1 Introduction

The life of academic disciplines is closely tied to that of their journals. This is very much so in the natural sciences, but also in the human and social sciences where journals constitute one of the most significant venues for published research. The discipline of philosophy of science is no different: alongside monographs and edited volumes, disciplinary journals have become an increasingly important vehicle for research outcome over the past decades. While the boundaries with the broader discipline of philosophy may be quite permeable (Wray 2010), some journals have come to play a most central role in the development of the philosophy of science up to what it is today. This is for instance the case of journals such as Erkenntnis, Synthese or Philosophy of Science for which the first issues all go back to the 1930s, but also of several other journals that have been founded since—including the recently established European Journal for the Philosophy of Science—and which show the vitality of the discipline. The broader historical context in which these journals became established—such as the emergence of the Vienna Circle in the early twentieth century followed by the war-time emigration of some of its leading figures to America—certainly matters to situating the intellectual content of publications, and so does also the smaller-scale context of editorship and editorial decisions (Howard 2003; Reisch 2005 chap. 5; Beisbart et al. 2019). Importantly, the published articles and their content play a most central role when it comes to mapping out the diversity of research questions the philosophy of science has come to address over the years. As a complement to more traditional exegetic and historical approaches, we propose to apply computational text-mining approaches to examine the overall content of these journals with the objective of offering a large-scale and data-driven perspective on the research agenda of the philosophy of science as it has unfolded since the 1930s.

With such approaches, it indeed becomes possible to investigate the content of large corpora of full-text documents in a comprehensive and quantitative way, and from synchronic as well as diachronic perspectives. Initially developed in linguistics and natural language processing, computational textual analyses have been applied in many areas, including the natural sciences (as illustrations, see e.g., Valente et al. 2018; Luiz et al. 2019) and—as concerns us here—in history and philosophy (e.g., Mimno 2012; Murdock et al. 2017; Peirson et al. 2017; Barron et al. 2018; Pence and Ramsey 2018). While the journal Philosophy of Science has recently been characterized through topic-modeling analyses (Malaterre et al. 2019), we examine here a much broader and comprehensive corpus composed of the full-text content of 15,897 articles published by eight philosophy of science journals from 1934 up until 2017. These journals, which are among the most recognized journals that publish general philosophy of science research in English language, include: the British Journal for the Philosophy of Science, the European Journal for the Philosophy of Science, Erkenntnis, International Studies in Philosophy of Science, Journal for General Philosophy of Science, Philosophy of Science, Studies in History and Philosophy of Science (Part A), and Synthese.

To investigate the major research themes that interested philosophers of science from the 1930s up to now, we mobilized computational approaches—notably topic modeling algorithms—that could provide information about the semantic content of the entire corpus and its evolution over time. The results of our analyses document, in particular, the shift in topics in the philosophy of science, with a relative decrease of philosophy of language, logic and philosophy of physics articles over 80 years, and an increase in epistemology, philosophy of biology and of mind. Such quantitative data offer an empirical basis for what might otherwise be informal claims about the discipline and its evolution in the past eight decades as reconstructed from the perspective of its major journals. They are also the type of data that may trigger novel discussions about the directions the field might take.

In what follows, we explain the text-mining methodology that structured our analyses and describe the corpus that we used (Sect. 2; note that this section can be skipped by readers more interested in the results themselves). The topic-modeling results are then presented, with a discussion about their content, their relationship to one another and what can be inferred in terms of the diversity of research work published in philosophy of science journals (Sect. 3). We complement these results with diachronic analyses in order to examine how topics have evolved over time (Sect. 4). We then zoom in on journals and map their topical profile (Sect. 5). As a conclusion, we briefly discuss text-mining methodologies and the broader context of their application (Sect. 6).

