Causal Attributions and Corpus Analysis

Abstract

Although philosophers have often held that causation is a purely descriptive notion, a growing body of experimental work on ordinary causal attributions using questionnaire methods indicates that it is heavily influenced by normative information. These results have been the subject of sceptical challenges. Additionally, those who find the results compelling have disagreed about how best to explain them. In this chapter, we help resolve these debates by using a new set of tools to investigate ordinary causal attributions—the methods of corpus linguistics. We apply both more qualitative corpus analysis techniques and the more purely quantitative methods of distributional semantics to four target questions: (a) Can corpus analysis provide independent support for the thesis that ordinary causal attributions are sensitive to normative information? (b) Does the evidence coming from corpus analysis support the contention that outcome valence matters for ordinary causal attributions? (c) Are ordinary causal attributions similar to responsibility attributions? (d) Are causal attributions of philosophers different from causal attributions we find in corpora of more ordinary language? We argue that the results of our analyses support a positive answer to each of these questions.

Full text

1
Causal Attributions and Corpus Analysis1
Justin Sytsma, Roland Bluhm, Pascale Willemsen, and Kevin Reuter
Abstract: Although philosophers have often held that causation is a purely descriptive notion, a
growing body of experimental work on ordinary causal attributions using questionnaire methods
indicates that it is heavily influenced by normative information. These results have been the
subject of sceptical challenges. Additionally, those who find the results compelling have
disagreed about how best to explain them. In this chapter, we help resolve these debates by using
a new set of tools to investigate ordinary causal attributions—the methods of corpus linguistics.
We apply both more qualitative corpus analysis techniques and the more purely quantitative
methods of distributional semantics to four target questions: (a) Can corpus analysis provide
independent support for the thesis that ordinary causal attributions are sensitive to normative
information? (b) Does the evidence coming from corpus analysis support the contention that
outcome valence matters for ordinary causal attributions? (c) Are ordinary causal attributions
similar to responsibility attributions? (d) Are causal attributions of philosophers different from
causal attributions we find in corpora of more ordinary language? We argue that the results of our
analyses support a positive answer to each of these questions.
1. Introduction
Most studies in experimental philosophy have used questionnaires involving vignettes. There are
good reasons for the prevalence of questionnaire methods in experimental philosophy, including
that these methods are fairly easy to use and are well-suited to investigating many of the
philosophical questions that have been asked. As the present volume amply illustrates, however,
questionnaire methods are not the only methods available to experimental philosophers, nor are
they the only ones that experimental philosophers have used. In this chapter we will offer a brief
introduction to a powerful set of non-questionnaire methods that can aid experimental
philosophers in investigating a wide range of questions—methods of corpus linguistics.2
1 To appear in Methodological Advances in Experimental Philosophy, edited by Eugen Fischer and published by
Bloomsbury Press. We would like to thank an anonymous reviewer for the many insightful comments, as well as
Eugen Fischer for his philosophical and administrative support.
2 Although a handful of philosophers have used different tools of corpus linguistics (for a brief overview, cf. Bluhm
2016), and although the interest in the approach seems to have increased recently (e.g., Reuter 2011, Bluhm 2012,
2013, Hahn et al. 2017, Fischer and Engelhardt 2017, Sytsma and Reuter 2017, and the contribution by Mejia et al. in
this volume), the use of corpora is still marginal to philosophy.

2
Our primary goal in this chapter is to introduce experimental philosophers to working with
corpora, to survey some of the tools available, and to demonstrate how these tools can
complement the use of questionnaire methods. Toward this, we will put some of these tools to use
in an area of research that has seen a flurry of interest in recent years—investigations of the effect
of norms on ordinary causal attributions. Specifically, we focus on four questions: (a) Can corpus
analysis provide independent support for the thesis that ordinary causal attributions are sensitive
to normative information? (b) Does the evidence coming from corpus analysis support the
contention that outcome valence matters for ordinary causal attributions? (c) Are ordinary causal
attributions similar to responsibility attributions? (d) Are causal attributions of philosophers
different from causal attributions we find in corpora of more ordinary language? We argue that
the results of our analysis provide evidence for a positive answer to each of these questions.
Here is how we will proceed. In Section 2, we will briefly discuss recent work in
experimental philosophy on ordinary causal attributions, highlighting our four questions. In
Section 3, we introduce corpus linguistics. In Section 4, we bring corpus analysis methods to bear
on our target questions. In Section 5, we use methods of distributional semantics to support the
previous analyses. We conclude with some general methodological advice concerning the
integration of corpus analysis techniques into experimental philosophy and philosophy as a whole.
2. Ordinary Causal Attributions and Injunctive Norms
Philosophical discussions of causation are often concerned with what has been termed actual
causation. Actual causation is usually contrasted with general causation. A general causation
statement describes a law-like relation between two types of events that stand in a causal relation,
such as ‘smoking causes cancer’ or ‘throwing rocks at windows causes them to break’. An actual

3
causal statement, in contrast, describes the relation between two event tokens, such as ‘Peter’s
smoking caused his lung cancer’ or ‘Jenny’s throwing the rock caused the window to break’. For
both general and actual causation, most philosophers assume that the concept of causation is a
purely descriptive notion, refering to a relation in the world. As a consequence, a causal
attribution such as ‘A caused B’ is true if and only if the relation of causation holds between A
and B. Such an understanding of causation, however, means that normative considerations are
irrelevant to causal attributions. The basic idea here is that whether or not an action is permitted
by morality or convention simply does not matter for purposes of assessing whether that action, or
the entity carrying it out, caused the outcome. Similarly, whether an action causes a morally good
or bad outcome is irrelevant for causal considerations. Call this the standard view on causation.
Against the standard view, a growing body of empirical findings indicates that ordinary
causal attributions are sensitive to normative information, prominently including injunctive norms
(Hilton and Slugoski 1986, Alicke 1992, Knobe and Fraser 2008, Hitchcock and Knobe 2009,
Sytsma et al. 2012, Reuter et al. 2014, Kominsky et al. 2015, Livengood et al. 2017a).3 Injunctive
norms include both prescriptive norms, which tell people what they should do, and proscriptive
norms, telling people what they should not do.4 Moral norms, such as the impermissibility to kill
or hurt others, are a prime example of injunctive norms, but there is also a variety of non-moral
norms that have similar action-guiding functions, such as social rules and regulations, etiquette
norms, and so on.5
3 These results do not directly challenge the standard view. Rather they put pressure on it insofar as philosophers are
committed to what Livengood et al. (2017a) call the folk attribution desideratum. The folk attribution desideratum
asserts that a key measure of the acceptability of an account of actual causation is that the verdicts it issues about
specific cases line up with ordinary causal attributions about those cases. And there is reason to think that many,
perhaps most, philosophers working on causation are committed to this desideratum.
4 It should be noted that in the recent literature in experimental philosophy of causation, ‘prescriptive norm’ is often
used indiscriminately to refer to both prescriptive norms and proscriptive norms as we understand them.
5 Injunctive norms can be distinguished from descriptive norms (or ‘statistical norms’). While there is an ongoing
debate among experimentalists about whether descriptive norms have an independent effect on ordinary causal

4
Here are a couple of the empirical findings that have received much attention in the
literature. Knobe and Fraser (2008) presented people with a story in which a secretary keeps her
desk stocked with pens, and both administrative assistants and faculty members help themselves
from this stock. However, faculty members are not supposed to take pens. One day, both
Professor Smith and the administrative assistant take a pen. Later that day, the secretary has no
pen left to take an important message. Who caused the problem? In this case, Professor Smith and
the administrative assistant performed symmetric actions (each took a pen), jointly leading to a
bad outcome. The key difference between them is that while Professor Smith violated an
injunctive norm (faculty members are not supposed to take pens), the administrative assistant did
not (administrative assistants are allowed to take pens). Despite the two agents performing
symmetric actions, participants were significantly more likely to agree that Professor Smith, the
norm-violating agent, caused the problem than that the administrative assistant did.
To make matters more interesting, in a follow-up study, Sytsma et al. (2012) tested what
happens if you remove the injunctive norm from the Pen Case, such that both Professor Smith and
the administrative assistant are allowed to take pens. They found that participants now tended to
disagree that Professor Smith caused the problem, while continuing to deny that the
administrative assistant caused the problem.
Livengood et al. (2017a) used a computer case scenario, for which they found the same
effects. More specifically, their studies revealed that participants were significantly more likely to
agree that an agent who violated a norm caused a bad outcome, compared to the norm-conforming
agent. Agreement that the norm-conforming agent caused the bad outcome was significantly
attributions (e.g., Knobe and Fraser 2008, Sytsma et al. 2012, Livengood et al. 2017a), we will focus on injunctive
norms in this chapter.

