Levinson 2003 Natural

‘NATURAL CONCEPTS’ IN THE SPATIAL TOPOLOGICAL
DOMAIN—ADPOSITIONAL MEANINGS IN CROSSLINGUISTIC

PERSPECTIVE: AN EXERCISE IN SEMANTIC TYPOLOGY
STEPHEN LEVINSON SÉRGIO MEIRA THE LANGUAGE AND
Max Planck Institute for Leiden University COGNITION GROUP
Psycholinguistics Max Planck Institute for
Psycholinguistics
Most approaches to spatial language have assumed that the simplest spatial notions are (after
Piaget) topological and universal (containment, contiguity, proximity, support, represented as
semantic primitives such as IN, ON, UNDER, etc.). These concepts would be coded directly in
language, above all in small closed classes such as adpositions—thus providing a striking example
of semantic categories as language-specific projections of universal conceptual notions. This
idea, if correct, should have as a consequence that the semantic categories instantiated in spatial
adpositions should be essentially uniform crosslinguistically. This article attempts to verify this
possibility by comparing the semantics of spatial adpositions in nine unrelated languages, with
the help of a standard elicitation procedure, thus producing a preliminary semantic typology of
spatial adpositional systems. The differences between the languages turn out to be so significant
as to be incompatible with stronger versions of the UNIVERSAL CONCEPTUAL CATEGORIES hypothesis.
Rather, the language-specific spatial adposition meanings seem to emerge as compact subsets of
an underlying semantic space, with certain areas being statistical ATTRACTORS or FOCI. Moreover,
a comparison of systems with different degrees of complexity suggests the possibility of positing
implicational hierarchies for spatial adpositions. But such hierarchies need to be treated as succes-
sive divisions of semantic space, as in recent treatments of basic color terms. This type of analysis
appears to be a promising approach for future work in semantic typology.*
1. ADPOSITIONS AS A DISTILLATION OF HUMAN SPATIAL COGNITION. In studies of spatial

language, a standard line or set of orthodox assumptions has arisen, along the following
lines:
1. The simplest spatial notions are topological—concepts of proximity, contiguity,
containment (Piaget & Inhelder 1956).
2. Such notions can be taken to be either primitive, so that we have conceptual
primes like IN, ON, UNDER (Jackendoff 1983), or near-primitive, so that, for
example, IN is decomposed in terms of at least partial INCLUSION (Miller &
Johnson-Laird 1976).
3. These concepts are more or less directly coded in spatial language, above all in
the closed-class spatial relators like prepositions and postpositions, which have
(comparatively) simple semantics (Talmy 1983), largely universal in nature since
they correspond to elements of our neurocognition (Landau & Jackendoff 1993).
Consequently, ‘we can develop a fairly comprehensive idea of the spatial relations
expressed in language by focusing on spatial prepositions’ (Landau & Jackendoff
1993:223).
4. Hence, the topological adpositions are among the earliest linguistic concepts
learned by children (Johnston & Slobin 1979), and in learning them children map
* Colleagues in the Language and Cognition Group who provided crucial data are: Jürgen Bohnemeyer,
Angela Terrill, Raquel Guirardello, Nick Enfield, Iraide Ibarretxe-Antuñano, Felix Ameka, David Wilkins,
and Carlien de Witte. We do not hold them to the theoretical views developed here on the basis of their
data. We are grateful to Melissa Bowerman for intensive discussions, and the influence of her ideas will be
evident throughout. Many helpful suggestions were gratefully received from Brian Joseph, Adele Goldberg,
and two anonymous referees.
485
486 LANGUAGE, VOLUME 79, NUMBER 3 (2003)
prelinguistic universal spatial concepts directly onto words (H. Clark 1973, E.
Clark 1974), suggesting that we have rich innate concepts in this field (Li &
Gleitman 2002).1
If these ideas are correct, they would be important clues to the general relation
between semantics and cognition. They would support the idea that semantic categories
are basically projections of universal conceptual categories and thus are essentially
uniform across languages (Pinker 1994, Li & Gleitman 2002).
