Volume 8 Issue 1 (2010)
Note: Linguistic Discovery uses Unicode characters
to represent phonetic symbols. Please see Optimizing Display
for requirements to accurately reproduce this page.
What Do Semantic Maps Tell Us?
Comment on ‘Semantic Maps and Mental Representation’
by Sonia Cristofaro (2010)
Cristofaro (henceforth C) challenges the hypothesis
advocated by many users of the semantic map model in typology, including myself,
that the structure of conceptual space in the semantic map model reflects
conceptual similarity relationships between situation types that are part of the
mental representations of speakers. C arrives at two general conclusions about
the relationship between the semantic map model and mental representations.
First, she argues that the putative similarity relationships between the
situation types expressed by linguistic forms which are captured in semantic
maps are not actually similarity relationships present in a speaker's mental
representation of those situation types. Second, she argues that semantic maps
‘pertain to the principles that govern the creation of novel constructions
at the diachronic level, independently of synchronic grammatical representation
in a speaker’s mind’. In this commentary, I argue that the semantic
map model does in fact reflect conceptual similarity relationships which are
part of a speaker’s knowledge.
C’s second conclusion implies a much sharper division between
synchrony and diachrony than actually exists. There is not really a
‘diachronic level’ independent of speaker’s knowledge about
their language. Diachronic processes begin with speakers making linguistic
choices in the process of verbalizing their experiences in utterances (Croft
2000). These choices cannot be made without a speaker using her knowledge about
her language, no matter how fragmentary that may be (as in the case of limited
second language speakers). If the semantic map model represents something about
diachronic processes, then they represent something about a speaker’s
mental representation, i.e. her knowledge about her language.
It is of course possible that it is not similarity relationships which
are represented. C’s first conclusion is that the supposed similarity
relationships in the semantic map model are not that. Instead, they are the
result of processes that do not involve conceptual similarity. C gives two
examples of such processes. The first is the hypothesis that metonymic processes
drive certain semantic changes such as grammaticalization and
“generalization” (the examples that C gives all appear to involve
metonymy, and “generalization” in the usual sense of that term
actually does not play any role in the process). The second is that processing
and/or token frequency determines certain typological universals such as the
animacy hierarchy, not conceptual similarity relationships among the elements in
an implicational hierarchy.
I will start (again) with C’s second argument, regarding
implicational hierarchies such as the animacy hierarchy. In
(Croft 2001), I proposed that typological markedness
patterns, including certain implicational hierarchies, could be captured by
aspects of the structure of the conceptual space. However, I realized that was
incorrect by the time I wrote the new edition of
Typology and Universals
(Croft 2003). In the latter book, I argue that typological markedness patterns
were a function of properties of the semantic maps—that is, not the
conceptual space itself but the language-specific categories that are mapped
onto the (presumably universal) conceptual space. Typological markedness is a
consequence of the token frequency of the linguistic categories in question.
For example, nominative case is less marked with respect to structural
coding than accusative case because the token frequency of the category mapped
by a nominative form (the combination of A plus S roles in the conceptual space
of grammatical relations) is higher than the token frequency of the category
mapped by an accusative form (P in the conceptual space). The absolutive is also
less marked in structural coding than the ergative, and that is because the
combination of roles mapped by the absolutive (S plus P) has a higher token
frequency than the roles mapped by the ergative (A). What matters is that A+S
outnumbers P in token frequency, and S+P outnumbers A; and this is a property of
the semantic maps, not of the conceptual space itself. More precisely, it does
not have to do with the relations between situation types represented by the
links in the conceptual space. It is those links that are intended to reflect
similarity relations between the situation types in the conceptual space. It
does, of course, have to do with the (combined) token frequencies of the
individual situation types in the conceptual space.
