C’è lì un posto, lo ha lasciato tuo padre.
Non dovrai che restare sul ponte
e guardare le altre navi passare
le più piccole dirigile al fiume
le più grandi sanno già dove andare.
Fabrizio De André, La canzone del padre
1. Introduction
Philosophical and scientific ideas can manifest and be instantiated in multifarious forms, from individual intuitions to social practices. A core approach to investigating the evolution and unfolding of these ideas, at least during the early modern period, is also to look at how they manifest themselves through language and, more specifically, in written discourse. It is almost a reflex for a historian wishing to learn about what Newton thought to go and read what Newton wrote. As we enter the arena of discourse analysis, and particularly that of written discourse analysis, it becomes apparent that discourse looks more like a dense polyphonic structure than a sequence of otherwise disconnected solo performances. Any linguistic articulation is inevitably part of a conversation, a dialogue, which is open on all fronts to a potentially infinite number of interlocutors. Reading Newton only does not yield a complete picture of his thought. One would also need to understand why he was thinking what he was thinking (and writing), which sort of problems were more urgent for him, and which ones might have been more remote, if not entirely inconceivable.
To deepen such an investigation, the natural move is to read more texts, possibly somehow connected in relevant ways with Newton’s work, so as to better understand the context of Newton’s ideas. This is a fairly intuitive and commonsensical understanding of what a context is. Sometimes, the notion of context can be taken in a much broader sense—to include anything that was happening around (and arguably somehow influencing) the author or work that is the focal point of one’s research. In a narrower sense, though, as we look at the ways in which ideas are embodied by texts, the context we look for is a certain set of other texts that can be more or less relevantly connected with the main source (both the criteria for relevance and their weight can be the object of further discussions). By taking this broader contextualist view, we can see, for instance, which kinds of ideas were more widespread and thus less unique to a single author, which were instead more idiosyncratically dispersed through the context, or, taking time into account, how these distributions and relations evolved during a given period of time.
However, as rewarding as this contextualist investigation may promise to be, it is not obvious that historians of early modern philosophy and science are willing to wholeheartedly embark on it. Reconstructing the context of any source might become an endless task, given the potentially indefinite scope of the context itself, which can always be expanded by taking into account further relations and more remote sources. This may be just one of the many pragmatic reasons why, especially in the Anglophone scholarship since the second world war (see discussion in Beiser Reference Beiser, Cappelen and Hawthorne2016), a prominent trend in the field has favoured the genre of “rational reconstruction,” in which the historian tends to rebuild the views and ideas of their source based on criteria of internal logical consistency, and as they should have been presented and articulated to make them relevant for today’s audience. This approach dramatically cuts down the need to embark on a painstaking exploration of historical contexts, but at the cost of missing perhaps one of the greatest rewards of historical research proper, which, in Richard Rorty’s words, “helps us to recognize that there have been different forms of intellectual life than ours” (Rorty Reference Rorty, Rorty, Schneewind and Skinner1984, 51).
The current debate on the methodology of the history of philosophy and science, however, tends to favor the need to take the historical context into serious consideration and avoid a purely anachronistic exercise of rational reconstruction (Roux Reference Roux2011, 224–230; Smith Reference Smith, Laerke, Smith and Schliesser2013; Vermeir Reference Vermeir, Laerke, Smith and Schliesser2013; Laerke Reference Laerke, Laerke, Smith and Schliesser2013 and Reference Laerke2015; Garber Reference Garber2015). It has even been argued that contextualism has “won the fight” against rival approaches like pure rational reconstruction (Mercer Reference Mercer2019). This latter claim may sound exaggerated, given that the flow of publications aimed at providing seemingly rational reconstructions of their sources has not stopped in leading journals and with influential presses. To reconcile this point, one might argue that today’s historians seem to endorse as normative a demand concerning the need to take the historical context into serious consideration, as a fundamental prerequisite for any engagement with historical sources. In practice, however, it has been noted that rational and contextual reconstructions can (and perhaps should) be blended to various degrees, depending on researchers’ aims (Beiser Reference Beiser, Cappelen and Hawthorne2016), and this might provide a justification for keeping the contextualist demand at its bare minimum in many concrete cases. While accounting for the simultaneous survival of the genre of rational reconstruction in a climate that, methodologically and normatively speaking, seems more and more inclined toward contextualist research, this sort of compromise disguises a more fundamental pragmatic problem with the contextualist approach itself.
Contextualism requires making two decisions: first, what the primary focus is and, second, how to pick its relevant context. Neither of these decisions is straightforward. Historiographical canons are created precisely to facilitate these decisions. The canon provides both a list of authors and works that are worth studying as the main focus of attention and a number of other authors that could be used to put each item in some context. Nevertheless, the past few decades of historical research have made it abundantly clear that canons provide a context that is too selective and ultimately non-representative for a proper contextualist understanding of even canonical authors.Footnote 1 Much recent research is motivated by a commitment to expanding the existing canon (e.g., Shapiro Reference Shapiro2016; Beaney Reference Beaney2018; Sangiacomo et al. Reference Sangiacomo, Tanasescu, Donker and Hogenbirk2021a) and introducing new authors and works (e.g., O’Neill Reference O’Neil and Janet1998, Mercer Reference Mercer2017). But with this growth in the number of sources available, it is much less obvious why a scholar should focus on this or that particular author as the main focus of their research. In other words, if one leaves the reassuring narrowness of canons behind, then there are plenty of possibilities to explore, and perhaps even too many possibilities to decide on. Moreover, canons offer preestablished patterns of recognition, which provide not only significant exemplars of past ways of articulating ideas, but also normative paradigms about what, how, and who should take part in the game of ideas. An important political dimension behind the recent dissatisfaction with established canons lies precisely in the way in which the patterns they establish are systematically biased against certain groups (women or non-Europeans, for instance). However, simply expanding canons or creating alternative ones will not suffice. A canon is effective in providing a proxy for contextualist research precisely insofar as it is limited and thus helps the researcher not to sink in the endlessness of potential sources. Alternative (but somehow still limited) canons aimed at the recognition of neglected figures might remedy certain specific biases, but they offer no guarantee of not introducing new ones.
What this overview suggests is that today’s historians of philosophy and science are torn between, on the one hand, a normative methodological desideratum that urges them to take historical contexts into serious consideration for the examination of any source, and, on the other hand, the practical difficulty of implementing this demand, either because of the factual limitations of existing canons or because of the difficulty of just exploring the mare magnum of potentially relevant sources.
In this paper, we argue that a wise integration of computational approaches with more traditional methods can contribute to alleviating this problem. A simple but powerful first step in this direction consists in looking at contexts as networks, or structures of related nodes (texts). The use of network analysis for the study of past ideas and documents is receiving growing attention (Davis et al. Reference Davis, Gardner and Gardner1941; Kerschbaumer et al. Reference Kerschbaumer, Linda von Keyserlingk and Düring2020; Valleriani et al. Reference Valleriani, Federau and Nicolaeva2022), although its implementation in the fields of the history of early modern philosophy and science is still scarce (see exceptions in Bourke Reference Bourke2017; Sangiacomo and Beers Reference Sangiacomo and Beers2020; Rossini Reference Rossini2022). This move helps us not only to operationalize the basic demand of looking for the context of a given text (namely, all the other texts that might be related to it based on certain criteria), but also the idea that networks (like contexts) are patterned by the dynamics of recognition, conflict, and competition, which profoundly shape their structure, and which constitute one of the most interesting aspects to uncover in historical research.