Studies by our research group suggest that in fact there are many mistaken steps in
this argument. Quite precise and complex axial geometry seems to be involved in so-
called topological concepts (Levinson 1994); notions like IN or ON do not seem to be
primitive holistic concepts (Brown 1994, and the current paper),2 many languages seem
to make alternative kinds of distinctions, which are learned just as early (Bowerman
1996, 2003); in many languages topological concepts are wholly or partially expressed
in contrastive locative verbs (Ameka & Levinson 2003). So far, many of these counterar-
guments of ours have been mounted by looking at different ways languages code spatial
relations, for example in verbs. But here we focus on the central claims—we explore
just what kinds of notion are in fact coded crosslinguistically in spatial adpositions,
concentrating on those topological notions concerning nonprojective relations, like dif-
ferent kinds of contiguity and coincidence, which have been the central subject of
debate.
2. METHODS FOR COLLECTING DATA. First, we need an operational definition of the

formal class, adposition. Of course, the crosslinguistic isolation of comparable form-
classes is fraught with difficulties—ultimately, as Greenberg 1966 pointed out, one
has to resort to meaning. Nevertheless, we have found the following working definition,
which combines semantic and syntactic criteria, good enough for current purposes: a
spatial adposition is any expression that heads an adverbial phrase of location in the
BASIC LOCATIVE CONSTRUCTION (answers to where-questions). This definition is not
designed to exclude SPATIAL NOMINALS, since they so often gradually develop into ‘true’
adpositions that boundary problems would plague a comparative exercise of this sort.
Local cases, in languages for which they form a complicated system (e.g. Finnish,
Hungarian, Tamil) should also be considered in a wider sample (though it so happens
that our sample contains no language with a complex locative case system); we do not
expect them to pattern significantly differently from locative adpositional systems with
respect to our elicitation procedure (cf. Bowerman and Pederson (1992, 2003), who
do consider such languages). Given the distributed nature of spatial semantics (Sinha &
Kuteva 1994), there are other means of expressing topological relations; locative/posi-
tional verbs are one example that has already received some attention (Ameka & Lev-
inson 2003, Brown 1994, Levinson 2000a). We will call these various form classes
involved in coding topological relations TOPOLOGICAL RELATION MARKERS —or TRMs
for short. However, considering the history of concentration on adpositions and the
theoretical presumptions to be found therein, we believe it is important to set the
crosslinguistic record straight on this narrower topic by concentrating on adpositions.
1
See, incidentally, our response to Li and Gleitman in Levinson et al. 2002.
2
We use English terms like ON in full caps as a very loose metalanguage for central meanings of the
relevant sort. The tradition uses these as hypothetical semantic primitives (see e.g. Jackendoff 1983), but as
will become clear, crosslinguistic comparison shows that any semantic primitives will have to be at a much
finer level of discrimination.
‘NATURAL CONCEPTS’ IN THE SPATIAL TOPOLOGICAL DOMAIN 487
The relations between the semantics of locative adpositional systems and complex
locative verb systems (such as the ones found in Mayan languages) remain a target for
future research.3
The fine-grained meaning differences we are interested in cannot be extracted from
normal dictionaries or grammatical descriptions. Instead, direct fieldwork is required.
Nor can this be done by unstructured elicitation if one is to obtain strictly comparable
results across languages. Instead, what is minimally required is structured elicitation
using a standard set of stimuli, an ETIC GRID which can be used to calibrate responses
across languages. In this we follow the pioneers of scientific comparative semantics,
as for example in the color work of Berlin and Kay (1969), which we think yields
much more reliable results than attempts to compare senses across languages without
carefully controlling reference (as, for example, in the work in the NATURAL SEMANTIC
METALANGUAGE tradition; Wierzbicka 1980). It is true that this method is open to certain
obvious objections—first and foremost (as Lucy 1997 has pointed out) the choice of
etic grid can ensure a false sense of familiarity, since one may be inclined to choose
a grid that makes just the kind of distinctions to be found in one’s own language. We
think this danger can be minimized by successive piloting and the construction of such
a grid by teams of fieldworkers who have extensive experience of the languages they
intend to investigate, but in any case the method is sounder than any of the available
alternatives.
The elicitation tool we used in this study was first developed by Melissa Bowerman,
and then extended in collaboration with Penelope Brown and especially Eric Pederson,
on the basis of experience with a number of non-Indo-European languages. Bowerman
and Pederson (2003) have, in work on forty languages, established a number of interest-
ing findings that we mention below (§4; see the summary in Bowerman & Choi 2001).
