1.1 A working definition of the corpus

Current perspectives on historiography see a continuum between the archive, the document, the source, and the corpus, and are all connected with the notion of “documents produced by the actors of the history under study” (Offenstadt 2011, 68). Since historians generally have a clear understanding of what a corpus is, one could expect there would be an effort towards defining, clarifying, and systematizing the notion. That has not been the case: for example, the use of the term corpus among French medievalists is widely spread, but there has not been any reflection whatsoever about its meaning, nor any effort at conceptualization (Magnani 2017a). Before delving into the history of corpora in historiography, we therefore propose to rely on a simple definition of the term in the context of historiography to avoid any ambiguity: a corpus is a set of documents assembled with a specific purpose (a set may naturally be a singleton, so that a corpus may only contain a single document).¹ In linguistics, the notion of corpus has been an operating concept for many decades, at the core of corpus linguistics. Approaches to corpora in the field are geared towards methods intended to make analyses more systematic (McEnery and Brezina 2022, 4–28). The linguistic corpus is a collection of texts in an electronic database that can be used for linguistic analysis and description (Kennedy 1998, 1–3). In fact, the corpus is seen as a system whose very internal consistency is the subject of study (McEnery and Brezina 2022, 74–78). The use of the notion of corpus is hence narrower in linguistics than in history, where any collection of documents may be a usable corpus, because the logic that defines its construction comes from the historian rather than from the documents themselves. For example, a series of ancient coins in a specifically localized find may constitute a valid historical corpus (Nicolet-Pierre 2002, 58–59), but it may not have any internal consistency beyond the fact that the coins were found together. The difference between mentioning a corpus and mentioning a more generic “set of documents” resides in the former’s more direct association with historical methodology. Writing of a corpus conjures up the logic that prevails in determining which documents are included or not.

1.2 The corpus through history

The notion of document and the writing of history have been associated for a long time. Constituting a corpus to carry out historical analysis is part and parcel of historical work and has even been considered to sufficiently define the very fact of writing history. In 1934, in a violent critique of a monograph on royal accounting, which triggered an intense historiographical debate, Lucien Febvre insisted that analysis was more important than erudition. He still recognized that putting together a corpus defined the writing of history, “the most exacting attention to procuring usable documents, unfold them, file them, and sort them in a rational order” (Febvre 1934, 149).

How has history writing come to rely on constituting corpora as the centerpiece of its methodology, “la centralité de l’archive” as Jean Boutier coined it (Boutier 2014, 10–11)? As Jose Carlos Bermejo Barrera pointed out in his historiographical study, there could be no notion of a corpus or document in ancient historiography, because the very witnessing of events served as the foundation for history writing (Bermejo Barrera 2001, 193). Relying on one’s direct witnessing of events naturally implies that history writing was mainly concerned with contemporary history, and texts were held as inferior witnesses (Boutier 2014, 12–14). Boutier also distinguishes the historians, who wrote about contemporary events, from antiquarians such as Varro, who relied on a close study of text and artifacts to examine prior customs and ways of life (Boutier 2014, 15), and both endeavours remained distinct until the end of Antiquity.

It is only when the history of the Church became a domain of inquiry that historians began to consider the existence and necessity of a text corpus, according to Bermejo Barrera. This body of text had a materiality, just like sacred relics did, and the direct testimony from the contemporaries of Christ could not carry as much importance as personal testimony could have in the ancient Greek tradition (Bermejo Barrera 2001, 194). Close reading of the ancient texts developed into hermeneutics, which eventually led to the foundation of a historical method. The first university seminar focusing on the examination of primary sources took place at the university of Göttingen in 1766 (Boutier 2014, 19). According to Gunther Pflug, this prevalence of the document at first prevented any form of deductive operation, and in the eighteenth-century thinker Pierre Bayle’s perspective: “The scholar’s goal consisted of surveying the factual data, penetrating the historical givens, without attempting to impose any order unless it were for mere purposes of clarity” (Pflug [1954] 1971, 5).

