When people talk about uncertainty in a historical context in digital humanities, most of the time they talk about questions such as the exact date of birth of a person, whether two names refer to one or two persons, what geographical location a place name refers to, or the location of a person at a specific point in time. These are important questions and it is important to find ways to computationally model the associated uncertainty. However, history is ultimately not about drawing exact maps or time lines, even if those can certainly help: history is about causality. In this talk, I would like to reflect on some of the issues on the level of historical interpretation, i.e., historiographical rather than historical uncertainty.
In the last couple of years there has been an increasing interest in handling uncertainty in DH. I don’t think it’s really mainstream yet, but there’s definitely a growing awareness of the issues.
This is not surprising, or at least it shouldn’t be. When we’re talking about historical research, many, if not most, things are to some extent uncertain.1 Of course y’all know this, that’s why you’re here… But just to give some common examples, dates of birth are often lacking, we may not know to what person a name refers, whether two occurrences of a name refer to the same person or to two different persons, or to what location a place name refers, when a charter was issued, and whether it’s authentic or a forgery, and so on.
All of this (and then some) is usually subsumed under the heading of “uncertainty and vagueness.” Uncertainty and vagueness are related but not the same, but “vagueness” is itself a vague concept. When I’m uncertain whether some person was a member of some group, this may be due to a lack of information, but it could also be due to a vague concept of membership.
In this talk I’ll focus on what I believe to be uncertainty, so I don’t want to go into details, but I’d also like to point out that there are further related phenomena, such as
- context sensitivity—what does “here” or “now” or, for that matter, “I” refer to?
- underspecification—if someone tells you they’re going to visit their aunt, they underspecify whether it’s their mother’s sister, their father’s sister, or their father’s brother’s wife (and which one, if there are several) whom they’re going to visit.
- generality—the sentence “I ordered a steak” doesn’t really specify whether the steak is cooked or raw: in a restaurant you expect it to be cooked, but not at the butcher’s.
- sense and reference transfer, such as “I’m the steak,” which really means that you’re the person who ordered a steak. This also includes pars pro toto, totum pro parte, and similar figures of speech. You may think of rhetorics here, but it actually also happens a lot when referring to geography: for example, using “Holland” for the Netherlands (pars pro toto) or “America” for the United States (totum pro parte).
- and then there’s also (linguistic) ambiguity, of course. The best known type is probably lexical ambiguity with standard examples such as “bank” or “bat,” or “gericht” in German (meaning meal and court), but there are even wilder cases such as “hôte” in French, which can mean both guest and host.
While these are all very real phenomena, they usually don’t pose too many problems in everyday communication, at least to humans: we can usually infer from the context—and our world knowledge—what is meant; and if it isn’t, we can generally seek more information, e.g., by asking the speaker.
History, however, is concerned with “decoding what the sender did not want to transmit,” as Manfred Thaller put it (2018), i.e., the sources used by historians weren’t intended as a messages to them. The problem is precisely that we are often missing much of the context, that we have only limited knowledge of the world of the past, and that we can’t ask for clarification.
So far, with respect to uncertainty, computational humanities is, I think, mostly concerned with the identification of the referents of linguistic expressions—named entities and dates—or things like manuscript readings. The primary goal is usually to reduce uncertainty as much as possible in order to obtain something unambiguous—such as geographical coordinates—and precise (to some precision) that’s amenable to computational processing.
You could say that in that respect computational humanities faces challenges similar to science and engineering (actually: metrology): the “true” quantities (length, weight, voltage, etc.) are unknowable, so all measurements are subject to uncertainty.
Having more data by doing repeated measurements can provide an estimate of the true value of the quantity that generally would be more reliable than an individual measured value. A measurement result is therefore only considered complete when it is accompanied by a statement of the associated uncertainty, such as the standard deviation. Metrology has developed sophisticated approaches for working with measurement uncertainty and to avoid or at least control the propagation of errors.
Now, ideally we’d have the same rigorous methods for dealing with uncertainty in computational humanities, right?