Bruno Latour’s interpretation of network analysis has emphasized the dynamic nature of networks and how they are constituted by conflict among their actors.Footnote 2 The network is interpreted as an associative unit, which is brought together by the way in which its constituents relate to one another (through links). Yet, this unit is based on conflict (controversy), as a network’s agents acting in common is grounded on the tension between program (a goal, or a course of action) and anti-program (the opposite of a goal or of a course of action). In our case, defending and disseminating a certain approach (say, a “Scholastic” way of dealing with natural philosophy) would constitute a program, while rival attempts (such as those inspired by Cartesian and Newtonian ideas) would count as antiprograms.Footnote 3
However, what is a program and what is an antiprogram “is relative to the chosen observer,” (Akrich and Latour Reference Akrich, Latour, Bijker and Law1992, 261) who, in our case, is the historian of philosophy. In Latour’s view, action within networks may take place at a distance, which makes the unit of action take note of its own heterogeneity—it relocates and distributes (and thus, translates) the action to various places that each have their own features, but share the same objective or program (in our case, natural philosophy or whatever else we want to emphasize). Thus, the boundaries of networks are permeable, open to exploration; they are the expression of collectives’ needs to broaden their unity based on the tension between program and anti-program. Exploration is the way the collective moves through new associations. There are no pre-established associations, or groups. It is the way such associations are formed, the way groups are born, that matters. In our case, thinking about contexts in terms of network analysis helps us both to lay the foundation for a quantitative study of historical contexts and to emphasize how any such context will be characterized by divides, oppositions, or by what we shall henceforth call “rifts” between different trends and tendencies.
The main goal of this paper is to present a computational method that can be used to “parse”—that is, to divide and analyze—a large corpus of works based on their use of language, alongside some preliminary information that can be retrieved from metadata. The goal of the method is to use the available information about the corpus to create broad groups that can work as sub-contexts for better understanding different sorts of works and authors. In turn, this makes it possible to better profile each group and identify its most distinguishing linguistic features, by thus signaling where the potential rifts between them lie. Once these features are clarified, it will eventually become possible to also identify what the most representative works and authors are in each group and which of them could be worth exploring in greater detail. Our method seeks to provide a computational, quantitative scaffolding for the preliminary work of discerning, within a broad corpus of otherwise undifferentiated texts, which ones have the potential to be more relevant for a specific research question and provide a more appropriate context of study. It is important to stress once again that this scaffolding works at the preliminary level of research: the integration of computational tools needs to be wise, because computation needs interpretation and it is only when a human reader looks at the results that data become meaningful (and debatable). Part of our interest in this study is thus to show how such a computational method should be implemented, and to map its potential shortcomings.
2. Method
As previously mentioned, a fairly general way of understanding the context of any given source is to think of it as a structure of more or less closely related other texts, which could be deemed relevant for the understanding of the departing source. The methodological challenge consists in operationalizing this rather vague understanding and implementing it in a productive way.
One relatively straightforward solution is to look at how historical texts connect with one another (e.g., Laerke Reference Laerke, Laerke, Smith and Schliesser2013 and Reference Laerke2015; Sangiacomo et al. Reference Sangiacomo, Tanasescu, Donker and Hogenbirk2021b). Correspondences and controversies are perhaps the most extensively studied and discussed cases. But the relative idiosyncrasy of this kind of corpora shows the urgency of extending this model of linkage beyond explicit references. On the one hand, not all historical sources are linked in this explicit fashion, while, on the other hand, historians might be interested in studying connections that hold between texts regardless of whether or not they are explicitly referring to one another. Explicit links are made by authors themselves and are thus based on (and to some extent limited by) their local and historical knowledge of other sources available to them.
But the historian cannot adopt this authorial point of view. Explicit references to authorities might be the result of cultural, social, and political circumstances and do not necessarily entail an actual endorsement of them, nor are early modern authors always committed to full disclosure or even sincerity in their references. More generally, the historian is situated in an entirely different time and place, and has access to sources that might not have been available to the historical actors. In other words, the historian can more easily take a broader, bird’s-eye perspective on the historical landscape, and exploit its heuristic potential to gain insights that might have been unavailable to the historical actors themselves. The methodological challenge is to adopt an approach that could potentially take stock of any preexisting explicit linkage between the sources themselves, while also extending this criterion so as to handle any corpus, regardless of whether its constituents are explicitly linked.
One way of rephrasing this challenge is by considering the envisaged method as a tool for putting counterfactual constraints on the way in which the context of a given text is reconstructed. Instead of determining the most likely connections, a contextualist method could also be used to rule out the most unlikely connections, or rather to parse the corpus by distinguishing different groups of sources within it, and thus indicating where one should look in the search for likely allies or likely foes.Footnote 4 Traditional historians of philosophy and science already tend to do this by using different ideological labels. For instance, calling a certain author a “Scholastic” or a “Cartesian” carries implications for the expectations that one will then nurture about the likely allies or foes of that author. But these categorizations are usually based on the use of a few selected paradigms that are taken to be representative of a whole group of sources. We argue that this qualitative approach can be fruitfully strengthened and combined with a more quantitative approach based on computational semantic network analysis, which provides ways of measuring the proximity between various potentially relevant sources based on preestablished criteria. We suggest that the method illustrated below contributes precisely to facilitating this measurement.
In order to illustrate the workings of our method, we consider a corpus of 124 early modern printed books (containing just over 1 million words), written in Latin. All these books are either textbooks of natural philosophy or they provide equivalently encompassing and systematic treatment of what “natural philosophy” is about. Chronologically, the corpus spans from 1587 (Nicholas Abraham’s Methodicae Institutiones) to 1832 (François Jacquier’s Institutionum Philosophicarum Synopsis). Most of the texts were published in the Dutch Republic, France, and England. All these works are also available in the public domain. Thanks to a specific OCR procedure we developed and implemented (Sangiacomo et al. Reference Sangiacomo, Hogenbirk, Tanasescu, Karaisl and White2022b), we were able to use an enhanced digital transcription of the texts (with a minimum of 90% of word accuracy per page).Footnote 5
Two preliminary worries might need to be dispelled about the choice of this corpus. First, it should be noted our focus on systematic treatises (textbook-like) does not necessarily mean that we are ignoring the most original fronts of the ongoing debate in early modern natural philosophy. While these sorts of work might to some extent be regarded as relatively “conservative,” covering a sufficiently broad collection of them, spread over a sufficiently long period (as we do), should be a sufficient guarantee that any relevant ideas that gained traction at some point will eventually be integrated or discussed somehow in more systematic works as well. The data we present below confirms this expectation.Footnote 6
Second, since textbooks in particular often include natural philosophy among other fields of study, one might wonder what is the relevance of this corpus for natural philosophy proper. On the one hand, this paper is illustrative in nature, in the sense that we aim to use this corpus to illustrate our method for computational parsing, more than for introducing a new interpretation of the evolution of natural philosophy per se. On the other hand, textbooks are in fact relevant for better understanding this evolution, since they do provide one context (the curriculum of disciplines in which natural philosophy was part), and for better framing the place of natural philosophy in the field of knowledges cultivated during the early modern period (Ariew Reference Ariew2014, 41–105).
The results presented in this paper were derived from a corpus containing a flattened map of intellectual movements, with few pre-established expectations on the part of the observers (researchers). In fact, the corpus includes authors and texts that were largely unfamiliar. From a chronological point of view, it is sensible to say that Cartesian and Newtonian scholars were “anti-programs” (to use Latour’s terminology) to Scholastics. However, the computational approach presented here offers hypotheses and results based only on the degree of textual similarity between any two texts, leaving room for further debate about how best to interpret them.Footnote 7 As we explain in the following, our network understanding of the corpus in question offers a different vista on the programs and antiprograms of early modern natural philosophy, one that both confirms and contradicts our initial assumptions.