Voltaire and later Turgot pulled history away from straight facts towards scientifically inspired analysis, relying on reason and common sense, thereby making the notion of pure document-based facts less central. Still critical and essential, the corpus now functioned with the application of reason and inserted itself in the context of the question asked by the historian (Pflug [1954] 1971, 9–12, 20–21). As the writing of history became professionalized in the nineteenth century, the methodology of source critique converged towards current practice (Offenstadt 2011, 70). At that juncture, historical knowledge acquired “a new configuration thanks to the introduction of two notions: that of document […]; and that of the scientific method” (Bermejo Barrera 2001, 198). This perspective effectively established the document and the aspiration to a scientific approach as two facets of the same coin. Indeed, in a Foucauldian approach, analysis that is specifically historical, as well as a more generalized form of analysis common in the social sciences at large, both stem from the same source: “the questioning of the document” (Foucault 1969, 13). Bermejo Barrera stresses the primordial place that the corpus holds in current historiography: “History builds its object starting from the constitution of its documentary corpora; it then develops different methods of reading and interpreting the texts, methods that are sometimes contradictory and that are not reducible to a common factor” (Bermejo Barrera 2001, 204). This view is largely connected with Foucault’s perception of historiography’s position with respect to the document: it seeks not to interpret it, but to work it from the inside and elaborate it; hence the document should not be seen as inert material. Foucault defines the writing of history as the way a mass of documents is organized (Foucault 1969, 14).

1.3 Questioning the corpus in history

It hence follows that the creation of a coherent corpus is one of the salient issues in historical methodology: determination of the documents to include or not, determination of the level of detail of the analysis that is required, determination of an analytic method for the quantification of various aspects from the documents (Foucault 1969, 19). Recognizing the influence that the presence or absence of a particular document may have on the writing of history, Boutier raised an important issue: one may ask which document should be used for which inquiry, but one should also ask to what extent the historian’s questions will drive the gathering of a corpus (Boutier 2014, 10). The use of a corpus in historiography, for Foucault, is strongly related to a necessary serial perspective: it is constituted following a particular and systematic methodology and lends itself to quantitative analyses (Foucault 1969, 19). Collecting similar or comparable elements, by construction, creates information that can be processed in a serialized form.

Some historians have recently sought to put their practices into question as they pertain to the notion of corpus. While there is no question that the document is central in historiography, the issue is raised to precisely establish what a corpus is, and how one should make one. Implicitly following Foucault, Cécile Treffort stresses that a corpus has meaning, an anima inspired by the historian, and in that sense would differ from a simple set of documents; it “emanates from, translates and illustrates the researcher’s thought” (Treffort 2014; my translation). At the same time, the corpus is tasked with aiming at comprehensiveness or at least representativity for the question under study. It therefore appears that one could distinguish the most comprehensive corpus containing every available document, a universal corpus, from the one specifically formed as a subset of the larger corpus to address a specific question, an oriented corpus. This distinction is largely related with that between a corpus with a collective aim (made available online, for example) and a corpus with a personal aim (for the researcher), drawn by Treffort. Magnani pointed out that the use of the term by French medievalists was very often qualified with a possessive (for example, “mon corpus,” “notre corpus,” see Magnani 2017a), which stresses the personal and specific way in which the corpus is constituted, from the standpoint of the historian.

To combine perspectives on the corpus in linguistics, literature, and human sciences, Damon Mayaffre proposed a conception of the corpus that reflected the ongoing interaction between the final interpretative act of a corpus and the original act of its constitution (Mayaffre 2002), which in effect internalizes the tension that Boutier pointed out. First, one needs to realize that creating a corpus necessarily implies a serial view, as Foucault had already mentioned. Second, the corpus, which in this context must be an oriented corpus as a research object, is an arbitrary construction whose worth only comes from the questions it raises and the answers it provides (Mayaffre 2002, 3). (We could also express this idea by stressing that it is the corpus’s orientation that carries meaning, more than the corpus itself: the corpus’s purpose in its definition is its key attribute.) Therefore, the corpus in history must be seen as something flexible and evolving, whose limits are dependent on the historical question and on the researcher’s own discretionary decisions. As a mechanical consequence, there cannot be a unique methodological approach to creating a documentary corpus for historical research.

Considering this historical perspective, we conclude by stressing the crucial importance of the dialogue between manufacturing a corpus and effectively exploiting it, and on the fact that this ongoing dialogue renders the making of a corpus heavily dependent on the question at hand. As the power of digital processing advanced, it is natural to wonder how, and to what extent, digitization in general has impacted the making and the handling of corpora.

2.1 The broader question of hermeneutics in the humanities

In the humanities, hermeneutics have long dealt with the problem of how meaning is produced, interpreted, and transmitted through layers of historical context. Considering the evolution from philology to modern textual criticism and, more recently, to digital humanities, it appears that interpretative practices have been shaped by methodological factors rather than epistemological breakthroughs, unlike in the hard sciences. Hence, the underlying question is not whether digital methodologies alter hermeneutic engagement with texts, but how they reconfigure the conditions under which interpretation takes place.