5
below the neutral point, while agreement for the norm-violating agent was significantly above the
neutral point. Similar results were found using a between-participants design.
While the Pen Case and the Computer Case are probably two of the most prominent
examples in the literature, similar effects have repeatedly been found in subsequent research, and
they seem to be robust for different causal structures (Kominsky et al. 2015, Sytsma et al. ms,
Livengood and Sytsma under review), for both actions and omissions (Henne et al. 2015,
Willemsen & Reuter 2016, Willemsen 2016), and across multiple test queries (Livengood et al.
2017a, Livengood and Sytsma under review).
Several explanations of the relevance of injunctive norms have been proposed in the
literature. The most fundamental dispute regards the question of whether the observed effects
reveal a real effect of injunctive norms on causal attributions. In two recent papers, Samland and
Waldmann (2014, 2016) have denied this. According to their alternative explanation, when
participants answer the question in studies like those noted above, they do not read the questions
as being about causation, but about a related notion such as accountability or responsibility.
The researchers who are convinced that norms do affect causal attributions have offered a variety
of explanations of why and how this effect occurs. In the following, we will focus on two specific
explanations, the norm violation and the responsibility account, as they make empirical
predictions that we believe can be effectively tested with help of corpus analyses.6
The norm violation account put forward by Hitchcock and Knobe (2009) holds that the
effects of norms in cases like we saw above are best explained in terms of the cognitive processes
that lead to causal attributions being sensitive to norms. According to Hitchcock and Knobe’s
6 While we will focus on these two accounts here, it should be noted that these are not the only two plausible
explanations in the literature nor are they the only two worth discussing. Just a few notable examples are the work of
Alicke 1992, Cushman 2013, Malle et al. 2014, and Reuter et al. 2014.

6
account, causal judgements serve to identify suitable points for intervention in a system, and
norms come into play because they affect which counterfactuals are most salient for determining
the suitability of different intervention points. The basic idea is that in considering a situation,
people think about how the outcome could have been prevented. But they do not consider every
way in which the outcome might have been prevented; rather, they focus on those aspects of the
situation in which something abnormal (i.e., counter-normative) has occurred. As such, while the
norm violation account holds that the evaluation of norms is a crucial component in causal
cognition, this does not mean that the concept of causation at play in ordinary causal attributions
is a normative concept. Instead, norms come into play when people attempt to identify suitable
intervention points in normatively-laden situations.
The responsibility account put forward by Sytsma et al. (2012) contends that the concept
of causation at play in ordinary causal attributions is itself normative. Thus, the cognitive process
of making causal attributions starts off from a normatively-laden concept. Instead of making a
purely descriptive causal judgement that is later tainted by norms, the causal judgements are
already normative. Sytsma and Livengood (2018, 7–8) express the idea this way:
Saying that an agent caused an outcome… typically serves to indicate something more
than that the agent brought about that outcome or that the agent’s action produced that
outcome. Rather, it serves to express a normative evaluation that can be roughly
captured by saying that the agent is responsible for that outcome or that the agent is
accountable for that outcome, whether for good or for ill.
The norm violation account and the responsibility account make a number of diverging
predictions. One such prediction is that while Hitchcock and Knobe hold that normative
considerations have the same effect on causal attributions regardless of the outcome valence
(regardless of whether the outcome is good or bad), the responsibility account allows that
outcome valence might often make a difference.

7
As we noted above, Hitchcock and Knobe explain the influence of norms on causal
attributions exclusively in terms of shaping the counterfactuals that are considered. And whether
the outcome is good or bad would not seem to be directly relevant to assessing whether a
counterfactual in which a candidate cause did not occur was more normal than what actually
happened. Thus, Hitchcock and Knobe argue that to assess Alicke’s (1992) competing account,
which they read as explaining the influence of norms in terms of the desire to assign blame for the
outcome, what is needed are cases in which a norm is violated and yet where no one is assigned
blame because the outcome is good. They write that for such cases their account ‘suggests that the
impact of normative considerations should remain unchanged (because people still see that a
norm has been violated)’ (2009, p. 603; emphasis added).
The responsibility account, by contrast, does not make a direct prediction about the role of
outcome valence in ordinary causal attributions; rather, it makes a prediction when coupled with a
plausible prediction about responsibility attributions—that people are more likely to assert that an
agent is responsible for a bad outcome than a good outcome.7 If this is correct, and if causal
attributions are relevantly akin to responsibility attributions, then we would expect that outcome
valence will often make a difference.
One way to make progress on the issues concerning the role of norms on ordinary causal
attributions that we have surveyed in this section would be to run still more questionnaire studies.
However, we want to suggest that there is also benefit in approaching the problem from another
angle. What we aim to demonstrate in this chapter is that there is another source of evidence that
7 We find this plausible because we expect that people are generally more concerned with assigning blame than
praise, as a number of philosophers have noted. For instance, Prinz (2007, 79) writes: “We blame someone for
stealing, but we don’t issue a medal when he refrains from stealing. We don’t lavish the non-pedophile with praise for
good conduct. In other words, we tend to expect people to behave morally.” While we will not argue for the veracity
of this prediction here, it does find some support in the corpus analyses detailed below.

8
can be brought to bear on these questions, namely corpus linguistics, and that its methods can
both complement and enhance the use of questionnaire studies. After offering a brief introduction
to corpus linguistics and applying some of its methods to the domain of ordinary causal
attributions, we will return to the use of questionnaire studies and discuss potential shortcomings
of such studies and how they can be alleviated through the use of corpus analysis.
3. The Basics of Corpus Analysis
Corpus linguistics is a discipline of linguistics that is defined by its use of methods of corpus
analysis. In its most basic sense, the term ‘corpus’ simply refers to ‘a collection of texts’
(Kilgarriff and Greffenstette 2003, p. 334) and ‘analysis’ to the process of looking at the linguistic
data that the corpus contains and assessing it for some research purpose.8
There are, briefly summarized, three sources of this method (cf. McCarthy and O’Keeffe
2010a). Its historical roots can be traced back as far as the Middle Ages, when concordances of
Biblical words and phrases in context were compiled to assist exegesis. One of the basic functions
of present-day corpora still is—particularly important for qualitative assessment of corpus data—
to provide listings of queried linguistic expressions in context, in much the same mode of
presentation as the one that medieval concordances used. The second important source of corpus
analysis is the recognition of the importance of data representing actual use of a language, as
opposed to data generated by the linguists themselves.9 The third factor driving the development
of corpus linguistics—particularly important for quantitative analyses of corpus data—is the fast-
paced development and spread of computer technology and the Internet in the late 20th century.
8 Helpful overviews of the discipline are given by Biber et al. 1998, McEnery and Wilson 2001, and McCarthy and
O’Keeffe 2010b. For a quite comprehensive collection of articles on many aspects of corpus linguistics, see Lüdeling
and Kytö 2008–2009.
9 See Leech 1992.

9
Thus, while all corpora used in corpus linguistics basically are ‘collections of texts’, present-day
corpora typically are collections of digitalized texts that are accessed with computers.
Paradigmatic examples of well-known, large and freely accessible English-language
corpora are the British National Corpus (BNC) and the Corpus of Contemporary American
English (COCA).10 The World Wide Web also offers a rich repository of digital texts, and while it
is somewhat problematic to simply use the web as a corpus (termed WaC in the literature), there
are some ways to tap into the Web’s wealth of data by using extracts of it for building a corpus
(termed WfC in the literature). We will make use both of COCA and a WfC approach below.11
One of the aspects that make a corpus out of a mere collection of texts is the decision to
look for evidence of some use of linguistic expressions in it. The access to the data granted by the
search engine is therefore not at all marginal to corpus analysis. Every run-of-the-mill search
engine can do a full text search and handle wildcards, that is, it is possible to execute a query with
a search string (a sequence of alphanumeric characters) in which some letter or letters are
substituted with a variable. For example, ‘cause*’ will not only find all instances of the use of
‘cause’ as a verb and a noun in the corpus, but also tokens of ‘causes’, ‘caused’, and more
unexpected words such as ‘causeway’ and ‘causer’. A more sophisticated corpus is pre-analysed
and annotated with linguistic information, allowing, for example, to specifically search for the
lemma ‘cause’. This means that all instances of the root word ‘cause’ regardless of inflection. It
further allows to find tokens by grammatical category, e.g., only instances of the verb ‘causes’ in
its third person singular form, instead of the noun in its plural form; or co-occurrences of
expressions, e.g., the lemma ‘cause’ together with some noun (within a specified distance). Such
10 The number and variety of corpora compiled by linguists is ever-growing. Xiao 2008 and Lee 2010 give an
overview of extant corpora, usefully sorted by type.
11 The useful terms ‘WaC’ and ‘WfC’ go back to de Schryver 2002. For a brief introduction to the rapidly growing
field of ‘web linguistics’, you may turn to Bergh and Zanchetta 2008.

