guidelines emphasise that addressing bias is not an all-or-nothing proposition—even resourceconstrained projects can implement basic bias-aware practices. This scafolded approach helps prevent “bias paralysis" by making intervention accessible regardless of project scale or resources.
Our approach reconceptualises bias not as noise to be eliminated but as parameters that reveal how knowledge is constructed. By making bias visible and analysable, we transform it from a technical problem into a productive analytical lens that enhances both the integrity and reuse potential of digital humanities research. The Bias-Aware Framework provides a structured vocabulary, makes visible critical intervention points in the research process, and ofers actionable strategies adaptable to various resource constraints. Rather than eliminating bias—an impossible task—we aim to establish bias awareness as a fundamental aspect of scholarship comparable to citation practices or methodological transparency. 1. Disciplinary Expansion and Empirical Validation: We aim to extend the framework beyond colonial archives to other humanities domains while validating its efectiveness through diverse case studies. This parallel expansion and validation will test the framework’s flexibility, identify domain-specific adaptations, and document how bias awareness transforms research outcomes across project types. 2. Formalising Vocabulary and Knowledge Structures: A critical next step involves formalising the Bias Thesaurus, which will capture the relationships between diferent bias expressions, their manifestations across the data lifecycle, and appropriate mitigation strategies. 3. Theoretical Foundations: Further research will justify and reflect on digital humanities’ unique position at the intersection of computational methods and humanistic inquiry. This work will contribute to ongoing debates about how computational approaches can be informed by critical humanities perspectives, particularly regarding knowledge representation and classification systems. 4. Sustainable Infrastructure Long-term maintenance of the framework requires developing sustainable infrastructure through community governance and versioning systems. We envision creating a collaborative platform where researchers can contribute examples, adaptations, and extensions to the framework, ensuring it evolves alongside changing research practices and emerging technologies. 5. Implementation Formats: To maximise accessibility and adoption, we will explore various presentation formats from open-access platforms to downloadable templates, interactive tools, and integration with existing data management frameworks.
The authors are the main researchers of the Combatting Bias project10, which is a collaborative initiative based at the Huygens Institute and International Institute of Social History in Amsterdam, Netherlands, funded by the NWO via the Thematic Digital Competence Centre Social Sciences and Humanities11. We thank the reviewers for their critical feedback.
10https://combattingbias.huygens.knaw.nl/ 11https://tdcc.nl/about-tddc/ssh/ [8] M. Luthra, K. Todorov, C. Jeurgens, G. Colavizza, Unsilencing colonial archives via automated entity recognition (2023). doi:10.1108/JD-02-2022-0038. [9] M. Ehrmann, A. Hamdi, E. L. Pontes, M. Romanello, A. Doucet, Named entity recognition and classification in historical documents: A survey (2023). doi: 10.1145/3604931. [10] E. F. T. K. Sang, F. De Meulder, Introduction to the CoNLL-2003 shared task: Language-independent named entity recognition (2003) 142–147. URL: https://aclanthology.org/W03-0419. [11] G. Doddington, A. Mitchell, M. Przybocki, L. Ramshawa, S. Strassel, R. Weischedel, The automatic content extraction (ACE) program - tasks, data, and evaluation, in: M. Teresa Lino, M. Francisca Xavier, F. Ferreira, R. Costa, R. Silva (Eds.), Proceedings of the Fourth International Conference on Language Resources and Evaluation (LREC’04), European Language Resources Association (ELRA), 2004. [12] C. H. Chu, S. Donato-Woodger, S. S. Khan, R. Nyrup, K. Leslie, A. Lyn, T. Shi, A. Bianchi, S. A.