In his history of the humanities, Bod (Bod 2013) reduces the field to a general form of structured reasoning and argues that the search for underlying principles and structures has historically driven humanistic inquiry, much like in the sciences: “The transition from early modern to modern disciplines is usually seen as a conceptual break. […] However, […] contrary to the conventional division between natural sciences and humanities, we find a continuous interaction and even methodological similarities” (Bod 2013, 347). Hence, “the strongest interaction between the sciences and humanities is currently happening in the upcoming field of digital humanities” (Bod 2013, 347). Digital humanities, in this perspective, are a simple extension of the humanities because they rely on the same underlying scientific principles (Bod 2013, 91–92).

Seeing digital humanities as an extension of the humanities is not problematic here, but we take exception about the characterization of this extension. Indeed, one needs to go back to the roots of the humanities. Philology, which dominated the field from Antiquity to the nineteenth century, does not merely consist in the study of texts but in the broader investigation of language and textual transmission. Turner describes it as “the multifaceted study of texts, languages, and the phenomenon of language itself” and traces its development into modern humanities (Turner 2014, 3). He argues that disciplines such as literary studies, history, and linguistics all emerged from the core methods of philology. However, while philology seeks to reconstruct text and uncover its historical meaning, textual scholarship has increasingly recognized that meaning is not merely extracted but also produced through interpretative acts. Indeed, interpretation is an extension of “meaningful action,” so that meaning is not simply retrieved from texts, but actively produced through engagement with information (Ricoeur 1971, 211–213).

The traditional dichotomy between lower criticism, focused on establishing reliable readings of texts, and higher criticism, engaged in interpretative analysis, becomes blurred in modern textual scholarship, particularly in digital humanities, where the construction of a digital corpus requires interpretative choices at every stage: “the essential critical components of selection, evaluation, emendation, and annotation of texts still need to be emphasized” (Greetham 2013, 16–17). The assumption that a corpus is a neutral repository of texts is countered by the fact that encoding, metadata structuring, and selection criteria are themselves hermeneutic decisions. Blanke and Hedges (Blanke and Hedges 2013) also argue that these infrastructures encode epistemic constraints into digital environments, implicitly directing scholarly engagement through preconfigured analytical frameworks. This perspective aligns with the argument that digital textual methodologies should not be assessed merely in terms of their computational capabilities, but in how they mediate interpretative engagement with texts (Aledavood 2024, 14). Aledavood calls for a mixed-method approach that acknowledges both the benefits of systematic processing and the inherently interpretative nature of corpus creation. This suggests that digital humanities should be seen not as a rupture with traditional textual studies but as an adaptation of their underlying hermeneutic principles to digital environments. The notion of “scholarly primitives” describes research activities such as annotating, comparing, and referencing that remain consistent across different methodological paradigms (Unsworth 2000; Pacheco 2022). Digital methodologies transform the scale and efficiency of these activities, but they do not fundamentally change their epistemological status.

At stake in these debates is the broader question of how interpretation in the humanities is reshaped by changing methodological paradigms. If philology provided the original framework for textual interpretation, and textual scholarship redefined how texts were edited and analyzed, then digital humanities introduce new layers of interpretative mediation. Yet the fundamental hermeneutic challenge remains the same.

2.2 Impact of the digital on the chaîne opératoire of historiography

Today, putting together a corpus in historiography does not in general imply the physical gathering of original manuscripts, archives, or artifacts. It does usually imply, however, gathering paper copies of such material, or notes about them, and organizing them in some manner. In the physical instantiation of a reflective corpus, for example, one may find photocopies of journal articles or book chapters written on the subject of interest, organized in folders along some relevant taxonomy, and with handwritten remarks in their margins. One may also find a recent edition of a primary source filled with numerous and colourful Post-it notes, with some commentaries on them. (Interestingly, Kiewra 1989 provides empirical evidence that active engagement in text organization enhances comprehension and recall, as there are cognitive benefits in note-taking and other manual forms of information processing.) This practical way of making a corpus is only scalable to a point. As the size of the corpus reaches the hundreds of thousands of pages, one must add more and more meta-information and structure, and, for example, maintain lists of lists of references and extracts. In this process, digitization intervenes at several levels: document access, document exploitation, document organization, and storage and organization of the historian’s own work. First, the documents themselves may be identified, accessed, or read in electronic form, from which they still may be printed and become a physical instantiation of a corpus. The electronic format also permits the electronic use of the documents by allowing word searches or the automatic extraction of relevant text excerpts, for example, in the case of textual data (one may also, for example, filter a large corpus of digitized images based on their meta information). The electronic exploitation of documents extends in fact well beyond that of text, and includes all the domains that are related to history, in particular, paleography (see, for example, the range of meta-information that can be added to manuscripts with digital analysis in Andrews and Macé 2015), art history (see, for example, Näslund and Wasielewski 2021), as well as archaeology (the turn towards digital methods in archeology took place early; see, for example, Evans and Daly 2006). The gathering of the documents and their logical organization may be also done in electronic format, using the logic of computer file directories, in its simplest form. The notes, whether on primary or secondary sources, can all be entered and stored into an electronic form as well. Finally, the historian’s production itself will typically be carried out in a word processor. When all these steps take place virtually rather than physically, we do not necessarily observe that more information gets produced about how their logic is articulated. This creation and use of corpora in a historical perspective does not contain in itself a systematic description of their construction.