This is a bit the idea behind projects like the famous Venice Time Machine. Don’t get me wrong, I think this is (or was) a great project for developing methods and tools for supporting historical research through mass digitization and information extraction, but unfortunately its claims go far beyond this. In the VTM project, the famous dictum of Langlois and Seignobos, “L’histoire se fait avec des documents” (Langlois and Seignobos 1898, 1) (↗ Topolski  1976, 112) was not only taken literally, but understood as “l’histoire se fait seulement avec des documents.” In fact, Langlois and Seignobos speculated:
On peut penser qu’un jour viendra où, grâce à l’organisation du travail, tous les document auront été découverts, purifiés et mis en ordre, et tous les faits dont la trace n’a pas été effacée, établis. — Ce jour-là l’histoire sera constituée […]. (Langlois and Seignobos 1898, 277)
This is precisely the understanding of history underlying VTM and the new project Parcels of Venice (which aims to also digitize all the historical catasters of Venice): once we have digitized all documents, we can digitally reconstruct the past—and we will then have established history, and consequently there will no longer be any uncertainty.
I’m not sure whether this “digital positivism” is the ultimate realization of 19th-century positivism or even goes beyond it.
Both the idea that history needs to be based on sources and the idea that the more numerous the observations, the more reliable the research (↗ Topolski  1976, 112) are not completely wrong, of course. But there are two basic problems that are not accounted for by digital positivism.
First, historical sources aren’t like measurements: the more I have, the closer I get to the “truth.” Sources don’t follow a Gaussian distribution. The historian’s work is really more like that of a detective:2 if one witness describes the suspect as tall with long hair and another describes him as short and bald, you can’t infer that the actual person is of medium height and has, I don’t know, a crew-cut. In addition, even if there are ten witnesses who describe the suspect as tall and long-haired, it is still possible that they were bought or that they actually saw someone else.
Second, the past is not the same as history, and the detective’s task doesn’t merely consist in gathering all the facts—a dead body, a gun, a broken window—the goal is to determine the causal relation between them; and it is only then really that the facts become evidence.3
Similarly, when I know that a particular house in Venice in 1648 belonged to a Signor Rossi, who paid this amount of taxes—well, that’s interesting. For a historian, these can obviously be very useful pieces of information about the past, but it’s not yet history.
What is the point of history, then? Well, this is likely to draw flak, because it’s unavoidable to simplify in a 30-minute talk, but anyway. Let’s put it this way: the foremost output of historical research are narratives. And what’s a narrative? Ultimately a reflection of a causal model.4 The causal model itself is usually not directly accessible, because it exists only in the historian’s mind, that’s why I say the narrative is a reflection of this model. A narrative goes beyond a mere chronology of facts in that it aims to describe causal relationships. It’s not merely: “first A happened, then B, then C”—remember: post hoc, sed non propter hoc. What we want to know is whether A caused B, which then led to C, or whether A and B independently contributed to C, and so on. (↗ Topolski ( 1976, 239f), Hewitson (2014))
If we want to learn anything from history, we must understand why something happened. And only if we can determine causality, we can establish more abstract patterns (↗ Bod 2015).
The challenge is then to build a causal model on the basis of historical facts. At this point—roughly speaking—the historical facts turn into historiographical facts. In other words: evidence for a narrative. (↗ Topolski  1976, 223–24)
When we’re talking about uncertainty—and in particular, when we want to model uncertainty—this distinction is essential.
Let me say a few more words about it. Again, this is necessarily a simplification. In fact, it’s more of a continuum than a binary distinction.
So, things happen. Things happened in the past. We don’t have a direct access to these past events, but we can reconstruct them to some extent using various kinds of sources. We are creating a model of the past, so these reconstructed historical facts have all of Stachowiak (1973)’s fundamental properties; in particular, the reduction property, which means that they do not capture all attributes of the original they represent, but only those that the model creators and/or model users deem relevant, and the pragmatic property, which means that they are not per se uniquely assigned to their originals; they fulfill their function as a replacement for the original a) for particular—cognitive and/or acting—subjects that use the model, b) within particular time intervals and c) restricted to particular mental or actual operations. (Stachowiak 1973, 131f)
So, for example, we know with reasonable certainty, from military records, etc., that a Polish soldier called Wojciech Najsarek, was killed at 0450 hours on the Westerplatte peninsula on September 1, 1939. Was he killed in World War II?
The straightforward answer is “yes,” as the salvos fired by the German battleship Schleswig-Holstein at about 0445 hours (there’s some uncertainty around the exact time, it may have been at 0447 or 0448 hours) are commonly considered to mark the beginning of World War II.5
Since you’ve been primed you may say, ah, World War II is a vague concept, and this is a classic borderline case.