Rahimi, A. Grenier, Age-related bias and artificial intelligence: a scoping review 10 (2023) 1–17. doi:10.1057/s41599-023-01999-y, publisher: Palgrave. [13] A. Ortolja-Baird, J. Nyhan, Encoding the haunting of an object catalogue: on the potential of digital technologies to perpetuate or subvert the silence and bias of the early-modern archive1 37 (2022) 844–867. doi:10.1093/llc/fqab065. [14] E. M. Bender, B. Friedman, Data statements for natural language processing: Toward mitigating system bias and enabling better science, Transactions of the Association for Computational Linguistics 6 (2018) 587–604. doi:10.1162/tacl_a_00041. [15] M. K. Scheuerman, A. Hanna, E. Denton, Do datasets have politics? disciplinary values in computer vision dataset development 5 (2021) 317:1–317:37. doi:10.1145/3476058. [16] M. van Rossum, Labouring transformations of amphibious monsters: Exploring early modern globalization, diversity, and shifting clusters of labour relations in the context of the dutch east india company (1600–1800) 64 (2019) 19–42. doi:10.1017/S0020859019000014. [17] I. Kizhner, M. Terras, M. Rumyantsev, V. Khokhlova, E. Demeshkova, I. Rudov, J. Afanasieva, Digital cultural colonialism: measuring bias in aggregated digitized content held in google arts and culture 36 (2021) 607–640. doi:10.1093/llc/fqaa055. [18] P. R. Lobo, E. Daga, H. Alani, M. Fernandez, Semantic web technologies and bias in artificial intelligence: A systematic literature review, Semantic Web 14 (2023) 745–770. doi:10.3233/ SW-223041. [19] J. Buolamwini, T. Gebru, Gender shades: Intersectional accuracy disparities in commercial gender classification, in: Proceedings of the 1st Conference on Fairness, Accountability and Transparency, PMLR, 2018, pp. 77–91. URL: https://proceedings.mlr.press/v81/buolamwini18a.html. [20] H. Suresh, J. lobo, A framework for understanding sources of harm throughout the machine learning life cycle, in: Equity and Access in Algorithms, Mechanisms, and Optimization, ACM, 2021, pp. 1–9. doi:10.1145/3465416.3483305. [21] A. Søgaard, B. Plank, D. Hovy, Selection bias, label bias, and bias in ground truth, in: Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Tutorial Abstracts, 2014, pp. 11–13. URL: https://aclanthology.org/C14-3005.pdf. [22] S. L. Blodgett, S. Barocas, H. Daumé III, H. Wallach, Language (technology) is power: A critical survey of “bias” in NLP, in: D. Jurafsky, J. Chai, N. Schluter, J. Tetreault (Eds.), Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, Association for Computational Linguistics, 2020-07, pp. 5454–5476. URL: https://aclanthology.org/2020.acl-main. 485. doi:10.18653/v1/2020.acl-main.485. [23] W. Modest, R. Lelijveld, Words matter: an unfinished guide to word choices in the cultural sector, 2018. URL: https://amsterdam.wereldmuseum.nl/en/about-wereldmuseum-amsterdam/research/ words-matter-publication. [24] DE-BIAS - detecting and cur(at)ing harmful language in cultural heritage collections, 2024. URL: https://pro.europeana.eu/project/de-bias. [25] H. Alkemade, S. Claeyssens, G. Colavizza, N. Freire, J. Lehmann, C. Neudecker, G. Osti, D. v. Strien,