10
queries obviously are much more powerful than mere full text search. They are indispensable if a
pertinent linguistic expression cannot be specified by a definite search string, or if, as in our case,
the relevant linguistic phenomena include phrases such as ‘responsible for’ or ‘caused’ followed
by some noun.
4. Exploring Causal Attributions with Corpus Analysis
We have already hinted at some very basic search options offered by common corpus search
engines. Generally speaking, there are two approaches to using corpora (cf. Biber 2010): Corpus-
driven research uses corpora to generate theories on linguistic phenomena from bottom up.
Accordingly, the corpus is approached with minimal hypotheses as to the linguistic forms relevant
to a given research question, e.g., searching for tokens of ‘cause’ as a starting point to develop an
understanding of causal attribution language. Corpus-based research, on the other hand, uses
corpora to verify or falsify extant hypotheses about the use of language based on available
theories about linguistic forms, e.g., trying to confirm that ‘cause’ has some specific collocations.
In practice, these approaches overlap to some extent, with researchers switching back and forth
between them in their research process.
Similarly, corpora can be approached in a qualitative manner, that is, focusing on
interpreting corpus findings with respect to meaning, or a quantitative manner, focusing on
analyses based on countable objects and statistical facts. Both the methods used in this section and
the next would count as quantitative on this definition. However, despite relying on frequency
counts, the methods employed in this section have a somewhat more qualitative aspect to it, while
the methods employed in the next are decidedly more quantitative, as will be apparent.

11
All the approaches to corpora we have hinted at depend on the existence of linguistic
phenomena that can be traced with the help of available search engines. But in studying the
language of causation, broadly construed, it is not immediately clear which linguistic expressions
to look for. Linguists have identified an astonishing number and variety of ways that arguably are
used to express causal relations.12 Apart from the verb ‘to cause’ and the noun ‘cause’, as well as
(partially) synonymous expressions, there are conjunctions for the subordination of clauses like
‘because’, ‘since’, ‘as’ (cf. Altenberg 1984, Diessel and Hetterle 2011), but also causative verbs,
adverbs, adjectives and prepositions (cf. Khoo et al. 2002), and no exhaustive list of such means
to express causal connection is available. There are also linguistic means to express a causal
relation without lexical means, e.g., the coordination of sentences and text organization (cf.
Altenberg 1984, Achugar and Schleppegrell 2005). In consequence, it is only possible to find
some, but not all instances of causal language in a corpus with the help of a search engine.
Moreover, most of the expressions mentioned above do not only serve to express causal relations,
but may also be used differently. To give but one example, ‘cause’ may also refer to a concern or
purpose, as in ‘her cause was just’.
With these caveats in mind, it is, of course, possible to access some of the causal language
contained in a corpus. In our present context, we are interested in similarities and differences of
the language of causal attributions and the language of responsibility attributions, and it seems
plausible to approach our linguistic study with a focus on ‘cause’ as a relational verb and the
phrase ‘is responsible for’, matching the type of phrases used in questionnaire studies to elicit
12 We need not pass judgement at this point on whether such utterances really are intended to express or really do
refer to a causal relation of some kind, let alone whether they express the specific relation at issue for causal
attributions as we have defined them.

12
causal attributions (e.g., ‘Lauren caused the system to crash’, ‘Marcy is responsible for the death
of the bystander’).13
In the second section of this chapter, we surveyed some recent studies on ordinary causal
attributions. These studies suggest that injunctive norms play a substantial role. Various theories
have been proposed to account for this effect. We focused on two of these—the intervention
account and the responsibility account. According to the intervention account, while ordinary
causal attributions are influenced by norms, the underlying concept of causation is descriptive and,
thus, diverges notably from the concept of responsibility. Importantly for our research purposes,
only norms, but not the valence and the severity of the outcome are said to affect causal
attributions. From that we can infer the empirically testable prediction that the language of causal
attributions and the language of responsibility attributions should be clearly distinguishable. In
contrast, the responsibility account holds that the language of causal attributions and the language
of responsibility attributions are quite similar. Moreover, the responsibility account predicts that
the valence of the outcome will often have a notable effect on causal attributions.
If the ordinary concept of causation was a purely descriptive concept, then we would have
no a priori reason to expect it to be used disproportionately in contexts with any particular
valence. Rather, we would expect ‘cause’ to be used indiscriminately in the contexts where the
outcome is good, in contexts where the outcome is bad, and contexts where the outcome is
neutral. And if this was the case then we should see a good mix of positively and negatively
connotated causal expressions without one of these types of expression dominating people’s use
of ‘cause’. This is not what we find, however.
13 The phrase ‘is responsible for’ is used to cue responsibility attributions in studies in Sytsma and Livengood (2018).

13
In order to investigate the nature of the terms most commonly used when expressing a
causal statement, we looked at the most frequent nouns appearing after the causal phrase ‘caused
the’ using the Corpus of Contemporary American English (COCA). The ten most frequently used
nouns (numbers in brackets indicate the number of hits) are ‘death’ (103), ‘accident’ (87), ‘crash’
(80), ‘problem’ (79), ‘explosion’ (47), ‘fire’ (46), ‘collapse’ (27), ‘injury’ (26), ‘damage’ (24),
and ‘loss’ (23). Independent raters classified all of these terms as negative.14 Of the top 50 nouns,
30 were classified as negative, 19 neutral, and only one positive. This data supports the results
from questionnaire studies indicating the relevance of norms to ordinary causal attributions, as
discussed in Section 2. Furthermore, the commanding presence of negative terms strongly
suggests that ‘cause’ is not only in part normative, but also primarily directed at negative
outcomes. In other words, the results of our analysis strongly indicate that outcome valence has a
substantial effect on ordinary causal attributions.
A key objection against our interpretation of this corpus data might be raised at this point.
First, one important source of the corpus we have used is newspaper articles. And newspaper
articles are notorious for focusing on negative events. Accordingly, it would not be surprising to
find statements about causal relations for which the outcome is often negative. However, when
we limited our search to sources of other types, like fiction, as is possible in COCA, no
differences were found. E.g., the top ten list of nouns following the phrase ‘caused the’ for spoken
language only were: ‘death’ (41), ‘accident’ (40), ‘crash’ (38), ‘fire’ (30), ‘problem’ (30),
14 Three independent raters were given a prompt—‘Please classify each of the following items based on whether you
think instances of this type are most often positive, negative, or neutral?’—followed by 780 items for classification.
Items were the top 50 hits for ‘caused the’, the top 50 hits for ‘responsible for the’, and the top 20 hits for each of the
eight synonymous expressions used below. This produced a list of 260 items that were then presented to each rater in
three randomized orders. For the ten most frequently used nouns just reported, there was 99.4% agreement across the
classifications with 169 out of 170 occurrences of these items being classified as negative. Overall, inter-rater
agreement was high with a Kendall’s tau of 0.751 and Spearman’s rho of 0.803 averaging across the values for each
possible pair of raters. For subsequent classifications we treated a term as negative (positive) if it was classified as
negative (positive) at least two-thirds of the time.

