The History of Ideas is presently enjoying a certain renaissance after a long period of disrepute. Increasing quantities of digitally available historical texts and the availability of computational tools for the exploration of such masses of sources, it is suggested, can be of invaluable help to historians of ideas. The question is: how exactly? In this paper, we argue that a computational history of ideas is possible if the following two conditions are satisfied: (i) Sound Method. A computational history of ideas must be built upon a sound theoretical foundation for its methodology, and the only such foundation is given by the use of models, i.e., fully explicit and revisable interpretive frameworks or networks of concepts developed by the historians of ideas themselves. (ii) Data Organisation. Interpretive models in our sense must be seen as topic-specific knowledge organisation systems (KOS) implementable (i.e. formalisable) as e.g. computer science ontologies. We thus require historians of ideas to provide explicitly structured semantic framing of domain knowledge before investigating texts computationally, and to constantly re-input findings from the interpretive point of view. In this way, a computational history of ideas maximally profits from computer methods while also keeping humanities experts in the loop. We elucidate our proposal with reference to a model of the notion of axiomatic science in 18th-19th century Europe.
There is a conceptual, interdisciplinary approach to intellectual history, according to which it is the historian’s
task to trace the history of “certain ‘unit-ideas’ as they find expression in a wide range of cultural fields from
philosophic systems to literature, the other arts, the sciences, and social thought”
There is more than a passing indication that one of the reasons why ‘big history’ studies are witnessing a new
legitimacy is the growing availability of massive amounts of digital textual sources, as well as the growing
availability of computational tools and best practices for the exploration of such masses of sources
As practitioners in the field ourselves, we consider these new developments extremely valuable and exciting. But to what extent are they suitable to enact a real computational turn in the history of ideas and related fields? We say that there is an important sense in which the particular kind of computational practices mentioned above, however valuable, contrasts strikingly with the reality of research in the history of ideas and related fields such as history of science. Studies such as Michel et al. 2011 apply generic computer methods, simple n-gram analyses (i.e. detection of sequences of a number n of words), and shallow Natural Language Processing (NLP) tools to historical textual material. Yet researchers in the history of (philosophical or scientific) ideas typically apply painstakingly finegrained analyses to diverse textual material of extremely high conceptual density. The contrast between computational practice and scholarly reality in the humanities forms thus an easy target for skeptics towards digital humanities. Skeptics will see two problems.
Unorganised Data. Having generic, simple and shallow bottom-up analysis of a massive amount of diverse, ‘long’ and complex data cannot help in reconstructing the history of an idea such as science through centuries. Feeding a computer masses of diverse and complex texts can only yield masses of unorganized details, at least in this field.
Faulty Method. Even if it were possible to make sense of such data within a computational enterprise,
there are fundamental problems with the very method of the history of ideas. The notion of one idea
traceable through centuries of thought is illusory: for ideas cannot be studied in isolation from their context,
and their meaning is in constant flux
Can these two problems be overcome? In other words, is creating the new field of the computational history of ideas possible? Most of all, what should its method be? Our question is not whether it is valuable to apply distant reading in order to trace ‘big ideas’ (it is), but whether there is a way to scale up the mix of close reading, conceptual analysis and historical contextualisation typical of the historian of ideas. We maintain that what is needed to create a computational history of ideas is certainly more than distant reading, and in this paper we lay out what we take that more to be.
In a nutshell, we argue that a computational history of ideas is possible on condition that the two problems above are tackled as follows.
Sound Method. A computational history of ideas must be built upon a sound theoretical foundation for its methodology, and the only such foundation is given by the use of models, i.e., fully explicit and revisable interpretive frameworks or networks of concepts developed by the historians of ideas themselves. This requires construing ideas as complex relational frameworks (models) that combine both stable parts (continuities) and variable parts (discontinuities) (Betti 2014: 14; Betti and van den Berg 2014.)