2.1 Network representation
In order to investigate the corpus as a whole, we need to have a relational representation of it. A network is a very common form of mathematical representation that stresses the relational structure of a given reality (Barabási Reference Barabási2016; Newman Reference Newman2018). Networks are made of two main groups of elements: nodes (vertices) and edges (links). Nodes represent the individual items we want to represent (the books in our corpus as digitized files), while the edges represent the way in which they relate to each other according to specific criteria. A preliminary difficulty we have to overcome is that the inventory of our corpus comes without any consistent information about direct links between works or authors. Therefore, we have to find a way of establishing these links, a common feature that links all of them to a higher or lower extent.
Our approach consists in linking works based on a degree of linguistic similarity (for a more detailed discussion of this methodology see Sangiacomo et al. Reference Sangiacomo, Tanasescu, Donker and Hogenbirk2022a). In other words, we establish a link between two works insofar as they share a common feature in their use of language. This captures a layman’s intuition according to which two books saying similar things in a similar way can be taken to be similar to one another. This similarity might or might not be further examined in terms of historical or intellectual influences among the works under scrutiny. The intuition is relatively easy to operationalize through computational methods, since the field of natural language processing (NLP) offers very well-established methods to compare linguistic similarity between text documents from different perspectives (Aletras and Stevenson Reference Aletras and Stevenson2014; Delamaire et al. Reference Delamaire, Juganaru-Mathieu and Beigbeder2019).
The two most common and widespread techniques are topic modelling (TM) and text frequency-inverse document frequency (tf-idf). Topic modelling is an unsupervised statistical model which identifies a number of abstract topics (or a latent semantic structure) in a large collection of texts. A “topic” is a collection of keywords that tend to be used in the same context together. The Latent Dirichlet Allocation (LDA) algorithm (Blei, Ng, and Jordan Reference Blei, Ng and Jordan2003) generates several topics, with each topic containing a given number of keywords that are relevant for a certain subset of works and less relevant for other subsets. For instance, a topic model containing keywords like pars, corpus, video, pars, moveo, locus is counted as different from a topic formed by keywords like moveo, locus, pars, deus, forma, causa. The two topics share an overlap in certain keywords (e.g., locus, pars), but these keywords have a different weight in each topic, thus differentiating how well each topic applies to any given work. In order to use topics to create links between works, we need to generate relevant lists of keywords for the whole corpus (or subgroups in it, depending on what we study), and then assess how well each of these topics is associated with each book (node) in our network. Any time two works use the same topic, we create a link between them, and depending on the strength of this similarity, we can add a weight (or a value) to the link.Footnote 8 This gives us a way of assessing the similarity between different nodes and relating them (creating an edge between them) based on that similarity expressed as a weighted link by the correlation score between any two document vectors (Sangiacomo et al. Reference Sangiacomo, Tanasescu, Donker and Hogenbirk2022a).Footnote 9
Topic modelling takes into account all the words included in a text. This approach might be too coarse grained for capturing more specific discursive similarities connected with specific philosophical approaches; it is conceptually opaque. It seems likely that all philosophers share a common dictionary of words that are always used in a non-specific and non-sectarian way (like corpus or locus), while they differ in the use of words that are more salient or relevant for articulating a specific approach. In order to capture this aspect and use it to strengthen our similarity analysis, topic modelling is usually best complemented by another computational method, tf-idf (Lee Reference Lee2019). The latter helps us assess how unique or idiosyncratic each work is in its use of language. The tf-idf algorithm devalues the most common words in the corpus and groups together works that use the same specific words. For instance, if two works both use a very common word like video, they will not establish a particularly strong connection (because this is a common word), but if they both use the more specific word substantialis, then they will. This method offers yet another way of linking the books in our corpus, thus resulting in a second network. Since these two networks share the same nodes (books) but have different sorts of edges (generated by the two different computational methods), the resulting combination is best interpreted as a multiplex network, namely, a multi-relational network made of several layers in which the nodes are the same (Lee, Min, and Goh Reference Lee, Min and Goh2015).
Overall, in combining these two methods we observe that the corpus is fairly homogeneous and all works are more or less similar or connected with all the other works. This is entirely expected and corroborates the validity of the methods we used: both topic modelling and tf-idf are employed to assess document clustering and we applied them to a corpus made entirely of systematic treatises on the same topic, written in a relatively fixed time and space. However, the fact that the resulting network is homogeneous does not mean that all works are seen as equivalent to one another. What we have described so far is just the underlying methodology and the preliminary step that allows us now to look more specifically for possible ways of tracing differences and parsing the network.
2.2. A procedure for parsing sources
Similarity is a relational property that is best assessed through comparisons. In order to make comparisons, we need to parcel our network in such a way so as to be able to assess the relative similarity or difference between various books. To do so, we need to find a profitable way of parsing the network (and the corpus) that provides a heuristic point of reference for such comparisons. One obvious way of slicing the corpus is chronological, by grouping works by publication date. However, publication dates are not distributed evenly and relying on them alone runs the risk of creating significant imbalances in how the corpus is divided. More importantly, publication dates might be correlated with different trends and discussions in the subject matter, but this is not necessarily the case, so a chronological parsing of the corpus would tell us very little about the homogeneity or heterogeneity of each chronological group in terms of what they actually discuss.
Instead of taking a chronological approach, we factor in one further piece of information that we can derive from the metadata associated with the inventory of our corpus, namely, from the titles of our works. We observe that while the great majority (87%) has a fairly general title that concerns the discipline of natural philosophy or one of its core topics, a significant 13% include in their titles a direct reference to what we recognize as a canonical authority.Footnote 10 By far, the most instantiated authorities (12% of the total Latin corpus) are Scholastic, among whom Aristotle, Aquinas, and Scotus are the most common. Note that for the purposes of this paper, we do not distinguish the orientation (pro, contra, or other) of the work towards the authority acknowledged in the title, but simply observe that a certain authority is acknowledged and flagged in the title. Our discussion of authority acknowledgments is thus neutral with respect to whether the authors using these acknowledgments intended to argue in favor, against, or in some other way about the authorities they mentioned.
Existing scholarship has stressed that no historical group of philosophers or school was particularly homogeneous (Garber Reference Garber, Robert and Daston2016) and appealing to the same flag or authority allows for a wide range of individual deviation and differentiation. ‘Non-Scholastic’ authors are perhaps even less homogeneous, as they tend to be unified only by a reference to what they aim to oppose (Garber Reference Garber, Garber and Roux2013). From a methodological point of view, our purpose here is not to define what Scholastic or non-Scholastic (Cartesian or Newtonian, for instance) natural philosophy looks like, but rather to derive from the information embedded in our corpus a valid, heuristic way to parse its contents further. We derive this information from explicit mentions in the titles of the works themselves. We then further integrate this information with background information concerning their authors, which we derive from the available bio-bibliographical Dictionaries of seventeenth- and eighteenth-century philosophers.Footnote 11 The integration of this latter information is required in order to create a group that is sufficiently large to generate meaningful computational results.
This allows us to introduce a first operational distinction between Scholastic (45.6% of the total) and non-Scholastic works (the remaining 54.4%). Note that the use of background information derived from existing scholarship (the Dictionaries) increases the group of Scholastic works by 33.6%. However, as documented below, we do not take this figure entirely for granted, since our method will help us corroborate (or falsify) whether authors originally included in the Scholastic group are indeed homogeneous, and whether there are some hybrid or “fuzzy” cases.
We are thus in a position to use our network representation to address the following two research questions:
-
(1) To what extent are the Scholastic and non-Scholastic groups homogeneous?