Datasheets for digital cultural heritage datasets 9 (2023) 17. [26] A. Chilcott, Towards protocols for describing racially ofensive language in UK public archives, in: V. Frings-Hessami, F. Foscarini (Eds.), Archives in a Changing Climate - Part I & Part II, Springer Nature Switzerland, 2022, pp. 151–168. doi:10.1007/978-3-031-19289-0_10. [27] M. Luthra, M. Eskevich, Data-envelopes for cultural heritage: Going beyond datasheets, in: I. Siegert, K. Choukri (Eds.), Proceedings of the Workshop on Legal and Ethical Issues in Human Language Technologies @ LREC-COLING 2024, ELRA and ICCL, 2024, pp. 52–65. URL: https: //aclanthology.org/2024.legal-1.9. [28] A. Masschelein, F. Truyen, S. Taes, J. van Mulder, A. Stynen, R. Pireddu, Report on research into bias types and patterns, including a typology applied to europeana use cases and a vocabulary co-created with communities, 2023-12-31. [29] M. K. Scheuerman, K. Spiel, O. L. Haimson, F. Hamidi, S. M. Branham, HCI guidelines for gender equity and inclusivity, Maryland Shared Open Access Repository, 2020. doi:10.13016/ M2NW1F-P0JX. [30] J. O’Sullivan, The bloomsbury handbook to the digital humanities, Bloomsbury Publishing, 2024. [31] N. L. Peluso, Whose woods are these? counter-mapping forest territories in kalimantan, indonesia,
Antipode 27 (1995) 383–406. [32] C. W. van Galen, B. Quanjer, The wolf, the island and the sea: truancy and escaping slavery in curacao (1837–1863) 29 (2024) 262–279. doi:10.1080/1081602X.2024.2340542, publisher: Routledge. [33] A. L. Stoler, Along the Archival Grain: Epistemic Anxieties and Colonial Common Sense, Princeton
University Press, 2010. [34] G. C. Spivak, The rani of sirmur: An essay in reading the archives 24 (1985) 247–272. doi:10.
2307/2505169, publisher: [Wesleyan University, Wiley]. [35] M. Foucault, Archaeology of Knowledge, 2 ed., Routledge, 2002. doi:10.4324/9780203604168. [36] J. W. Scott, Gender: A useful category of historical analysis 91 (1986) 1053–1075. doi:10.2307/ 1864376, publisher: [Oxford University Press, American Historical Association]. [37] T. Gebru, J. Morgenstern, B. Vecchione, J. W. Vaughan, H. Wallach, H. D. III, K. Crawford, Datasheets for datasets 64 (2021) 86–92. doi:10.1145/3458723. [38] E. S. Jo, T. Gebru, Lessons from archives: strategies for collecting sociocultural data in machine learning, in: Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency, FAT* ’20, Association for Computing Machinery, 2020, pp. 306–316. doi:10.1145/3351095. 3372829. [39] W. Orr, K. Crawford, The social construction of datasets: On the practices, processes and challenges of dataset creation for machine learning, 2023. doi:10.31235/osf.io/8c9uh. [40] R. Brate, A. Nesterov, V. Vogelmann, J. van Ossenbruggen, L. Hollink, M. van Erp, Capturing contentiousness: Constructing the contentious terms in context corpus, in: Proceedings of the 11th Knowledge Capture Conference, K-CAP ’21, Association for Computing Machinery, 2021, pp. 17–24. doi:10.1145/3460210.3493553. [41] Chimamanda Ngozi Adichie, The danger of a single story, 2009. URL: https://www.ted.com/talks/ chimamanda_ngozi_adichie_the_danger_of_a_single_story. [42] C. Kring, K. KU Leuven, DE-BIAS, Face/surface. metamorphosis of colonial perspectives, 2024.
URL: https://kadoc.kuleuven.be/3_onderzoek/33_onzeonderzoeksoutput/tentoonstellingen/2024/ tt_2024_bias. [43] L. V. Belle, In the place of shadows, 2022. URL: https://www.lavaughnbelle.com/home-1# /in-the-place-of-shadows/. [44] E. Sitzia, Multiple narratives and polyvocality as strategies of inclusive public participation:
Challenges and disruption in the history museum 10 (2023) 51–63. doi:10.7202/1108037ar. [45] J. W. Scott, Gender: Still a useful category of analysis? 57 (2010) 7–14. doi:10.1177/ 0392192110369316. [46] M. Foucault, R. Hurley, The history of sexuality. Volume 1, The will to knowledge, Popular Penguins, Penguin, 2008.