On classifying connectives and coherence relations 
Liesbeth DEGAND* 
Department of Experimental Psychology, University of Louvain 
Place du Cardinal Mercier, 10 
1348 Louvain-la-Neuve, Belgium 
degand @exco.ucl.ac.be 
Abstract 
This paper tackles the methodological 
question of how coherence relations and their 
surface cues, e.g. causal connectives, may be 
categorised into different discourse domains. 
In this context, the central issue is the need for 
explicit and objective criteria that can be 
applied by different analysts to all kinds of 
natural texts. Strengths and weaknesses of 
different methods that are currently used in 
corpus studies are passed in review. 
Introduction 
In this paper, I would like to address a 
methodological question that starts from the 
following observation: There is a growing 
number of corpus studies in the area of the use 
of coherence relations and discourse markers in 
natural texts. Most of these studies try to 
establish a relationship between the coherence 
relations found in the text and the discourse 
markers used to signal them linguistically, very 
often inter-clausal connectives (Degand 1996, 
1998; Grote, Lenke & Stede 1997; Knott 1996; 
Mann & Thompson 1992; Oversteegen 1997; 
Pit, Pander Maat & Sanders 1997; Pander Maat 
& Sanders 1995, Sanders 1997). However, the 
results of these studies are hardly ever 
straightforward to compare. The main reason 
for this is that the methods used to classify both 
the coherence relations and their linguistic 
markers are very often divergent and in the best 
cases only partially overlapping. So, in order to 
enable comparison of the different studies, there 
is a need for clear and explicit classification 
criteria. In this paper, I will try to set out a 
number of such (operationalisable) elements that 
could function as tools in the categorisation task. 
The goal will be reached if the categorisation 
criteria are formulated in such an explicit way 
that they can be applied by different analysts to 
all kinds of natural texts. 
1 Coherence relations 
A classical approach to coherence relations is to 
classify them into relational or discourse 
domains: semantic vs. pragmatic (van Dijk 
1979; Sanders 1992; Pander Maat & Sanders 
1995), content vs. epistemic vs. speech-act 
(Sweetser 1990), external vs. internal (Halliday 
& Hasan 1985, Martin 1992), subject matter vs. 
presentational (Mann & Thompson 1988), 
ideational vs. interpersonal vs. textual (Degand 
1996, Maier & Hovy 1993), ideational vs. 
rhetorical vs. sequential (Redeker 1991). These 
classifications all account for the fact that there 
are different types of coherence relations that 
exist at several "levels", "strata", "planes" or 
"domains" of the discourse. It is striking to note 
that there is a great amount of agreement 
between the authors with respect to the different 
general types of coherence relations. They can 
be grossly characterised as follows: 
® Ideational / semantic / content / external 
/ subject matter relations establish a link 
between discourse units that denote 
states of affairs in the world described in 
the discourse. Examples are: cause, 
sequence, contrast, etc. 
Interpersonal / pragmatic / internal / 
presentational / rhetorical relations 
establish a link between the 
illocutionary meaning of one of the 
discourse units with the Iocutionary 
meaning of the other. Examples are: 
* Senior research assistant at the Belgian National Fund for Scientific Research (F.N.R.S.). 
29 
justification, motivation, evidence, 
conclusion, etc. Furthermore, a 
distinction can be made between (i) 
epistemic (knowledge of a fact leads to 
conclusion of a belief), and (ii) speech- 
act relations (a fact leads to the utterance 
of a speech-act). 
• _Textual / sequential relations establish a 
sequential, text organising relation 
between ideationally and interpersonally 
only loosely or indirectly related 
adjacent discourse segments. Examples 
are: list, joint, etc. or any other 
ideational or interpersonal relations 
when they are used as comment, 
digression, excursus, etc. 
While the characterisation of these relational 
domains is general enough so as to account for 
all different kinds of discourse structures, this 
generality poses a number of problems as soon 
as one starts to work with natural texts. Most of 
the time there is agreement on proto~.'pical 
examples (very often built-up examples for 
purposes of illustration), but in actual text 
analysis one gets rapidly confronted with an 
important number of ambiguous cases. That is, 
when one tries to establish what type of 
coherence relations are present in the text, it 
becomes rapidly impossible to classify found 
relations on the basis of these definitions alone. 