-
(2) What are the most central works in each group?
Note that both questions are asked from the point of view of the linguistic features of the corpus that we can study with computational methods and hence do not immediately or necessarily concern the philosophical or conceptual nature of the works themselves. But it seems plausible to surmise that a degree of linguistic homogeneity might be positively correlated with a degree of philosophical homogeneity, and linguistic centrality with philosophical centrality (and this hypothesis is falsifiable).
The first question can be used to assess whether and to what extent a certain group constitutes a relatively uniform “context” of works reasonably similar and well-connected with one another. As mentioned, existing scholarship has often stressed the diversity within broader groups, by emphasizing that different Scholastic or Cartesian authors (for instance) might have produced significantly different outputs. But does this mean then that such school labels should be rejected altogether and that they constitute nothing but a conventional umbrella-term for an otherwise heterogeneous set of texts? We argue here that looking at this issue from a linguistic point of view provides a new approach to tackling it.
The second question concerns the problem of selection, or directing attention within a large data set to those particular cases that might be worthier of close scrutiny. In terms of network analysis, this is most commonly done by analyzing centrality measures, namely, various metrics that rank the nodes in a network based on their structural position in the network (i.e., their role in constituting said network). Since centrality is a structural property that results from the way in which the network is shaped, studying centrality measures as a proxy to identify relevant cases for further scrutiny provides a bottom-up approach that can be helpfully integrated with the more top-down attitude of focusing on well-established case studies in existing literature or canons.
In building a network where links between nodes (books) represent similarity relations (either based on TM or on tf-idf), we have two ways of further interpreting these relations. On the one hand, we can simply consider the strength of any connection between any two books, which means how similar they look from the point of view of topic or tf-idf vectors. But the fact that two books are strongly connected (strongly similar) does not entail that they are also central in the network. In the following discussion, we shall take into account two standard centrality measures: eigenvector centrality, and betweenness centrality (Newman Reference Newman2018). In this context, we interpret eigenvector centrality as expressing the likelihood for a work to be reflective of a topic that is represented in multiple other works. Betweenness centrality expresses instead the ability for a node to create links between different usages of certain topics. From these working definitions it should be clear that centrality cannot be conflated with correlation strength (or degree of similarity), since the latter applies to any two books regardless of their structural position in the network, while centrality depends precisely on that structural position. Distinguishing similarity strength and centrality helps provide different perspectives on the overall network. In the most straightforward way, we can use the average similarity strength as an indicator of homogeneity (research question 1) and centrality as an indicator of structural relevance (research question 2).
However, both aspects are tackled from the point of view of similarity measures (or the degree to which works in a certain corpus correlate among each other), which means that our discussion is not necessarily going to spot what traditional scholarship has singled out as the most original and outstanding works or authors. On the contrary, these measures will rather emphasize those works and authors that contributed mostly to creating a shared conversation, bridging different approaches and mediating between programs and anti-programs (to use Latour’s phrasing). To give an example, although Newton’s Principia is included in our corpus and it is surely one of the most important works in early modern natural philosophy, it does not appear at all among our top-scoring results, while a relatively more obscure Dutch Newtonian, like Willem ’s Gravesande (1688-1742), does. We reflect on this methodological point and on its implications as our discussion unfolds.
However, in order to fully address both our research questions, we need to refine our initial grouping. Obviously, the fact that a work does not include a direct reference to a Scholastic authority in its title, or the fact that we do not know enough about it or its author, does not entail that such a work is not similar to other works we flagged as “Scholastic.” This means that our initial parsing based on metadata needs further distillation. To do so, we implement the following three-stage procedure, which will be explained and discussed throughout the following sections.
Stage 1: Analysis of the departing groups. Here, we simply take our two groups (“Scholastics” and “non-Scholastics”), generated on the basis of metadata only, and use network analysis to gain a better understanding of some of their features. We also use this analysis to test what the initial answers to our two research questions would look like if we were to limit our analysis to this stage. Since our network has two layers (one based on topic modelling, henceforth TM, the other based on tf-idf), our analysis will have to proceed on two parallel, but complementary tracks for each group.
Stage 2: Computational parsing. After the preliminary analysis in stage 1, we identify which works in the non-Scholastic group are most similar (based on their network properties) to the works in the Scholastic group. We then move these works to the Scholastic group, leaving the rest in a more distilled non-Scholastic group. This computational parsing process, however, comes with a few unexpected results, which we document below.
Stage 3: Analysis of the updated groups. We repeat the same analysis in stage 1 for the two updated Scholastic and non-Scholastic groups, in order to assess how the computational parsing affected our initial results and what answers to our two research questions we can derive from the revised groupings. In going through this process, we also draw attention to a number of works that emerge as markers for differences and similarities among different groups. We take this last remark as an important, serendipitous finding, which we expand on in our conclusions.
3. Implementation
3.1 Stage 1: Scholastics and non-Scholastics
We begin by splitting the corpus into two groups, Scholastic and non-Scholastic authors, as explained in the previous section. Figure 1 provides a NetworkX visualizationFootnote 12 (spring layout) of the way in which the two groups appear from the point of view of Stage 1. A file label number was automatically assigned to each text in the corpus.Footnote 13
Network of books in the Latin corpus (1587-1832) cf. topic vectors (red = Scholastic authors; grey=non-Scholastic authors).
Starting with TM, we generate four topics that are particularly relevant for the Scholastic group in its initial configuration (i.e., based on direct acknowledgments in the titles, or information derived from the bio-bibliographical Dictionaries). We set the number of relevant topics to four based on a computationally determined coherence value, which establishes the most relevant number of topics for a certain collection of documents (Röder, Both, and Hinneburg Reference Röder, Both and Hinneburg2015). The keywords pertaining to these four purely Scholastic topics are presented in Table 1, while the network configuration of the group is presented in Figure 2.Footnote 14
Topics in the Scholastic flag group (with Topic 2s being the most representative)
Network of Scholastic authors writing in Latin (1587-1832) cf. topic vectors.
The keywords reflect the subject matter of natural philosophy (pars, corpus, ratio, moveo, materia), but also the specific framework that is common in Scholastic philosophy (deus, forma, species, virtus) and the fact that, in this context, natural philosophy is usually dealt with within more encompassing systematic discussions of other subjects as well (hence, keywords like homo, bonus, voluptas, malus). However, not every topic has the same relevance to each work.
We observe that all works in this group are very strongly correlated with one another, all between 0.9999 and 0.9861, out of a maximum of 1.00, meaning that they are reflective of the same topics to the same extent. This suggests that this departing Scholastic group is internally cohesive in terms of language usage. The two most strongly correlated works are Cursus Philosophicus (1656) by Jean Lalemandet (1591–1647) and Philosophia Scoti (1690) by Jean Gabriel Boyvin (1605-1681). Both authors are French Scholastics and neither tends to be considered as particularly original in his own right.Footnote 15
Table 2 plots the top-scoring works in both eigenvector and betweenness centrality in the two network layers.Footnote 16 In this respect, it is worth mentioning that there are only a few titles in each corpus presenting non-trivial betweenness centrality and, therefore, the likelihood to have acted as bridges between certain pairs of works.