14
‘explosion’ (26), ‘plane’ (11), ‘damage’ (9), ‘recession’ (9), ‘collapse’ (8).15 And a similar list
obtained when the corpus was restricted to academic texts.
It might be offered in rejoinder that a focus on the negative is simply a part of the human
condition. As such, it might be suggested that terms that are arguably synonymous with ‘cause’
will also tend to be used in negative contexts. If that were true, then we could not conclude that
we have discovered a specific characteristic of the language of causal attributions, but rather a
more general phenomenon, for which an entirely different explanation would seem to be required.
In order to investigate this objection, we posited the following null hypothesis:
Synonymy Effect: There is no significant difference in the normative use between ‘cause’
and synonymous expressions.
If ‘cause’ is indeed specific in being used in a predominantly normative way, then we should be
able to falsify Synonymy Effect. To test the hypothesis, we executed a corpus search with the
eight terms that are listed by the English Thesaurus as being often used synonymously with the
verb ‘to cause’: ‘create’, ‘generate’, ‘induce’, ‘lead to’, ‘make’, ‘precipitate’, ‘produce’, and
‘provoke’. We inserted the phrase ‘Φed the’ and noted the 20 most frequent nouns that appear
after that phrase for all eight synonymous terms. (Table 1 lists those terms that were rated
negatively for ‘caused the’ as well as the synonymous phrases.)
15 Only ‘plane’ was classified as neutral, with 160 out of 170 occurrences of these ten terms being classified as
negative by our raters.

15
phrase number of
negative terms
(out of 20)
negative terms
caused the 16 death, accident, crash, problem, collapse, injury,
damage, loss, crisis, destruction, decline,
extinction, harm, demise, explosion, fire
created the 3 proble
m
, need, illusio
n
generated the 3 waste, wa
r
, killin
g
induced the 4 coma, panic, defendant, oppositio
n
led to the 7 death, arrest, collapse, demise, loss, firin
g
, en
d
made the 2
m
istake, cut
precipitated the 13 crisis, war, attack, decline, conflict, collapse,
downfall, fight, invasion, violence, demise, end,
split
produced the 1 plutoniu
m
provoked the 9 anger, fight, violence, murder, resignation,
rebellion, strike, crisis, evacuatio
n
Table 1: Most frequent negatively connotated nouns after phrases synonymous to ‘caused the’
Considering only the 20 most frequent nouns, we calculated whether there was any
significant difference in the use of ‘caused the’ compared to synonymous expressions. Pearson’s
chi square tests revealed that only ‘precipitated the’ was not significantly different (χ2(1.13, 1) =
0.288). All other comparisons were highly significant: ‘provoked the’ (χ2(5.23, 1) = 0.022), ‘led
to the’ (χ2(8.29, 1) = 0.004); p < 0.001 for all other phrases.
The results are noteworthy in a couple of respects. (a) The searches demonstrate that the
COCA corpus is not—at least not strongly—tilted towards texts that feature negative events. (b)
The Synonymy Effect is likely to be false. Seven out of eight synonymous expressions of the
form ‘Φed the’ are not only significantly different in their most frequent uses compared to ‘caused
the’, most of the synonymous terms seem to be used mainly in a neutral fashion. This means that
the effect we recorded for ‘cause’ seems to be rather specific.
We have seen that Sytsma and Livengood propose that causal attributions are inherently
normative, being used to assign responsibility. If this is correct, then we would expect

16
responsibility attributions to be similar to causal attributions, including that they should also tend
to occur more often in negative contexts. To test this expectation, we ran the same corpus search
as before, this time entering the phrase ‘responsible for the’ and recorded the most frequent nouns
that occur after that phrase. The ten most frequent are: ‘death’ (130), ‘attack’ (46), ‘murder’ (44),
‘actions’ (43), ‘safety’ (42), ‘development’ (42), ‘loss’ (35), ‘design’ (34), ‘decline’ (31),
‘violence’ (31).16 Of the 50 most frequent nouns occurring after ‘responsible for the’, 19 terms
were rated negative, 17 neutral, and 14 positive. It should be further noted, however, that many of
the positive terms seem to belong to an alternative sense of ‘responsible’ from the responsibility
attributions we are after—a sense indicating the duties involved in a role (e.g., ‘content’,
‘creation’, ‘design’, ‘implementation’, ‘safety’, ‘security’).
Nonetheless, the results show that ‘responsible for the’ has a similar environment to
‘caused the’ in being normatively laden and more often directed at negative events than positive.
And by looking in greater detail at the respective numbers of hits for various terms, we found
more striking similarities. For many terms like ‘death’, ‘decline’, ‘damage’, ‘destruction’, ‘crisis’,
the use of responsibility language is roughly as frequent as causal language, suggesting that at
least for some terms, both phrases might be used interchangeably (Table 2 below lists the number
of hits for these terms as well as the frequency ratio). This, of course, would provide further
support for the hypothesis that the causal attributions and responsibility attributions are often used
to express the same state of affairs.
However, other comparative results between ‘responsible for the’ and ‘caused the’ do not
quite match, which might suggest that we only cherry-picked the data that fits our hypothesis.
16 Every occurrence of six of these terms was classified as negative by our raters. Every occurrence of ‘safety’ and
‘design’ was classified as positive. Classifications for ‘actions’ and ‘development’ were split, although they were
positive overall. In total, 154 out of 198 occurrences of these ten terms were classified as negative.

17
Table 2 lists two terms (‘attack’, ‘murder’) which are far more commonly used with responsibility
language than causal language, e.g., people seem to be far more likely to say ‘she is responsible
for the attack’ than ‘she caused the attack’. An explanation is easy to give: when we want to
express a causal relation between a person and an attack or a murder, we would usually just rely
on the causative aspect of the verbs and say that ‘s/he attacked’ or ‘s/he murdered’. In other
words, the English language has a simpler means to express causal language when it comes to
attack and murder. The opposite result was found for the terms ‘problem’ and ‘accident’. The
corpus analysis revealed that ‘caused the problem’ is far more frequent than ‘responsible for the
problem’. If both concepts are akin, should we not expect that their uses are similarly frequent?
Here, a closer look at the search hits is helpful.
word responsible for the caused the ratio
death 130 103 1.26:1
decline 31 15 2.07:1
damage 24 24 1:1
destruction 18 16 1.13:1
crisis 11 17 0.65:1
attack 46 3 15.33:1
murder 44 2 21:1
problem 13 79 0.16:1
accident 11 87 0.13:1
Table 2: Hits for some of the most frequent nouns after the phrase ‘responsible for the’ and
‘caused the’ and the ratio between them.
What we find is that speakers often raise questions like ‘what caused the problem’ leaving
it open that it was not an agent but rather an event that caused the problem. In contrast,
responsibility language is mostly used in relation to agents.17 Thus, it is relatively rare that people
17 There is some quite strong evidence coming from further corpus analyses that support such a view. When entering
the phrases ‘what caused’, COCA delivers 1,189 search hits compared to only 250 hits for ‘who caused’. The situation

18
make claims such as ‘the malfunctioning brakes are responsible for the accident’ but rather speak
of malfunctioning brakes causing accidents. This in turn suggests that the semantic similarity
between the language of causal attributions and responsibility attributions might be most
pronounced for agent causation. A further investigation into this possibility is, however, beyond
the scope of this paper.
It might be objected that the similar frequency in use of ‘responsible for the’ and ‘caused
the’ for many nouns are merely coincidental and do not reveal any semantic similarity between
these phrases; other phrases may be just as frequent. To counter this objection, we further
examined which verbal phrases occur most frequently before nouns such as ‘death’, ‘decline’ and
‘destruction’, for which we have observed the same frequencies. The word ‘death’ was most
frequently preceded by the verbal phrases ‘caused the’ and ‘responsible for the’.18 For the term
‘decline’, only ‘contributed to’ was a more common phrase than ‘caused the’ and ‘responsible for
the’. And for the term ‘destruction’ only ‘stopped the’ and ‘prevented the’ were more common
than ‘caused the’ and ‘responsible for the’. This data shows that the similarity in use between the
language of responsibility attributions and the language of causal attributions is unlikely to be a
mere matter of coincidence. Rather, corpus analysis indicates that these languages are highly
similar in meaning. This also puts pressure on Samland and Waldmann’s (2014, 2016) alternative
account of the effect of normative information on ordinary causal attributions: it does not seem to
be the case that participants read questions in vignette studies to be about a related notion such as
responsibility. Instead, the data suggests that the notion of causation is in itself inherently
normative.
is reversed for responsibility language. A search on COCA lists 412 hits for ‘who is responsible for’ but only 16 hits
for ‘what is responsible for’.
18 In fact, ‘seek the’, ‘face the’ and ‘get the’ were even more common, but occurred not together with ‘death’, but
mostly with the fixed expression ‘death penalty’.