2 Corpus and methodology

We usually do not use words at random, but mobilize them in specific combinations that not only respect the syntactical rules of the languages we adopt but also enable us to express ideas and convey meaning to others. Over and over, these combinations of words give rise to repeated patterns. Hence the intuition that examining these patterns may provide insights into the semantic content of the texts in which they occur. As the linguist John R. Firth used to say, “you shall know a word by the company it keeps” (1957, p. 11). This intuition is found at the basis of text-mining methods that mobilize computational tools to quantitatively examine the patterns of occurrences of terms within digitized sets of texts (e.g., Srivastava and Sahami 2009; Aggarwal and Zhai 2012). Because words are usually used in a coherent fashion across texts and authors, such approaches have shown to be very effective. One of them—called “topic-modeling”—makes it possible to identify sets of words that are often used together in similar circumstances across varying documents. Through careful examination of the meaning of their terms and the documents in which they occur, these sets of words can be interpreted as meaningful topics which, in turn, provide a topical perspective on the semantic content of the corpus being studied.Footnote 1 By considering metadata such as publication years or journals, additional analyses can be conducted that offer, for instance, diachronic views on the evolution of the corpus content over time or journal profiles in terms of topical content. Here, we applied these approaches to a set of major philosophy of science journals to identify their topical content and the evolution of that content over the past 80 years. The topic-modeling itself was carried out with an algorithm based on the well-established Latent Dirichlet Allocation (LDA) model (Pritchard et al. 2000; Blei et al. 2003) which is part of a broader family of statistical algorithms for topic discovery in texts (e.g., Griffiths and Steyvers 2004; Blei and Lafferty 2009; DiMaggio et al. 2013). Such approaches have two noteworthy characteristics: not only do they make it possible to gain semantic insights about corpora that would have been too large to investigate manually, but they also function in an unsupervised data-driven way, meaning that topics are identified on the basis of the word content of the corpus, without a priori knowledge of which topics might indeed be present (though such knowledge may prove useful to interpret results, as we will see below).Footnote 2 With LDA modeling, each topic is a probability distribution over the lexicon of the corpus—the lexicon being the set of all word types present across all documents of the corpus—and, in turn, each document is modeled by a probability distribution over topics. Any random document can then be thought of as being the result of multiple probabilistic word drawings according to the word probability distributions of topics, these topics also being drawn according to their own probability distribution within said document. The overall methodology we used here can be decomposed in five major stages:

(1) Corpus preparation. We constructed our corpus with the full-text content of the following 8 general philosophy of science journals: the British Journal for the Philosophy of Science, the European Journal for Philosophy of Science, Erkenntnis, International Studies in the Philosophy of Science, the Journal for General Philosophy of Science,Footnote 3Philosophy of Science, Studies in History and Philosophy of Science Part A and Synthese (see Table 1 for an overview; more details about each journal can be found in Appendix A). One may comment that philosophy of science is also published in other journals, be they more general philosophy journals (e.g. Mind), more specialized philosophy of science journals (e.g. Studia Logica, Hyle, Biology and Philosophy), or even science journals (e.g. Bioscience). Philosophy of science is also published in many non-English languages (e.g. Principia, Epistemologia, Philosophia Scientiae) and in numerous books and edited volumes. Needless to say, our ambition is not to cover the entirety of what has ever been published in philosophy of science. However, by focusing on the 8 major general philosophy of science journals in English language, we hope to provide more than a representative and useful perspective on the thematic content of philosophy of science and its evolution over the past 8 decades. The corpus that we settled on comprises 15,897 full-text articles from 1934 up until 2017, totaling over 62 million word occurrences (for an average word count of about 3912 words per article).Footnote 4 Whenever journals were published in several languages, we retained only those articles that were written in English due to algorithmic linguistic constraints.Footnote 5

Table 1 List of the philosophy of science journals included in the corpus

(2) Preprocessing of textual data. This preprocessing aims at both removing elements that might create noise and optimizing the size of the data retained for analysis. As is considered best practice, we removed stop-words such as determinants, prepositions or pronouns with the assistance of a part-of-speech (POS) tagging algorithm that makes it possible to identify the morphosyntactic category of every word of the corpus. Here we used the Penn TreeBank POS tagging algorithm (Marcus et al. 1993) and only kept nouns, verbs, modals, adjectives, adverbs, proper nouns, and foreign words. On the basis of this POS tagging, we also lemmatized the textual data in order to reduce the number of word variants.Footnote 6 This was done using the TreeTagger algorithm (Schmid 1994). We furthermore filtered out words that occurred in fewer than 50 sentences in the corpus. These operations resulted in a lexicon of 22,958 distinct words distributed among over 1 million sentences of the 15,897 articles.

(3) LDA topic modeling. To carry out the topic-modeling itself, we used the Latent Dirichlet Allocation (LDA) algorithm, following (Blei et al. 2003) and (Griffiths and Steyvers 2004).Footnote 7 As mentioned above, LDA is a generative probabilistic computational method that models topics as probability distributions over words, and documents as probability distributions over topics. Given a hypothesized number K of topics, the algorithm starts with random probability distributions and iteratively adjusts these distributions (constrained by sparse Dirichlet priors) up until a convergence criterion is met. As with any other well-established topic-modeling approach, LDA assumes K is fixed beforehand. In the present case, the corpus included articles from 8 different journals over some 8 decades. This meant that, to present meaningful results both synchronically and diachronically, as well as overall and per journal, a fairly low value of K had to be chosen. After running different models and assessing them, we chose the value K = 25 topics.Footnote 8 Implementing the LDA algorithm on that basis thereby resulted in 25 topics and their probability distributions in each one of the 15,897 articles of the corpus.Footnote 9

