=Paper=
{{Paper
|id=Vol-1681/Wierst_et_al_Philo_at_scale
|storemode=property
|title=Phil@Scale: Computational Methods within Philosophy
|pdfUrl=https://ceur-ws.org/Vol-1681/Wierst_et_al_Philo_at_scale.pdf
|volume=Vol-1681
|authors=Pauline van Wierst,Sanne Vrijenhoek,Stefan Schlobach,Arianna Betti
|dblpUrl=https://dblp.org/rec/conf/dhlu/WierstVSB13
}}
==Phil@Scale: Computational Methods within Philosophy==
https://ceur-ws.org/Vol-1681/Wierst_et_al_Philo_at_scale.pdf
Phil@Scale: Computational Methods within
Philosophy
Pauline van Wierst*, Sanne Vrijenhoek‡, Stefan Schlobach‡, Arianna Betti†
* Scuola Normale Superiore, Pisa, Italy
p.v.wierst@gmail.com
‡ The Network Institute, VU University Amsterdam
sannevrijenhoek@gmail.com, k.s.schlobach@vu.nl
† The Network Institute / Institute of Logic, Language and Computation, University of Amsterdam (corresponding
author)
ariannabetti@gmail.com
Abstract
In this paper we report the results of Phil@Scale, a project directed at the development of
1
computational methods for (the history of) philosophy. In this project, philosophers and computer
scientists together created SalVe, a tool that helps philosophers answering text-based questions. SalVe
has been tested successfully on the Wissenschaftslehre (1837), an extensive work by the Bohemian
polymath Bernard Bolzano (1781-1848). Bolzano was a philosopher, mathematician and theologian
whose work has been of fundamental importance for the development of Western logic and the
foundation of sciences such as mathematics and computer science. The testing of SalVe on the
Wissenschaftslehre reveals that with respect to certain questions within philosophy valuable
contributions are obtained by applying even rather simple, well-known computational techniques. We
conclude that there is definitely a future for computational methods within text-based philosophical
research.
We explain how SalVe can be used within philosophical research that relies on textual sources. We
will start out with an explanation of our aims in developing SalVe and give a short description of
SalVe’s functionalities, followed by a technical description of the tool. Then we will give a concrete
example of how SalVe aids philosophical research. We conclude the paper with an evaluation of the
potential of Digital Humanities tools for philosophy, and the challenges that face us if we wish to
continue this development further.
1
Phil@Scale was part of the 2012 Academy Assistants project from the VU Network Institute, which is directed at combining expertise from
ICT and the social sciences and humanities in order to create innovative technologies. In Phil@Scale, Sanne Vrijenhoek, a master student in
Artificial Intelligence, developed the text-mining tool SalVe together with Pauline van Wierst, a master student in Philosophy, under the
supervision of Arianna Betti (Philosophy) and Stefan Schlobach (Computer Science). To the project Hein van den Berg, Stefan Roski, Jeroen
Smid (all Philosophy), Iris Loeb (Philosophy/Mathematics), Mariya Koleva (Computational Linguistics), Robin Brons (Liberal Arts &
Sciences, University College Roosevelt), and Aron de Jong (Lifestyle Informatics) all contributed. See the Phil@Scale project at
http://www.networkinstitute.org/academy-assistants/aaprojects/.
Copyright held by the author(s).
1 Our goal
Research in philosophy as it is traditionally done consists for a large part in conceptual analysis and close
reading of philosophical texts. Researchers in philosophy come up with hypotheses and confirm or disconfirm them
on basis of their analysis and interpretation of the text. This method of research is particularly valuable in the history
of philosophy, where researchers read closely very complex, sometimes very long texts that contain concepts far
removed from contemporary ones. The method has two main characteristics, which we aim to change for the better.
First, slowness, which limits philosophical research to a small scale, i.e. research based on a small corpus of a few
selected texts. Second, strong reliance on the analysis and interpretation of a researcher working in isolation, which
makes philosophical work a qualitative and, in some important sense, subjective area of research. The subjective
aspect of the work of solo researchers in philosophy is mainly due to the fact that the interpretive assumptions
forming the background of the research in question are rarely made explicit. Our aim is to put computers to use to
make philosophical research faster and larger-scale first, and, second, more quantitative and objective.
