1. Introduction
Our understanding of science is being broadened by the digitization and automated analysis of the various outputs of the scientific process, such as scientific literature, archival data, and networks of collaboration and correspondence. These technological changes are laying the foundation for new types of problems and solutions in the philosophy of science. The purpose of this article is to provide an overview and guide to some of the novel capabilities of digital philosophy of science.
To best understand the reasons why digital philosophy of science lets us ask a new class of questions, let us consider how it differs from more traditional approaches. For example, consider how we might draw conclusions about articles in the journal Nature. It has published over 360,000 articles since its founding in 1869, meaning that one would have to read 10 articles a day for 100 years to work through the complete archives of this journal alone. Of course, the standard response in the philosophy of science is to favor depth over breadth and closely read a much smaller number of articles. While there is certainly much we can learn about science in this way, some broad questions about the nature and history of science—questions, for example, about how theories arise and become established in the literature as a whole—would remain unanswerable without a way to glean information from hundreds of thousands or even millions of journal articles. Much the same argument holds for scientific images or information about the collaboration, communication, training, or citation connections between researchers.
The question, then, is to what degree we can learn from the vast scientific literature without having to read every article closely—to instead do what is called distant reading (Moretti Reference Moretti2013). With distant reading, we input a large body of literature into a computer and use it to do the “reading” for us, extracting large-scale patterns that would be invisible or impractical to find otherwise. In the philosophy of science in particular, this process has been aided by a number of large digitization efforts targeted at the outputs of the scientific process. One crowning achievement of these efforts is the nearly complete digitization of the academic journal literature. This content is thus now accessible in ways that it has never been before.
Digital approaches to the philosophy of science contrast with traditional methods involving close reading—intently reading a narrow body of literature within a focal area. With close reading, a philosopher will have an impressive command over a limited domain. He or she closely reads a select set of documents from the scientific literature or analyzes the experimental, training, or collaborative records of a small group of researchers to attempt to extract the structure of a scientific theory or to understand the meaning of its terms.
We should stress that the close reading–based traditional philosophy of science and distant reading–based digital philosophy of science are not in competition. Instead, they are complementary. If, for example, a researcher wants to know how the meaning of a particular term has changed over time, he or she could use automated textual analysis tools to locate instances of the term, find hot spots in which the term is used frequently, and quickly see which words it is associated with and how these word associations have changed over time. In conjunction with digital analysis, performing close reading of key texts will be invaluable. The close reading may then spur further digital inquiries, and so on. Thus, traditional and digital philosophy of science work in tandem, each supporting the other.
The remainder of this article will canvass a number of significant issues that must be dealt with in order to develop a digital philosophy of science research program. We hope that this overview will be helpful to researchers who are interested in moving forward with digital tools but are not certain where or how to begin.
2. Getting Started
Because digital philosophy of science is a relatively new field, not only is there no set of standard tools, it is often unclear what sorts of questions can be answered by the extant tools. Thus, let us begin by considering some of the new kinds of questions one can address.
One of the most significant advantages of distant reading comes from the ability to engage with corpora significantly larger than those usually treated by philosophers and historians of science. For example, Murdock, Allen, and DeDeo (Reference Murdock, Allen and DeDeo2017) were able to analyze large-scale patterns in Darwin’s reading by accessing the full text of every book that we know him to have read over a period of decades. These kinds of analyses simply would not be possible without the aid of technology. Answering research questions that leverage broad (yet still circumscribed; see sec. 3.2) sets of data is thus likely to be a fruitful use of digital tools. For example, one could track concepts over the entire print run of a journal, the collections of books published in the Biodiversity Heritage Library (Gwinn and Rinaldo Reference Gwinn and Rinaldo2009), or the PubMed Open Access Subset of contemporary biomedical journal articles (Roberts Reference Roberts2001). These kinds of investigations allow us to explore the conceptual landscape of a field through distant reading, by offering (at least in some cases) an exhaustive analysis of an area.
Another advantage comes from the ability of analytical algorithms to parse texts in ways that even well-trained close readers cannot. For example, fine-grained patterns of language usage, such as the shift in a term from a noun use to a verb use, or a shift from referring to science as a one-person activity to a group activity, could be traced in the literature with a level of exhaustiveness, objectivity, and care that would simply be impossible for a single reader. Automated tools can analyze sentence structure, word order, or parts-of-speech usage in a way that would try the patience of any scholar (Manning et al. Reference Manning, Surdeanu, Bauer, Finkel, Bethard and McClosky2014).