(4) Topic interpretation and clustering. As mentioned above, the topics that result from an LDA topic modeling are probability distributions over the words of the corpus lexicon. By examining the words that have the highest probabilities, it is usually possible to identify the semantic content that these words are meant to convey in the corpus. This semantic content can be corroborated by examining the articles in which given topics are the most expressed. One may also look at topic-to-topic similarity (as measured by the distance between their respective probability distributions) and topic-to-topic correlation in documents as heuristics to help with the interpretation. Ultimately, interpreting a word-probability distribution results in giving it a meaningful label. Though manual, this stage is in practice highly constrained both by the sets of words that express topics and by the sets of documents in which they appear. In a second step, we ran clustering analyses to group topics into broader clusters, so as to facilitate their discussion. We did this by running a modularity analysis on the network graph of topic-to-topic (positive) correlations—following the approach of (Blondel et al. 2008; Lambiotte et al. 2008) as implemented in (Bastian et al. 2009)—which resulted in the identification of 8 clusters comprising between 2 and 5 topics each (we describe them below in Sect. 3).Footnote 10

(5) Topic analysis by time-period and by journal. The topics were further analyzed in light of two types of metadata: article publication years and journals. In order to provide a view on the temporal evolution of the relative frequency of topics over time, we segmented the corpus into 21 periods of 4 years each, from 1934 up until 2017, and computed, for every topic and time-period, the average probability of finding that topic in all articles of that time-period.Footnote 11 In parallel—and with a view to shed light on the specific topical profiles of the journals—we segmented the corpus per journal and computed, for every topic and journal, the average probability of finding that topic in all articles of that journal.Footnote 12

3 A topical perspective on philosophy of science

The topics that resulted from our topic-modeling analyses cover a broad range of themes that naturally make sense in the philosophy of science. Table 2 provides the list of the 25 topics, with their top-10 words. The 25 topics were grouped into 8 clusters following the modularity analyses we conducted on the graph of topic correlations in documents (as depicted in Fig. 1). For each topic, the top-10 articles in which the topic has the highest probability were retrieved so as to provide a representative basis; they are listed in Appendix B.

Table 2 List of all 25 topics with top-10 keywords, sorted by cluster (topic labels include cluster letter, topic name and topic ID)
Fig. 1
figure 1

Graph of the 25 topics grouped into 8 clusters (nodes represent topics that have been colored depending on cluster; node areas are proportional to topic probabilities in the corpus; thickness of edges represent topic correlation within documents; node labels include cluster letter, topic name and topic ID); visualization and community detection tool: Gephi (Bastian et al. 2009), with Multigravity ForceAtlas 2 for layout rendering)

The first cluster–cluster A—relates to philosophy of logic and philosophy of language. It is the largest cluster and includes five topics. Two topics—Language and Sentence—are characteristic of philosophy of language themes, with such keywords as “language”, “sentence” or “meaning”, yet also with terms that denote philosophy of logic (such as “logical” or “predicate”). Representative articles include work on Carnap and translational indeterminacy, antinomies and designation, intensional/extensional definitions, as well as essays on demonstratives, indexicals, or reference among others. The topic Truth is closely related to research themes that belong to logic and philosophy of logic (with such keywords as “truth”, “proposition”, “logic”). This topic covers work on logical truth, modal logic, the meaning of logical constants or on vagueness, to name a few. The other two topics—Mathematical and Formal—appear more mathematical in scope (with keywords such as “mathematics”, “proof”, “axiom” or “set”, “function”, “theory”). They relate to research topics that concern the foundations of mathematics, arithmetics and geometry, Hilbert’s program, as well as set theory, topology, and how measurement relates to mathematical principles.

Cluster B includes three topics that denote research interests in scientific knowledge and, more generally, epistemology. Topic Knowledge includes such words as “belief”, “knowledge”, “epistemic” or “justification”, and is very much concerned with research in the theory of knowledge, as exemplified by articles on skepticism, infinitism, self-knowledge, doxastic justification, epistemic luck or testimony. Topic Scientific-theory appears more specifically centered on scientific knowledge and realism (with keywords such as “theory”, “scientific”, “theoretical”, “empirical” or “realism”). Some of its most representative articles include work on Kuhn, Laudan and scientific rationality, as well as on relativism, structural realism, and more generally scientific realism. The third topic—Arguments—is an ambivalent topic that may very well capture a real research interest on argumentation while at the same time denote a set of very commonly used terms in philosophy of science articles. Keywords indeed include what may look like very usual terms in the field, such as “argument”, “claim”, “say” or “question” (explaining, in part, the relatively high probability of the topic across the whole corpus, as can be seen by the relative size of the topic nodes on Fig. 1). Yet at the same time, the topic also denotes a specific research interest in argumentation, with representative articles that are about one-sided arguments, knockdown arguments, verbal disputes, counter-examples or non-argumentation.