Different elements may account for these 
problems: (i) coherence relations are not always 
marked linguistically, (ii) there is no one to one 
mapping between coherence relation and 
linguistic discourse marker, and (iii) there is no 
general consensus on the type of linguistic 
constructions that may function as discourse 
markers (connectives, prepositions, adverbials, 
intonation patterns, cue phrases, etc.). 
In the literature, connectives are generally 
considered to be prototypical linguistic markers 
of coherence relations. In this paper, I will then 
concentrate on the problems that arise when one 
tries to relate to each other the discourse 
domains as described above, the coherence 
relations and the connectives that may signal 
these coherence relations. This will be 
exemplified on the hand of causal constructions. 
2 From connective to coherence 
relation 
While it is clear that coherence relations are not 
always signalled linguistically, and that they 
may be signalled by other linguistic means than 
connectives, it seems plausible to accept that 
inter-clausal connectives always mark (at least) 
a coherence relation. As a matter of fact, this is 
what appeared from my corpus analyses: any 
causal connective linking two (minimally 
propositional) segments always functioned as a 
linguistic marker of a coherence relation. This 
means that it is justified to use a bottom-up 
approach to study coherence relations. The 
question to answer then is: How does one know 
what type of coherence relation is signalled by 
the connective? And corollary to this: In what 
discourse domain is the connective used? 
Answering these types of questions implies to 
combine both a top-down approach in which 
broad general semantic categories are defined 
with a bottom-up approach in which the 
meaning and function of a number of linguistic 
markers, such as connectives, are described. The 
combination of these two approaches should (i) 
guarantee the linguistic validity of the discourse 
domains and the coherence relations and (ii) 
help to reach a parsimonious and organised 
classification of the connectives that function as 
linguistic markers of these higher-level semantic 
categories. Working with more than one 
language in this context may be helpful to reveal 
the degree of idiosyncrasy of the different types 
and categories. 
Since I will restrict my investigations here to the 
coherence relations that are linguistically 
marked by connectives, I will only address one 
of the three above mentioned problems. Namely, 
the observation that there is no one to one 
mapping between coherence relation and 
discourse marker. This entails that a coherence 
relation is not always marked by the same 
connective, and that a given connective does not 
always signal the same coherence relation. For 
example, in the causal domain, the ideational 
relation NON-VOLITIONAL CAUSE may be 
marked by the Dutch connectives omdat and 
doordat ('because'). At the same time, the 
connective want can signal the coherence 
relations VOLITIONAL CAUSE, NON-VOLITIONAL 
CAUSE, and CLAIM-ARGUMENT, in either of the 
ideational or interpersonal discourse domains. 
30 
From what has been mentioned so far it should 
be clear that there is a need to disambiguate the 
(sometimes complex) relationship between 
coherence relations and their surface cues. 
Objective and explicit methods are needed that 
operationalise the type of above-mentioned 
definitions. 
Different methods that are used in the field will 
pass the review, including (i) the meaning of the 
connective itself, (ii) the substitution test, (iii) 
the paraphrase test, (iv) the underlying meaning 
of the relation, and (v) the role of other linguistic 
markers such as perspective indicators. In my 
opinion, the combination of different tests 
together with corpus analyses opens up 
interesting perspectives with respect to the 
usability and applicability of these methods in 
real field work. 
2.1 The role of the linguistic marker itself 
In agreement with the functional approach to 
language I consider that different surface cues 
reflect different functions and meanings. In the 
field of coherence relations, this functional 
stance leads to the following assumption : If one 
generalised meaning, e.g. a causal relation, can 
be expressed by different linguistic expressions, 
e.g. because, for, so, therfore, and, etc., it is 
highly probable that these different linguistic 
markers are (at least) slightly divergent in 
meaning. The working hypothesis is then that 
every connective has its own specific meaning 
and that there are no true synonyms. It follows 
from this that one should be (at least) suspicious 
when different linguistic markers end up in a 
same category. In those cases, finer distinctions 
might be necessary. A way to test this is to 
make use of a substitution test. 
2.2 Substituting one connective by 
another 
In a substitution test a given connective is 
substituted by another one (cf. Knott & Dale 
1994; Knott 1996 for a systematic study). 