For Philippe Rygiel, historians are indeed “hypertextual polygraphs, who dissimulate most of the inscriptions they produce” (Rygiel 2011, 32). He stresses the fact that most of the historian’s work is not visible from the results or the analyses they publish. The practical instance of a corpus we described above would effectively be mostly hidden and could only be marginally inferred from the resulting history work. In Rygiel’s view, the historian’s annotations, essentially in textual form, constitute the core of their work, their production. In this perspective, the historical inquiry becomes the delineation of a corpus, augmented with these annotations (Rygiel 2011, 34). The production of annotations is organically linked with the material corpus off which it is based, and we can see how this would logically end up in a possessive attribution, as Magnani pointed out: the corpus’s orientation is part and parcel of the historian’s work. How does the digital impact this framework? It appears the introduction of digital tools has only had a limited impact on historians’ practices: Rygiel observes that the historian’s production, the annotations, are mostly made in a digital framework and hence should ideally be made accessible to all, but they are generally not. One reason for this, he argues, is that if this production was systematically made public by providing all with the same raw material, it would end up raising the bar of expectations among all historians (Rygiel 2011, 38). Gibbs and Owens also lamented the fact that the way data is used by historians is rarely well documented and made accessible (“Despite some recent methodological experimentation with data, historians have not been nearly as innovative in terms of writing about how they use it,” Gibbs and Owens 2013, 163). Focusing on many of the tools that are nowadays available to historians for textual analysis (such as Google keyword searches or newspapers electronic archives), Tim Hitchcock pointed out that “academic historians did not ask for these resources, and nor for the most part have they been directly responsible for their creation” (Hitchcock 2013, 10). The tools in question, as well as many of the generic tools we mentioned earlier about the practical creation of a corpus, are indeed not specific to the field of history. More recently, Siebold and Valleriani pointed out the fragmentation of historians’ approaches: “A close look at digital databases, network analysis, and ML shows that a very diverse and highly specialized set of new research practices has evolved, many of which are still being further developed. However, it must also be noted that the current situation is marked by a certain heterogeneity as to how these new practices are carried out” (Siebold and Valleriani 2022, 174). Crymble provides an overview of the various digital tools that historians are generally trained on, whether through their institutions or through self-learning, and points out that the tools in question are extremely diverse (see Chapter 4, “Building the Invisible College,” of Crymble 2021).

Rygiel and Hitchcock’s perspectives seem to establish that the evolution towards digital frameworks, tools, and analyses has taken place despite, rather than thanks to, historians. Hitchcock’s observations recoup those of Rygiel: he notes that the way historians carry out their work has not changed substantially from the 1980s. Some have argued that the very act of publishing could massively benefit from the new tools afforded by a purely digital framework: Nawrotzki and Dougherty illustrated this idea quite efficiently by using a collaborative and web-based publication process for their book on the subject (Nawrotzki and Dougherty 2013). They stress that, in their experience, the opening and sharing of resources did not increase competition between scholars, but rather led to more collaboration. However, their project did not include putting in common a corpus in electronic form, but rather their own production, that is, what Rygiel described as the historians’ annotations.

It appears so far that history aligns rather well with the hermeneutic framework of the humanities even in its extension to the digital: the methods stay seemingly the same, and evolutions take place at the margin.

3.1 Making digital corpora for history

As Gibbs and Owens stated, “historical scholarship increasingly depends on our interaction with data” (Gibbs and Owens 2013, 159), and indeed such a statement would probably match most historians’ intuition.

In her review of the use of the notion of a corpus by medievalists, Magnani took digital processing as a given, and in his definition of the notion of reflective corpus, albeit from a linguistics perspective, Mayaffre proposed that it be in practice structured as hypertext, whereby each text in what we called the oriented corpus would be linked to the its parent texts, in particular using standard XML encoding to account for these connections, implicitly stressing the necessity of an electronic representation of the corpus, in order to make its internal structure apparent (Mayaffre 2002, 8). Corpora and digitization, from the perspective of historians, are more and more linked together.