More generally, in
Typology and Universals I argue that there are
typological generalizations which have the effect of constraining semantic maps,
over and above the constraints imposed on possible semantic maps (i.e. possible
linguistic categories) by the links between situation types in a conceptual
space. The constraints imposed on possible linguistic categories by the links
between situation types are summarized in the Semantic Map Connectivity
Hypothesis: that a possible linguistic category (semantic map) is constrained to
be a connected region in the conceptual space. The explanation behind the
Semantic Map Connectivity Hypothesis is based on the hypothesis that the links
represent similarity relationships: situation types are grouped under a single
linguistic category only if they are similar enough to each other in the
relevant semantic dimensions. But these are not the only constraints on possible
semantic maps. That is, the Semantic Map Connectivity Hypothesis is insufficient
to describe possible linguistic categories.
For example, in the animacy hierarchy illustrated in (1), not only is it
the case that the Semantic Map Connectivity Hypothesis holds (categories must be
connected in the linear conceptual space), but that a linguistic category such
as plurality must always include the “highest” function in the
space, namely 1st/2nd person:
1st/2nd person — 3rd person — human N — animate N
— inanimate N
One of Cristofaro’s examples also illustrates the need
for further constraints and explanations for that constraint. She cites
König and Siemund’s observation that no language in their survey
includes the derived intransitivity function under the same marker as reflexives
if that marker is also used for the intensifier function, despite the fact that
all three are linked in a grammaticalization path/conceptual space as in
Intensifier — Reflexive — Derived intransitivity
These further constraints call for explanation, of course.
(There is a plausible explanation for König and Siemund’s
observation, incidentally: the process that extends reflexives to the derived
intransitivity function is much slower than the process that replaces old
intensifiers—which have acquired a reflexive function—with new
intensifiers; this latter process is known to be relatively rapid.) But the fact
that there are further constraints on possible linguistic categories/semantic
maps does not invalidate the Semantic Map Connectivity Hypothesis or the
similarity relationships that it embodies. For example, it is true that the
functions found at the “upper end” of the animacy hierarchy have a
greater token frequency: we talk more about ourselves than others, and more
about people than nonhuman entities, and so on. For this reason, plurals are
more likely to occur in categories at the “upper end” of the
animacy hierarchy (that is, such categories have a greater behavioral potential,
thanks to their greater token frequency; see Croft 2003). But the frequency
effect does not explain why one does not find a language with plural marking for
personal pronouns and nonhuman animate nouns but not for human nouns, even
though the personal pronoun + animate noun category would have a higher token
frequency than the human noun category. In other words, we still need the
animacy conceptual space structure, and the similarity relationships it
reflects, in order to explain the linguistic categories we find, even if we need
other things as well.
I now turn to the more challenging question of what the semantic map
model has to say about the sorts of categories that C describes which appear to
involve metonymy. The first question I would like to raise is whether the
semantic map model should be expected to apply to these cases at all.
The semantic map model, and more sophisticated models to capture the
same phenomena such as multidimensional scaling (Croft and Poole 2008), presumes
that the data in question are similarity data. The assumption behind the
applicability of the semantic map model and of multidimensional scaling to
typological linguistic data is that crosslinguistic category data do provide
evidence of similarity relations. Here, the hypothesis is that if a language
encodes two situation types under the same linguistic form, then there is some
conceptual similarity between those two situation types in the mind of a
speaker. This hypothesis seems quite reasonable for many of the linguistic
phenomena for which the semantic map model has been used. For example, the
implicational hierarchies such as the animacy hierarchy presuppose that the
values in question (i.e. the functions) form a set of paradigmatic alternatives
(Croft 2003). Haspelmath’s semantic map analysis of indefinite pronouns,
and the multidimensional scaling analyses of spatial adpositions by Levinson et
al. (2003) and Croft and Poole (2008) also involve paradigmatic alternatives.
The same is true of the grammatical relations space in Croft (2001); they are
all participant roles. In some cases, two related, paradigmatically-defined
conceptual dimensions are crossed: in the parts of speech space in Croft (2001),
propositional act functions are crossed with lexical semantic classes, and in
the tense-aspect space in Croft and Poole (2008), time reference is crossed with
One could argue that in the examples which C presents, the hypothesis
that formal similarity implies conceptual similarity simply does not hold. The
metonymic shifts lead to jumps from one semantic dimension to another, and there
is no paradigmatic relationship between the two, for example between body parts
and spatial relations, or between time and contrastiveness (adversative). Hence,
one could conclude that the semantic map model should not be applied in those
cases. In at least one case that C discusses, a semantic map model would in fact
not reveal the pattern in question. This is the case of the exaptation or
hypoanalysis of the imperfective as a subjunctive. As C notes (citing Bybee et
al. 1994:230-36 and Haspelmath 1998; see also Croft 2000:127-28), this results
when a progressive has taken over main clause contexts; the imperfective is
restricted to subordinate clause contexts and is reanalyzed as a subjunctive.