19
5. Corroborating the Findings with Distributional Semantics
In addition to the somewhat qualitative approach to corpus analysis taken in the previous section,
there is also a more mathematical way of exploiting corpus data by using an array of
computational methods for investigating word meaning. One prominent approach is based on the
‘distributional hypothesis’, which follows Firth’s dictum that ‘you shall know a word by the
company it keeps’ (Firth 1957, p. 11; see also, Harris 1954). Accepting this, word meaning can be
explored by using computational methods to look at the distribution of words across a corpus.
One set of such tools are distributional semantic models (DSMs). The typical DSM represents
terms as geometric vectors in a high-dimensional space that can be compared to give a
quantitative measure of similarity. This is typically done by taking the cosine of two vectors, with
a value of 1 being taken to indicate identical meanings while a cosine of 0 would indicate
completely dissimilar meanings.19, 20
There are a number of different ways of carrying out distributional analyses. Unfortunately,
the details of these different approaches can get quite daunting, especially for researchers who are
new to the area. That said, we believe that even the more accessible techniques for distributional
analyses are of value. As such, we encourage experimental philosophers to begin employing these
tools, and to tackle their more complex aspects and sophisticated varieties in due course. We will
begin with some tools that any experimental philosopher could employ, then expand the analysis to
tools that require greater familiarity with programming.
Perhaps the most prominent type of distributional analysis is Latent Semantic Analysis
(LSA; Deerwester et al. 1990), and this is the method we will begin with in this section. LSA
19 The cosine can also take on a negative value. It is at best unclear how negative values should be interpreted,
however, and these are generally just treated as being 0.
20 For more detailed discussions of DSMs see Baroni et al. 2014a, Erk 2012, and Turney and Pantel 2010.

20
begins with the texts of a corpus being broken down into pre-defined documents, such as
paragraphs of text. The frequency of each term in the corpus is then counted for each document to
produce a term-by-document matrix. It should be noted that this matrix does not include
information about the relative location of terms in a document. Because of this, LSA is sometimes
referred to as a ‘bag-of-words’ approach. And in this, LSA is perhaps most markedly different
from the approach used in the previous section, which specifically looked at the relative position
of terms in a sentence.
While LSA has had a good deal of empirical success, one should be mindful of the
limitations of the bag-of-words approach and recognize that other approaches are available. In
LSA, the context for a target word is the rest of the document. Alternatively, window-based
methods use the terms surrounding the target word as context (while this can be thought of as a
bag-of-words, it is a relatively small bag of words). For instance, a window of size 5 would take
the two terms to either side of the target word as context. Another option is to use the words that
stand in a particular syntactic relation to the target word as the context.21 In contrast to these
approaches, the ‘new kids on the distributional semantics block’ are what Baroni et al. (2014b)
term context-predicting models. Instead of counting the terms occurring in a given context around
a target word, these models use artificial neural networks to set vector weights that ‘optimally
predict the contexts in which the corresponding words tend to appear’ (238).22
The easiest way to begin using LSA is to query a premade semantic space. One option is
the LSA website from the University of Colorado Boulder.23 This website allows users to run a
21 For a comparison of these approaches, as well as a number of other parameters involved in constructing vector-
based DSMs, see Kiela and Clark 2014.
22 Baroni et al. (2014b) conduct an extensive comparison between context-predicting and count-based models. To
their surprise, they ‘found that the predict models are so good that… there are very good reasons to switch to the new
architecture’ (245).
23 See http://lsa.colorado.edu/

21
number of different types of queries for a range of semantic spaces. To illustrate, we used the
pairwise comparison tool for the General Reading up to 1st Year College space24 to look at cosine
values for ‘cause’ and ‘caused’ as compared to four terms relevant to assessing causal attributions
and their relation to outcome valence—‘responsibility,’ ‘blame,’ ‘fault,’ and ‘praise’. As
predicted on the basis of our previous analyses, both terms show a notable similarity to
‘responsible’. Further, in line with our previous analyses we found that both terms show a notable
similarity with the negative terms ‘blame’ and ‘fault’, but essentially no similarity with the
positive term ‘praise’.
responsible blame fault praise
cause 0.32 0.28 0.30 0.05
caused 0.32 0.31 0.26 0.01
Table 3: Cosine values for term comparisons for the General Reading up to 1st Year College
space on the LSA website from the University of Colorado Boulder.
It would be nice to be able to say something absolute about the degree of similarity
indicated by a given cosine value. Unfortunately, this is complicated by differences in the sizes of
LSA spaces. As a result, the values should be thought of as relative measures. One option for
getting a sense of the relative values for a space is to test some comparison terms. For instance,
the value we found for ‘cause’ and ‘responsible’ is slightly higher than the value we find for ‘dog’
and ‘wolf’ (0.30), while the value for ‘dog’ and ‘animal’ (0.15) is half that, and the value for
‘dog’ and ‘sandwich’ is slightly higher than we found for ‘cause’ and ‘praise’. While such
comparisons can help you get an initial sense of the space, it is dependent on the terms that you
select and might be misleading. A more systematic approach is to look at a predefined list of
24 This space is built from a corpus of 37,651 documents and covers 92,409 unique terms.

22
comparisons. One option is to use a list that is part of a benchmark, such as MEN (Bruni et al.
2013). MEN includes a test set of 1,000 comparisons whose relatedness has been assessed by a
large sample of human judges. We can run each of these comparisons, then look at the pairs of
terms that have a similar cosine value to the pairs we want to assess.
Another tool available through the LSA website of the University of Colorado Boulder is
to search for the nearest neighbours of a given term. This provides the terms closest to the target
term in the semantic space. When we did this for ‘cause’ and ‘caused’, we found that the nearest
neighbours, excluding terms sharing the same word stem, generally have a negative cast (e.g.,
‘damage’ (0.66), ‘symptoms’ (0.61); ‘disease’ (0.69), ‘infections’ (0.67)). Again this is in keeping
with our previous findings. When we turned to ‘responsible’, however, we found that many of the
nearest neighbours for this term are of a different sort. For instance, we found that ‘duties’ (0.57)
is the nearest neighbour, followed directly by ‘supervision’ (0.56) and ‘personnel’ (0.55). This
suggests that the responsibility attributions we are after are getting drowned out by the alternative
usage of ‘responsibility’ noted in the previous section—that of the duties associated with a role.
In addition to getting a sense of degree of similarity indicated by a cosine value in a given
space, we also assessed whether it is doing a good job in capturing the semantic relatedness of terms.
To do this we employed the MEN list, mentioned above, and used the list of cosine values for the test
set to analyse how well this correlates with the scores from the human judges. When we did this for
the General Reading space, we found that it does a relatively good job: we got a Spearman’s rho of
0.67. For comparison, Kiela and Clark (2014) report values of 0.66 to 0.71 for the spaces they
compared in Table 6, while Baroni et al. (2014b) report in Table 2 a top value of 0.72 for the best
count-based model tested and a value of 0.80 for the best context-predicting model tested.

23
It is also possible to build an LSA space oneself. While the details that go into the
construction of an LSA space are complicated, a number of tools are available to facilitate the
process. We will begin with tools that can be used through the statistical software package R.
While there are several benefits to using R in the present context,25 it also suffers from some
limitations, as we will see.
The easiest way to get started with LSA in R is to use the LSAfun package to import a
premade semantic space (Günther et al. 2015). To illustrate, we used the EN_100k_lsa space to
further explore the relationship between ‘cause’/‘caused’ and ‘responsible’. This space was
created from a corpus of some two billion words combining the British National Corpus, the
ukWaC corpus, and a 2009 Wikipedia dump. We began by looking at the same set of
comparisons that we performed above. Again we see that ‘cause’/‘caused’ show a notable
similarity to ‘responsible’, and that both terms are much more similar to ‘blame’ and ‘fault’ than
to ‘praise’. Next, we looked at the nearest neighbours for ‘cause’, ‘caused’, and ‘responsible’. The
results were in line with what we saw previously, with ‘cause’ and ‘caused’ being close to a
number of negative terms (e.g, ‘excessive’ (0.78), ‘suffer’ (0.78); ‘damage’ (0.82), ‘fatal’ (0.71)),
while ‘responsible’ was close to a range of terms related to duties associated with a role (e.g.,
‘overseeing’ (0.76), ‘supervising’ (0.63)). Like the previous space, the EN_100k_lsa space
performs well on the MEN benchmark with a Spearman’s rho of 0.67.26
25 One is that R supports a large range of statistical analyses of use to experimental philosophers. Another is that the
only book-length guide to the practice of experimental philosophy currently available (Sytsma and Livengood 2016)
uses R as its preferred statistical program, and Chapter 10 provides a general introduction to the use of R.
26 There is also a window-based space available from the same corpus—the EN_100k space—that performs slightly
better on the MEN benchmark (Spearman’s rho of 0.71). Another option are the spaces available from the
COMPOSES Semantic Vectors website, which provides the best models from Baroni et al. (2014b) as text files.
Their context-predicting model performs especially well with a Spearman’s rho of 0.80 on the MEN benchmark.
Both spaces paint a similar picture to what we saw for the EN_100k_lsa space, both in terms of the comparisons in
Table 4 and the nearest neighbors of ‘cause’/’caused’ and ‘responsible’.