The large-scale digitization of historical material does not systematically require the formal modelling of complex links between the elements that compose the corpus; the effort sometimes mainly consists of ensuring the quality of the resulting electronic text or data, and in the storage of all relevant metadata pertaining to the original documents. Such electronic corpora are usually constructed to be “simple, multiple, open and free of access,” as is the case in the example of a medieval charter corpus (Magnani 2017b, 64). Nevertheless, when the underlying information is more complex, or the orientation of a corpus requires more effort, historians approach the creation of a digital corpus specifically under the guise of an ontology (in the sense of information theory), that is, a formal representation of knowledge, typically in the form of a database, which forms a structured view of the data and of some of its relations. To build a corpus for her dissertation, Ansley Erickson created a database as a way of keeping track of her notes relating to particular sources or material (Erickson 2013). This led her to reflecting upon the role of categorization in storing data for a historiographical use. Attaching attributes to the data that was thus created (as opposed to, for example, a simple alphabetical organization) allowed for more flexible thought processes pertaining to the matter at hand. Erickson’s work produced data (her notes related to various documents), but the underlying data did not have an electronic representation, and, in this instance, the resulting database was effectively disconnected from the sources.

In numerous cases, the sources themselves already are in electronic form, thanks to a prior digitization, and additional data, whether textual, iconographic, or quantitative, resulting from analysis or from other sources, can be attached to it. An example combining electronic sources and historians’ annotations is the Homer Commentary in Progress, where scholars can contribute detailed analysis attached to any word or sentence in the Iliad and the Odyssey, so that both the underlying text and the annotations can all be queried, searched, and compared (Crane et al. 2016). This type of relationship between historical documents, a universal or oriented corpus, and historical work associated with them, is indeed conceived of as a set of annotations. This approach is far from systematic among historians, but there are many instances of such practices, in particular in classics. For example, Mugelli and colleagues defined an annotation logic to isolate and categorize references to sacrifice in Greek tragedy (Mugelli et al. 2017). The underlying corpus, in this case, is the Greek text in electronic form, and the annotations inserted within the text contain all the elements contributed by the historians: the specific location in the text where a sacrifice is identified, the categorization of the sacrifice, or the disambiguation of the discourse about the ritual from the ritual itself, for instance. By processing the annotated text, one can generate a database of ritual events. Barker and colleagues (Barker et al. 2013) used a very similar framework to study spatial references in Herodotus. They relied on the English translation of the Greek text and added annotations tracking the geographical information relevant to them; a copy of these annotations with the surrounding text were then stored in a database.

In a recent methodological paper, Hoekstra and Koolen generalized these notions of organized data for historical research using data scopes, which they defined as “the process through which different views on research data are created that are relevant to a specific research question” (Hoekstra and Koolen 2019, 80). In their perspective, a corpus for historical research should be created with a general data structure in mind, where different related sources of information can be easily paired to better gain new insights. In effect, this corresponds to transforming a general corpus (from the sources that may be available) into an oriented corpus. In their analysis of the interpretation of data by historians, Gibbs and Owens did not focus on a particular way in which data should be represented but concentrated on the general idea of sharing data processing methodology, which can be understood as a more general perspective than data scoping. Siebold and Valleriani (Siebold and Valleriani 2022) also called for a degree of unification of formats to promote data transparency and sustainability, which stresses that the current situation may not be transparent nor sustainable.

In historiography, the creation of a digital corpus hence appears to be a complex exercise for which, to the dismay of many of the researchers whose work we mentioned above, there is no clear and unique epistemological framework or methodology. Further, the amount of detail provided to readers of historical research, when such work is carried out, varies greatly.

3.2 The digital processing of a corpus

When a digitized corpus is exploited in a way that leverages its electronic form, we may wonder if there is a generally agreed-upon method, a systematic manner to carry out this process. If it appears that it is desirable to make the processing of a corpus automatic, to what extent is this automation done in a unified manner? In the field of classics, it is possible that the relative paucity of textual sources, by making the processing of the entire corpus a reachable objective, encouraged digitization efforts, which led to the widespread use of electronic text search tools. As Barker and Terras (Barker and Terras 2016) pointed out, it has become quite common in ancient history to rely on detailed text searches, and there are a variety of dedicated tools for this purpose. Relying on electronic sources such as the Thesaurus Linguae Graecae (TLG) or Perseus (Pantelia 2020; Crane 2012), for example, is standard practice in Greek history. In these cases, while historians effectively use the algorithms that power these search tools, they do not encode the series of steps they follow in a formal way. The way in which the electronic nature of the underlying corpus is leveraged by the historian through their use of textual research in these examples is not made explicit: the orientation of the corpus, in effect, cannot be exactly reproduced in a systematic fashion.