The resulting elicitation tool is a booklet of seventy-one line-drawings or pictures
(the TOPOLOGICAL RELATIONS PICTURE SERIES or TRPS for short), each representing a
topological spatial relation, covering a large range of spatial relations that would be
coded in English using such prepositions as on, in, under, over, near, and against, as
well as complex prepositions like inside, on top of, in the middle of, and such like.
Each picture has a designated FIGURE (or theme or trajector) colored yellow, and a
GROUND object (or relatum or landmark), and the researcher uses the pictures to set up
a verbal scenario as close as culturally possible to that depicted, and asks the consultant
to answer a question of the form: ‘Where is the [Figure]?’ (given the sketched scenario).
For some of the languages investigated, the Western cultural objects depicted (books,
tables, lights) did not have local counterparts, and some replacement parallel scenario
had to be verbally sketched. Figure 1 illustrates a few of the pictured scenarios, and
shows how, for example, they sample a range of intermediate types of spatial relation
between a clear containment relation at one end and a relation of contiguous spatial
superposition at the other.
3
Current research suggests two main types of contrastive locative verb (Ameka & Levinson 2003): small
contrastive sets of posture verbs (often glossing ‘sit’, ‘stand’, ‘lie’, ‘hang’, or the like), and large sets of
contrastive positional verbs (specifying exact disposition of the figure and the precise relation between figure
and ground). These have rather different semantic properties, so that, for example, the small sets have
classifying or presuppositional uses (the figure need not be in the canonical posture), while the large sets
have primarily assertional uses. It seems that the positional systems correlate with a relatively impoverished
adpositional/locative case system, while the small-set postural systems can coexist with fully elaborated
adpositional (or local case) systems (as in Yélı̂ Dnye, mentioned below; for the postural verb system see
Levinson 2000a). Postural systems are typologically common, positional systems typologically quite rare.
We touch on the relevance of this for our particular sample below.
FIGURE 1. Standard comparative stimulus material.
This elicitation tool does not attempt to cover the ‘projective’ meanings of adposi-
tions, that is, those that involve specifying an angle with respect to a ground object
and projecting a search-domain for the referent from that landmark object (e.g. English
behind in behind the town-hall). Projective concepts belong to a different conceptual
subdomain, where coordinate systems or frame of reference are necessary—see Lev-
inson 2001, 2003, for a sketch of the semantic typology here.
The responses to the stimuli are sentences of the kind: ‘It (the cup) is on the table.’
Where languages have contrastive spatial predicates, they may be of the sort ‘The cup
is sitting/standing/lying on the table’. Some languages may, in addition to adpositions,
deploy a series of spatial nominals, so that one obtains responses of the kind ‘The cup
is on the top/surface of the table’. Here we abstract away from these additional codings
of spatial discrimination in order to concentrate on the adpositions proper. We do not
think that this is wholly legitimate—serious comparative semantics must take into
account the full range of discriminations wherever and in whatever form-classes they
are made. However, the idealization is warranted as a response to the orthodox thesis
outlined above, namely the claim that the closed-class adpositions yield a specific
kind of abstract, universal spatial semantics. Our aim is to investigate whether this is
actually true.
The responses give us extensional maps: we can, for example, look at all the scenes
that elicit the preposition on in English, versus all those that elicit in or under, or other
prepositions. Most of this paper is devoted to what can be inferred directly from such
extensional maps. But obviously an analysis of the meanings or intensions that project
those extensions is also in order, although we cannot do more than sketch that here.