Data Organisation. Interpretive models in our sense must be seen as topic-specific knowledge organisation systems (KOS) implementable (i.e. formalisable) as e.g. computer science ontologies. We thus require historians of ideas to provide explicitly structured semantic framing of domain knowledge before investigating texts computationally, and to constantly re-input findings from the interpretive point of view. In this way, a computational history of ideas maximally profits from computer methods while also keeping humanities experts in the loop.
Here below we elucidate our proposal with reference to a model (in our sense) of the notion of axiomatic science in 18th-19th century Europe. In the first two sections we discuss and give a solution to the problem of Faulty Method. Here we draw on our previous Betti and van den Berg 2014 to make the present paper self-contained for a digital humanities audience.1 In the rest of the paper we tackle the problem of Unorganised Data and propose a way-out.
Three remarks before we start are in order. First, the two problems just mentioned are both methodological, and
importantly related, for without solving the second (Faulty Method) there can be no hope of solving the first
(Unorganised Data);2 still, while the second problem is more of an ideological one, and requires no reflection on
new computational tools to be tackled, the first one does require such reflection. Second, our proposal as to the
second problem comes with a neutral, non-ideological philosophical stance on what ideas are: we are neutral on the
ontological or metaphysical nature of ideas, and we take this neutrality to be a fundamental presupposition for a
successful and workable operational proposal. Third, we should say right away that the word ‘model’ might be
confusing to the reader, since in computational contexts what is meant by ‘model’ is normally a computational
model. What we mean by ‘models’ are instead non-computational conceptualisations within a (pure,
noncomputational) humanities domain; in this sense, models are technical artefacts. The reader might most fruitfully
identify our models with very complex, articulated working definitions of a concept, conceptual tools that are
provided by historians of ideas themselves in order to, say, ‘chart their landscape’ to their own interpretive goals. In
principle, models as we intend them are (to be) used and constructed within history of ideas without any reference to
a computational enterprise.3 This said, in this paper our point is specifically the following. The only kind of history
of ideas that can successfully take a computational turn must rely on models in our sense for its own methodology; if
historians of ideas wish to take a computational turn, they must turn the knowledge they possess about the ideas they
want to study into models of those ideas in our sense before any computational exploration begins.4
1 Portions of section 2 and 3 have been previously published in Betti, Arianna & van den Berg, Hein 2014, Modelling The History of Ideas,
British Journal for the History of Philosophy 22 (4), 812-835. Published online 28 Aug 2014, copyright © BSHP. Reprinted by permission of
the publisher (Taylor & Francis Ltd, www.tandfonline.com, on behalf of BSHP). The authors paid 64.80 USD to reuse material of their
original article.
2 This point in particular distinguishes our approach from the one assumed in
Lovejoy characterized the history of ideas as being concerned with unit-ideas, simple entities that retain their
meaning through time. The historian must isolate certain unit-ideas (e.g. the idea of truth) and trace their history in
various contexts, that is, periods and intellectual settings
Holism. The meaning of concepts is contextual and always in flux. Hence, the notion of one unit-idea traceable through time is illusory.
Scope. It is difficult to define the scope of unit-ideas. If, for example, we want to write a history of the concept of force in the modern period, we must define it in a way that allows us to attribute this concept to both Descartes and Newton. However, a narrowly Newtonian concept of force cannot be attributed to Descartes, and if we define the concept too broadly we will miss important differences between historical actors.
Arbitrariness and biases. Lovejoy’s method is (i) arbitrary, because it abstracts from the intentions of authors and hence does not yield a proper understanding of historical texts, and (ii) biased, because it interprets these texts by necessarily applying preconceived schemas.
Defenders of Lovejoy have formulated promising counterarguments against (b) and (c). Defenders, in particular
Jouni-Matti Kuukkanen, give up the tenet that ideas are simple: ideas can change exactly because they are complex.