It should be stressed that close reading of texts will always be a substantial part of doing philosophical research.
It is inconceivable that computational methods will ever completely replace close reading. SalVe is therefore
developed in order to facilitate analysis by philosophy researchers. It does so, for example, by allowing researchers
to determine quickly which parts of the textual corpus are most relevant for their research. This connects to our first
aim of making philosophical research faster and larger-scale. Furthermore, SalVe approaches text as data, and has
several options to analyze these data in a quantitative manner. Results from computational analysis can be used to
test (and create) hypotheses about texts, and to strengthen the evidence in favor of a particular interpretation of the
latter. As such, it connects to our second aim of making philosophical research more quantitative and objective.
SalVe is based on word counts and has several functionalities that deliver information about a textual corpus and
its parts. The functionalities are (1) advanced searching for words within a corpus; (2) calculating the similarity
between parts of the corpus; (3) displaying the co-occurrence of words (term window); and (4) listing the words that
seem most relevant for particular parts of the corpus. Here below we give a technical description of SalVe.
2 Technical description
SalVe is a (desktop) Java program based on version 3.6 of Apache Lucene (http://lucene.apache.org/core/), an
open source text search engine. SalVe analyzes and indexes text files specified by the user. The input text files
should preferably be first properly preprocessed, so that they (i) only contain standard characters, and (ii) are split
2
into predefined units (SalVe’s ‘documents’). These units must be determined a priori by the user. They can be
single sentences, paragraphs (when the scope is slightly bigger), or even full books when analyzing a large corpus.
SalVe is also usable on non-preprocessed input, but the quality of the input will influence the quality of the results.
SalVe’s core action is indexing, during which documents are broken down into terms, the position of each term
in each document is then determined and efficiently stored in an index file. Indexing requires a preprocessing step
2
We assume here the text files are more than perfectly OCRed copies of the printed originals. By this we mean texts that are not only free of
OCR (Optical Character Recognition) errors but also free of any other typographical elements that can interfere with text-mining results,
such as end-of-line splits, page numbers, etc. Note that philosophy and other humanities disciplines based on fine-grained textual analysis
require exactly this kind of more-than-perfectly-OCRed texts. So, if the text files at disposal aren’t of such quality, preprocessing should also
aim at improving the files in this sense. Arianna Betti’s CLARIN-NL project @PhilosTEI (2013-2014) and follow-ups aim at building
workflows for philosophers yielding more-than-perfect TEI-encoded digital copies of printed originals.
Copyright held by the author(s).
relying on analyzers, programs that normalize the input text. Normalization is a process used to improve
comparability of documents. The process transforms texts according to certain norms or standards by reducing
certain strings to other strings, e.g. all numerals (‘27’) to number names (‘twenty-seven’). Normalizing a text can
mean different things, as the process may be more or less intrusive depending on the purpose at hand. SalVe’s users
can choose among four different analyzers: Whitespace, Standard, English, and German. The least intrusive analyzer
is the Whitespace analyzer, which just tokenizes the text, i.e. chops it off into strings enclosed in white spaces. The
Standard analyzer also lowercases all words and removes punctuation in addition to tokenizing the text, so that e.g.
‘Word’, ‘word,’ ‘word.’, and ‘Word,’ are treated as four occurrences of the same term. The Standard analyzer might
be useful when dealing with texts in multiple languages. The English and the German analyzers do all the Standard
analyzer does, but in addition also filter stop words and stem the words in the text according to language-specific
rules. Stop words filtering removes very common words in a language (such as articles and prepositions), while
stemming reduces inflected words to their stem, so that e.g. by using the German analyzer ‘analytisch’ and
‘analytische’ are treated as two occurrences of the same term - which also means that searching for either will yield
the same results. No semantic analysis is applied in SalVe.