Do more advanced, or more technical, uses of electronic corpora translate into a fuller documentation of their processes? The use of advanced tools, such as machine learning and text mining, and the tools of big data in general, have made their way into the hands of historians (Graham, Milligan, and Weingart 2015). Some use cases of text corpora effectively involve advanced computational methods, programmed to operate one after the other. Such quantitative analysis of literary texts is not recent, and it has reached a certain maturity (Hoover 2013). The systematic exploitation of textual data has in fact become widespread and standard enough that there are programming manuals focused on this type of exercise, covering all the standard operations one might need. Some domains require specific technical expertise, such as the analysis of the networks stemming from literary texts. These analyses resort to various technical domains that are not part of traditional linguistics (Kenna, MacCarron, and MacCarron 2017). The analysis of intertextuality through the methods of computational biology is another example (see, for example, Barbrook et al. 1998; Howe et al. 2001; Chaudhuri and Dexter 2017). Still, the tools now available allow their users to program all forms of lexical and semantic analyses within a unified framework, in a process conceived of as a pipeline, comparable to the data analytics suites used in the hard sciences. Jockers and Thalken (Jockers and Thalken 2020) cover the practical approaches to this technique. In the case of classics, P. J. Burns (Burns 2019) describes a series of useful operations on classical texts following the same logic. A more general treatment of pipelines from the standpoint of data analytics, can be found in Wickham, Çetinkaya-Rundel, and Grolemund (Wickham, Çetinkaya-Rundel, and Grolemund 2017), pages 261 to 268 in particular. In contemporary French history, Magali Guaresi resorted for instance to logometrics and sentiment analysis, and applied factor analysis on a large corpus of speeches by members of the French Parliament in order to characterize the evolution of these discourses (Guaresi 2019). Classicists and computer science specialists have also collaborated on an empirical study of the quality of automated sentiment analysis in Greek tragedy and shown that automatic processing yielded good results compared with humans (Yeruva et al. 2020). In these examples, one cannot in fact speak of methodological unity, because, while the analyses in question were discussed in these research articles, the full details of the actual tools and algorithms that were used were not made available. If the analytic pipeline is not made explicit, then the logic that prevailed in the assembly of the corpus cannot be made explicit.

It appears that, while various tools and technical approaches have converged towards a form of conceptual pipeline (if not algorithmic), thanks to general technical progress, there is no centralized perspective about how textual corpora may be processed, and there rarely is detailed information on the actual steps undertaken to create the oriented corpus and carry out the analysis, in relation with each other. Then, although the notion of a corpus in historiographical work, stemming from a long history, as a collection of documents, is effectively manipulated by historians, the conjunction of that notion with the massive expanse of digital methods has not resulted in a clear and unified epistemological framework, which would leverage the electronic format in order to make the construction of a corpus and its processing clearly identifiable.

4.1 Corpora and algorithms

It is fundamentally beneficial for the design of a corpus and the processing of its information to be automated and leverage its electronic nature when possible. One can indeed make an argument linking the reproducibility of research, gained by digital processing, to its scientificity. For McGillivray and colleagues, collecting and processing historical material with computational methods “would be a science if we could learn to automate it” (McGillivray, Wilson, and Blanke 2019, 53). In their view, historians should clearly delineate between what they define as “evidence” and what they define as “claims,” so that in a positivist perspective one may separate evidence-based findings from other statements, thanks to the systematic analysis of evidence. Example-based analysis would be, in their view, kept separate from quantitative evidence (see the chart in McGillivray, Wilson, and Blanke 2019, 55). Distinguishing evidence and claim in historiography is no trivial affair, as Febvre insisted (Febvre [1952] 1992, 115).

Without trying to reduce history to a simplistic dichotomy between “claims” and “evidence,” we can still recognize that there is something to be gained from being able to precisely understand the documents and analysis supporting any statement, when they can be made explicit. Making processes automatic, hence, may not, per se, serve a fundamental purpose, but it would have the advantage of making the creation of a corpus and some of its processing reproducible, and open it to critique. Reproducibility makes any statement falsifiable, which is a fundamental feature of any scientific statement according to the Popperian logical framework (see “Falsibiability,” Popper [1934] 1992, 78 sq.). We are not arguing for scientificity for the sake of it, but rather as a form of transparency. This would also recoup with the objective laid out by Gibbs and Owens, by inherently providing detailed documentation on the use and processing of a corpus. Hence, keeping in mind the ability of others to reproduce parts of an analysis could help make the method itself more open to critique, without reducing or limiting historians’ work. Should the humanities, because they can become digital, strive for the kind of reproducibility generally associated with the natural sciences? Digitization does not negate the hermeneutic challenges attached to textual analysis, so this notion of scientificity must effectively be adapted. Instead, the emphasis should be on transparency and interpretative flexibility, which ensures that the process of digital corpus creation and manipulation is open to critique.