Top ranking nodes in eigenvector and betweenness centrality in the topic and tf-idf layers
Top-ranking nodes in terms of eigenvector centrality are almost all seventeenth-century French Scholastics, strongly associated with topic 2s, which is the one that includes almost no keywords directly relevant to the description of natural phenomena as such. We also observe that the top-ranking works (by Boyvin and Columbus) are of Scotist orientation, suggesting that this might be a dominant trend among French Scholastics. In this context, natural philosophy is presented as a parcel of a broader curriculum, most commonly within multi-volume works that also include sections on metaphysics, logic, and ethics. This means that we have a group of French Scholastic works acting as an influential pole in both disseminating and reiterating a certain way of presenting and discussing natural philosophy during the period. If this is correct, then topic 2s might also be used for detecting more Scholastic works in our corpus, as we shall illustrate below.
Betweenness centrality helps us better understand how the cluster of French Scholastics might relate to other parts of the Scholastic group. French Scholastics (Semery, Boyvin, Piny) are still very present among the most high-raking authors in betweenness. But we also find non-French Scholastics, such as Case (one of the few British Scholastic authors) and Burgersdijk (one of the forefathers of Dutch Scholastics), or French anti-Scholastics such as Sébastien Basson (b. 1573), who engaged with Aristotelian natural philosophy, but for the sake of criticizing it (cf. Lüthy Reference Lüthy1997). This reminds us that our initial Scholastic group does not necessarily encompass only authors that advocated Scholastic natural philosophy, but also those who explicitly criticized it, while retaining most of its jargon and way of discussing the subject. In the case of Basson, we can correlate this observation with the fact that the most relevant topic for his discussion is 1s (one of the most relevant for a description of natural phenomena). For Burgersdijk and Case, topic 2s (the flagship topic of French Scholastics) is decisively less relevant. This corroborates the intuition that differences in the relevance of a topic for a particular author might be positively correlated with their role in the overall network and the specificity of their philosophical project.
The tf-idf network layer provides a different, and yet complementary picture about the corpus. French Scholastics appear here on a par with Dutch authors, suggesting that in their manner of writing they were comparatively similar. Combined with the topic analysis mentioned above, this indicates that French Scholastics might have been the most direct source or reference for Dutch authors, despite the religious divide between the two groups. In terms of eigenvector centrality, the top-ranking author is still a French Scholastic, though a later one, namely Claude Frassen (1620-1711), who is “generally considered as the foremost seventeenth-century French Scotist” (in Foisneau 2015, 748), while Lalemandet still ranks sixth. The emerging picture is thus that of a tightly connected and homogeneous group (research question 1), dominated mostly by French Scotists (research question 2). This also confirms that our initial move of creating the Scholastic group by relying on both authority acknowledgements in the titles and information derived from secondary scholarship (Dictionaries) does produce a homogeneous group.
Moving on to the non-Scholastic group of authors, we generated eight other topics that map their works thematically, as shown in Table 3.
Topics in the non-Scholastic group (with Topics 3ns and 8ns being the most representative)
While there is some overlap in keywords among different topics in the Scholastic and non-Scholastic groups, the non-Scholastic topics are overall quite diverse. The list of keywords in the latter group tends toward a stricter natural philosophy lexicon (including elements like corpus, natura, motus, aqua, ignis, or more technical terms like tempus, aequal, radius, ratio, velocitas).
In the TM layer, the first six most correlated works (mostly correlated among themselves) are by eclectic authors with some Scholastic background or influence (Le Clerc, Du Hamel, Cally, Bootius). Immediately after come the works by Dutch Newtonians (Swinden and ‘s Gravesande). This suggests that while Newtonians do cluster together, the rest of the non-Scholastic group tends to be relatively heterogeneous and the most related works are by authors that do not necessarily fit a clear-cut category. This can either reflect a property of the works themselves, or (as we will discuss in step 3 below) be partially due to the fact that our grouping of the “non-Scholastics” needs further refinement.
In the top-scoring works for eigen centrality (Table 4) we can easily discern that Descartes plays a powerful role in the non-Scholastic group. Out of the ten highest ranking works, Descartes and his interlocutors or followers (Geulincx, Clauberg, DeRaey, Bayle) take six spots. The others can be identified as Scholastics.
Top ranking nodes in eigenvector and betweenness centralities in the departing non-scholastic group
Remarkably, we also observe a complete convergence in terms of topic towards 8ns. This needs to be interpreted cautiously, since the same topic is prominent among works conversant with Cartesian ideas and among some Scholastics. We suggest that some of the Scholastic works strongly associated with this topic might be regarded as performing a bridging role, between Cartesian and non-Cartesian conversations.
The betweenness ranking shows a rather heterogeneous group, both in terms of topics and philosophical orientations. A play of brokerage between new ways of conceiving of natural philosophy and more Scholastic forms becomes more apparent. Bacon shows up here as a relevant author, as historians of philosophy and science are accustomed to consider him a forefather of the new anti-Scholastic approach that evolved in the seventeenth century. But we also have works by Gregory (a British Newtonian) and Nicolas-Louis de Lacaille, who was a deacon but mostly devoted to experimental observations in astronomy. This suggests a marked interest in a more autonomous discussion of natural phenomena, which might well have constituted a common territory for the interplay between Scholastic and non-Scholastic interpretations. The presence of Du Hamel and Pourchot in the top-ranking betweenness supports this intuition, since both authors were committed to hybridizing otherwise traditional Scholastic philosophy with new emerging approaches (especially Cartesian). Pourchot wrote a widely used textbook of Scholastic philosophy, but with ample openings to Cartesian and more recent discussions. In natural philosophy, for instance, his book draws from the atomism of Gassendi and defends the Copernican system (in Foisneau 2015, 1426). Du Hamel is also known for his work as “a mediator between supporters and foes of Aristotelian philosophy” (in Foisneau 2015, 601).
As far as the tf-idf layer is concerned, we find the strongest correlation between the works by Musschenbroek (a leading Dutch Newtonian) and Le Clerc (99.99% correlation). In terms of eigenvector ranking, Cartesian authors (Le Grand, Clauberg) are at the top of the score list. We also notice here a certain mélange between Dutch and French authors. By contrast, when we look at betweenness centrality, the top-ranking works are by authors belonging to religious orders, but active or involved in discussing new trends in natural philosophy. These include Antoine Séguy (fl. 1730?) who, like Pourchot, was active in discussing Cartesian natural philosophy, the already mentioned de Lacaille, and François Jacquier (1711-1788), an influential Jesuit commentator of Newton active mostly in Italy and Spain. Another interesting author that emerges is Nicholas Abraham, sieur de La Framboisière (1560–1636), who was a physician singled out in recent scholarship as witnessing “that Aristotelianism was on the way out in early modern medicine, just as it was in the physical sciences of the period” (Shapiro Reference Shapiro2003, 426). So, unlike the eighteenth-century figures mentioned above, Abraham represents a trend similar to that of Basson (i.e., early critics of Scholastic philosophy) and parallel to the more famous case of Descartes.
This preliminary analysis of the non-Scholastic group thus reveals that it is fairly heterogeneous (research question 1): it contains a number of works that could be classified as Cartesian or Newtonian in orientation, but also a variety of important works (from a network point of view), which play a prominent bridging role between Scholastic and non-Scholastic orientations. Those works that stand out as marking the above-mentioned distinctions are precisely the ones that would be worth investigating more closely (research question 2).
3.2. Stage 2: Computational parsing
In order to proceed further with a computational parsing of the groups we constituted during stage 1, we must first study the whole corpus in order to identify which of the works currently tagged as “non-Scholastic” relate more closely to those we tagged as “Scholastic.”
Starting once again with TM, we generate a new list of topics that apply to the whole corpus, as illustrated in Table 5.
Topics in the whole corpus (with topic 8w the most representative marked in dark gray, followed by topics 5w, 3w lighter gray, and 2w light gray)
While there are some overlaps between the keywords in various topics compared to the lists of topics generated in stage 1, we observe that, overall, the keywords included in the topics generated for the whole corpus blend both Scholastic-specific terminology (deus, quaestio, genus, forma) and more technical terminology (pars, corpus, ratio).