But the typological data for a semantic map model would not pick this up: the
imperfective acquires its subjunctive function after it has lost its
imperfective function to the progressive, so one would not find languages with
both functions synchronically. The synchronic mutual exclusivity of meanings is
the usual case in exapation/hypoanalysis.
There are certainly other typological phenomena to which the semantic
map model (or multidimensional scaling) should not be applied. For example, word
order universals do not imply any similarity relation between different word
orders (Croft and Poole 2008:10, fn. 1). And perhaps we should be cautious about
assuming that subsuming two functions under one form implies that the functions
are always similar. But I doubt that C is really making that argument. For
example, her examples of “generalization” could be analyzed as
members of a paradigmatic set, of modality (albeit divided into deontic and
epistemic modality). But C argues (following Traugott and Dasher) that the
semantic process leading to the grouping of different functions under a single
form are due to metonymy, not similarity: the different uses are associated in a
particular context of use, and the meaning in the context is transferred to the
form that expressed the other meaning. Typological studies of lexical polysemy
(Brown 1979; Evans 1992; Croft et al. 2010) demonstrate that some sort of
metonymy is a common motivation for the subsumption of different meanings under
a single lexical item. For example, the word for ‘moon’ very
frequently also means ‘month’: this is a metonymic relationship
between a celestial object and a temporal cycle associated with that
object’s behavior. Also, the word for ‘night’ very frequently
also means ‘evening’: this is a metonymic relationship between a
celestially-defined time period and the beginning phase of that period.
So we are back at the ancient opposition between association and
similarity, between metonymy and metaphor. The semantic map model assumes the
latter, while C’s putative counterexamples are claimed to be examples of
the former. I have suggested above that precisely because they are not examples
of similarity, then there might be reason to say that the semantic map model
does not and should not apply to them. C’s putative counterexamples
certainly do not invalidate the many cases where similarity is a very plausible
explanation for the processes of semantic extension which are presumably behind
the typological patterns, just as C’s observation that similarity is
not enough does not invalidate the role that similarity does play in grammatical
patterns. However, I would like to suggest that there is in fact not such a
strict dichotomy between association and similarity.
Semantic map models generally use very broadly-defined functions, such
as “property predication”, “ability modality”, or
“non-specific indefinite pronoun”. These categories are in fact
generalizations over large classes of specific occurrences of these meanings in
situation types produced in utterances. In some studies, however, the functions
are much more fine-grained. Dahl’s tense-aspect typological study (Dahl
1985), on which the tense-aspect spatial model in Croft and Poole (2008) was
based, consists of tense-aspect sentences elicited as translations of sentences
in a questionnaire. Levinson et al.’s typological study of spatial
adpositions, referred to above, uses sentences elicited in a picture description
task. These functions are basically utterances produced in a controlled context.
These studies point to the logical conclusion of the process: the conceptual
space should consist of points representing actual situation types verbalized by
linguistic forms in utterances (Croft, to appear). These instances of language
use are the exemplars on which semantic knowledge is constructed by the speaker.
The speaker’s conceptual space consists of all of the situation types she
has experienced, related to each other conceptually in ways that are reflected
by the choices of linguistic expressions used to verbalize those
For example, a conceptual space for participant roles might contain two
roles defined in a coarse-grained manner, ‘instrumental’ and
‘comitative’. These roles are certainly linked; a common
grammaticalization path leads from the latter to the former. They can also be
thought of as paradigmatic alternatives. But in fact, particular examples
suggest not only that a more fine-grained division of participant roles would be
desirable, but experimental evidence suggests that the
‘comitative’/‘instrumental’ roles can be ordered in a
one-dimensional conceptual space (Schlesinger 1979:310, cited in Heine et al.