Semantic relations—antonymy, contrast, entailment—between terms tells one much
about their sense or intension. In general, we tend to think of these relations as simply
one of contrast: John is in the truck contrasts with John is on the truck, and the two
sentences can be thought of as semantic incompatibles (in and on contrast like, say,
oak and beech). But clearly John is outside the house is a contradictory of John is
inside the house, for if John is not inside the house is true, then John is outside the
house must also be true (assuming John exists as a physical body, of course). Other
less obvious relations may also hold: if The train is at the station is true, then The
train is near the station must be true, even though we would not describe it that way
for good Gricean reasons (see Levinson 2000a, 2000b:96). A full account of the meaning
of adpositions clearly requires much intensional analysis of this sort. What we have
found is that languages with large sets of spatial adpositions often seem to have taxo-
nomic relations between them. We can illustrate this with the adpositions of Tiriyó (a
Cariban language spoken in Brazil and Surinam), which can plausibly be arranged in
a taxonomic tree as in Figure 2. Here, subordinate terms are more specific: they have,
if one likes, additional features missing from their superordinate or more general
terms—thus, arguably, English inside is a more specific kind of in relation, namely
one in which enclosure (or at least convex closure) is complete. These hierarchical
relations have been completely ignored in the literature as far as we know, and they may
be of some importance in understanding crosslinguistic patterns, which such subordinate
categories help to obscure.
FIGURE 2. Hierarchy in Tiriyó adpositions.
A further important factor is the need to distinguish semantic from pragmatic factors.
We have already alluded to Gricean factors involved in the understanding of adpositions,
and we think that implicature plays an important role here. Consider for example the
taxonomic relations just mentioned—these introduce privative oppositions between
superordinate and hyponym, where the latter has (if one likes) additional semantic
features. Given this, the use of the superordinate implicates that the speaker is not in
a position to use the more informative expression (else he or she would be in breach
of Grice’s second maxim of Quantity, or Levinson’s 2000b Q-principle). Hence, other
things being equal, There’s a nail in the door suggests a nail not INSIDE the door, but
FIGURE 3. Choices between adpositions (Yélı̂ Dnye); 4 consultants, 7 scenes.
rather projecting from the plane of the door. The result is a pragmatic contrast found
exactly where one has semantic compatibility.
We have found that there are some rough-and-ready fieldwork methods that can be
used to detect such pragmatic patterns. Figure 3 shows the pattern of responses to seven
scenes by four consultants, native speakers of the Papuan isolate Yélı̂ Dnye (see also
Levinson 2000a). Two adpositions are in play, p:uu and ‘nedê, and one can see that
all consultants agree on some scenes, but that they are split on others. If one now looks
at the patterns of preferred responses versus those deemed acceptable but offered only
later, the picture clarifies further. Figure 4 shows the pattern of preferred responses for
the middle scene, clothes pinned on a line, where two consultants gave as their first
choices p:uu and two gave ‘nedê.4 The preferred responses show that consultants who
offer ‘nedê will back down to p:uu, but those who offer p:uu will not escalate to ‘nedê.
All this is compatible with an analysis whereby ‘nedê is more specific, a hyponym of
p:uu, such that the use of p:uu implicates that ‘nedê is not applicable. P:uu seems to
mean just ‘attached to’, while ‘nedê seems to mean ‘attached by spiking’ (more on
‘nedê below). A similar pattern can be observed with respect to first (spontaneous) and
second (checking) answers in Tiriyó for the postpositions juuwë ‘on top of’ and rehtë
‘on (summit of)’ (see Meira 2003). For two scenes, almost all (100% for one scene,
80% for the other) spontaneous answers contained rehtë; however, when asked if juuwë
was also possible, all speakers agreed that it was. The reverse was not necessarily true:
4
The figures represent, left to right and then top to bottom, papers on a spike, apple on a skewer, coat
on a hook, clothes on a line, pendant on a necklace, mud on a knife, stamp on an envelope. Here as elsewhere
the arrows indicate the figure object, whose location is to be described.
FIGURE 4. First and second choices for single scene (Yélı̂ Dnye).
juuwë scenes were not automatically compatible with rehtë. Again, this situation is
compatible with the idea that rehtë is ‘more informative’ than, that is, a hyponym of,
juuwë. In fact, rehtë seems to contain additional information about the ground; it must
be ‘hill-like’ or have a clear summit on its surface. We believe that patterns like these
found in Yélı̂ Dnye and in Tiriyó are symptomatic of hyponymy relations.
Although we believe that working out the full intensional relations between terms,
and distinguishing these from the pragmatics, is crucial for a proper semantic analysis,
we here have to abstract ourselves away from these details and operate on a coarser
level of generalization. We are concerned in particular with the EXTENSIONAL patterns,
making the rough and ready presumption that such extensional patterns are closely
related to intensional distinctions. For our current purposes, we believe this idealization
is good enough.