Acceptability of the substitution is dependent on 
the degree of similarity between the two 
connectives. This leads us to various 
possibilities: from connectives that would be 
totally overlapping in meaning ~ to partially 
overlapping, and exclusive ones. While this 
substitution test can bring us closer to a 
linguistically motivated classification of 
connectives because it groups together surface 
cues that are close in meaning, the main problem 
is that the whole classification is based on the 
analysts' judgement. And it is clear that these 
judgements may diverge between different 
analysts, especially when the connectives are 
largely overlapping in meaning. Analysts' 
intuitions must certainly be completed by other 
analyses. 
2.3Paraphrasing the meaning of the 
connective 
The use of paraphrase tests is a very frequent 
method to analyse the meaning of connectives 
(Pander Maat & Sanders 1995, 1997: Degand 
1996; Pit, Pander Maat & Sanders 1997; Sanders 
1997). They help to determine what the 
meaning of the segments is on their own, i.e. 
when the influence of the connective itself is left 
out. The advantage of a paraphrase test is that it 
forces to be explicit about the interpretation of 
the meaning of the connective under 
investigation. Once the connective is left out, the 
analyst checks whether the segments can be 
related by the given paraphrase(s) without 
changing the reading of the overall fragment. Of 
course, the analyst's intuition is still the ultimate 
criterion to classify the connective as expressing 
one or another meaning, but the paraphrase test 
makes the different interpretations explicit, so it 
helps the analyst to systematically check his/her 
intuitions. This is illustrated below on the basis 
of the following causal paraphrases: 
• Ideational, volitionali this action is the 
consequence of the following fact : 
• Ideational. non-volitional: this situation 
is the consequence of the following 
fact : 
• Interpersonal, epistemic: one may 
conclude this on the basis of the 
following situation : 
• Interpersonal, speech-act: this utterance 
is motivated by the following situation : 
With the help of these paraphrases, we can 
classify the use of because as being epistemic in 
Note that this in fact opposite to the basic functional 
31 
stance which aims for an ideal "one form - one 
meaning" correspondence. 
example (1) and ideational, volitional in 
example (2). 
. John loves her because / one may conclude 
this on the basis of the following situation he 
came back. 
precise and explicit way so that it can be applied 
unambiguously. Second, the examples also show 
that it is not always clear whether the paraphrase 
vehicles the meaning of the connective alone. 
On the basis of the paraphrase test, example (5) 
should be rejected as being odd. 
. John came back because / this action is the 
consequence of the following fact he loved 
her. 
Nevertheless, although this paraphrasing gives a 
number of indications concerning the type of 
relation that is expressed by the connective, it is 
not entirely satisfactory. Consider the following 
examples: 
5. ?He's handsome because he loves pasta. 
But other linguistic elements may influence the 
reading of the overall fragment. This is 
especially the case with the perspective 
indicators. Adding such a perspective indicator 
to (5) seems to make it already a little more 
acceptable to an epistemic reading (see Section 
2.5). 
. D: De rivier zal .wel hoog staan want her 
heeft weer de hele nacht geregend. 
D: De rivier zal wel hoog staan omdat het 
weer de hele nacht heeft geregend 
E: The river is surely high because it has 
been raining all night. 
. D: Het zal weer de hele nacht geregend 
hebben, want de fivier staat erg hoog. 
D: Het zal weer de hele nacht geregend 
hebben, *omdat de rivier erg hoog staat. 
E: It must have been raining all night, 
because the river is high. 
On the basis of the above-mentioned 
paraphrases, the two examples are best 
categorised as being epistemic in meaning. This 
would mean that both the connectives want and 
omdat can express an epistemic relation and that 
they should be mutually substitutable in this 
context (cf. example (3)). Nevertheless, this 
seems not to be the whole story. In example (4), 
for instance, only the use of the connective want 
is acceptable. On the basis of the paraphrase 
alone one should then accept that the connective 
omdat sometimes can and sometimes cannot 
express an epistemic relation. From a functional 
point of view this is not satisfactory. In the 
following section, I will show that the 
underlying meaning of the causal relation also 
plays a restrictive role on the selection of these 
causal connectives. 
In addition, this example illustrates another 
problem linked to the use of paraphrases. First, 
the paraphrase itself should be formulated in a 
6. He really must be handsome, because he 
loves pasta. 