SalVe calculates similarity between documents by relying on a mathematical representation of textual
information based on the so-called Vector Space Model (VSM). In particular, SalVe calculate document similarity
using the simple cosine similarity metric, which is defined as:
Here below we explain the formula and the basics of the Vector Space Model. First we explain what vectors are.
V (a)⋅V (b)
sim(a, b) =
V (a) ⋅ V (b)
A vector is an object that has both a magnitude and a direction, for example, a force of a certain strength
(magnitude) is a vector.
Figure 1: Vector
“Geometrically, we can picture a vector as a directed line segment, whose length is the magnitude of the vector
and with an arrow indicating the direction. The direction of the vector is from its tail to its head.” (Frank & Nykamp
2015). Vectors can also be represented in a system of coordinates in an n-dimensional space, such as the Cartesian
coordinate system in 3-dimensional space. So a 3-dimensional vector is represented as V = (x, y, z) – where the three
coordinates (called ‘elements’ or ‘components’ of the vector) represent the point assigned to the head, while the
origin of the axes is assigned to the tail of the vector. For example, v = (0, 0, 1), a so-called standard unit vector.
Multiple vectors can be represented together as a matrix, e.g. this one, where our v = (0, 0, 1) vector occupies the
first row.
! 0 0 1 $ first row
# &
m (for matrix) = # 1 0 0 &
# 0 1 0 &
" %
Copyright held by the author(s).
Recall now SalVe’s indexing. When SalVe creates an index file, it creates in fact a vocabulary, an index of each
term in all the documents: it counts how many (different) terms appear in the documents, and assigns them a
number. Say you use the English analyzer (which removes stop words) and your collection of documents or
document set contains the following two (very short!) documents for a total of three terms:
Document a: The cat is on the mat
Document b: The bat is on the cat, and the cat is on the mat
SalVe first creates an index for the terms such that cat = v1, mat = v2 and bat = v3 (there are no more terms to
count, since the stop words ‘The’, ‘is’, ‘on’, and ‘and’ have been removed by the English Analyzer). Then it
converts the two-document set into a vector space with a number of dimensions equal to the number of terms in (all)
the documents. Since we have three terms in this case, SalVe works here with a 3-dimensional vector space. SalVe
then measures the term frequency for each document, i.e. the frequency of a certain term in a certain document. In
this way, SalVe obtains a 3-dimensional vector for each document (note that since cat = v1, it corresponds here to the
first vector element, bat to the second, and mat to the third):
𝑣"#$%&'() + = 1, 0, 1
𝑣"#$%&'() 0 = 2, 1, 1
Indeed, Document a has 1 occurrence of cat, 0 of bat and 1 of mat, while Document b has 2 occurrences of cat, 1
of bat and 1 of mat. We can also put the two vectors altogether in a matrix:
1 0 1
𝑀=
2 1 1
We can now see how SalVe calculates similarity between two documents. Let’s go back to our formula:
The formula says: the similarity between document a and document b, i.e. (sim (a, b)), is equal to the product of
V (a)⋅V (b)
sim(a, b) =
V (a) ⋅ V (b)
the vector representation of document a (V(a)) and b (V(b)), divided by the product of the (Euclidean) magnitude of
the vector representation of a (||V(a)||) and b (||V(b)||). To understand what the magnitude of a vector is, recall the
representation of vectors as arrows: the (Euclidean) magnitude of a vector is simply the distance between the
vector’s tail and its head. Now we’re almost done: what the formula calculates, is the difference in orientation of the
two vectors a and b (more technically, the formula yields the cosine of the angle between the vectors). By measuring
the vectors’ orientation instead of their magnitude, we avoid that different document lengths unjustly influence the
similarity values. For instance, if document a contained cat three-hundred-fifty times and document b contained cat
two times, on account of the first’s being much longer than the second, the magnitudes of the vectors would be quite
different, but their orientation would instead be quite similar. The result of the formula is a value between 0 and 1,
where 0 indicates no similarity, and 1 an identical document.
Copyright held by the author(s).