24
responsible blame fault praise
cause 0.32 0.46 0.50 0.23
caused 0.34 0.47 0.57 0.18
Table 4: Cosine values for term comparisons for the EN_100k_lsa space.
R also offers tools for creating corpora and building LSA spaces. To illustrate, we used the
RCurl package to scrape the text for all entries in the Stanford Encyclopedia of Philosophy and
the Internet Encyclopedia of Philosophy from their websites. The result was a corpus including
2,378 entries split into 136,946 paragraphs (the documents for our analyses) and comprised of
over 149 million words and 115,644 unique terms. We then used the koRpus package, to annotate
the documents with lemma information. The tm package was used to convert this into a corpus,
which was fed into the lsa package to generate the term-by-document matrix and create the
semantic space. It is in this final step that we ran into the limitations of R noted above.
Specifically, in R the data for analysis is held in RAM, which places severe limits on the size of
the matrix it can process on a typical home computer. One option is to reduce the size of the term-
by-document matrix by removing infrequently occurring terms before creating the semantic
space. For the philosophy corpus, we needed to reduce the matrix to the 8,821 most frequently
occurring terms.
Given that philosophers have often treated the ordinary concept of causation as being a
purely descriptive concept and that many have expressed either surprise or outright skepticism
toward the results surveyed in Section 2, we would not expect to find the valence effect for the
philosophy corpus that we saw in our previous investigations. With regard to the relation between
‘cause’ and ‘responsible’, one might predict that these terms would also be largely unrelated.
Alternatively, one might note that many philosophers hold that causing an outcome is a

25
prerequisite for being responsible for that outcome. As such, one might expect to see a notable
similarity between these terms in the philosophy corpus. What we found is that ‘cause’ showed
virtually no relation to ‘blame’, ‘fault’, or ‘praise’, and that it showed virtually no relation to
‘responsible’.27 The space performed better than expected on the MEN benchmark, with a
Spearman’s rho of 0.48. Because the term-by-document matrix was significantly reduced,
however, the correlation was only calculated on 269 comparisons.
responsible blame fault praise
cause –0.0003 0.0004 –0.0011 –0.0051
Table 5: Cosine values for term comparisons for the philosophy corpus LSA space.
Given the degree to which the term-document-matrix was reduced, the results for the
philosophy corpus LSA space should be taken with a hefty grain of salt. To further test these results,
we switched to the Gensim toolkit implemented in Python, which is designed to efficiently handle
large corpora and is able to carry out a wide variety of distributional analyses, including LSA and the
context-predicting models using word2vec that performed best in Baroni et al.’s (2014b)
comparisons. We exported the processed philosophy corpus from R into Gensim, then analysed it
using word2vec with recommended parameters, including a five-word context window. The results
were quite different from what we found for the LSA space. Most notably, we found a much higher
cosine value for ‘cause’ and ‘responsible’. We also saw a moderate relation between ‘cause’ and
‘blame’ or ‘fault’, but no relation between ‘cause’ and ‘praise’. Further, the nearest neighbours of
‘cause’ and ‘responsible’ were quite distinct.28 While these results are more like what we’ve seen for
27 We only ran the comparisons for ‘cause’ since we lemmatized the text and ‘cause’ and ‘caused’ belong to the same
lemma.
28 Five nearest neighbours for ‘cause’: ‘proximate’ (0.64), ‘efficient’ (0.60), ‘effect’ (0.59), ‘volition’ (0.51), and
‘necessitate’ (0.50); five nearest neighbours for ‘responsible’: ‘accountable’ (0.69), ‘blameworthy’ (0.64),
‘attributable’ (0.58), ‘culpable’ (0.56), and ‘negligent’ (0.55).

26
the general corpora, they continue to suggest that the philosophical usage diverges from the ordinary
usage, as will be spelled out below. The space performed comparably on the MEN benchmark, with
a Spearman’s rho of 0.48 on a much higher number of comparisons.29
responsible blame Fault praise
cause 0.41 0.18 0.17 -0.04
Table 6: Cosine values for term comparisons for the philosophy corpus word2vec space.
These same tools can be applied to other corpora that are available for download,
including COCA. To facilitate comparison to philosophical usage, we excluded academic texts.
Since COCA comes with lemma information, we did not need to annotate the documents. Other
than this we followed the same procedure detailed for the philosophy corpus to generate a
word2vec space. The results were in keeping with what we saw for the other general corpora
above, with there being a notable similarity between ‘cause’ and ‘responsible’, between ‘cause’
and the negative lemmas ‘blame’ and ‘fault’, and no similarity between ‘cause’ and the positive
lemma ‘praise’. As expected, the space performed extremely well on the MEN benchmark with a
Spearman’s rho of 0.80.
responsible blame fault praise
cause 0.41 0.56 0.48 -0.06
Table 7: Cosine values for term comparisons for the non-academic COCA corpus word2vec space.
With access to a full corpus it is also possible to more carefully target causal attributions and
responsibility attributions by directly comparing multi-word expressions. To do this we replaced the
phrases ‘caused the’ and ‘responsible for the’ with single tokens before lemmatizing and processing
29 944 of the 1,000 comparisons were used (25 terms were missing from the corpus).

27
the non-academic COCA corpus. We then analysed it using the same predictive model as above.
The cosine value between the causal attribution token and the responsibility token was quite large,
and notably larger than for the previous comparison between ‘cause’ and ‘responsible’. Further, each
token was one of the five nearest neighbours of the other. The nearest neighbours for each token
included a number of terms with a negative cast (e.g., ‘catastrophic’ (0.68), ‘fatal’ (0.63); ‘culpable’
(0.62), ‘complicit’ (0.58)), including that ‘blame’ was one of the five nearest neighbours for the
responsibility attribution token. In addition, none of the nearest neighbours for this token indicated
the notion of duties associated with a role that marred our previous results.
responsible
for the blame fault praise
caused the 0.63 0.55 0.42 –0.10
responsible
for the 0.61 0.35 0.16
Table 8: Cosine values for term comparisons for the non-academic COCA corpus word2vec
space with causal attribution and responsibility attribution tokens.
To better compare philosophical usage with ordinary usage, we tokenized the philosophy
corpus and repeated the analysis. We found that the causal attribution token was quite similar to
the responsibility attribution token. In line with the alternative prediction noted above, this might
reflect that many philosophers hold that causing an outcome is a prerequisite for being
responsible for that outcome. Although the cosine values for the two tokens are similar for both
the philosophy space and the COCA space, when we look deeper we find evidence that the causal
attributions of philosophers are quite different from the causal attributions of more ordinary
language. Thus, while there was a strong relation between the two tokens in the philosophy space,
the causal attribution token was much less similar to ‘blame’ and ‘fault’. This stands in marked
contrast to what we saw for the COCA space. Further, the same contrast holds for responsibility

28
attributions. The results suggest that the ordinary usage of both causal attributions and
responsibility attributions has a negative cast that the philosophical usage lacks.
responsible
for the blame fault praise
caused the 0.55 0.16 0.17 0.03
responsible
for the 0.10 0.07 0.00
Table 9: Cosine values for term comparisons for the philosophy corpus word2vec space with
causal attribution and responsibility attribution tokens.
Overall the results of our latent semantic analyses nicely line up with the results of the
analyses in the previous section, with the two approaches providing a consilience of evidence.
Looking across the two sets of analyses, we find compelling evidence for a positive answer to
each of the questions we opened this chapter with: our results provide independent support for the
thesis that ordinary causal attributions are sensitive to normative information; they provide
support for the contention that outcome valence often matters for ordinary causal attributions;
they indicate that causal attributions are similar to responsibility attributions;30 and they suggest
that philosophers use the language of causal attribution differently from lay people.
6. Concluding Remarks: Corpus Analysis as a Method for Experimental Philosophy
Causation is one of the most contested concepts in philosophy. Recent questionnaire studies have
produced some rather surprising insights into how we use that concept. Most importantly, they
30 It could be objected here that while our results indicate that causal attributions are similar to responsibility
attributions, they do not indicate that causal attributions are themselves normative. For instance, it might be suggested
that they are close in semantic space because causing an outcome is a prerequisite for being responsible for that
outcome. This would not explain the results of our analysis in Section 4, however, or the valence effect observed for
causal attributions in the semantic spaces. While this could be investigated further using DSMs to assess the
Synonymy Effect hypothesis, space prevents us from doing so here. Alternatively, expanding on Alicke’s view, it
might be argued that the desire to blame biases both causal attributions and responsibility attributions. While we
cannot rule this out based on the present results, we hold that the responsibility view offers the simpler explanation.