Maybe, but that’s not what I’m trying to get at. This becomes perhaps clearer when you think about the question when World War II ended: on May 8, 1945? Or rather May 9? Or perhaps August 15? Or only on September 2? And if we consider the war in Europe and in Asia and the Pacific to be part of a single event, didn’t it really already begin in July 1937 with the second, larger-scale, Japanese invasion of China (following an earlier campaign in 1933)?
My point here is not that World War II is a vague term but that it’s something different from the death of the soldier: it’s not a historical but a historiographical fact.
And this is thus where we get into historiographical uncertainty. Historical facts delimit or restrict the possible historiographical interpretations. It would, for example, not be reasonable to arbitrarily set the end of World War II to August 31, 1949, just because it makes it last a nice, round 10 years. However, one can reasonable argue that it makes sense to consider World War I and World War II as a “Second Thirty Years War,” in particular from a Eastern European perspective, as fighting continued in what is commonly referred to as the “interwar period,” and one might even consider to extend it to include the Cold War (see, e.g., Pennell and Meneses 2019). Similarly, the conception of World War II as a “European Civil War” starts from the same historical facts but arrives at a different historiography.
We see that uncertainty is quite different in these cases: concerning the soldier we may be uncertain about, for example, the exact date or the exact location of the event. These are questions where the uncertainty could—in principle—be reduced (and thus “the truth” better approximated) by more data.
Asking for the “exact extent” of World War II, however, is meaningless. It also doesn’t make sense to do a survey among historians to find out. As a historiographical fact it depends on your causal model, the narrative you want to tell.
It would also be fruitless to try and define it quantitatively: you could look at the excess mortality and argue that that’s when the war begins, and that it ends when the mortality is back to normal. This may remind you of all the graphs and charts you’re seeing during our current pandemic, and in fact, pandemic could be said to be a historiographical fact, too.
So, you can’t define it quantitatively, but it heavily influences any quantitative analysis. “How many people died in World War II” sounds like straightforward counting, but it can only be answered if you’ve addressed the qualitative problem of historiographical uncertainty. The number you arrive at can then have very real effects, such as reparation claims.
This is an example of what we call historiographical uncertainty, and it is clearly different from the historical uncertainty, as it fulfills a different role in the causal model and the corresponding narratives.
Even though I’m stressing this difference, it’s important to remember that this is a continuum: it may be tempting—especially from a computational perspective—to say, “we’re just dealing with the historical facts, the rest is interpretation and thus doesn’t concern us.”
Historiographical uncertainty obviously also concerns not only time, but also space. In our World War II example one may consider the question of where the war took place. Took it place in, say, Finland? This depends on whether you consider the Winter War and the War of Continuation to be part of World War II—which is commonly done, even though one could argue that they were ended by completely separate peace treaties between Finland and the Soviet Union. Such questions are certainly highly relevant for computational modeling and illustrate that what’s called the “spatial humanities” can’t just restrict themselves to putting points on a map.
Here’s one way to think about it. Past events are temporally and spatially located, and thus time and space are also the foundations of the historical narrative. The space in which we move, in which we find objects, and in which events take place can be said to be structured on three levels. (Topolski 1997, 36–37)
First, we have a natural structure: mountains, rivers, deserts, climate, and so on, which structure or subdivide space in various ways.
Second, we have artificial, man-made structures such as towns, villages, borders, roads, mines, and so on. In the historical narrative, the artificial structures generally absorb or incorporate natural structures. History is only concerned with natural structures in that they serve a particular function, for example, a river as a means of transport, a border, or a source of water or energy.
Third, we have what Topolski (1997, 37) calls “subjective structures”: the structures that historians create in constructing their models and through which they tell their narratives. Narrative structures may be a better name, as they should at least aim to be intersubjective. In any case, these are abstract categories, independent from the natural and artificial structures, which they absorb in turn. Their construction is a fundamental step in what Topolski (1997) calls the transfiguration of space.
One example may be the idea of what’s called the Iron Curtain. The term already existed before the war, but it was given its current meaning by the well-known British historian (yes, he also happened to be prime minister) Winston Churchill in a speech in 1948.
The Iron Curtain wasn’t a tangible object. It relates to borders—artificial structures—but absorbs them, and thus also the differences between different stretches of the constituent borders: in some places there was just a chain-link fence, in others walls with mine fields and automated firing devices. For the historiographical concept, this doesn’t play a role, and neither does the exact geographical location of the borders.