The ability of digital tools to increase the breadth of a research question is also important. If one has a hypothesis drawn from a particular domain (e.g., maximization or optimality inferences in biology), this hypothesis could be tested in other separate domains (economics, psychology, sociology) with only a modest further investment of resources.
While digital tools can aid in answering existing research questions, these tools also open the possibility of framing new questions without a clear analogue in the predigital world. For instance, work by Miller and colleagues (Reference Miller, Beard, Laubichler and Bliss2015) applies dynamic network analysis to our understanding of scientific conceptual development. The questions they ask arise in conjunction with the digital tools and in dialogue with digital humanities researchers in other disciplines.
3. Choosing the Right Tools
Now that we have a sense of the advantages of digital analysis, let us consider the currently available tools and corpora of data relevant to the philosophy of science. To begin, we should draw attention to the central repository of digital humanities tools, known as the DiRT Directory, accessible at https://dirtdirectory.org/ (for more on its construction and predecessors, see Dombrowski [Reference Dombrowski2014]). There are nearly as many digital humanities tools as there are digital humanities researchers, and the landscape of contemporary software changes rapidly. For nearly any kind of analysis, the directory will include some tool that performs it—the most important question will be whether the data available can efficiently be converted into the format required by that tool.
3.1. Basic Tools
There are a number of tools that may be used immediately by researchers, as they do not require that one collates a set of documents of interest in advance. Perhaps the most famous of these is the Google Ngram corpus (Brants and Franz Reference Brants and Franz2006; Michel et al. Reference Michel2011), accessible at https://books.google.com/ngrams. This corpus contains the entirety of the scanned Google Books project, current as of 2012, with frequency data for single words as well as pairs and longer sequences (so-called bigrams, trigrams, and, more generally, n-grams).
Obviously, the Ngrams project does not exclusively contain scientific or philosophical content, and hence a number of queries that might interest philosophers of science will simply not be meaningful when queried against the Ngram Viewer. For example, the scientific usage of the term “evolution” will be completely masked by the broader cultural use of the term, and hence philosophers interested in the use of this term are unlikely to be able to uncover interesting data. There are also a number of worries about the statistical representativeness of the Google Ngram corpus, even when judged as a measure of broader cultural usage or popularity (Morse-Gagné Reference Morse-Gagné2011; Pechenick, Danforth, and Dodds Reference Pechenick, Danforth and Dodds2015).
Much more precise search and analysis may be performed by using JSTOR’s Data for Research (DFR) project (Burns et al. Reference Burns, Brenner, Kiser, Krot, Llewellyn and Snyder2009), available at http://dfr.jstor.org/. This tool allows users to perform searches and analyses against the entire corpus of JSTOR journals. Researchers may search for articles by journal, publication date, author, subject, and more, allowing for careful control over the set of articles to be analyzed. These articles may then be queried for word frequencies (and n-gram frequencies), as well as automatically extracted “key terms,” which are words common in the selected articles but uncommon in the corpus as a whole (computed using the tf-idf [term frequency–inverse document frequency] score). The frequency scores from JSTOR DFR may also be used as an input to a variety of the tools described below.
3.2. Gathering a Corpus
The more advanced tools are set apart primarily by not coming with a preloaded corpus of material to study. This means that the challenge of obtaining data falls to individual researchers. As mentioned above, we find ourselves in a particularly fertile period for data availability in the philosophy of science. Much of the journal literature, in some cases back into the nineteenth century, is available online in PDF or HTML form. Comprehensive online projects are available that focus on the works, life, and correspondence of figures like Darwin (Secord Reference Secord1974; Whye Reference Whye2002), Newton (Iliffe and Mandelbrote Reference Iliffe and Mandelbrote1998), Poincaré (Walter, Nabonnand, and Rollet Reference Walter, Nabonnand and Rollet2002), Einstein (Mendelsson Reference Mendelsson2003), and others (Pouyllau et al. Reference Pouyllau, Blondel, Wronecki, Wolff and Usal2005; Beccaloni Reference Beccaloni2008; Mills Reference Mills2011). A number of discipline-specific archives have also been constructed, such as the Embryo Project Encyclopedia, an open-access, digital repository covering the history of embryology and developmental biology (Maienschein et al. Reference Maienschein, Laubichler, Ranney, MacCord, Elliott and Colonna2007). To this may be added the digital collections now increasingly available from a wide variety of museums and libraries. With an appropriate collection of data obtained for a researcher’s private use, it becomes possible to leverage a much wider variety of analytical tools. (These data must also be carefully curated and safely preserved; we will return to the question of data management in the next section.)