Cluster C is about confirmation theory and comprises two topics. The first topic, Confirmation, includes terms such as “hypothesis”, “evidence”, “test” or “inference” and is most strongly found in several articles about Hempel’s and Goodman’s paradoxes, and, more generally, confirmation, induction, projection or entrenchment. The second topic, Probability, reflects the use of probability theory in certain philosophy of science articles. With keywords such as “probability”, “measure”, “chance”, or “distribution”, the topic is most characteristic of research work on such problems as Dutch book arguments, sleeping beauties, and Bayesianism more generally.

A nearby two-topic cluster concerns rational choice theory and related questions (cluster D). It includes topic Agent-decision (with keywords “action”, “agent”, “decision”, “choice”, “rational”, “moral”, “utility” among others) which is highly probable in articles that concern the prisoner’s dilemma, cooperation, coordination, free-riding, exploitation and related issues. The other topic—Game-theory—is closely related, with such terms as “model”, “game”, “strategy”, “player” or “equilibrium”, and notably denotes research on social interaction and communication in learning and knowledge formation.

Cluster E is more biological in nature and concerns work on evolutionary theory, mental states and the neurosciences. It includes three topics labeled Evolution, Mind and Neurosciences. Topic Evolution is denoted by such terms as “selection”, “population”, “organism”, “evolutionary” or “gene”, and is found with high probability in articles that concern the problem of the units of evolution, the question of biological individuality, heredity, fitness or the notion of species and gene, among others. This topic is clearly about philosophy of biology. The other two topics of the cluster are more similar and concern mind, cognition, psychology and the neurosciences in general. With keywords such as “experience”, “object”, “state”, “mental”, “perception” and “color”, topic Mind is characteristic of articles on representationalism, internalism, conceptual and nonceptual content and, more generally, on mental states. On the other hand, topic Neurosciences tends to characterize research on the cognitive architecture, on functional modularity, neural mechanisms, neurobiological activities, and the like that often mobilize such terms as “system”, “information”, “function”, “process”, “mechanism” or “cognitive”.

More general philosophy of science topics are found in cluster F, with Causation, Explanation and Property. Topic Causation is straightforwardly about causation, with such keywords as “causal”, “cause”, “effect”, “event” or “causation”, and is found in articles that investigate different aspects of causation, including probabilistic, counterfactual, interventionist accounts of causation, as well as questions about causal asymmetry, causal relata or contrastive causation. Likewise, Explanation is also a topic with a straightforward interpretation. Its top terms include “explanation”, “law”, “explain”, “account”, “phenomenon” and the topic is found in articles that examine the concept of scientific explanation in relationship, for instance, to the concepts of law of nature, deduction, mechanism, theoretical unification, or causal regularities. The third topic of cluster F is Property. With top-terms that include “property”, “object”, “relation”, “physical”, or “kind”, the topic concerns research that dwells on ontology, physicalism and related questions such as supervenience and emergence.

Philosophy of physics topics are found in cluster G. That cluster includes topic Particles which is characterized by such keywords as “theory”, “energy”, “experiment”, “chemical”, “atom”, or “electron”. Some of the key articles in which the topic is found concern research on certain particles such as positrons, neutrinos, electrons, and by ways of consequence on related phenomena—electrostatics, electromagnetism—as well as on energy and chemistry. With top-terms such as “state”, “quantum”, “system”, “particle” or “measurement”, the topic Quantum-mechanics picks out a class of research interests that clearly focus on quantum mechanics and related issues, with characteristic articles on the interpretation of quantum mechanics, on the questions of quantum decoherence, locality, the EPR paradox and quantum measurement. Topic Relativity—with keywords such as “space”, “time”, “field”, “relativity” or “Einstein”—is found in articles that investigate different aspects of special and general relativity theories, including questions about isotropy, synchronization, and also space–time. That topic is complemented by the very nearby topic Time (see Fig. 1), with top-terms such as “time”, “event”, “future”, “present” or “temporal”. Time appears to relate both to the physics of time—hence its proximity to Relativity—and to the metaphysics of time, with articles on McTaggart’s temporal series or on time travel.

The last cluster–cluster H—gathers three topics that have a more social or historical nature. With keywords that include “body”, “motion”, “force” but also “newton”, “galileo”, “aristotle”, “descartes” and “god”, topic Classics concern a broad range of work in the history of science, for instance on Galileo’s observations, on Descartes’ physics or on Newtonian mechanics. Topic Philosophy appears to characterize work that borders the domain of traditional philosophy or history of philosophy. The topic includes quite general terms such as “concept”, “nature”, “knowledge”, “science” as well as “world”, “idea” or “kant”, and is found in a broad variety of articles that range from Kant to Hegel and Whitehead. Finally, topic Social has a more science-studies connotation, as is apparent from some of its top-words: “science”, “social”, “research”, “work” or “study”. Related articles include research on the social organization of science, on scientific disciplines or on science museums, among many others.