Ideas have parts (or subconcepts), and what changes are an idea’s parts. Some parts of an idea belong to its stable
core (the features of the concepts that persist over some amount of time), whereas some other parts belong to its
variable margins, its context-specific features
However, Kuukkanen’s talk of core and margins of a concept cannot counter (a) and (d) (these objections are due in particular to holist theorists and to Quentin Skinner). It is also unclear how the core/margin terminology should be implemented. In the next section, we expound an implementable methodology for studying the history of ideas which is able to also counter adequately both the holists’ (a) and Skinner’s objection (d). 3 A Sound Method: Neolovejoyanism, or The Model Approach to the History of Ideas
How should we specify the core and the margins of a concept? To do this, we say, we must represent many (shifting) relations among the parts of complex ideas, which will often form a network. And we say that such representations are best accomplished by using (what we call) models: explicitly stated conceptual frameworks or perspectives developed with the overt intention to achieve insightful interpretive goals. Models are abstract relational structures or networks of (sub)concepts.5 In this section we propose a method for the history of ideas based on models. In our ‘model approach to the history of ideas’ concepts or ideas are construed as (parts of) models.
To illustrate our model approach we will discuss de Jong and Betti’s Classical Model of Science (CMS). The CMS is articulated in seven conditions that capture a concept of proper science (according to an axiomatic view of science) adopted throughout history by different philosophers and scientists. These conditions, for any system S of propositions and concepts (terms), are the following: (1) (2a) (2b) (3a) (3b) (4) (5) (6) (7)
All other concepts (or terms) occurring in S are composed of (or are definable from) these fundamental concepts (or terms).
All other propositions of S follow from or are grounded in (or are provable or demonstrable from) these fundamental propositions.
All propositions of S are known to be true. A non-fundamental proposition is known to be true through its proof in S.
All concepts or terms of S are adequately known. A non-fundamental concept is adequately known through its composition (or definition). (de Jong and Betti 2010: 186, our emphasis)
Conditions (1)-(7) are very abstract. Together they are meant to capture the core of the concept of proper science
according to the axiomatic ideal, the basics of an idea held by multiple thinkers in various periods. The concept of
proper science articulated here is also highly complex: its core is described as a relational structure of seven
conditions containing, in turn, many (sub)concepts. In terms of the core/margins distinction, De Jong & Betti’s
claim is that whereas many thinkers adhered to the core of the concept of proper science as fixed by (1-7), this
concept may differ in different thinkers in its variable margins (terms in italics), that is, authors may apply or
5 ‘Any well-articulated idea would qualify as a model of its subject’
We said that in applying the CMS to individual thinkers by specifying conditions (1-7) as to their margins, we are able to contextualize it. In this way we can maintain that various authors, say Bolzano and Kant, have adhered to the same conception of proper science and still allow for variation. To make this clearer, consider the following fragment of CMS in its most stripped down version, one that reveals only its core: (1) […]
(3a) Some xs in S are fundamental (call them Fxs). […] Whereas in bold you see features of the core, which is stable, the variables x, y, which are placeholders, indicate the margins of the idea of axiomatic science, i.e. elements that can be determined or specified differently. Take, for example, condition (3a). Whereas Kant thought that a science should contain fundamental judgments, (so Kant’s ‘xs’ are Urtheile) Bolzano thought that a science should contain fundamental propositions-in-themselves (so Bolzano’s ‘xs’ are Sätze an sich). This difference concerns the margins of the concept of proper science. Insofar as both Kant and Bolzano still adhered to conditions (1)-(7), which specify the core of the concept of proper science, they can nevertheless be said to adhere to the same concept of proper science. In this way, we can trace both continuity and differences in the history of thought.
The model approach can counter Skinner’s arbitrariness & biases objections (objection (d) in the previous section) in the following way. De Jong and Betti maintain that the concept of science according to (1)-(7) is accepted by the historical actors they study, and that no other condition plays a more or equally significant role in shaping the notion of proper science. In addition, they claim that when historians investigate how the model is determined by historical authors, they describe the author’s intentions. Models thus allow one to formulate empirical research hypotheses that are open to falsification: Skinner’s arbitrariness objection is thus met. Skinner’s biases objection is met instead by maintaining that the best way to avoid biases in the history of ideas is to make them explicit. Models allow just that. Models such as CMS simply are interpretative biases made explicit: whoever employs models like the CMS to reveal (dis)continuities in a certain concept, makes her interpretive biases fully explicit and revisable. By explicating presuppositions in models, and by allowing models to be revised if empirical research shows them to be falsely applied, we minimize the risk of interpretative biases.