A researcher gathering a corpus must consider how and to what extent the data should be annotated. Minimal annotation—for example, leaving content as plain text with only bibliographic data for tracking—allows for the rapid creation of a large corpus and lowers the future burden of maintaining and updating the annotations. But more significant annotation—such as marking up textual data in a format like that described by the Text Encoding Initiative (TEI; Ide and Véronis Reference Ide and Véronis1995)—allows for more complex, fine-grained, and accurate analyses. This annotation can take a variety of forms. For textual data, TEI allows users to indicate the locations of various parts of the document (pages, paragraphs, chapters, indexes, figures, or tables) or the various kinds of references made by a piece of text (dates, citations, abbreviations, names of persons or institutions, etc.).
This process of cross-referencing documents may be aided by the use of external ontologies—in the sense (not the one common in philosophy) of collections of standardized verbs and concepts that allow for the same term to refer unambiguously across multiple documents. In philosophy, the Indiana Philosophy Ontology project (InPhO; Buckner, Niepert, and Allen Reference Buckner, Niepert and Allen2007), available at https://inpho.cogs.indiana.edu/, allows standardized references to concepts such as “sociobiology” or to particular philosophers. A number of such ontologies also appear in other areas of the sciences, and a document may be marked up with multiple ontologies to add further semantic richness.
With a heavily annotated document, significantly more complex analysis may be applied, as the computer now “knows” where particular concepts are mentioned, how they are used, and how they relate to other ideas. While the use of such methods is relatively untested in philosophy, the biomedical field has made significant strides in this direction in recent years—for example, analysis of the usage of gene and chemical concepts in the scientific literature has actually enabled the extraction of novel relationships (previously unpublished by researchers but discernible from the body of literature as a whole) and even the generation of novel hypotheses about future drug development (Cohen and Hersh Reference Cohen and Hersh2005).
The question of the representativeness of one’s sample of data is also a significant one with which researchers must engage. As we noted above, even in the largest corpora, such as Google’s Ngram collection, there are still problems with the statistical significance of the sample (Morse-Gagné Reference Morse-Gagné2011), with biases in temporal availability of data (more data tend to be available closer to the present, as the relevant outputs were “born digital”; Michel et al. Reference Michel2011) and systematic sources of error such as that introduced by optical character recognition (Hoover Reference Hoover2012). These concerns are somewhat alleviated when using a curated corpus known to be complete (such as databases of historical correspondence), but even in these instances, researchers must remain constantly vigilant against statistical bias.
3.3. Advanced Tools
With a corpus in place, there are a variety of options for users interested in performing analyses impossible with the basic tools described above. First, there are a number of tools designed to aid researchers in presenting their material as an easily navigable, searchable, categorized public resource—a public digital archive or museum. The most popular of these is Omeka (Cohen Reference Cohen2008), available at https://omeka.org/. Omeka is a free, open-source software product that allows users to construct online archives and museum exhibitions, to add catalog information and metadata to digital items, and to attractively present all of this material to the public at large. Deploying a website such as this is a nice way to garner some immediate, public-facing payoff from the difficult work of obtaining and curating a digital collection.
One alluring feature of large digital data sets is the possibility of analyzing the networks found within them—whether these are networks of collaboration drawn from experimental archives or lab notebooks, networks of correspondence drawn from digitized letters, or citation networks extracted from the journal literature. Such network analysis can often allow us to see patterns in the overall structure of a field that would be otherwise difficult to discern. One of the most user-friendly network analysis tools available is Gephi (Bastian, Heymann, and Jacomy Reference Bastian, Heymann and Jacomy2009), available at https://gephi.org/. Gephi allows users to import graphs in a number of formats (including ones as simple as comma-separated value [CSV] spreadsheet data) and to perform a variety of analyses and visualizations. The network may be broken into clusters (using a standard measure known as modularity; Blondel et al. Reference Blondel, Guillaume, Lambiotte and Lefebvre2008), the degree of connectivity of individual nodes may be easily explored, and the results can then be rendered graphically for presentation.