1991:104; I am not ruling out the possibility that a higher-dimensional spatial
model might be more appropriate in this case):
The pantomimist gave a show with the clown.
The engineer built the machine with an assistant.
The general captured the hill with a squad of
The acrobat performed an act with an elephant.
The blind man crossed the street with his dog.
The officer caught the smuggler with a police dog.
The prisoner won the appeal with a highly paid lawyer.
The Nobel Prize winner found the solution with a
The sportsman hunted deer with a rifle.
The hoodlum broke the window with a stone.
This observation is familiar to those who do corpus-based
studies of the semantics of particular words or construction types: many
actually occurring instances do not easily fit into categories defined a priori,
but the instances are semantically related and sometimes almost form a
continuum, as in the examples in (3a-j). In this exemplar-based approach to
semantics, meanings are mapped onto a very fine-grained conceptual space. As a
result, the meaning of a word such as
with is actually deeply embedded or
intertwined with its context, both linguistic (as in the constructed examples in
3a-j) and nonlinguistic (in actual usage).
I now return to C’s examples with the perspective of this
exemplar-based, context-sensitive representation of the meaning of individual
linguistic forms. Consider the examples of the simultaneity and adversity uses
while in the history of English cited by C, from Hopper and Traugott
(2003:91; I have added the final invented Modern English example from Hopper and
ðaet lastede þa [xix] winttre wile
Stephne was king
‘That lasted those 19 winters while Stephen was king’
(ChronE [Plummer] 1137.36)
Whill others aime at greatnes boght with blod,
bee great thou stryves, bot to bee god
‘While others aim at greatness that is bought with blood,
you strive to be not great but good’ (1617, Sir W. Mure,
While you like peaches, I like nectarines.
Simultaneity and adversity appear to be very different
meanings. But as C notes, following Hopper and Traugott, events reported in
discourse as simultaneous are often in contrast as well, and this has led to the
extension of the meaning of
while from simultaneity to contrastivity. C
interprets this as a metonymic extension, because simultaneity and adversity are
associated in a particular utterance such as (4b). But it can be looked at in a
different way. Each of the utterances in (4a-c) represents a whole situation
type. Even if we focus solely on what the connective contributes to the
situations expressed in (4a-c), we see that there is indeed conceptual
similarity between (4a) and (4b), and between (4b) and (4c). Most crucial for
the semantic map model, there is a lower degree of similarity between (4a) and
(4c); this is crucial because what makes the semantic map model valuable is the
ability to posit different degrees of similarity. The similarity is found in the
real world, where simultaneity and adversity occur together in the construal of
paired situations for communicative purposes. But it is similarity
The example of English
while is just one data point. A
typological study would reveal just how close these situation types really are.
I would predict with some confidence that a typological study would show that
the semantic relationship between clauses in (4a) and (4c) is less close than
that between either of those situation types and the situation type in (4b).
These three types would not exhaust the conceptual space. Hopper and Traugott
(2003) for example mention that in German,
weil has become causal (and
only very recently has it become concessive, via the causal path according to
Hopper and Traugott). But my point here is that one must take a more holistic
view of the situation types being represented in a conceptual space, and then
the metonymic examples may turn out to represent a kind of similarity after
Even the lexical semantic examples I mentioned above may be amenable to
a more holistic approach. ‘Moon’ and ‘month’ seem very
different semantically. But in the frame-semantic or semantic domain approach to
meaning introduced by cognitive linguistics (see especially Fillmore 1982, 1986;
Langacker 1987, ch. 4; a survey is found in Croft and Cruse 2004),
‘moon’ and ‘month’ are in the same semantic frame, or
more precisely, the same complex matrix (Langacker) of frames/domains which
includes the celestial object itself and its periodic appearance and movement in
the sky over time. I have argued that metonymy involves a shift in which a
domain in a matrix (or a facet of a word’s meaning, to use Cruse’s
terminology) is highlighted and in which a concept in the frame is profiled
(Croft 1993/2002). The semantic domain matrices for ‘moon’ and
‘month’ are therefore very similar, differing only in which domain
or facet is highlighted, and therefore also which concept is profiled. This is
only part of the story. We do not have an explanation for which concepts are
profiled when a different domain is highlighted. We also do not have an
explanation for why some metonymies are more common than others (e.g.