Data, as the etymology tells us, is what is given, and cannot be worked out otherwise. Operations are applied to data to transform it into something more usable or practical. Operations orientate the data. Hence, data is information that cannot be derived from other information using logical operations and requires a human input. In computer science, this distinction between data and operations is better known as that between data and algorithm. Since each needs the other to exist, they are naturally intimately related, but are fundamentally and conceptually different (Wirth 1976). A database, or an ontology, is not simply data, because it combines data and algorithms that describe the way the data may be exploited. Any database can be represented as a combination of raw structured data (typically as tables of numerical or categorical data, because any multi-dimensional dataset can naturally be represented as a two-dimensional array) and algorithms that define potential relationships between the columns in these tables. Seen as a whole, the database may obfuscate the interactions between data and operations, and makes it seem as though it only contained the resulting dataset, organized as its designer intended.

All the examples of digital corpora we have discussed so far effectively are combinations of data and operations: they never were the straight input as provided by a person. These corpora were nevertheless considered as things, as data, focusing on the result of a series of operations rather than on the operations themselves. The precise descriptions of these results, without explaining the series of operations, were therefore incomplete and superficial. If we realize that a corpus is not data, but rather a set of operations, it may solve many of the issues we have noted so far. In addition, an operation could at the limit effectively contain any dataset: an algorithm may simply possess the list of values that constitute the data. In this sense, a process is superior to data: it is denser in informational content. We have mentioned the notion of a pipeline earlier, in the context of modern data analytics, or in computational linguistics. In such a pipeline, the raw input at each stage is transformed in place and serves as the input for the next stage. From this perspective, an oriented corpus should be seen as a pipeline, that is a series of operations, not as a dataset. The definition of this corpus is the computer code that builds it, whatever the language in which it may be expressed. This code can be analyzed, and it can be run in whole or in part by anyone. Purely seen as a dataset, a corpus cannot be properly analyzed from an external perspective, but once its construction is made entirely explicit, then this process can be fully subject to critique. In practice, historians often follow such a pipeline: for example, obtain some data (in a spreadsheet), transform it, save the clean version, compute some aggregates. However, only the result survives, and the details of all the steps are lost to everyone else. If this entire process is coded as a pipeline or as an algorithm more generally, every single assumption, explicit or not, becomes visible.

Corpus construction often involves nested layers of selection rather than a single sub-setting operation. Consider, for example, a study of certain interjections in nineteenth-century dialogues (Gauthier 2025): one moves from the complete corpus of digitized texts to a subset of novels and plays, then to direct speech, then to interrogative sentences, and finally to those containing interjections. Each stage structures the corpus in a way that preconditions analysis, making the creation of the corpus itself a dynamic and relational process rather than a static object. One might argue that corpus processing pertains more to analysis than to its construction, vindicating the notion of a corpus as an object. However, corpus formation itself is a structured, iterative process that encodes methodological choices at every stage. Selection, categorization, and filtering are not neutral operations but interpretative acts that define analytical possibilities, even more so in a digital environment.

Hence, once defined as computer code, or as a pipeline, the corpus contains and makes explicit all the decisions, small and large, made by the historian in cleaning, filtering, completing, or arranging the raw information, in the most concise manner possible. By construction, this perspective on the corpus as anima, that is, the process which orientates it, also directly expresses the dialogue between the research question and the setup of the corpus. Modern data analysis and text edition tools are converging, so that there is not such a strong distinction between the two anymore: the text a researcher produces (the historian’s annotation work in Rygiel’s words) and the computer code that gathers and processes data can now exist in the same document. Considering the data organization, processing, analysis, and write-up of a historical corpus as a continuum in a seamless process naturally leads to reproducible research. The RMarkdown language, for example, combines the data and statistical modelling infrastructure of the R language with the edition capabilities of the Markdown syntax and LaTeX system; it has been suggested as a good framework for reproducible research (Calero Valdez 2020).

We can relate this view of the corpus back to the Annales’ co-founder’s definition of history: for Lucien Febvre, history may be viewed not as a science, but as the scientifically elaborated narrative of the activities and creations of humankind (Febvre [1952] 1992, 19). Febvre’s perspective is in line with that of Raymond Aron, for whom a philosophy of history cannot be simply positivist.² We argue that the notion of scientificity here, at the core, is about transparency and openness to critique. The idea of carrying out historical research not as a science, but with a scientific aspiration, is clearly furthered by making the historiographical process more replicable, at least for the part that concerns the setup of a corpus.