4 Diachronic perspective (1934–2017)

Diachronic topical evolution can be studied by examining the probability of finding specific topics in articles that were published within specific temporal windows. Here, we segmented the corpus into 21 consecutive time-periods of 4 years each, from 1934 till 2017 (84 years in total) and quantified the relative probability in finding each one of the 25 topics in a given time-period (see Fig. 2). Since topics reveal words that tend to be used in similar patterns by philosophers of science in their essays throughout the corpus, topics also mirror the research interests of these same philosophers of science and the relative significance of these research interests within each time-period. At the topic-cluster level, maybe one of the most striking results of the analyses is the strong probability decrease of cluster H (that includes topics of a more social and historical nature) over time as well as of cluster A (philosophy of language and logic), while clusters that relate to knowledge and epistemology (cluster B), to the philosophy of biology and of the neurosciences (cluster E) and to explanation/causation (cluster F) tended to increase in probability in the corpus over time.

Fig. 2
figure 2

Overall diachronic evolution of topic probability between 1934 and 2017 (colored bars, left-side y axis; legends include cluster letter, topic name and topic ID) and corresponding number of articles (dotted line, right-side y axis)

Within cluster A, it is mostly the topic Language that decreased in probability: while it was one of the most significant topics overall up until the 1960s, it continuously decreased from the 1970s onward. This trend is very much in line with what is well-known about the history of philosophy of science and the significance of logical empiricism in its early stages. Epistemology-related topics—cluster B—have generally increased in probability over time. In particular, topic Scientific-theory—which is about scientific knowledge and realism—has tended to increase from the 1970s, with a stabilization in the 2000s. Topic Knowledge itself—in the sense of theory of knowledge or epistemology—has seen a recent increase since the 2000s. As for the topic Arguments—which is a fairly dominant topic that can relate both to specialized research on argumentation and to the generic usage of “philosophical jargon”—one notes its increase since the mid-1950s and a leveling off from the 1980s onward.

Within cluster C—which is about confirmation theory broadly construed—, topic Confirmation had its heydays in the 1960s, whereas Probability has seen a more significant increase since the 2000s. This also shows in the publication dates of the topic most-closely related articles: whereas 6 of the top-10 articles for Confirmation were published before the 1970s, the trend is reversed for Probability which has 7 of the its top-10 articles published in the 2000s. It is as if research on these themes started in the 1950s–1960s by being framed in terms of confirmation problems or even paradoxes, and then turned to probability theory as a means of addressing such questions.

The evolution of the topics of cluster D—which concerns rational choice theory and related questions—is contrasted, though the cluster as a whole has been on the rise ever since the 1930s: whereas topic Agent-decision appears to signal a philosophical interest that has been more or less constant over time, topic Game-theory denotes a more recent field of investigation that really gained in popularity in the 2000s (all top-10 articles of that topic are posterior to 2000).

The overall probability of cluster E—about the philosophy of biology and the neurosciences—has been slightly increasing from the 1930s up to now, but with fluctuations between topics. For instance, Evolution was somehow present in the 1940s, then decreased in the 1950s–1970s, and started again to increase in the 1980s up untill today. This is consistent with (Byron 2007; Malaterre et al. 2019; Malaterre et al. 2020), who showed that philosophy of science was much concerned with biology-related questions quite early on, in the first half of the twentieth century, and then again starting from the 1980s onward with the establishment of a dedicated field of research. Topic Mind also appears to have followed a similar patter, though its increase in the more recent decades is not as significant, especially compared to topic Neurosciences that has strongly gained in momentum, especially since the 2000s, denoting a strong interest in the neurobiological mechanisms of mind and different mental phenomena.

Cluster F—which includes explanation/causation topics—also appears to have increased in probability over the eight decades of the study. Explanation had its heydays in the 1950s–1980s, and started picking up again in the 2000s. This is consistent with what is known of the field, for instance through Salmon’s Four decades of scientific explanation (Salmon 1989) and more recent work on mechanistic approaches to explanation and on reduction (as denoted by some of the top-10 articles of this topic). Interest in Causation developed in the 1980s—typically on counterfactual accounts and issues about causal asymmetry, possibly in the wake of problems encountered by non-causal models of explanation at that time—and continued to spark interest throughout the 2000s, especially on questions about probabilistic and interventionist accounts. While being somehow present in the corpus from the 1960s to the 1980s, topic Property really increased in probability throughout the 1990s and 2000s: 9 of its top-10 articles were published in the 2000s, in particular on issues about physicalism and supervenience.