The calculation also incorporates the principle of Term Frequency Inverse Document Frequency (TFIDF or
TF/DF). TFIDF divides for each term the frequency in the document by the total number of documents the term
appears in. This way terms that only appear in a few documents are more distinctive for that document or cluster of
documents than those that appear in a large number of documents.
The results obtained by SalVe can be exported to a csv file and used as a basis for further analysis using various
programs, such as the graph platform Gephi, which can be used to visualize the similarity relations between terms in
documents.
In the following we will show how SalVe’s functionalities can be used to answer a concrete philosophical
3
research question.
3 A concrete application
To show a concrete application of SalVe to philosophical research, we focus on two important notions that are
treated in Bolzano’s Wissenschaftslehre: analyticity and grounding (see for an extended treatment van Wierst 2013).
Roughly described, analyticity is the characteristic of a claim thanks to which this claim is true (or false)
4
independently of the meaning of the terms used in it. Grounding (Abfolge) is a relation between two (clusters of)
claims such that one is the proper scientific explanation of the other.
Analytic truths in Bolzano’s view are, for example, truths of the form ‘A is A’, such as ‘A triangle is a triangle’.
Bolzano calls all truths that are not analytic synthetic. An example of a synthetic truth is ‘Triangles have angles that
together equal two right angles’. Previous research maintains that Bolzano saw the distinction between analytic and
synthetic truths connected with his notion of grounding (which he calls Abfolge), although he nowhere explicitly
mentions this. The claim is this:
De Jong’s thesis According to Bolzano, every analytic truth is grounded in a synthetic truth. (de
Jong 2001: 346)
In other words, according to de Jong, Bolzano held that every analytic truth stands in the relation of grounding to
a synthetic truth, which comes down to saying that the former is explained by the latter. In the following we will
show how SalVe helps to determine whether de Jong’s thesis is correct.
We will start out by asking ourselves what we would have done to confirm or disconfirm de Jong’s thesis by
philosophical research as it is traditionally done. Close reading and conceptual analysis by a philosopher is often
guided by implicit research questions. Using SalVe requires that we make these research questions explicit. In our
case, the questions that guide our research would for example be: What exactly did Bolzano write about analyticity,
syntheticity and the grounding relation? How do these notions fit in the bigger picture of the Wissenschaftslehre?
Which evidence is there that Bolzano intended to link up the analytic-synthetic distinction with grounding (as de
Jong says)?
3
For more introductive information on the vector space model, see http://blog.christianperone.com/?p=2497
4
Note that, being analyticity one of the most debated, elusive and technical concepts in all of philosophy, including its history, this
characterization should be seen as capturing only a specific variant of this concept, to wit the Bolzanian notion as it appears in the text we
explore. See Rey (2013). Something similar holds for grounding.
Copyright held by the author(s).
Now that our research questions are explicit, we reformulate them so that we can use SalVe’s functionalities to
answer them. For example, the question ‘What exactly did Bolzano write about analyticity?’ can be reformulated as:
‘In which paragraphs does the word ‘analytisch’ appear most frequently?’; ‘Which paragraphs are most similar to
the paragraphs in which ‘analytisch’ appears frequently?’; ‘Which words co-occur with ‘analytisch’’? These
questions can be answered by means of three of SalVe’s functionalities: (1) advanced search for the term
‘analytisch’, (2) similarity calculation between the paragraphs in which ‘analytisch’ appears frequently and the other
paragraphs of the Wissenschaftslehre, and (3) the term window of ‘analytisch’.
Advanced search for the terms ‘analytisch’ and ‘synthetisch’ gives the following result:
Illustration 1: Advanced search for analytisch and synthetisch
In the left upper corner, we see a hierarchy of the paragraphs in which these two terms occur most frequently.
Only paragraphs are listed in which each of the terms that we searched for occurs at least once. The number of
occurrences of the terms we searched for taken together is given in the column ‘Number of Hits’. When we click at
a paragraph in the left upper corner this paragraph appears in the lower part of the screen, and the terms that we
searched for will be highlighted. This allows philosophers to quickly identify the parts of the paragraph where the
terms that she is interested in occur, and close read the text around the occurrences of these terms.