If the data to be analyzed are text, a popular choice is Voyant Tools (http://voyant-tools.org/). Once a corpus of text is uploaded to Voyant, the user is immediately presented with a wide variety of options: a word cloud, a cross-corpus reader, a tool for tracking word trends through the text, and a short snippet concordance are among the immediately available tools, and a variety of other more complex analyses and attractive visualizations may be performed using plug-ins. Voyant may also be used to save online corpora for future use, which facilitates classroom usage of textual analysis.
Another challenging problem likely to be faced by philosophers of science interested in the scientific literature is the analysis of a large number of journal articles, a kind of analysis not often performed in traditional digital humanities, which often focuses on book-length source material. To solve these problems, one of us has created a software package, RLetters (Pence Reference Pence2016), available at http://www.rletters.net. (One public installation of this software, containing a corpus of journals in evolutionary biology, is available at http://www.evotext.org and described in Ramsey and Pence [Reference Ramsey and Pence2016].) This is a web application, backed by a search engine and database, which may be deployed by anyone wishing to analyze a corpus of academic journal articles. It includes a variety of analysis methods (sharing many of those described for Voyant), including an especially powerful word frequency analyzer.
Finally, should all of these tools fall short, the statistical computing language R (R Core Team Reference Team2017), available at http://www.r-project.org/, has become a very popular base for constructing novel analyses in the digital humanities (Jockers Reference Jockers2014). R combines a comprehensive set of standard statistical analyses (such as principal component analysis and dendrogram or tree clustering) with an extensive collection of user-contributed packages that may be used to perform complex tasks such as querying Google Scholar or Web of Science. This power comes at the cost of significant complexity, however, as R operates like a programming language rather than a graphical application.
3.4. Copyright Issues
One of the most common pitfalls that users are likely to encounter when building corpora of digital data is copyright and licensing issues. While much material pertaining to figures like Newton or Darwin is available in the public domain, a confusing legal landscape besets all work created after 1923 (the date of “public domain” for published works in the United States). A number of recent court decisions (most significantly Authors Guild v. HathiTrust, 755 F.3d 87 [2d Cir. 2014]) have begun to clear the legal landscape in the United States, indicating that scholarly textual analysis and other sorts of digital-humanities work are likely to fall under the US “fair use” provision. This, however, does nothing to simplify obtaining copyrighted materials, nor does it help scholars in other countries, many of which lack an analogue to fair use. It also may well be cold comfort to litigation-sensitive universities.
Increasingly, however, publishers are recognizing the demand for digital analyses of their materials. Elsevier has deployed a text- and data-mining policy that applies to all of its journals and will allow researchers to access and analyze articles as part of any institutional subscription (Elsevier 2014). Under the auspices of JSTOR’s DFR project, researchers may request access to full-text articles, if their university subscribes to the appropriate JSTOR collections. We also have had some degree of personal success negotiating access contracts for closed-access journal articles with their publishers, including with Nature Publishing Group, which was very receptive to the possibilities opened by digital analyses. We anticipate that this trend toward increased ease of access will only continue.
4. Data
The academic process relies on the ability of other researchers to access, verify, and reproduce the results of analyses such as these. We next consider how to publish and archive data and how make public the tools and techniques used to achieve the results.
4.1. Data Management
Philosophers are not, as a rule, accustomed to producing large amounts of data as part of our research. When using digital tools, we find ourselves faced with many of the same questions our scientific colleagues have dealt with for some time—how do we document, store, and preserve the data that our research generates? We cannot offer comprehensive answers to these questions here; we raise them only to emphasize that problems of metadata, documentation, and archiving have been discussed extensively in other contexts and should not be neglected. Early engagement with these resources will prevent significant problems from arising in the long term (York Reference York2009; Michener Reference Michener2015).
4.2. Reproducibility
If digital analyses are to serve as elements of the permanent research record along with journal articles, then we must take care to make those analyses reproducible in the future. This is a multifaceted problem that has, in recent years, received significant attention from the scientific community (Munafò et al. Reference Munafò, Nosek, Bishop, Button, Chambers, du Sert, Simonsohn, Wagenmakers, Ware and Ioannidis2017). For most digital philosophy projects, there are three key components to reproducibility.