Topolski certainly has a point when he calls such structures “subjective,” since even on the basis of the same source-based and non-source-based historical data, the resulting model of the past “cannot, by definition, be the same in the case of all historians, since historians differ from one another be it alone by their non-source-based knowledge.” (Topolski  1976, 224) This means that historiographical uncertainty is unavoidable because historiographical facts depend on a model that’s only partially accessible to us. This also means that it’s ultimately unresolvable.
The computational humanities are concerned with the construction of formal models in the humanities. Since historiographical uncertainty cannot be resolved, it has to be taken into account in our models. In fact, I believe that unless our models take into account the uncertainty inherent in the research objects of the humanities, they cannot realize their full potential.
One of the main challenges of the theoretical computational humanities is thus to devise metamodels that allow scholars to build models that can incorporate historical and historiographical uncertainty in a way that enables computational reasoning.
Bod, Rens. 2015. A New History of the Humanities: The Search for Principles and Patterns from Antiquity to the Present. Oxford: Oxford University Press.
Evans, Richard J. 1997. In Defence of History. London: Granta.
Hewitson, Mark. 2014. History and Causality. London: Palgrave Macmillan. https://doi.org/10.1057/9781137372406.
Langlois, Charles Victor, and Charles Seignobos. 1898. Introduction Aux études Historiques. Paris: Hachette.
Marcus, Gary, and Ernest Davis. 2019. Rebooting AI: Building Artificial Intelligence We Can Trust. New York, NY, USA: Pantheon.
Pennell, Catriona, and Filipe Ribeiro de Meneses, eds. 2019. A World at War, 1911–1949: Explorations in the Cultural History of War. History of Warfare 124. Leiden/Boston: Brill. https://doi.org/10.1163/9789004393547.
Russell, Bertrand. 1948. Human Knowledge: Its Scope and Limits. New York, NY, USA: Simon; Schuster.
Stachowiak, Herbert. 1973. Allgemeine Modelltheorie. Wien, New York: Springer.
Thaller, Manfred. 2018. “Decoding What the Sender Did Not Want to Transmit. Information Technology and Historical Data; or Something.” In Proceedings of the Workshop on Computational Methods in the Humanities (COMHUM 2018), edited by Michael Piotrowski. CEUR Workshop Proceedings. http://ceur-ws.org/Vol-2314/abstract3.pdf.
Topolski, Jerzy. (1973) 1976. Methodology of History. Translated by Olgierd Wojtasiewicz. Synthese Library 88. Dordrecht: Reidel. https://doi.org/10.1007/978-94-010-1123-5.
———. 1997. Narrare La Storia: Nuovi Principi Di Metodologia Storica. Edited by Raffaello Righini. Milano: Mondadori.
White, Hayden. 1978. Tropics of Discourse: Essays in Cultural Criticism. Baltimore, MD: Johns Hopkins University Press.
Winks, Robin W., ed. 1969. The Historian as Detective: Essays on Evidence. New York, NY: Harper & Row.
See Hewitson (2014) summarizing Evans (1997): “‘A historical fact is something that happened in history and can be verified as such through the traces history has left behind’, he writes in his revision of Carr’s What is History? and his reply to ‘post-modernism’, In Defence of History (1997): ‘Whether or not an historian has actually carried out the act of verification is irrelevant to its factuality’. By contrast, ‘evidence’ – ‘that is, facts used in support of an argument’ – does depend on explicit deployment by scholars: ‘here, theory and interpretation do indeed play a constitutive role’. However, the theoretical and interpretative decisions made by historians concern the treatment of facts (the light in which they are viewed and the means by which they are linked to each other), not their selection in the first place: ‘historians are seldom if ever interested in discrete facts entirely for their own sake; they have almost always been concerned with what Ranke called the “interconnectedness” of these facts’.” (Hewitson 2014, 101) (emphasis mine)↩︎
We understand causality in a broad sense here, as “any kind of knowledge about how the world changes over time” (Marcus and Davis 2019, 166). Obviously, we reject narrativist and postmodernist views of historiography, which claim that history is nothing but “verbal fiction, the content of which are as much invented as found and the forms of which have more in common with their counterparts in literature than they have with those in the sciences” (White 1978, 82). This doesn’t mean we take a determinist view or postulate the existence of “laws of history” analogous to the laws of physics; we rather agree with Hewitson (2014), who argues that “the identification and evaluation of causes, explaining why one set of events or state of affairs came into being and not another, is a – perhaps the – fundamental task of historians” (Hewitson 2014, 2).↩︎
Despite Hitler famously claiming: “Seit 5:45 Uhr wird jetzt zurückgeschossen!”↩︎