Sometimes a philosopher is writing about a concept without using his or her technical term (label) for it. This
could happen for example when a philosopher writes about a technical concept before the label for it is introduced,
or when different labels are used to designate the same concept. In our case, we want to make sure that we do not
miss any paragraphs that are about analyticity, but in which the word ‘analytic(ity)’ does not occur. Our hypothesis
here is that, in comparison to all other paragraphs of the Wissenschaftslehre, such paragraphs contain relatively
many of the same words as the paragraphs that are about analyticity and in which the term ‘analytic(ity)’ does occur.
SalVe has a similarity calculation functionality in order to trace these paragraphs. A similarity calculation for
paragraph 148, which we identified above as the one in which ‘analytisch’ and ‘synthetisch’ occur the most, shows
that paragraphs 447, 197, and 367, from all the paragraphs in the Wissenschaftslehre are the most similar to 148.
Copyright held by the author(s).
Illustration 2: Similarity calculation between paragraph 148 and the whole
Wissenschaftslehre with a threshold of 0.3
Note that 447 and 367 did not appear in the top 5 results of the advanced search for ‘analytisch’ and
‘synthetisch’. It seems hence that these paragraphs are exactly the ones that we wanted to trace with the similarity
calculation: paragraphs in which Bolzano writes a lot about analyticity and syntheticity, without using these terms
(or not often). That this application of SalVe was successful in this case is shown by the fact that the four paragraphs
148, 197, 367, and 447 are exactly the ones that de Jong found relevant for Bolzano’s conception of analyticity (de
Jong 2001: 337).
The results of a similarity calculation can be quickly visualized in a program such as Gephi. In the following
visualization of all the paragraphs of the Wissenschaftslehre with each other, we see that the paragraphs that are
about analyticity are nicely grouped together. (Recall that paragraph 294 was the 4th most similar paragraph to 148
in the above similarity calculation.) The thicker the line between two dots that represent the paragraphs, the higher
the similarity between those paragraphs, and since in this case we set a threshold of 0.3 on the results, paragraphs
that do not have a similarity value of 0.3 or higher with another paragraph do not have any lines connecting them
with others.
Copyright held by the author(s).
Illustration 3: Gephi visualisation of the similarity between all paragraphs of the Wissenschaftslehre with each other. Zoom:
Cluster of paragraphs that are about analyticity.
The advanced search and similarity calculation functionalities are useful to determine quickly which paragraphs
are useful to read closely. We might also be interested in the number of occurrences of specific terms, and the terms
that they are co-occurring with. For this, we can use SalVe’s term window functionality. How this can be useful is
illustrated below by the comparison of the term windows of analytisch and synthetisch throughout the
Wissenschaftslehre.
Copyright held by the author(s).
Illustration 4: Term windows (of 5 words) of analytisch and synthetisch in the whole Wissenschaftslehre
This is the result of searching for analytisch respectively synthetisch using a term window size 5. This means
that SalVe counts all words within a distance of five words to the left and to the right of all occurrences of analytisch
respectively synthetisch throughout the Wissenschaftslehre (stop words omitted). The comparison of these term
windows shows that the word bloss (‘mere’) co-occurs 26 times with analytisch, and only 6 times with synthetisch
(where the co-occurrence of bloss with synthetisch could also be simply a byproduct of the fact that analytisch and
synthetisch often co-occur with each other, i.e. it might be that one or more of the 6 cases of co-occurrence of
bloss/synthetisch is included within the 26 cases of co-occurrence of bloss/analytisch). We could see this as an
indication that Bolzano saw analytic truths as less important than synthetic truths, which in turn would be
quantitative evidence in favor of de Jong’s thesis that according to Bolzano, every analytic truth is grounded in a
synthetic truth. If analytic truths are grounded in synthetic truths namely, then they depend on synthetic truths, and
in this sense synthetic truths will be more important than analytic ones.