So the underlying assumption of a paraphrase 
test is that the meaning of a connective is 
compatible with the meaning of the two related 
segments but that it cannot itself impose a 
relational meaning. In other words, it is not 
enough to put a causal connective between any 
of two segments to obtain a causal meaningL 
2,4 The nature of the underlying relation 
In causal relations the underlying meaning 
reflects the causal inference to be made. This 
causal reasoning can proceed in three ways: 
deductively, inductively or abductively. It is 
my claim that the type of causal reasoning that 
underlies the causal relation plays a restrictive 
role on the selection of connectives (at least in 
Dutch). 
In general terms, a causal reasoning process 
consists of three elements: rule (general law), 
circumstance (specific case), and result, like in 
the following example: 
Rule: Every time when it rains (a lot), the river 
is high. 
Circumstance: It has rained a lot. 
Result: The river is high. 
2 But, Degand (1998) shows that the connective itself 
may sometimes force a given reading. E.g. the 
subjectifying effect of the connective want 
('because/for'), and the objectifying effect of the 
connective omdat ('because'). 
32 
Deduction is reasoning from circumstance and 
rule to the result: It has rained a lot (and every 
time when it rains (a lot), the river is high.), so 
the river is high. In a backward deductive causal 
relation the consequence (result) in the first 
segment (SI) is related causally to the 
cause/reason (circumstance) in the second 
segment ($2). In Dutch, this relation can be 
expressed both by want and omdat (see example 
(7)). 
. D: De rivier staat hoog want het heeft de 
hele nacht geregend. 
D: De rivier staat hoog omdat het de hele 
nacht geregend heeft. 
E: The river is high because it rained all 
night. 
Induction is reasoning from (several) results and 
circumstances to a rule: The river is high (and it 
has rained al lot, and last week when the river 
was high, it also had rained a lot .... ), so when it 
rains a lot, the river is high 3. 
Abduction, finally, is reasoning from result and 
rule to circumstance: The river is high (and 
every time when it rains (a lot), the river is 
high.), so it has been raining a lot. In a 
backward abductive causal relation a 
cause/reason (circumstance) in S 1 is inferred on 
the basis of a consequence (result) in $2. This 
inference can only take place if there is indeed a 
real causal link between the two segments 
(Mackie 1975; Lagerwerf & Oversteegen 1994). 
If the two segments are substituted by one 
another, one should again obtain a purely 
deductive relation. This reversibility of the 
reasoning pattern is a good way to operationalise 
the deductive-abductive distinction. 
In my opinion, a (backward) abductive relation 
can exclusively be expressed by the connective 
want (see example (8)). 
. D:Het heeft veel geregend, want de fivier 
staat hoog. 
D: *Het heeft reel geregend, omdat de 
rivier hoog staat. 
E: It has rained a lot, because the river is 
high. 
Now, the question that rises is to what extent the 
underlying reasoning pattern may contribute to 
the classification of the causal relation into the 
"classical" ideational and interpersonal 
discourse domains. One could argue that a 
deductive reasoning pattern is ideational in 
nature because it is close to the real cause- 
consequence process. This would then mean 
that example (2) above expresses an ideational 
relation (marked in Dutch by want or omdat), in 
spite of the presence of perspectivising elements. 
When these are left out as in (7), the ideational 
nature of the causal relation appears very 
clearly. With respect to the abductive reasoning 
pattern one could argue that it is interpersonal, 
and more specifically epistemic, in nature 
because it reflects a causal reasoning in the 
mental space of the speaker ~. Here, presence or 
absence of perspectivising elements does not 
change anything to the situation (see example 
(8)). If we take this deductive-abductive 
distinction to be indicative for the ideational- 
interpersonal distinction, then the discourse 
domains would have a restrictive role on the 
selection of connectives. Namely, want and 
omdat can both express an ideational (deductive) 
relation, while an interpersonal (abductive) 
relation can only be expressed by want. 
However, it might be counterintuitive and 
contrary to the general domain definitions to 
classify examples (3), (6), or (9) as being 
ideational in nature only because they express a 
deductive reasoning pattern. A closer look 
should be taken to these other elements which 
seem to influence the categorisation, namely 
perspective indicators. 
. D: Door al die milieuwetten mag je het riet 
niet meer verbranden of in een sloot 
'afdrijven'. Dat is allemaal onzinnig, want 
riet is een natuurlijk produkt. (NRC, 1994) 
E: Because of all these environment laws 
you're not allowed anymore to burn the reed 
or to drive it 'off' the ditch. That is all 
nonsense, because/for reed is a natural 
product. 