4.2 An example in ancient history

To illustrate this logic, let us consider an example with ancient Greek theater, where we would want to study the occurrence of decisions. This would require the creation of a dataset of these decisions, as a subset of the entire set of theater texts, so that they can be closely read and examined.³ The traditional way of tackling this project could benefit from the fact that these texts are available in electronic form: one could easily copy and paste all the Greek text of interest, maybe accompanied with translations, into a word processing software. For each excerpt, one could give a categorization for the decision at hand (whether it is an acceptance or refusal, for example). If instead we followed a more digital-based logic, comparable to the one described by Mugelli and colleagues (Mugelli et al. 2017), we would first create a copy of the Greek text (from Perseus, for example) in some standard TEI format. Then, inside this text, we would add markers that designate decisions, with some specific coding logic to distinguish between different sorts of decisions. We may also correct the text, if there are issues in how it has been established. Processing the annotated text could then result into a database, containing all the text excerpts with their categorization, amounting to a much more structured and easily exploitable corpus than a list of quotes in a word processor. Researchers would presumably query the database in question, and obtain, for example, tables showing which kind of character expresses decisions the most often; or they may query all decisions expressed by male characters who are slaves. The resulting tables could be pasted into the researcher’s text. This later approach, nevertheless, by creating a new object, severs the link between the original textual source and the resulting corpus.

A process-driven, rather than object-driven, perspective on this corpus would consider the text source as an input that should not be changed in place or copied. If corrections are needed on the text, they are made explicit in the pipeline, as a set of overrides. The historian’s work in defining the corpus would be embedded in a separate dataset, simply containing the decisions’ location identification (in the form of line and word number, for example) and their categorization. From the text input and the categorization data, by merging the two, one can automatically generate the equivalent of the annotated text if needed. Any textual analysis on the text of the oriented corpus, the decisions themselves, can be easily compared with the same analysis applied to the rest of the text, excluding decisions. It may be useful to define categories as a function of the characters’ social status, gender, age, or any set of characteristics that may be relevant when examining decisions, and one may also distinguish sections such as the prologue, or sung parts in the text, for example. If one has indeed carried out this categorization, then all these taxonomies directly and automatically percolate through to the distinction of decision related in opposition to non-decision related text. If a researcher wanted to categorize certain parts of the text differently, leaving all else the same, this would be possible, and allow them to directly obtain the corresponding results by running the pipeline from beginning to end. In addition, if the underlying electronic Greek text’s edition is improved and some words are corrected, then the decision corpus immediately benefits from these improvements.

Figure 1 shows, as an example, the proportion of words pronounced and the proportion of wills or decisions expressed by different types of characters in a selection of ancient Greek tragedies. This can be commented and studied in many ways, but this is not our purpose here. This data, as is, cannot be further explored. A full algorithmic traceability of how this data is constructed is necessary. Figure 2 shows (for a handful of plays) one of the inputs, which can be independently examined and questioned: the manner in which each character is categorized. Figure 3 shows another one of the inputs, that is, the categorization of expressions of will, decisions, or acceptations/refusals attached to specific verses and word numbers in the verses, here for Philoctetes. Odysseus wants Neoptolemos to do various things, and Neoptolemos agrees or reformulates some of them. The precise references to verses have to be mapped to an electronic reference for the text itself, in the example here, Perseus’s text shown in Figure 4. Now, Figure 1 does not depend on a very close analysis of the text, so one could argue that the data from Figures 2 and 3 is almost enough by itself. Figure 5 shows another example of analysis that looks much more closely into the way decisions are expressed and relies on the full details of the text. in this case, one must closely map each decision to the words that are related to it. Every input in the process described here and every step in the algorithms that put together the data that can then be analyzed must be open to scrutiny. Just saying that the underlying data is the electronic text in Perseus is clearly far from sufficient. In addition, a simple “data dump” of the results would not allow any inquiry into these steps.

Figure 1

Relationship between character categorization, scenic presence, and expressions of will across a large sample of tragedies.

Figure 2

Greek theater character categorization example.

Figure 3

Categorization of decisions and expressions of will.

Figure 4

Sophocles’s Philoctetes on Perseus.

Figure 5

Comparison of syntactic forms used in the expression of decisions, relative to occurrences in entire tragic corpus (in French).

Disambiguating the universal corpus, in which one seeks excerpts or categorizations, from the added information produced by the historian, we open the entire process to a much better understanding and to document critique. The overriding principle is that any input by a historian related to the constitution of a corpus, whether in terms of raw data or processing logic, should be reflected as a step in an algorithm or pipeline, and not as a physical operation, such as clicking on some instruction on a piece of software. In Bod’s (Bod 2013) history of the humanities, hermeneutics were commingled with the hard sciences. Here, the perspective aligns much more closely with Turner’s (Turner 2014) history of philology, in that making the creation of digital corpora explicit allows a philological critique of this very process, which is impossible if one is just given an object.

Editorial