So far, we have a little evidence in favor of de Jong’s thesis, and none against it. But in paragraph 315, which we
found with the advanced search as the paragraph in which the term analytisch appears as the third most frequent, we
find Bolzano remarking the following:
(R) Some wesentliche Lehrsaetze of pure mathematics are analytic.
He gives an example of such a wesentliche Lehrsatz, namely:
(A) (a + b) + c = a + (b + c).
A natural translation of wesentliche Lehrsaetze would be ‘essential theorems’. In this context one normally
thinks of theorems as truths that allow us to prove or explain other truths. For example, (A) allows us to prove:
(E) (1 + 2) + 3 = 1 + (2 + 3).
But according to Bolzano’s definition of analyticity (E) is an analytic truth, just as (he says) (A) is: whenever we
replace both occurrences of 1 (a) and/or 2 (b) and/or 3 (c) with another number, then the result will be a true
statement. So the question is now: is the proof of (E) from (A) in Bolzano’s view an explanation? In other words,
Copyright held by the author(s).
does (A) ground (E)? Because if that is the case, then analytic truth (E) is grounded in analytic truth (A), which
5
means that de Jong’s thesis is disproved.
Here it is important to note that, for Bolzano – and this we know from research conducted with traditional
methods – the notion of explanatory proof reduces to grounding: the grounding chain for a specific claim is an
explanatory, i.e. scientific, proof for that claim. We also know that Bolzano distinguishes from grounding another
kind of proof, which is not explanatory. Such a proof makes it clear to us that something is the case, but does not
give us the reason why. For example, we can prove that it is warmer in summer than in winter by showing that well-
functioning thermometers generally indicate a higher temperature in summer than in winter, but this proof is not an
explanation according to Bolzano: it is not the case that it is warmer in summer than in winter because these
thermometers indicate a higher temperature in summer than in winter (but rather the other way around) (WL §198).
As Bolzano sees it, proofs in the second, non-explanatory sense relate truths in a subjective order, whereas
6
grounding proofs relate them in the objective order. So, the way to go to establish which of the two kinds of proof
the proof of (E) from (A) is according to Bolzano, is to figure out whether proofs from wesentliche Lehrsaetze in
general are grounding proofs. In other words: do wesentliche Lehrsaetze occur in the objective order of truths?
In answering this question using SalVe, we will make use of the knowledge (again, knowledge that we have from
traditional research) that Bolzano discusses the two ordines in two different parts of the Wissenschaftslehre (de Jong
2001: 329). What we call in the following the ordo essendi is the part of the Wissenschaftslehre in which Bolzano
discusses the objective, or grounding order of truths; what we call the ordo cognoscendi is the part in which Bolzano
discusses the subjective or non-grounding order of truths. Bolzano uses the term ‘(wesentliche) Lehrsatz’ in
alternation with ‘(wesentliche) Lehre’, so in the queries that we did for answering the current question, we search for
both terms (see e.g. WL§592).
On the basis of this previous qualitative knowledge obtained with traditional methods, we now make the
quantitative step to SalVe. In the following line chart, the word Lehre is represented as it occurs throughout the
Wissenschaftslehre. The black line distinguishes the ordo essendi (left), from the ordo cognoscendi (right). (The
utmost left part should be left out of consideration, since this is the introduction and Bolzano discusses both ordines
here.)
5
In this paper we deliberately disregard (and keep for another occasion) an alternative, purely qualitative line of argument based on a
suggestion by de Jong to the effect that (A) is, in fact, not analytic. We also limit exegetical information and references to the philosophical
literature on Bolzano to the absolute minimum needed to expound the results of the Phil@scale project.