Table 6 shows the top-ranking works in eigenvector and betweenness centrality in the whole corpus, plotting them against the most relevant topics (from Table 5). We observe that topic 3w is shared by all top-ranking works in eigenvector centrality, while top-ranking works in betweenness centrality have a more mixed topic profile. Top-ranking works in eigenvector centrality tend also to be connected with Cartesian natural philosophy, while most of the top-ranking works in betweenness centrality have more a Scholastic orientation. Le Clerc is eclectic (and Newtonian) in orientation when it comes to natural philosophy, while Du Hamel, Regius and Geulincx are all conversant with Descartes. The only Scholastic-oriented authors included in the top ten are in fact Du Hamel and Pourchot, who we already encountered as eclectic Scholastics. This suggests that topic 3w can be used to profile potential “non-Scholastic” works.
Top-ranking works in eigenvector and betweenness centrality in the TM layer (whole network)
The case of top-ranking works in betweenness centrality is less clear cut, also because this measure is associated with a bridging activity between different orientations, and thus the use of topics is expected to be more mixed. Nonetheless, some regularity emerges: topic 2w is prominent only in ‘s Gravesande (the only Newtonian in the group, and the top-ranking author in betweenness), while topic 5w is shared by both some Scholastics (Crassot) and non-Scholastics (Abraham, Bacon), and topic 8w is associated with more traditional Scholastic works, and in particular French Scotists (Boyvin, Cauvin, Columbus, Crassot). This suggests that topic analysis can be an important proxy for better profiling the role and nature of the most relevant works in a corpus.
Looking at the same corpus from the point of view of tf-idf (Table 7) we once again observe both Scholastics and Cartesians as the most central works: Frassen (Scholastic), LeGrand and Clauberg (both Cartesians) are the top-ranking in terms of eigenvector centrality, while Jaquier and De Lacaille are the top-ranking in betweenness, suggesting again that natural philosophers belonging to religious orders might have acted as important brokers between Scholastics and non-Scholastics (especially Newtonians).
Eigenvector and betweenness centrality rankings in the tf-idf vector layer (whole network)
In order to further distinguish which of the works that we initially tagged as “non-Scholastic” might be most similar to the works we identified as “Scholastic,” we use a relatively simple algorithm: (i) identify, for each work in the non-Scholastic group, the strongest and weakest correlations in the TM layer and the strongest correlation in the tf-idf layer;Footnote 17 (ii) if the strongest correlation is to another work identified as “Scholastic” (and the weakest to a work not identified as such), then count that work as “Scholastic”; (iii) remove the newly identified “Scholastic” works from the “non-Scholastic” group and move them to the “Scholastic” group. Appendix 1 shows the results of this procedure for the works that this algorithm identifies as “Scholastic.”
There are two striking features in the results we obtained: the first concerns some of the authors that our algorithm identifies as “Scholastic,” while the other concerns the works used to establish this conclusion. Let us consider each feature in turn.
We can clearly identify two sub-groups in the list we obtained. One is constituted by relatively clear-cut cases of Scholastic authors, while another group (marked in grey color in Appendix 1) includes more eclectic works (some of which we already encountered) and some prima facie unexpected names.
At first sight, something seems to go wrong with the algorithm, since canonical novatores like Bacon and Hobbes are classified as “Scholastics.” This is not the only such case. We already encountered Nicholas Abraham, an early physician who moved away from Aristotelianism. Also, Pourchot was broadly working in the framework of Scholastic philosophy, though he eventually defended more Cartesian and atomist ideas in the context of natural philosophy. DuHamel is another author of broad Scholastic orientation who nonetheless engaged occasionally with Descartes and criticized him. Le Clerc represents yet another interesting case. Initiator of important seventeenth-century journals (Nouvelles de la République des Lettres, Bibliothèque universelle et historique), he is mostly a theologian, but in his more philosophical works, like the one included in this corpus, he presents Cartesian views, alongside those of Locke and Newton, explicitly. The fact that he is ranked as a Scholastic in this case might be due again to his way of writing rather than to the conceptual content of his works. Furthermore, his work is in four volumes: the first three devoted to more philosophical topics (logic, ontology, and “pneumatology” or philosophy of mind), while the fourth is dedicated to natural philosophy proper. In this case, the strongest correlation goes to the Dutch Newtonians explicitly, showing his non-Scholastic orientation in this domain. The same split happens in the case of Pourchot’s volumes.
We also observe that comparing TM and tf-idf results helps further balance the outputs based on TM. Hobbes and Bacon, for instance, correlate more strongly in the tf-idf layer with later Cartesians (Le Grand and Clauberg) rather than with more traditional Scholastics. This suggests that in the specific way in which they use language, they are more akin to later generations of non-Scholastics, or (inclusive) that these later generations drew from them more than from the more traditional Scholastics (both options are accepted in existing scholarship).
Generalizing, we can draw the following conclusion: In the early seventeenth century (Abraham, Bacon, Hobbes), the divide between “Scholastic” and “non-Scholastic” is hardly traceable in terms of topics or linguistic features. Authors engaging with natural philosophy tend to share a broad way of talking about the subject and their divisions emerge within this shared background. This background is not something to be left behind in the past, but actually something that remains in place throughout the period covered by our corpus and shows up in relatively late figures like Le Clerc and Pourchot, which sit in between the two “worlds.”
The fact that Bacon and Hobbes are strongly correlated with Scholastic works, especially in terms of the topics they use, should not come as a surprise at a second glance. Despite how vocal both authors are in their fight against Scholastic philosophy, our methods look mostly at the use of Latin in their works and, from this point of view, it appears that Bacon and Hobbes did not write very differently from later Scholastic authors (especially Scotist Columbus and the Dutch Senguerd). Given that the strongest correlated works are all later in terms of publication dates, this might also be interpreted as a hint at the fact that a certain way of writing specific to Bacon and Hobbes might have filtered into later Scholastic works.
Note that our purpose here is not to define a paradigm for what a “Scholastic” author looks like, but rather to use NLP methods in order to group together works based on their similarity in the use of language. From this point of view, our method smooths out the traditional stark contrasts between Scholastic and some of the most vocal non-Scholastic authors, showing that both groups spoke to one another in a fairly similar way. This approach thus lends support to the hypothesis that innovation, especially in the early seventeenth century, was not introduced via a new way of speaking or writing, but rather by reshaping from within what was, at the time, accepted academic (Scholastic) language.Footnote 18 And to take this intuition even further, we surmise that not only might have innovative ideas been clothed in otherwise traditional (Scholastic) language, but, more importantly, this use of a relatively traditional language to spread relatively new and non-traditional ideas positively fostered their dissemination. In other words, the use of Scholastic language made innovative ideas more intelligible for a larger educated audience (even when the authors’ intention might have been to criticize those ideas), thus contributing to their normalization in the long term.
From a more methodological point of view, we also stress that comparing between TM and tf-idf is particularly helpful, since it provides a way of double-checking associations created by one single method. More importantly, topic associations are themselves rather consistent, with topics 1w and 8w being the most relevant for determining any association with Scholasticism. But notice that while Topics 5w and 8w might well include keywords with a Scholastic ring, topic 3w seems entirely attuned to the subject matter of natural philosophy. Hence, the ideological connotation of any topic is something that emerges mostly at the level of interpretation, when we look at who the authors that used that topic mostly are.
The second feature worth commenting on, and perhaps the most striking and unexpected result of this approach, is that only a few authors consistently show up as the strongest or weakest correlations for all the others. Table 8 below presents them in detail, across the two layers.