29
suggest that normative considerations play a central role in ordinary causal attributions. However,
it is still an open issue how best to interpret and explain these results.
We believe that to make progress in deciding between the different accounts of the impact
of norms on causal attributions, it is fruitful to expand the set of empirical tools used by
experimental philosophers working on causation. Specifically, we believe that the study of
ordinary causal attributions can benefit from the tools of corpus linguistics. One reason for this
belief is that while questionnaire methods are powerful and often well-suited to investigating
philosophical questions, they also have limitations. And the questionnaire studies on ordinary
causal attributions that we have looked at in this chapter do suffer from some of those limitations.
To illustrate, we will focus on the first study we discussed in Section 2, Knobe and Fraser’s
(2008) investigation of the Pen Case.31
The basic kind of misgiving one might have about Knobe and Fraser’s study is that it
relies on an instrument they unintentionally designed in such a way that it would elicit the
suspected effect, not because norms actually do have an impact on ordinary causal attributions but
because it leads participants to misread the prompts. For instance, one might object to participants
being asked which one of the two agents, the administrative assistant or the professor, caused the
problem. While this is certainly a very natural way to ask the question of interest, we believe that
asking people who caused ‘the problem’, as compared to ‘the outcome’ or ‘the situation’, might
trigger an interpretation of the question in normative terms. Alternatively, one might worry that
by having participants rate two statements—one about the administrative assistant and one about
Professor Smith—and phrasing this in a way that suggests an all-or-nothing state of affairs (as
31 We would like to emphasize that even though we focus on Knobe and Fraser here, our worries apply to
questionnaire studies more generally, and the authors recognize that their own work is also liable to these
methodological concerns.

30
opposed, for example, to asking whether the agent was ‘a cause’ of the outcome) might prompt
participants to feel that they should agree with at most one of the two questions. Since the only
distinguishing feature between the two agents’ actions is that one violates a norm while the other
does not, participants might latch onto this as a relevant cue for fulfilling the task. Or one might
note that Knobe and Fraser asked participants whether an agent caused an outcome. Typically,
when philosophers talk about causation, they talk about events as the causal relata, not people.
Asking about the agent, rather than his or her activities, might therefore create another reason for
participants to believe that the researcher is asking about something normative.
All of the potential issues we just noted for Knobe and Fraser’s study could be addressed
through further questionnaire research. And, in fact, a good deal of work has subsequently been done
on the Pen Case, or cases like it, that varies these sorts of factors. But follow-up studies addressing
one potential confound run the risk of introducing others. This is simply one of the difficulties
inherent to this type of research. It does not mean, of course, that questionnaire studies should be
abandoned. Rather, the moral we should draw from it is that in the face of these risks we should
diversify our set of methods. Turning to corpus linguistics seems natural here, as one strength of
corpus analysis is that it is relatively immune from the pragmatic pitfalls we have just highlighted.
One of the motivations of corpus linguistics is the preference for ‘real’ language data over
examples of language use generated by the linguists themselves. While a corpus cannot strictly
speaking be representative for a language in its entirety, because possible utterances of that
language are infinite, linguistic corpora aim at balanced sampling from this impressive
population. Unless the research interest is focused on a specific genre, say, the usage of a given
term in academic texts, a balanced corpus contains a considered choice of texts of various types
and from different sources. For example, it does not only contain written, but also (transcribed)

31
spoken language, not only specialised (e.g., academic) language, but also its everyday variety, not
only fiction or other literary texts, but also mundane ones such as operation manuals. In this way,
a large corpus does present a meaningful sample of actual language use.
The preference for ‘real’ language sits nicely with precepts of experimental philosophy, in
that it emphasises the importance of empirical data over that generated by researchers relying on
their own intuition or judgement. An important advantage it has over data generated by
questionnaire studies is that the linguistic data of a general corpus usually has been generated
independently of the researcher and her specific research questions. The data thus is usually
uncued, in the sense that the utterances the corpus contains have not been produced in response to
some prompt of the researcher. It is then plausible to assume that a corpus is unbiased with respect
to the specific research question a philosopher approaches it with (cf. Schütze 2010).32 And for the
same reason, such a corpus can be considered free of the biases of experimental pragmatics.
Having said that, there are, of course, limits to corpus analysis. The linguistic data of a
corpus can only be evidence in relation to some philosophical issue to the extent that the actual
use of language is relevant to it. This relevance may be direct or indirect, because the linguistic
data in a corpus may well allow us to infer something about deeper structures of language. But,
clearly, if the observation of linguistic phenomena in actual use is irrelevant to a philosophical
issue, then so is corpus analysis. Moreover, if the pertinent phenomena are of the very particular
and subtle kind that is common for philosophical problems, even a large corpus may not yield
any, let alone many, examples of their use. By way of contrast, questionnaires can be constructed
32 It is, of course, possible to come up with examples of corpora that are dependent on the researcher and that contain
cued language use. Most simply, for example, in the case that the corpus consists of written responses to a vignette. It
is the choice of texts that determines whether the data contained in a corpus is indeed independent and uncued. If a
pre-existing general language corpus is used, this objection can be assumed to be moot.

32
to elicit informants’ responses to precisely worded prompts, and in doing so, the wording can be
varied easily to bring out and test subtle differences in language.
Therefore, it is clear that corpus analysis is best viewed as a fruitful addition to the
methodological toolbox of experimental philosophy. Not only can it be used effectively to explore
the actual use of linguistic expressions—something that is called for in philosophy on many
occasions. It can, more specifically, be used to complement experimental studies in several
helpful ways: to pre-test hypotheses that inform such studies, to help with the general construction
of questionnaires and the precise wording of their items, and, most importantly, we believe, to test
the findings from questionnaire studies, either giving them independent support from another
empirical source or providing evidence against them.
References
Achugar, M, and M. J. Schleppegrell (2005), ‘Beyond connectors’. Linguistics and Education, 16,
298–318.
Alicke, M. D. (1992), ‘Culpable causation’. Journal of Personality and Social Psychology, 63,
368–378.
Altenberg, B. (1984), ‘Causal linking in spoken and written English’. Studia Linguistica, 38, 20–69.
Baroni, M., R. Bernardi, and R. Zamparelli (2014a), ‘Frege in space’. Linguistic Issues in
Language Technology, 9, 241–346.
Baroni, M., G. Dinu, and G. Kruszewski (2014b), ‘Don’t count, predict! A systematic comparison
of context-counting vs. context-predicting semantic vectors’. Proceedings of the 52nd Annual
Meeting of the Association for Computational Linguistics, 238–247.
Bergh, G., and E. Zanchetta (2008), ‘Web linguistics’, in A. Lüdeling and M. Kytö (eds), Corpus
Linguistics. Berlin: de Gruyter, Vol. 1, pp. 309–327.
Biber, D. (2010), ‘Corpus-based and corpus-driven analyses of language variation and use’, in B.
Heine and H. Narrog (eds), The Oxford Handbook of Linguistic Analysis. Oxford: Oxford
University Press, pp. 159–191.