6
This is a bit quick. Although Ableitbarkeit is, in itself, by no means an epistemic notion relating to knowledge subjects, when Bolzano speaks
of objective vs. subjective order, he tends to associate ‘objectiv’ (as things stand) with grounding terminology and ‘subjectiv’ (as things are
known to us) with deducibility terminology. A search for subjectiv ableitbar in SalVe yields the following two passages: “Insonderheit also,
wenn wir ihn für den obersten Grundsatz der ganzen Wissenschaft in objectiver oder nur subjectiver Hinsicht erklären: so müssen wir
darthun, daß wirklich alle Wahrheiten, welche nach dem Begriffe unserer Wissenschaft in ihr Gebiet gehören, aus diesem Satze objectiv
folgen, oder subjectiv ableitbar sind.“ (WL §489); “weil sie aus einem und demselben Obersatze entweder nur subjectiv ableitbar sind, oder
auch wohl objectiv, d.h. sich wie die Folgen aus ihrem (Theil-) Grunde ergeben.” (WL §420).
Copyright held by the author(s).
Illustration 5: Line chart of lehre throughout the Wissenschaftslehre
We immediately see that Lehre occurs far more, both absolutely and relatively, in the subjective order of truths.
We see the same result when we consider the term window of Lehre:
Illustration 6: Term window of lehre
What we see here, furthermore, is that many of the words that co-occur with Lehre are terms that are related to
the way we come to know truths, hence to the subjective order of truths, such as gewiss, Buch, and Lehrbuch. The
same queries for Lehrsatz gave similar results, which we cannot all show here out of considerations of space. We
will lastly show the occurrences of Lehrsatz compared to these of Abfolge (i.e. Bolzano’s term for the grounding
relation) throughout the Wissenschaftslehre:
Copyright held by the author(s).
Illustration 7: Line chart of lehrsatz and abfolge
Clearly, Lehrsatz and Abfolge have a different distribution: the former occurs more frequently in the parts where
Bolzano discusses the ordo cognoscendi, whereas the latter occurs more where he discusses the ordo essendi (the
search functionality revealed that the absolute occurrence of Lehrsatz throughout the Wissenschaftslehre is 139, and
that of Abfolge is 113).
We feel justified from this to form the hypothesis that Bolzano saw the role of wesentliche Lehrsaetze primarily
within the subjective order of truths. And consequently, it seems that Bolzano did not hold that the analytic (a + b) +
c = a + (b + c) grounds (1 + 2) + 3 = 1 + (2 + 3). Bolzano’s remark that some wesentliche Lehrsaetze are analytic
hence does not seem to provide a counterexample to de Jong’s claim. All in all, the quantitative information about
the Wissenschaftslehre that SalVe gave us, allowed us to give evidence in favor of de Jong’s thesis.
4 Evaluation and future plans
What did the above application of SalVe to a concrete philosophical research question show about the use of
computational techniques in philosophy? We saw that the advanced search and similarity calculation allowed us to
determine quickly which paragraphs of the Wissenschaftslehre are relevant with respect to the question that we are
posing. This clearly contributes to our aim to make it possible to do philosophical research faster. Besides this, these
functionalities point researchers in philosophy to connections between words and parts of the corpus that might not
have come to the fore only by means of philosophical research done in the traditional way, that is, only by close
reading of the text. In our case we were considering a book of 2500 pages, so it is easily imaginable that some parts
of the text that are relevant to our research question escape our attention. Application of SalVe increases the chance
that we find all relevant passages. Hence, not only does the application of the SalVe to philosophy sources contribute
to the speed in which philosophical research can be done, but also to the thoroughness of the research.
The term window and (the line chart of) the word occurrences offer quantitative data that count as evidence for a
particular hypothesis. In this way SalVe contributes to our second aim to make philosophical research more
quantitative and objective. In conclusion, SalVe offers a welcome complementary improvement to the traditional
way in which philosophical research is done. SalVe makes it possible to do philosophical research faster and makes
this research more quantitative and objective, at least for corpora of a certain size. Importantly, SalVe facilitates
close reading of the text, and does not restrict in any way what is important for philosophical research. Therefore,
the application of SalVe to this case shows that computational techniques can be fruitfully applied within
philosophy.
We find that it is important for philosophy to develop computational techniques further. One philosophical
discipline for which these techniques can offer substantial improvement is the discipline that is called history of
ideas. This discipline investigates the development of (technical) concepts in philosophical texts throughout history.