Authors and works that establish the largest number of strongest and weakest correlations in the whole Latin corpus
Boyvin is a particularly interesting author, since the works strongly correlated with him and weakly correlated with ‘s Gravesande tend to show a Scholastic orientation. Dutch Scholastics with a Cartesian interest, like Burgersdijk and Regius, operate instead as potential markers for more eclectic Scholastic authors (like Pourchot). The situation is more fluid in the case of the strongest tf-idf vector correlations, which show greater variety. However, we can still discern a tendency among the more clear-cut cases of Scholastic works to correlate strongly with likewise clear-cut Scholastic works (like Frassen), while less-clear cut cases tend to correlate strongly with equally eclectic (or Cartesian) works (like Le Grand).
There is also something peculiar with ‘s Gravesande Institutiones, since this single title shows up consistently as a marker (negatively or positively) in both Scholastic and non-Scholastic groups (as illustrated below). One might wonder whether something is going wrong here, since this work deals mostly with what others would have called the “mathematical sciences” and has only a first introductory part dealing with conventional natural-philosophical topics. One might thus interpret this as a book by a professor of mathematics dealing mostly with issues belonging to mathematics. However, this is not a reason to dismiss the book, but instead offers an incentive to take a closer look at it (although doing so is beyond the scope of this paper). The fact that ‘s Gravesande’s Institutiones are not entirely focused on various topics usually included within natural philosophy is not different from what happens in other multi-volume works in our corpus. In fact, it is a recurrent feature of early modern natural philosophy at large to be discussed in the context of other subjects as well. Moreover, our method is sensitive not to the various concepts that are dealt with, per se, but to the way in which language is used to express them. Hence, the consistency with which ‘s Gravesande’s work shows up in our results points to the fact that ‘s Gravesande’s use of language in this work is particularly helpful in parsing different approaches to natural philosophy, even if that is not the primary aim of ‘s Gravesande himself as an author. This is one instance of the heuristic potential of “distant reading,” through which the computational method might detect features of a text that a human reader might not be able to recognize or even pay attention to.
In terms of our research questions, this fact suggests that network analysis has the potential to uncover a number of authors who operate as potential markers for different orientations within the whole network. We shall come back to this observation in our conclusions.
3.3 Stage 3: The updated groupsFootnote 19
When we remove the above-mentioned works (Appendix 1) from the non-Scholastic group, we obtain the updated version of the same group, which we present in Appendix 2, by showing also the strongest correlations in the TM and tf-idf layers.
This updated version of the non-Scholastic group now reveals a much clearer profile, showing the presence of two rival (and expected) trends: a rather consistent group of works conversant with Descartes’s philosophy (Regius, Cally, De Raey, Geulincx, Le Grand, Andala, Clauberg), and a group of works more connected with Newton (‘s Gravesande, Musschenbroek, van der Eyck, Swinden, De Lacaille, Gregory). Figure 3 shows the distribution of Scholastic and non-Scholastic works before and after the computational parsing.
Whole corpus before (left) and after (right) parsing (red = Scholastic, grey = non-Scholastic).
When we look at eigenvector centrality, it seems that the dominant group is the Cartesian one (Table 9), as we already noticed at stage 1, although the names and their scores have changed slightly with respect to the previous Table 4. Perhaps the most visible difference is that Descartes himself now tops the ranking.
Eigenvector and betweenness centrality in the remaining non-flag group cf. topic modelling
In terms of betweenness centrality, authors with a strong interest in astronomy and Newtonian natural philosophy seem to take the foreground (De Lacaille, Swinden, ‘s Gravesande), as do authors who engage critically and eclectically with the Scholastic heritage (Bootius and Fabri), including a declared Cartesian (Cally). The results here are not particularly different from those at stage 1 (Table 4).
Moving on to the tf-idf layer (Table 10), we observe that the presence of Cartesian authors is much more solid and apparent, and they top all rankings in both the TM and tf-idf layers. In this context, van Musschenbroek is the only Newtonian that emerges among the most central authors in the TM network.
Eigenvector centrality and degree in the remaining non-flag group cf. tf-idf vectors
Once again, we notice that only a small pool of works (illustrated in Table 11) operate as markers for delineating most of the others. These are not surprising authors, since they are also the same that show up consistently in our top rankings, suggesting that there is some direct correlation between the centrality of a work and its ability to mark a drift between different groups of other works.
Authors and works that establish the largest number of strongest and weakest correlations in the non-Scholastic corpus
If we now re-run our analysis described in stage 1 for the two updated groups, we can summarize the results in the following Table 12.
Top centrality scores in the updated Scholastic and non-Scholastic groups
This summary Table 12 suggests that the refined Scholastic group is in fact dominated by French Scholastics of a Scotist orientation, as we surmised at stage 1. The non-Scholastic group seems now more clearly dominated by the presence of Cartesian works, although some Dutch Newtonians (such as Swinden) remain very central.
In a sense, these results corroborate the sorts of divides that we would expect based on the standard narrative about the evolution of early modern natural philosophy. For instance, the importance of the French Scotist school (and its influence on Descartes, among others) has been noticed by using more traditional methods (Ariew Reference Ariew2011, 71–100), but our approach provides an independent corroboration of it. This also shows that seemingly more original and eclectic works, like those of Jesuit authors (Rodrigo de Arriaga, for instance) do not seem to acquire a particular prominence in our representation. But given that we tackle similarity in this network, this might just confirm the fact that these authors are in fact much less standard and ‘normal’ than what might be expected and, when considered from the structural point of view of the network we built, this translates in them being less central. In other words, our results can be interpreted as suggesting that while French Scotists contribute to sustaining a shared use of language in discussions concerning natural philosophy (especially among scholastics), Jesuits might be more eclectic, innovative, and original.Footnote 20
4. Limitations
Before concluding, we would like to briefly comment on two relevant limitations concerning the method described.
The first and most obvious limitation concerns our initial choice of the corpus. We focused on Latin textbooks connected with only three geographical areas, namely, France, Britain, and the Dutch Republic. While surely important during the early modern period, this selection cannot claim to be exhaustive or representative of the broader landscape in which debates were unfolding. A larger corpus, the inclusion of authors and works from different areas, or even a greater variety of genres will arguably affect the results we obtained, although foreseeing exactly how would be speculative at this point. However, while this limitation has a direct impact on the interpretative claims we made concerning our results (which are directly dependent on the composition of the corpus), it has a much lesser impact on the workings of the method we presented, the illustration of which has been the main purpose of this paper. More generally, this sort of limitation is something inherent to any computational analysis and draws attention to the importance of sound methodological criteria and preliminary work in the selection and composition of the corpus to which the computational tools will be applied. When historians decide to rely on the established canons for their contextual reconstructions, they can somehow dispense from this consideration, outsourcing it as it were to the canon itself. But as digital tools are used to move beyond the canonical barriers, this sort of consideration should become part of the ordinary routine of any computationally inflected research.
A second, more technical limitation concerns the degree to which our method can capture linguistic nuances and finer details. For instance, much of what discussed above relies on topic modelling, which basically extracts series of the most relevant keywords from texts based on their statistical occurrence. However, the same word can mean different things in different contexts, or when used by different authors. More simply, words can have synonyms that are not immediately captured by topic modelling, or be used in equivocal ways that should be distinguished. This seems even more important in a corpus such as the one presented here, in which many texts share a largely common background and technical vocabulary, and yet many of them present philosophically distinct and sometimes divergent views, which are sometimes very nuanced even for human readers to grasp.