Philosophy of physics topics—as defined by cluster G—have been quite prominent in the philosophy of science, yet with an overall slightly declining trend. Topic Particles denotes an interest in sub-atomic particles, chemistry, electromagnetism—among others—that was quite present in the 1950s–1960s and that dwindled thereafter (6 of the top-10 articles of that topic are from the 1960s and earlier). Philosophy of physics appears to have then focused more on relativity theory and quantum mechanics, as shown by the probability of topics Relativity and Quantum-theory. Interest in relativity theory was notably strong in the 1960s–1970s, with quantum mechanics picking-up more strongly in the 1990s (5 of its top-10 articles are from that period). Time, which is a topic highly correlated with Relativity (as shown in Fig. 1), appears to have been quite probable in the corpus early on in the 1950s, then decreased in probability from the 1970s onward and leveled out.

Cluster H—which is more social or historical in content compared to other topic clusters—appears to have been very significant in the first half of the twentieth century, and then decreased quite drastically. This trend is dominated by topic Philosophy—which we interpreted as denoting interest in traditional philosophy or history of philosophy—and, to a lesser extent, by topic Social, though the latter has seen a renewed interest more recently, especially from the 1990s onward (6 of its top-10 articles were published in the 2000s or after). By contrast, topic Classics—which notably concerns history of science figures or topics, among other more diverse themes—has remained relatively constant in probability all throughout the period.

5 The journals and their profiles

When analyzing the diachronic evolution of topics, one should bear in mind that the volume of publications—in the corpus—increased by a factor of 12 (from about 40 articles per year in the 1930s to about 500 in the 2010s).Footnote 13 This increase in publication volume resulted both from an increase in the number of philosophy of science journals and an increase in the yearly number of articles published by each journal (see Fig. 3). Erkenntnis, Philosophy of Science (PS) and Synthese were founded in the 1930s, followed by the British Journal for the Philosophy of Science (BJPS) in the 1950s, then the Journal for General Philosophy of Science (JGPS) and Studies in History and Philosophy of Science Part A (SHPSA) in the 1970s. In the 1980s, International Studies in the Philosophy of Science (ISPS) was created, and the 2010s saw the launch of the European Journal for Philosophy of Science (EJPS). Though Synthese now publishes two- to three-times more articles than either Erkenntnis or PS (likely due to the topical issues it also publishes), the three first-founded journals still accounted for about 2/3 of all articles in the 2010s. Historically, publications started mostly in German and Dutch, followed by English, in the 1930s. The advent of WWII led to a temporary decrease in publication volumes in the 1940s, followed by a relative stagnation throughout the 1950s. Publications then tripled from the 1960s to the 1980s, at the same time as the discipline of the philosophy of science underwent a strong professionalization (Howard 2003). This upward trend in publication volumes was however interrupted in the late 1980s and 1990s, before picking up again in the 2000s, while the volume of articles published in languages other than English dwindled down to zero. One possible explanation of this slight slump in the 1990s is the creation of specialized journals in several sub-disciplines of the philosophy of science, for instance in philosophy of physics (SHPS-Part B was launched in 1995), in philosophy of biology (Biology and Philosophy started in 1986, SHPS-Part C in 1998) or in other domains (e.g. History and Philosophy of Logic launched in 1980, Economics and Philosophy in 1985, Philosophy and Technology in 1988).Footnote 14

Fig. 3
figure 3

Number of articles published in Synthese, Philosophy of Science (PS), Erkenntnis, The British Journal for the Philosophy of science (BJPS), Studies in History and Philosophy of Science Part A (SHPSA), International Studies in Philosophy of Science (ISPS), Journal for General Philosophy of Science (JGPS), and the European Journal for Philosophy of Science (EJPS) in English language, with a specific ‘stream’ for articles published in languages other than English (total number of articles: 15,897; articles are sorted into 4-year periods; the width of each ‘stream’ is proportional to the number of articles; for guidance, the width of SHPSA in the 2010s is approximately 80 articles per year; the ‘streams’ are sorted by decreasing size from top to bottom at each time-slice, thereby showing the leading journals in terms of publication volume on top; visualization tool: RAWGraphs (Mauri et al. 2017))

The topical profile we computed for each journal is shown in Fig. 4, while Fig. 5 depicts the diachronic evolution of these profiles. One should bear in mind that not all journals have the same publication span (see Table 1 above and Appendix A): some started in the 1930s such as Erkenntnis, others more recently, some even in the past decade such as the EJPS. Reassuringly from a methodological perspective, the composition and general evolution of topics for PS is fully consistent with what has been obtained independently in another study that focused on the sole corpus of PS (Malaterre et al. 2019).