The model approach can counter the holist objection (a) in the following way. Recall that the holist maintains that concepts change meaning across contexts. We agree, and think this accords with our method. When we work with a model such as the CMS, we reconstruct specifications of some abstract framework that are at most similar to each other, similar with respect to, indeed (their function as fixed by) the CMS. Such is the case with Bolzano specification of the CMS, who, as we have seen, took proper sciences to consists of fundamental propositions, versus Kant’s specification of the CMS, who took proper sciences to consist of fundamental judgments. Hence, Kant and Bolzano ultimately had a different concept of proper sciences. But as to the latter point in particular we can say that both Bolzano’s fundamental propositions and Kant’s fundamental judgments have the same function (or purpose) and the same place in a complex network of concepts that is explicated by the CMS. There is thus a
significant amount of continuity in their views on proper science, and this continuity can be adequately traced by means of CMS.6
We have come to a solution of the problem of Faulty Method: to do research in the history of ideas in a sound way, you should employ models. If no such model is at hand for the concept(s) you wish to investigate, you should build your own.7 So far, we have not ventured beyond the inner methodological workings of the discipline of history of ideas itself. In the next section we approach the question of how historians of ideas can best take the computational turn in their discipline.
To our mind, the model approach as described above is not only the best defense of Lovejoy’s method known to us, it is also the most appropriate foundation to do computational research in the history of ideas. We substantiate our claim here below. First we review the current state of affairs in the methodology of text-based digital humanities and detect problems; then we elucidate how the use of models solves them.
To the best of our knowledge, text mining projects in (text-based) digital humanities tend to typically follow this procedure: (1) humanities experts indicate text corpus; (2) computational experts apply tools to the corpus bottomup; (3) humanities experts interpret the results. Now, it is far from our intentions to say that this procedure is always fruitless – especially because our group currently runs experiments according to that procedure, too;8 what we intend to say is this. Unless certain interpretive choices that are in fact - and inevitably so – at play in (1-3), and which are kept implicit in the process of applying the tools, are either made explicit, or else critically assessed, and built in the computational tools themselves, this procedure, as it stands, will deliver unsatisfactory results from the point of view of research in the history of ideas. That is, the procedure will not be able to scale up the historian of ideas’ traditional method – namely a combination of elements of close reading, historical contextualisation and conceptual analysis in an interdisciplinary, cooperative and non-departmental setting (cf. Lovejoy 1940).
The core risk is, so to speak, one of interpretive self-deception, for steps (1)-(3) are too dissimilar from how
humans read and interpret texts to be apt for massive data enterprises in the field we are concerned with. Research in
cognitive psychology and psychology of teaching since the seventies has shown that comprehending texts involve an
interaction between sensory input (such as recognition of words) and prior contextual knowledge. Such contextual
knowledge is a prerequisite for comprehending prose passages: it is an interpretive key without which texts are
6 One might want to push the point further, and end up saying that neither the core of the concept of proper science is strictly speaking the
same in two different authors. We intentionally disregard this point. We also intentionally bypass all sorts of philosophical discussions of
what meaning, ideas, the relation between abstract objects and their specifications ‘really’ are, whether ideas or concepts are literally
composed of parts, and so on. As stated in our introduction, our model approach is a theory-free methodology that is deliberately neutral on
such issues. As a result, it is compatible with e.g. virtually all positions on concepts. The literature on such topics in the philosophy of
language is immensely rich, and a variety of choices is on offer in case one wishes to defend any particular doctrine in this respect.
Philosophers of nominalistic persuasion might e.g. find it useful to associate our idea of ‘serving the same purpose’ with a broadly Quinean
view on language, meaning and translation, cf.