There are two main approaches to addressing this limitation. The first consists in checking the most important results obtained by computational tools with close reading of (at least samples of) the texts. The second consists in combining multiple computational tools that can somehow compensate for the reciprocal limitations. To some extent, we already did so in combining topic modelling vectors with tf-idf vectors, which are more sensitive to the use of rarer and more idiosyncratic terms. Moreover, as we attempted to show in another study (Sangiacomo et al. Reference Sangiacomo, Tanasescu, Donker and Hogenbirk2022a), these tools can be further integrated with collocation analysis, which focuses precisely on the use of words in their immediate context and thus is fitted to distinguish between equivocal cases and possibly to detect synonyms or other variations in language usage (Garcia and García-Salido Reference Garcia and García-Salido2019). Nevertheless, close reading and the combination of multiple computational approaches are not alternative to one another; they simply expand the way in which traditional analysis and automated text mining complement each other.
5. Conclusions
In this paper, we demonstrated how an initial corpus of early modern texts can be represented as a network based on a text data vectorization method and then parsed in different groups, using both available information (including metadata) and computational approaches. These tools focus in particular on the way in which language is used in the works and hence it is, at best, indicative of potential philosophical or conceptual differences among them. As documented above, we have observed nonetheless that topic modelling in particular can provide a reliable basis on which to profile different groups of works and authors.
The results we obtained are consistent with our expectations (and this corroboration is important to assess whether the method we implemented is sound). Coming back to our two research questions, we can now establish (in answer to question 1) that the “Scholastic” group is relatively coherent and homogeneous, although it does allow for a number of authors that engage more critically with Scholastic sources and thus act as brokers or mediators in widening the scope of engagement with other sources. The “non-Scholastic” group is prima facie more heterogeneous, but after computational parsing, it shows two leading profiles: Cartesian and Newtonian.
If we consider now which authors would be most interesting to study (research question 2), our suggestion is that they should be those with highest centrality in general (Table 11), but also those that act as markers for delineating the various groups (Tables 9 and 12). As mentioned, centrality depends on the number and strength of correlations, but two works that are strongly correlated are not necessarily high in centrality. Hence, combining all the authors in these three tables allows us to cover two complementary perspectives on the network (correlation strength and centrality). The authors in Tables 9 and 12 are also particularly interesting because the ways in which they establish correlations with others highlight how distinct groups of authors and works can be identified within the corpus. Let us elaborate on this point.
From the point of view of network analysis, the strongest and weakest correlations among various works (such as those we studied) can be expected to indicate similarity and difference, especially when complemented with metadata about the nature of the works themselves (as we did by tagging certain works as “Scholastic”). What we did not expect was to observe that only a relatively small pool of works operated consistently as the most common markers for similarity strength across our corpus. This result might depend on the peculiar constitution of our corpus and it is only by replicating this research on an entirely different one that we might discern further elements to judge how consistent this feature might be. However, we also observe that most of the authors that serve as “markers” for different groups within the studied corpus also have significantly high centrality (like Descartes and ‘s Gravesande). By combining centrality and correlation strength, these authors are thus particularly significant for parsing the corpus in question. This raises a few more methodological reflections.
One more qualitative and established way of thinking about “central” authors in the history of philosophy and science is by considering them somehow “paradigmatic.” This intuition is at the heart of Thomas Kuhn’s well-known interpretation of “scientific revolutions” (Kuhn 1962). In its simplest form, a paradigm is a template for how to handle a certain puzzle or solve a certain problem. A paradigm, thus, provides an exemplary model for practitioners who are confronted with similar puzzles and problems by also allowing them to develop the paradigm further and extend its heuristic potential to other related cases. Developing paradigms can be considered the core business of what Kuhn called “normal science.” Kuhn is perhaps most famous for his view about the discontinuous breaks between different periods in which older paradigms seem no longer viable and are replaced (in more or less radical ways) by new ones. However, these “revolutionary” discontinuities should not obscure the fact that, once the turmoil of challenging the established paradigms is over, science reverts to its normal form, which determines what most scientists do for most of the time (cf. Kuhn Reference Kuhn1977). Early modern natural philosophy has been famously used as a case for the study of Kuhnian revolutions. In fact, the profound transformations that scholars have recognized in seventeenth- and eighteenth-century scientific practices became themselves a paradigm in history of science and philosophy for conceiving of “scientific revolutions.” Much debate since the 1980s at least concerned whether early modern discontinuities are really so radical to deserve the label of “revolution” or not (Garber Reference Garber, Robert and Daston2016).
However, we would like to resist the temptation of interpreting centrality measures as indicators of Kuhnian paradigms. Paradigms have at least two remarkable features: they tend to be chronologically prior to their further declensions and specifications, and they tend to be normative or prescriptive about what these further declensions should look like. Thinking about “normal science” in terms of paradigms thus amounts to searching for chains of variations within a continuous domain.
The works we singled out in Tables 9 and 12 are not paradigms. They are not chronologically prior with respect to many of the other works to which they are strongly correlated (especially in the case of the Scholastic works), and they are not necessarily or clearly normative (in Kuhn’s sense) for other works either.Footnote 21 And yet, these works are consistently singled out from a computational view as encoding information about the relations of difference and similarity within the corpus. We suggest calling them “rifts”—to stress the idea that they constitute some sort of border within the corpus, which can be used to both separate different groups in terms of similarity and difference, but also as a point of arrival of a certain crystallization in the use of language.
In our account, rifts have two features that make them different from paradigms: they tend to be chronologically posterior to the other works with which they most strongly correlate, and they tend to be proscriptive, in the sense that they rule out options (differences), without entailing more positive constraints. The textbooks by Senguerd or Le Grand arrive relatively late in the history of Scholastic of Cartesian philosophy, and yet they are indicative markers for all other works that operate broadly in the same context. However, these two textbooks do not positively determine what all Scholastic or Cartesian works should entail or discuss, or how, but simply show that departing significantly from what they do will lead to a drift away from Scholasticism or Cartesianism. In this sense, they operate in a similar way to geographical rifts between different territories.
The constitution of a philosophical canon can be understood as the search for the most paradigmatic works for a given period or field of study. Partially, this view lurks in the background of Kuhn’s own account of paradigms, which makes it more explicit what a paradigm is and how it works in a given context. Hence, normal science is constructed in terms of manipulation of paradigms. Nevertheless, our method suggests an alternative account (perhaps compatible with the Kuhnian approach, but not reducible to it). Instead of looking just for paradigms, looking also for rifts can be effective for parsing a given corpus of works. By studying rift-works in more detail, one might gain a better understanding of how certain forms of discourse crystallized to such an extent that any further deviation would give rise to something different. This points to a fruitful way of enhancing the historian’s ability to “get things right” when exploring philosophical ideas in their linguistic context.
Competing Interests Declaration
The authors declare none.
Acknolwedgements
This article is part of the project “The Normalization of Natural Philosophy” that has received funding from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme [grant number 801653].
Appendix 1. Works parsed as ‘scholastic’ based on computational analysis
Appendix 2. Computationally updated non-scholastic group (cf. topics and tf-idf vectors)
Andrea Sangiacomo is associate professor of philosophy at the University of Groningen and professor by special appointment at the Erasmus University Rotterdam on the chair “The Actuality of Spinoza.” He has been PI of the ERC Starting Grant “The Normalization of Natural Philosophy” (2019-2023).
Raluca Tanasescu is a Postdoctoral Researcher at the University of Galway (Ireland), having previously held the same position on the project “The Normalization of Natural Philosophy” at the University of Groningen. Her interdisciplinary research spans translation studies, digital media, cultural analytics, and digital humanities.
Open access