Fig. 4
figure 4

Topic probability for each journal (calculated by averaging the topic probabilities of journal articles; legends include cluster letter, topic name and topic ID)

Fig. 5
figure 5

Evolution of topic probability for each journal (colored bars, left-side y axis) and publication volume (dark curve representing number of published articles, right-side y axis) per time-period (x axis)

One of the most striking results is that both Erkenntnis and Synthese have always had a relatively stronger content in topics related to philosophy of logic and philosophy of language—cluster A—compared to the other journals, with an overall probability of about 28% (Fig. 4). Even if the trend has been decreasing over the last decades—after a peak in the 1950s–1970s at about 40% for Synthese (see Fig. 5)—both journals continue today to publish articles that concern these topics at a level of about 20%. This could be explained by the logical empiricist mindset of the founding philosophers behind Erkenntnis—Carnap and Reichenbach in the 1930s, Hempel for the relaunching of the journal in the 1970s—and possibly, the logic-oriented specialty of Hintikka who served as editor of Synthese from 1965 till 2002. The overall trend across all journals has however been a decreasing one over the past decades, even very markedly so in the case of PS.

On the other hand, both Erkenntnis and Synthese display some of the smallest proportions of topics related to historical and social questions (cluster H), especially compared to such journals as ISPS, JGPS and especially SHPSA (for which these topics represent over 30% of the content). These results make sense given the broader positioning of these three journals: in addition to general philosophy of science, ISPS highlights an interest also for “philosophically informed history and sociology of science”, JGPS for “the social, historical and ethical dimensions of the sciences as the context for understanding current problems of philosophy of science” and SHPSA for “topical areas of historiography of the sciences, the sciences in relation to gender, culture and society and the sciences in relation to arts” (“International Studies in the Philosophy of Science: Aims & Scope | Taylor and Francis” 2020; “Journal for General Philosophy of Science: Aims and Scope | Springer” 2020; “Studies in History and Philosophy of Science Part A | Elsevier” 2020). Interestingly, the journals that have the lowest probabilities for these historical and social topics in the last decade (approx. 5%)—Erkenntniss, Synthese, PS and BJPS—displayed much higher probabilities for these same topics in the first half of the twentieth century (in the range of 25–35%). This marked trend was known in the case of PS (Malaterre et al. 2019), and has been explained, more generally, by such factors as changes in editorship, the professionalization of the discipline or even decisions made by funding agencies (Howard 2003; Douglas 2010; Vaesen and Katzav 2019). Note however that the present data show that journals such as ISPS, JGPS and SHPSA all have retained a significant proportion of these topics (between 19 and 31% in the last years). These findings thereby suggest that factors such as disciplinary professionalization or funding agencies policies—which should have affected research themes across the entire discipline, hence across all journals—did not play as significant a role as probably did editorial policies broadly speaking, including journal strategic positioning within the ‘academic market’ (as revealed by the participation of certain journals to specific academic conferences or by the journal presentation pages on publishers’ websites; see Appendix A), journal publication format (possibility of special issues, conference proceedings), topical choices made by editorial teams (determining, for instance through desk reject, which papers fall in and out of a given journal scope) or even reviewer choices by editorial teams (who may reproduce associative patterns).Footnote 15

Scientific knowledge and epistemology—as depicted by topics of cluster B—appear as the second most probable group of topics for both Erkenntnis and Synthese (with probabilities between 19 and 22%), after topics of cluster A in philosophy of language and logic. This is in line with the reputation of these journals as preferred venues for epistemology articles. Note however that these epistemic topics also have about the same high probabilities in ISPS, JGPS, SHPSA and in the recently launched EJPS (about 20%). Note also that the topics have tended to decrease in probability for PS and BJPS in the last 20 years, while increasing within Erkenntnis and Synthese.

Topics about probability and confirmation—cluster C—tend to be more present in the profiles of BJPS, PS and EJPS than in the other journals, with a probability of about 10%. The diachronic evolution shows a probability increase in the 1960s–1970s, even up to 15% for BJPS, followed by a slight decrease and relative stagnation thereafter.

Overall, topics of cluster D, which broadly speaking concern rational choice, appear to be comparatively more probable in EJPS than elsewhere (about 10%), followed by Erkenntnis and Synthese (as depicted in Fig. 4). The diachronic profiles (Fig. 5) show that these topic probabilities also tended to increase in other journals, notably in PS, in the recent years.

The cluster of topics that concern mind, the neurosciences and biology—cluster E—has a more marked presence in PS, EJPS and BJPS (at about 13%) than in the other journals. This is notably true for the biology-related topic for which the probability has been especially increasing since the 1990s. Yet, it is interesting to note that the topics of this cluster were also quite strongly present in the first half of the twentieth century, in PS, Synthese and Erkenntnis, and even in BJPS in the 1950s, which is consistent with other studies (Byron 2007; Nicholson and Gawne 2015; Malaterre et al. 2019). Note also that the creation of SHPS-Part C in 1998 could explain the relative drop in probability of this cluster of topics within SHPS-Part A in the late 1990s, from about 9% down to 5%, though that probability slowly rose back again to about 9% in the last decade.