3 1 did not find this kind of expressions in my corpus. 
Most authors who refer to epistemic relations 
illustrate these with constructed examples of the 
abductive reasoning type (e.g. Sweetser, 1990). 33 
2.5 Perspectivisation 
In the previous section, it appeared that presence 
or absence of perspectivising elements can play 
a role in the categorisation of coherence 
relations in the ideational or interpersonal 
(epistemic) domain. Perspective indicators are 
elements that have to do (i) with the 
responsibility for the truth of the utterance 
(verbs of saying and perception, explicit quotes 
or references .... ), (ii) with the logico-semantic 
properties of the utterance (negation, epistemic 
modals .... ), or (iii) with the evaluation of the 
utterance (predicates of judgement, evaluation, 
comparatives, etc.). At first view, perspectivised 
segments are best categorised within the 
interpersonal domain because they mostly refer 
to the opinion or point of view of the speaker (or 
another utterer), i.e. they have an illocutive 
character. 
But then, how do perspectivisation, surface cues, 
and reasoning pattern relate to each other? And 
which elements should be taken into account for 
the categorisation into discourse domains? 
Contrary to the causal reasoning pattern, 
perspectivisation does not play a restrictive role 
on the selection of connectives in Dutch. For 
instance, both want and omdat can occur in a 
perspectivised context, Moreover, as 
exemplified above, both abductive and 
deductive reasoning patterns can involve 
perspectivised segments. If we agree on the fact 
that perspectivising elements set up an 
interpersonal context, we then obtain a fairly 
heterogeneous interpersonal discourse domain 
that is only partially restrictive on surface cues 
(in Dutch). To account for this, I propose a 
triple subcategofisation for the interpersonal 
domain: epistemic-deductive, epistemic- 
abductive, and speech-act relations. This 
subcategofisation is justified by the fact that 
they share a number of general properties with 
the interpersonal domain (illocutive character), 
but that the epistemic-abductive and speech-act 
relations are far more restrictive with respect to 
the surface cues that may be used to signal 
them 5. 
A final question is then whether 
perspectivisation automatically leads to 
5 In Dutch, among the causal connectives only want 
('because') and dus ('so') may be used to signal an 
episternic or speech-act relation. 
interpersonal categorisation. This seems not to 
be the case. As a matter of fact, perspective 
switch (change of perspective between the first 
and the second segment), type of perspective 
and explicitness of the perspective also play a 
role. There seems to be a tendency for segments 
with perspective switch in which the perspective 
(implicitly) refers to the evaluation of the 
utterance to set up an interpersonal context. 
When there is a continuing perspective, with 
explicit perspective indicators referring to the 
responsibility for the truth of the utterance, the 
tendency is rather ideational. When the 
perspective has to do with the logico-semantic 
properties of the utterance the tendencies are not 
so clear. 
What is furthermore interesting to note is that 
the connectives themselves seem to add some 
meaning to the reading of the overall fragments. 
For example, even in absence of any 
perspectivising elements there is a tendency for 
the connective want to have a subjectifying 
effect (closer to the interpersonal domain). The 
relation appears to be one of justification or 
explanation rather than factual causal relation. 
This is not the case for the connective omdat 
which tends to impose a factualising reading 
with an objectifying effect even in presence of 
perspectivising elements (closer to the ideational 
domain). 
So, the situation appears to be fairly complex 
since the presence/absence of perspectives and 
the type of connective used in the fragment seem 
to influence the categorisation in the discourse 
domain. The problem is then to find a way to 
operationalise the different types of 
perspectives, since a same perspective can have 
a different effect according to the connective 
with which it is used. 
34 
Conclusion 
In this paper, I have tried to show that the 
classification of coherence relations and 
connectives into discourse domains is a complex 
matter in which different elements play a role: 
the connectives themselves whose meaning can 
be explicated on the basis of a substitution test 
and/or paraphrase test; the underlying meaning 
of the relation (deductive or abductive), which in 
the case of causals seems to play a restrictive 
role on the connectives; other features in the 
fragments like the presence or absence of 
perspectivising elements. Together these 
elements co-determine the categorisation of the 
coherence relation into one of the discourse 
domains. 
In my view, some of these elements have a 
higher weight than others in the classification: 
1) the distinction between deductive and 
abductive reasoning enables to sort out the 
epistemic-abductive class straightforwardly. 