Decock and Douven 2011) . 7 Models for other concepts are available e. g. the model of fact as fixed in chapter 1 of Betti 2015, expressly built to trace the concept of fact in two millennia of philosophy. 8 See van Wierst, Vrijenhoek, Schlobach & Betti (2016), this volume.
meaningless. 9 There is no such thing as first bottom-up-reading then top-down-interpreting. All reading is interpretation-laden. Applying these findings to computational text mining for humanities research in the field we are concerned with, and to the aims specified in this paper, we can say: a text mining procedure that first turns a massive corpus into data with computational tools, and then lets humanities scholars interpret the results, is bound to be either unsatisfactory, or illusory.
In particular, the procedure shifts the traditional problem of interpretation (the humanities´ question: how to interpret these texts?) from texts to (derivative) data (how to interpret this data?). Far from helping with addressing the traditional problem of interpretation, the shift from texts to data in the three-step procedure above might magnify it in at least three ways: by giving us more bias, more complexity, and making corpus selection more challenging.
I. More bias. Humanities experts tend to have little knowledge of the technology applied, and to be little involved in tool development or calibration. As a consequence, in such cases humanities experts are unaware of the extent to which the resulting data extraction might be technologically biased, even when shallow text mining tools are applied. That is, the tools applied might have been designed or calibrated in ways that might risk to jeopardize certain interpretations (think, for a trivial case, of a text mining tool designed to standardly remove spelling variants applied in the context of a humanities investigation in which such variants do, in fact, matter). Biases in data extraction are new in the humanities, and as such they are bound to be exponentially more difficult to spot and handle than the biases in text interpretation humanities scholars are familiar with.10
II. More complexity. The humanities are concerned with complex texts, i. e. texts licensing a large variety of interpretations. In the field of the authors of this paper, the history and philosophy of sciences such as biology, logic and mathematics, the texts involved are of exceeding conceptual complexity. Generation after generation of philosophy freshmen struggle to understand texts of such complexity, coming back years after years to the same passages again and again, to only discover to their amazement yet a new layer of interpretation. The reason why researchers in a field such as this tends to stick to the study of one period, one author, and even a small cluster of the author’s texts all their academic life, is that not only the answers to one’s interpretive questions are unknown, but that even the right interpretive questions to be asked are unknown. In fact, in these fields, choosing the right interpretive question in a certain context forms a fundamental part of the research. The reason is that the interpretive questions to be asked depend on vast and complex contextual information, often of a highly technical nature, such as
10 An anonymous referee has pointed out to us that work such as Moretti’s distant reading approach is supposed to reduce biases. Let us, to
keep things simple, disregard the point that computational tools might themselves introduce biases: even so, one should realise that work
such as Moretti’s takes place under interpretive conditions which are implicit but heavily infused with theory. Take Moretti’s work on the
British novel 1740-1850
Copyright held by the author(s).
theories of plant nutrition in the 18th
III. More challenging corpus selection. Suppose that a historian of ideas wants to investigate the transformations in the concept of axiomatic science in two-hundred years of the history of human thought as represented in texts. In principle, researchers should consider everything relevant to their research question. What is this ‘everything relevant’? WorldCat counts 17,843,437 books published between 1700 and 1900. Imagine to have solved all sorts of digitisation problems and actually have this corpus: a large-scale interdisciplinary, multi-script, multi-language corpus of extreme conceptual density and diversity.11 This is a historian of ideas’ dream: it’s like a demographer having maximally complete datasets for all populations, the ideal body of evidence to argue for or against interpretive hypotheses. But the task of doing anything sensible with it is daunting. If we first apply computational tools to this corpus in order to extract data, and then set out to interpret the data, we will be faced with an impossible task, for this mass of data so extracted will be well-nigh impossible to interpret – it’ll be pure noise. One solution seems to be to restrict the corpus. Researchers working with traditional methods have developed a quite effective selection method, one that enables them to focus on a corpus of only one to several works by one author – exactly the amount a skilled researcher can study in three-four years. But traditional corpus selection is strongly biased: peer criteria of an author’s ‘importance’ are implicitly endorsed which are infused with – again, implicit – interpretation. Progress and knowledge discovery is often achieved by researchers that challenge corpus selection biases by including little known or forgotten (albeit demonstrably influential) sources in the corpus for new and deeper insight. The computational turn in the humanities gives us finally the opportunity of getting rid of biased corpus selection and to operate with a universal corpus: in our thought-experiment we would consider everything. But applying the three-step procedure above won’t work in this situation. What we would need is a new principle of selection for universal corpora. How shall we proceed?