33
Biber, D., S. Conrad, and R. Reppen (1998), Corpus linguistics. Cambridge: Cambridge
University Press.
Bluhm, R. (2012), Selbsttäuscherische Hoffnung: Eine sprachanalytische Annäherung. Münster: mentis
Bluhm, R. (2013), ‘Don’t ask, look! Linguistic corpora as a tool for conceptual analysis’, in: M.
Hoeltje, T. Spitzley, and W. Spohn (eds) Was dürfen wir glauben? Was sollen wir tun?
Sektionsbeiträge des achten internationalen Kongresses der Gesellschaft für Analytische
Philosophie e.V. Duisburg: DuEPublico.
Bluhm, R. (2016), ‘Corpus analysis in philosophy’, in M. Hinton (ed) Evidence, Experiment and
Argument in Linguistics and the Philosophy of Language. Frankfurt am Main: Peter Lang, pp. 91–109.
Bruni, E., N. K. Tran, and M. Baroni (2013), ‘Multimodal Distributional Semantics’, Journal of
Artificial Intelligence Research 49, 1–47.
Cushman, F. (2013), ‘Outcome, and Value: A Dual-System Framework for Morality’. Personality
and Social Psychology Review 17 (3), pp. 273–92.
De Schryver, G.-M. (2002), ‘Web for/as Corpus’. Nordic Journal of African Studies, 11, 266–282.
Deerwester, S., S. Dumais, T. Landauer, G. Furnas, and R. Harshman (1990), ‘Indexing by Latent
Semantic Analysis’. Journal of the Society for Information Science, 41 (6), 391–407.
Diessel, H., and K. Hetterle (2011), ‘Causal clauses: A crosslinguistic investigation of their
structure, meaning, and use’, in: P. Siemund (ed), Linguistic Universals and Language Variation.
Berlin: Mouton de Gruyter, pp. 23–54.
Erk, K. (2012), ‘Vector space models of word meaning and phrase meaning: A survey’. Language
and Linguistics Compass, 6 (10), 635–653.
Firth, J. R. (1957), ‘A synopsis of linguistic theory 1930–55’, in: Studies in Linguistic Analysis.
Oxford: Blackwell, pp. 1–32.
Fischer, E., and Engelhardt, P. E. (2017), ‘Diagnostic experimental philosophy’, teorema, 36 (3),
pp. 117–137.
Günther, F., C. Dudschig, and B. Kaup (2015), ‘LSAfun: An R package for computations based
on Latent Semantic Analysis’. Behavior Research Methods, 47, 930–944.
Hahn, U., F. Zenker, and R. Bluhm (2017), ‘Causal argument’, in: M. R. Waldmann (ed), The
Oxford Handbook of Causal Reasoning. New York, NY: Oxford University Press, pp. 475–494.
Harris, Z. (1954), ‘Distributional structure’. Word, 10 (23), 146–162.

34
Henne, P., Á. Pinillos, and F. De Brigard (2015), ‘Cause by omission and norm: Not watering
plants’. Australasian Journal of Philosophy, XXX, 1–14.
Hilton, D., and Slugoski, B. (1986), ‘Knowledge-based causal attribution: The abnormal
conditions focus model’. Psychological Review, 93, 75–88.
Hitchcock, C., and J. Knobe (2009), ‘Cause and Norm’. The Journal of Philosophy, 106, 587–612.
Khoo, C., S. Chan, and Y. Niu (2002), ‘The many facets of the cause-effect relation’, in R. Green,
C. A. Bean, and S. H. Myaeng (eds), The Semantics of Relationships. Dordrecht: Springer
Netherlands, pp. 51–70.
Kiela, D., and S. Clark (2013). ‘A Systematic Study of Semantic Vector Space Model
Parameters’. Proceedings of the 2nd Workshop on Continuous Vector Space Models and their
Compositionality, 21–30.
Kilgarriff, A., and G. Greffenstette (2003), ‘“Introduction” to “the Web as Corpus”’.
Computational Linguistics, 29 (3), 333–47.
Knobe, J., and Fraser, B. (2008), ‘Causal judgments and moral judgment: Two experiments’, in:
W. Sinnott-Armstrong (ed), Moral Psychology, Volume 2: The Cognitive Science of Morality.
Cambridge: MIT Press, pp. 441–447.
Kominsky, J., J. Phillips, T. Gerstenberg, D. Lagnado, and J. Knobe (2015), ‘Causal superseding’.
Cognition, 137, 196–209.
Lee, D. Y. W. (2010), ‘What corpora are available?’, in M. McCarthy and A. O’Keeffe (eds),
Corpus Linguistics. London, New York: Routledge, pp. 107–121.
Leech, G. (1992) ‘Corpora and theories of linguistic performance’, in: J. Svartvik (ed.) Directions
in Corpus Linguistics. Berlin, New York: Mouton de Gruyter, pp. 105–122.
Livengood, J., J. Sytsma, and D. Rose (2017a), ‘Following the FAD: Folk attributions and
theories of actual causation’. Review of Philosophy and Psychology, 8 (2), 274–294.
Livengood, J., J. Sytsma, and D. Rose (2017b), ‘Following the FAD: Supplemental materials’.
https://link.springer.com/article/10.1007%2Fs13164-016-0316-1
Livengood, J., and J. Sytsma (under review), ‘Actual causation and compositionality’.
Lüdeling, A., and M. Kytö (eds) (2008–2009), Corpus Linguistics (2 vols.). Berlin: de Gruyter.
Malle, B., S. Guglielmo, and A. E. Monroe (2014), ‘A Theory of Blame’. Psychological Inquiry
25 (2), pp. 147–86.

35
McCarthy, M., and A. O’Keefe (2010a), ‘Historical perspective’, in M. McCarthy and A. O’Keefe
(eds), The Routledge Handbook of Corpus Linguistics. London; New York: Routledge, pp. 3–13.
McCarthy, M., and A. O’Keefe (eds) (2010b), The Routledge Handbook of Corpus Linguistics.
London, New York: Routledge
McEnery, T., and A. Wilson (2001), Corpus linguistics (2nd edn). Edinburgh: Edinburgh
University Press.
Prinz, J. (2007), The Emotional Construction of Morals. Oxford: Oxford University Press.
Reuter, K. (2011), ‘Distinguishing the appearance from the reality of pain’. Journal of
Consciousness Studies 18(9–10), pp. 94–109.
Reuter, K., L. Kirfel, R. van Riel, and L. Barlassina, L. (2014), ‘The good, the bad, and the
timely: How temporal order and moral judgment influence causal selection’. Frontiers in
Psychology, 5, 1336.
Samland, J., and M. R. Waldmann (2014), ‘Do social norms influence causal inferences?’, in P.
Bello, M. Guarini, M. McShane, and B. Scassellati (eds), Proceedings of the 36th Annual Conference
of the Cognitive Science Society. Austin, TX: Cognitive Science Society, pp. 1359–1364.
Samland, J. and M. R. Waldmann (2016), ‘How prescriptive norms influence causal inferences’.
Cognition, 156, 164–176.
Schütze, C. T. (2010), ‘Data and evidence’, in: K. Brown, A. Barber, and R. J. Stainton (eds), Concise
Encyclopedia of Philosophy of Language and Linguistics. Amsterdam et al.: Elsevier, pp. 117–123.
Sytsma, J., J. Livengood, and D. Rose (2012), ‘Two types of typicality: Rethinking the role of
statistical typicality in ordinary causal attributions’. Studies in History and Philosophy of
Biological and Biomedical Sciences, 43, 814–820.
Sytsma, J., D. Rose, and J. Livengood (ms), ‘The extent of causal superseding’.
Sytsma, J., and J. Livengood (2016), The Theory and Practice of Experimental Philosophy.
Peterborough: Broadview.
Sytsma, J., and J. Livengood (2018), ‘Intervention, bias, responsibility… and the trolley
problem’. http://philsci-archive.pitt.edu/14549/
Sytsma, J., and Reuter, K. (2017), ‘Experimental philosophy of pain’. Indian Council of
Philosophical Research. 34(3), pp. 611–628.
Turney, Peter and Patrick Pantel (2010), ‘From frequency to meaning: Vector space models of
semantics’. Journal of Artificial Intelligence Research, 37, 141–188.

36
Willemsen, P. (2016), ‘Omissions and expectations: a new approach to the things we failed to to
do’, Synthese 195 (4), pp. 1587-1614.
Xiao, R. (2008), ‘Well-known and influential corpora’, in: A. Lüdeling and M. Kytö (eds),
Corpus Linguistics. Berlin, New York: de Gruyter, Vol. 1, pp. 383–457.