2) Paraphrasing should separate the speech-act 
relations from the other interpersonal and 
ideational relations. 
3) The distinction between epistemic- 
deductive and ideational relations would be 
based on perspectivising elements 
(perspective switch, explicitness of the 
perspective and type of perspective). 
4) The substitution test should be useful once 
the categories have been established in 
order to come to finer distinctions. 
The suitability of these different tests will appear 
from their confrontation with real texts. 
Acknowledgements 
I would like to thank two anonymous reviewers 
and Yves Bestgen for very useful comments on 
earlier versions of this paper. 

References 
Degand, L. (1996). A Situation-based Approach to 
Causation in Dutch with some Implications for 
Text Generation. Doctoral dissertation, Universit6 
catholique de Louvain. 
Degand, L. (1998). Het ideationele gebruik van want 
en omdat: een geval van vrije variatie ? \[The 
ideational use of want and omdat : a case of free 
variation ?\] To appear in Nederlandse Taalkunde. 
Dijk, T. van (1979). Pragmatic connectives. Journal 
of Pragmatics 3,447-456. 
Grote, B., Lenke, N. & Stede, M. (1997) Ma\[r\]king 
concessions in English. Discourse Processes 24, 1, 
87-118. 
Halliday, M.A.K. & Hasan, R. (1985). Language, 
Context and Text: Aspects of language in a social 
semiotic perspective. Geelong, Victoria: Deakin 
University Press. 
Knott, A. (1996). A Data-Driven Methodology for 
Motivating a Set of Coherence Relations. Ph.D. 
thesis, University of Edinburgh. 
Knott. A. & Dale. R. (1994). Using linguistic 
phenomena to motivate a set of rhetorical relations. 
Discourse Processes 18, 35-62. 
Lagerweff. L. & Oversteegen, L. (1994). Inferring 
coherence relations through presuppositions. In: P. 
Bosch & R. Sandt (eds.). Focus and Natural 
Language Processing. Proceedings of a conference 
on the 10 ~ anniversary of the Joumal of Semantics. 
Germany: Schloss Wolfsbrunnen, 485-502. 
Maier, E. & Hovy, E. (1993). Organizing discourse 
structure relations using metafunctions. In: H. 
Horacek & Zock, M. (Eds.). New Concepts in 
Natural Language Generation. Pinter Publishers, 
London, 69-86. 
Mann, W.C. & Thompson. S.A. (1988). Rhetorical 
Structure Theory: toward a functional theory of 
text organization. Text 8. 243-281. 
Mann, W.C. & Thompson. S.A. (Eds.) (1992). 
Discourse Description: Diverse Analyses of a 
Fund Raising Text. John Benjamins, Amsterdam. 
Martin, JR. (1992). English Text: systems and 
structure. John Benjamins, Amsterdam. 
Oversteegen, L. (1997). On the pragmatic nature of 
causal and contrastive connectives. Discourse 
Processes, 24. 1, 51-86. 
Pander Maat, H. & Sanders, T. (1995). Nederlandse 
causale connectieven en het onderscheid tussen 
inhoudelijke en epistemische coherentie-relaties. 
\[Dutch causal connectives en the distinction 
between content and epistemic coherence 
relations\]. Leuvense bijdragen 84, 349-374. 
Pit, M., Pander Maat, H. & Sanders. T. (1997). 
"Doordat'. "omdat' en 'want'. Perspectieven op 
hun gebruik. \['Doordat', 'omdat' and 'want'. 
Perspectives on their use.\] Taalbeheersing, 3, 238- 
251. 
Sanders, T.J.M. (1992). Discourse Structure and 
Coherence; Aspects of a Cognitive Theory of 
Discourse Representation. Doctoral dissertation, 
Tilburg University. 
Sanders, T. (1997). Semantic and Pragmatic Sources 
of Coherence: On the Categorization of Coherence 
Relations in Context. Discourse Processes, 24, 119- 
148. 
Spooren, W. & Jaspers, J. (1990). Tekstoperaties en 
tekstperspectieven \[Text operations and text 
perspectives\], Gramma. tijdschrift voor taalkunde 
14 (3), 195-220. 
Sweetser, E. (1990). From Etymology to Pra~m-natics: 
Metaphorical and cultural aspects of semantic 
structure. Cambridge: Cambridge University Press, 
\[Studies in Linguistics 54\]. 