Ideally, researchers in this area investigate the works of several philosophers through long periods of time, or
Copyright held by the author(s).
7
compare the works of entire schools. This requires tools apt to proper exploration of massive amounts of machine-
readable historical texts in philosophy.
We have experienced that the realization of the latter endeavor comes with several problems that are
characteristic for this field. To begin with, at the present time there is little material digitally available. Furthermore,
the texts that a researcher in history of ideas is concerned with comprise multiple languages, several scripts, and are
8
unstandardized as to e.g. format, citations, and references. Finally, the concepts that are to be analyzed in this
research are of extremely high complexity, which requires particularly challenging semantic analyses. Addressing
all these problems in a satisfactory way will be vital to the growth of the emerging field applying computational
tools to philosophy.
References
Betti, A. & Van den Berg, H. (2015). Creating a Digital History of Ideas [conditionally accepted]. Transactions in
Digital Humanities, this volume.
Bolzano, B. (1969/1837). Wissenschaftslehre. In: J. Berg and E. Winter (eds.), Bernard Bolzano Gesamtausgabe,
Reihe 1, Bd. 11-14. Stuttgart Bad-Cannstatt: Frommann-Holzboog.
Jong, W.R. de (2001). Bernard Bolzano, Analyticity and the Aristotelian Model of Science. Kant-Studien, 92 (3),
328–349.
Rey, Georges (2013). The Analytic/Synthetic Distinction. In: The Stanford Encyclopedia of Philosophy (Fall 2013
Edition), Edward N. Zalta (ed.), retrieved from http://plato.stanford.edu/archives/fall2013/entries/analytic-synthetic/.
Wierst, P.M.A. van (2013). Salva Veritate – A master thesis on Bolzanian analyticity and computational methods
within philosophical research, MA thesis, Faculty of Philosophy, VU University Amsterdam.
Biographies of the authors
Arianna Betti is Professor of Philosophy of Language at the University of Amsterdam. After studying historical and
systematic aspects of ideas such as axiom, truth and fact (Against facts, MIT Press, 2015), she is now trying to trace
the development of ideas such as these with computational techniques.
Stefan Schlobach is Associate Professor at VU University Amsterdam. He focuses on using nonstandard techniques
and semantics for reasoning and querying, scalable reasoning through approximation and parallelization, and
reasoning services for ontology languages such as mapping, explanation, or abduction. He has been involved in
several research projects on computational methods in the Humanities, in particular in Philosophy and Social
History.
7
See Betti & Van den Berg (2015) in this very volume.
8
Currently, solving this problem is at the core of Betti’s project @PhilosTEI, see also footnote 2 above.
Copyright held by the author(s).
Sanne Vrijenhoek is currently graduating at the VU University Amsterdam in Artificial Intelligence with an
interest in Natural Language Processing. She focuses on applying well-tested techniques from her field of study in
the humanities, with the aim of increasing synergy between the two.
Pauline van Wierst graduated in August 2013 at the VU University Amsterdam with an experimental thesis on the
notion of analyticity in Bolzano and computational methods within philosophical research under Arianna Betti’s
supervision. She is currently a graduate student at the Scuola Normale Superiore, Pisa.
Acknowledgements
Work on this paper has been funded by the projects Phil@Scale (van Wierst & Vrijenhoek), Phil@Scale 2 (van
Wierst), SalVing GlamMap (Vrijenhoek), ERC Starting Grant Tarski’s Revolution #203194 (Betti, van Wierst).
Betti has been further supported by the ERC Proof of Concept GlamMap grant #324630 and CLARIN-NL-12-006
project @PhilosTEI. Van Wierst has been further supported by the Prins Bernhard Cultuurfonds, the Fundatie van
de Vrijvrouwe van Renswoude te ‘s Gravenhage and by a DHLU 2013 travel grant that enabled her to travel from
US to deliver the talk which is at the basis of this paper.
Copyright held by the author(s).