We maintain that by employing our model approach as a foundation to text mining in humanities fields
concerned with tracing ideas through texts, all the problems mentioned in the previous section are solved (or kept to
a minimum). We propose to ground a computational history of ideas on an ontology-based information extraction
procedure where the construction of initial ontologies is guided by and reflects the stable parts of models developed
by humanities experts. The (looping) procedure we recommend will roughly look as follows: (1) humanities experts
provide computational experts with a model, i.e. explicitly structured, shared (or shareable) - though not formal
semantic framing of domain knowledge about a certain concept; (2) the computational experts turn the core (stable
parts) of the model into an ontology (initial ontology), and adapt techniques of ontology extraction to the domain
and the corpus, all of this in close collaboration with the humanities experts
results obtained by traditional research. The latter results will form a gold standard to evaluate the tools. When the tools deliver satisfactory results on known findings, they are applied to research questions that are still open.
This procedure avoids the problem of adding an extra layer of complexity to the sources as it builds a specific interpretation in the computational tools explicitly and right away. The problem of interpretation remains thus at the level of the texts themselves, instead of being pushed to a new level. Or better: the procedure is a novel computational version of the traditional method: it scales it up but keeps as many as possible of its advantages. The procedure also solves the problem of corpus selection insofar as the selection is afforded by the framing of knowledge given by models: if computational tools are applied that are built upon models, the universal corpus is filtered in a thematically relevant way which is automatic, explicit and thus revisable, as well as, crucially, guided by domain experts. Finally, the procedure minimizes the problem of added technological biases, because the tools are built by humanities experts in close cooperation with computational experts, and constantly refined and monitored against a gold standard obtained with a traditional, non-computational method. By proceeding in the way sketched, a computational history of ideas maximally profits from computer methods, reduces biases to a minimum by making them explicit, and makes a case for keeping humanities experts in the loop.
It should be stressed that from a computational point of view this procedure looks rather natural, because models
in our sense can be seen as topic-specific knowledge organisation systems (KOS) similar to ontologies. However,
our proposal differs from ontology-approaches in the practice in computer science in at least two ways: first, the
nearest computational equivalent of models, ontologies, are usually built by computer engineers; second, they are
explicit and shared, like our models, but unlike the latter, they are formal constructs
Our Neolovejoyan model approach as described above improves on Lovejoy’s method and later improvements in
three ways: it abandons Lovejoy’s view of simple unit-ideas, addresses the problem of biases effectively, and it
integrates the objection made against Lovejoy by rival views (holism, Cambridge School) that the context in which
ideas appear is important. It is also a more general approach than Lovejoy’s because it is able to detect continuity of
ideas even when ideas have nothing else in common than their place in a conceptual structure serving as context
We maintain that Neolovejoyanism can be used as a foundation for a computational history of ideas for the following reasons. Models are complex and revisable frameworks that are able to capture shifting perspectives on the development of concepts. They can guide the construction of appropriate computer science ontologies because models frame knowledge in a similar way, that is, in an explicit and shared (or shareable), though non-formal, way. Models can thus inform the construction of computer tools that, when applied to huge masses of texts, are able to aid the extraction of relevant information in a way that approximates what a human researcher in the domain would do. Since they are not formal, humanities experts themselves can be expected to possess the skills to build models to frame adequately the domain knowledge they possess.
LOVEJOY, Arthur O., 1940, Reflections on the History of Ideas. Journal of the History of Ideas. 1940. Vol. 1, no. 1/4, p. 3–23.
Arianna Betti is Professor of Philosophy of Language at the University of Amsterdam. After studying historical and
systematic aspects of ideas such as axiom, truth and fact