EJPS, ISPS and Erkenntnis are the journals in which topics that concern causation, explanation and ontology—tend to be the most probable, at a level of about 12–14% (cluster F). These topics are also found in the other 5 journals, though, comparatively, at a slightly lesser level.

Journal profiles with respect to physics-related topics—cluster G—are more marked, with journals such as BJPS exhibiting a probability of over 20% in these topics, followed by EJPS, ISPS and PS at about 17%. By comparison, the probability of these physics topics is much less dominant within Erkenntnis or Synthese, at about 8% overall, even showing a slightly decreasing trend over the last decades. This trend is even more significant for SHPSA, with a marked step in the late 1990s when SHPS was split into SHSP-Part A (general philosophy and history of science) and SHPS-Part B (philosophy of modern physics).

6 Conclusion

Over the eight decades that span from its infancy as a discipline to its present day, the philosophy of science has incurred significant changes in its research agenda. The text-mining analyses we conducted on the full-text content of eight major general philosophy of science journals reveal the large-scale topical developments of the field as reflected by the publications in these journals. Broad trends are clearly visible, notably the decrease of philosophical work of a more social and cultural nature in the 1950s, then compensated by the rise of research themes that concerned philosophy of logic and language, these themes then decreasing in the 1970s–1980s, followed by topics about confirmation in the 1980s and philosophy of physics in the 1990s–2000s, though to a lesser extent. Meanwhile, other research themes have flourished in the last four decades, notably in domains that concern scientific knowledge and epistemology, rational choice, philosophy of mind and of biology, as well as topics that concern causation, explanation and ontology: accounting for about only 25% of topic probability in the 1930s, these research themes increased to a level of about 65% lately.

The analyses also show similarities and differences between journals, both in terms of their overall content and of their diachronic evolution. Of course, the number of publications has literally exploded in 80 years, increasing more than ten-fold, novel journals being also founded at several points in time. It is also likely that journals have become an increasingly significant avenue for communicating research to peers, especially compared to books in particular, due to rising concerns about publication speed. Journal history—and the history of their semantic content—has thus become an integral part of the history of the discipline, and much still remains to be investigated.

Of course, the results should always be interpreted in light of a proper understanding of what text-mining methods can and cannot do. As explained earlier (see Sect. 2), LDA topic-models are probability distributions calculated on the basis of how words are distributed within documents in the corpus: topics are probability distributions over words while documents are modeled as probability distributions over topics. The topical content of a document can thereby be understood as resulting from drawing topics according to the probabilities assigned by the topic-model. Yet topic-models remain silent over the reasons why specific topics are likely to be found together. They also cannot reveal the argumentative relationships that topics may entertain in specific articles, nor can they give explanations about why specific topics did change over time—for instance, due to their contextual relevance in light of scientific advances, the maturity of certain lines of inquiry, or simply editorial policies or other pragmatic factors. One should also bear in mind that text-mining results depend, to a certain extent, on choices among many equally valid modeling options, often balanced by expert-judgement. This is for instance the case with the chosen number of topics K as mentioned above, or with the delicate task of topic interpretation. Yet, topic modeling algorithms in general and LDA in particular are well-established approaches to textual analysis that have been tested in many occasions. LDA topic-modeling has notably been shown to be a very reliable algorithmic tool for identifying topics in large textual corpora (e.g., Griffiths and Steyvers 2004; Blei and Lafferty 2009; DiMaggio et al. 2013).

Indeed, one of the most significant characteristics of computational text-mining methodologies is the possibility to deploy them to analyze very large corpora of full-text documents that would otherwise have been out of reach if done by hand. Another is that, thanks to their being quantitative and unsupervised, they provide perspectives that are complementary to classical historical or exegetic methods. They also provide the type of data that contributes an empirical basis for what might otherwise be informal claims—in the present case, claims about the evolution of a field broadly speaking or of some of its specific research themes. In a sense, these methods provide novel ways of answering old questions. But they also open up novel ways of asking novel questions: they have a heuristic value in that they provide insights about trends and patterns to be further investigated, notably with the more classical historical and analytical approaches of the philosophical methodology. Topic-models reveal topical patterns in corpora—such as temporal variations in topic prevalence or specific journal profiles—that can in turn be questioned and formulated as explananda worth enquiring about. As we hope to have shown, text-mining approaches are powerful tools to investigate the broader history of the philosophy of science, how this field has developed and flourished over the course of the past eight decades and how its journals have contributed to this history. No doubt much remains to be uncovered.