First, software must be reproducible—that is, easily installed and run by those with the relevant technical expertise. To that end, the development and use of open-source software is laudable, as is using a readily accessible distribution platform such as GitHub.
Second, corpora must be reproducible. This can be a difficult challenge, particularly if one has negotiated access to a body of copyrighted material for analysis. It is often possible to negotiate access not just for an individual researcher or research team but also for any researchers accessing a public resource (Ramsey and Pence [Reference Ramsey and Pence2016] successfully negotiated such contracts for evoText). We encourage researchers to think very seriously about this challenge as they develop corpora.
Finally, the original forms of data must be—and remain—available. Open data repositories such as figshare (https://figshare.com/; Kraker et al. Reference Kraker, Lex, Gorraiz, Gumpenberger and Peters2015) or Zenodo (https://zenodo.org/) will accept raw data and make it citable. Researchers should also take care to upload data into these repositories in formats that are likely to remain readable indefinitely into the future, such as CSV format for spreadsheets or plain Unicode text or XML for textual data.
5. Integrating Digital Results into Philosophy of Science
The digital tools are powerful, and they have great potential for the philosophy of science. But digital results do not automatically translate into philosophical results. We therefore must consider how to integrate them with broader answers to philosophical questions.
5.1. Justifying Digital Results
A recurring problem with digital humanities results consists in how we can be certain that we have obtained genuine information supporting the conclusions we hope to draw. We can in part resolve this by proceeding in a “hypothesis-first” manner—forming clear hypotheses before performing analyses. All data sets are apt to contain chance patterns, and we should not be led astray by these patterns by basing our conclusions on them. And when we formulate hypotheses, we should be open to a range of possible conclusions, since approaching a statistical analysis system with an answer to one’s question already in mind tends to result in the cherry-picking of tools and methods to produce the desired result (Ioannidis Reference Ioannidis2005).
That said, it can be difficult, even having carefully formulated and tested a hypothesis, to be certain that one has in fact demonstrated it conclusively. Many analyses in the digital humanities lack statistical validation and have only a history of successful use as evidence in their favor (see, e.g., the discussion of validation in Koppel, Schler, and Argamon [Reference Koppel, Schler and Argamon2009]). Others require collaboration among experts in philosophy and statistics, computer science, or even electrical engineering (Miller et al. Reference Miller, Beard, Laubichler and Bliss2015). An important step in developing a digital research program, therefore, is to consider how to assess whether a project has succeeded or failed. This may involve validating the methods, producing standard kinds of analysis outputs, or, as we now consider, using digital research methods only as a first step in a broader program of philosophical research.
5.2. Digital Humanities as Research Generator
Because digital tools give us significantly increased breadth and depth, we have found that they are useful not just as research tools in and of themselves but as a compass, directing us toward questions that would be answered by traditional methods in philosophy of science. For example, Pence has recently combined existing work on an episode in the history of biology (Pence Reference Pence2011) with digital tools (Ramsey and Pence Reference Ramsey and Pence2016), to produce a more general hypothesis about debates over paradigm change, which is now ripe for a nondigital analysis (Pence, Reference Penceforthcoming). We anticipate that this workflow will, in fact, be quite common. As a digital tool shows us a provocative but not fully theorized result, this can provide us with an excellent working hypothesis, case study, or set of sample data for developing a philosophical thesis.
6. Conclusion
As scholars interested in studying the natural sciences, we cannot ignore the availability of digital data that might assist us in our research. It was once the case that the body of scientific literature was modest in size and represented only a narrow distillation of and reflection on the world. Now the literature has become so massive, complex, and diverse that it constitutes a world unto itself, one poised for scientific and philosophical analysis. Adding to this all of the digital traces of work not heretofore published—archival images, notebooks, and so on—we are confronted with an overwhelming, but incredibly rich, world of information. Philosophers are beginning to see how this information can bear on questions in the philosophy of science and can inspire new ones. But the profusion of sources and formats of data, on top of the assortment of available tools, some of which require considerable technical savvy, provides a barrier to the philosopher. In this essay, we have attempted to provide a window into digital philosophy of science, with both an overview of what is possible and some guidance in seeking data and analysis tools. We are excited about the prospects for future work in this field and hope that this article will help to spread our excitement.