‘sun’ is far less likely to be extended to ‘day’ than
‘moon’ is to ‘month’). But a major part of the
relationship is similarity in a holistic sense: both concepts share highly
similar domain matrices.
This is not to say that all cases of different meanings subsumed under
the same form can or should be interpreted as similarity. Frame-semantic
similarity constitutes a shared semantics of a different type than similarity of
concepts that are paradigmatic alternatives, or the intermediate case of
holistic, contextually-defined similarity that is found in invited inference in
grammaticalization. Nor is it to say that similarity is all that is going on in
semantic change. But similarity—conceptual similarity as represented in
the minds of speakers—does play a major role in typological patterns,
synchronic and diachronic.
Brown, Cecil H. 1979. A theory of lexical change (with examples
from folk biology, human anatomical partonomy and other domains).
Anthropological Linguistics 21.257-76.
Croft, William. 1993/2002. The role of domains in the
interpretation of metaphors and metonymies. Cognitive Linguistics 4.335-70.
Revised version printed 2002 in Metaphor and metonymy in comparison and
contrast, ed. by René Dirven and Ralf Pörings, 161-205. Berlin:
Mouton de Gruyter. doi:10.1515/9783110219197.2.161
-----. 2000. Explaining language change: An evolutionary
approach. Harlow, Essex: Longman.
-----. 2001. Radical construction grammar: Syntactic
theory in typological perspective. Oxford: Oxford University
-----. 2003. Typology and universals, 2nd edition.
Cambridge: Cambridge University Press.
-----. to appear. Exemplar semantics. To appear in a
volume edited by Seana Coulson. Stanford: Center for the Study of Language and
Information. Available online at
Croft, William, Clayton Beckner, Logan Sutton, Jon Wilkins, Tanmoy
Bhattacharya and Daniel Hruschka. 2009. Quantifying semantic shift for
reconstructing language families. Poster, 83rd Annual Meeting of the Linguistic
Society of America, San Francisco, California, January.
Croft, William and D. Alan Cruse. 2004. Cognitive linguistics.
Cambridge: Cambridge University Press.
Croft, William and Keith T. Poole. 2008. Inferring universals from
grammatical variation: Multidimensional scaling for typological analysis.
Theoretical Linguistics 34.1-37. doi:10.1515/thli.2008.001
Dahl, Östen. 1985. Tense and Aspect Systems. Oxford:
Evans, Nicholas. 1992. Multiple semiotic systems, hyperpolysemy,
and the reconstruction of semantic change in Australian languages. Diachrony
within synchrony: Language history and cognition, ed. by Günter Kellermann
and Michael D. Morrissey, 475-508. Frankfurt: Lang.
Fillmore, Charles J. 1982. Frame semantics. Linguistics in the
morning calm, ed. by The Linguistic Society of Korea, 111-137. Seoul:
-----. 1985. Frames and the semantics of understanding.
Quaderni di semantica 6.222-254.
Heine, Bernd, Ulrike Claudi and Friederike Hünnemeyer. 1991.
Grammaticalization: A conceptual framework. Chicago: University of Chicago
Hopper, Paul and Elizabeth Traugott. 2003.
Grammaticalization, 2nd edition. Cambridge: Cambridge University
Langacker, Ronald W. 1987. Foundations of Cognitive Grammar, vol.
1: Theoretical prerequisites. Stanford: Stanford University Press.
Levinson, Stephen, Sérgio Meira and the Language and
Cognition Group. 2003. ‘Natural concepts’ in the spatial topological
domain: Adpositional meanings in crosslinguistic perspective: An exercise in
semantic typology. Language 79.485-516.
Schlesinger, Itzchak M. 1979. Cognitive structures and semantic
deep structures: The case of the instrumental. Journal of
Linguistics 15.307-24. doi:10.1017/s0022226700016418