3. THE LANGUAGE SAMPLE. In typological studies, the importance of a large and
well-balanced set of languages has often been stressed (see e.g. Dryer 1989, 1992,
Croft 1990, Whaley 1997). Given the exploratory nature of this pilot study, however,
a large sample was considered impracticable: it would take years, probably decades,
to collect all the data with the methodology outlined in §2, if the language sample was
to meet the highest standards now current in the typology of morphosyntax. Semantic
typology is in its infancy, the methods and questions not yet well worked out, and
unlike morphosyntactic patterns which can often be gleaned from published sources,
semantic data are not available without specially designed fieldwork. Nevertheless,
even a small sample may be sufficient to disprove prevailing assumptions, and the
target of the present enterprise is the set of assumptions outlined in §1.
We therefore decided to look at a ‘convenience’ sample of nine languages, selected
according to (a) the accessibility of field researchers working on them and (b) their
genetic independence. Thus although we have data on a number of other languages,
we have included only a subsample of unrelated languages. Table 1 lists the languages
of the sample, their genetic affiliation and number of speakers, the researcher who
collected the data, and the number of consultants interviewed by the researchers.5
LANGUAGE AFFILIATION LOCATION DEMOGRAPHY CONSULTANTS RESEARCHER
Basque Isolate Europe 660,000 26 I. Ibarretxe
Dutch Indo-European Europe 20,000,000 10 D. Wilkins,
C. de Witte
Ewe Niger-Congo West Africa 3,000,000 5 F. Ameka
Lao Tai-Kadai Southeast Asia 3,000,000 3 N. Enfield
Lavukaleve Isolate Solomon Islands 1,150 1 A. Terrill
Tiriyó Cariban South America 2,000 10 S. Meira
Trumai Isolate South America 50 3 R. Guirardello
Yélı̂ Dnye Isolate Papua New Guinea 3,750 4 S. Levinson
Yukatek Mayan Mesoamerica 700,000 5 J. Bohnemeyer,
C. Stolz
TABLE 1. The language sample.
As can be seen, a first problem is the fact that, due to constraints on the various
field sites and on the projects of the individual researchers, it was not possible to obtain
the same number of consultants for every language. This means that the averages
calculated for every language are based on different numbers of consultants (see §5).
As expected, the locative adpositional systems of the languages in this sample show
considerable diversity both in their internal organization and in the range of supporting
spatial distinctions elsewhere in other form-classes. In some languages (e.g. Tiriyó),
adpositions are the only topological relation markers (TRMs); in others (e.g. Basque,
Trumai), spatial nouns are the most important element (with or without a locative case);
in yet others, positional verbs influence locative descriptions (e.g. Dutch, Ewe, Yélı̂
Dnye), sometimes to a large extent (e.g. Yukatek, or, even more overwhelmingly so,
Tzeltal; see Brown 1994). Table 2 gives a first overview of the situation.
The differences between the systems create certain obvious problems for crosslinguis-
tic comparisons:
5
A reviewer asked about the monolingualism or otherwise of our sample, wondering whether Indo-
European influences may be present through bilingualism. In fact, apart from our Dutch and Basque consult-
ants, capacity in an Indo-European language was fairly restricted in the sample (effectively nonexistent in
the Lao or Tiriyó consultants, very limited among the Yélı̂ Dnye consultants; all Yukatek consultants used
Yukatek as the home language and had limited Spanish, while two thirds of the Trumai consultants were
bilingual in Portuguese, and all Ewe speakers had at least some English; the Lavukaleve consultant had
Solomon Island Pijin as a second language, but the Indo-European character of this is questionable). Bilingual-
ism in other indigenous languages, however, is another matter. To the extent that bilingualism in an Indo-
European language did have an effect, it could be expected to lower the divergences from Indo-European
semantic patterns which represent one of the main findings of this paper.
SPATIAL LOCATIVE POSITIONAL/

LANGUAGE ADPOSITIONS NOMINALS CASE LOCATIVE VERBS NOTES
Tiriyó 100Ⳮ, 30Ⳮ 3–5 (less none none (single fine distinctions
spatial important) locative verb) (‘aquatic’;
‘astraddle’)
Yélı̂ Dnye 50Ⳮ, 25Ⳮ 3–5? (less none 3 (‘sit, stand, fine distinctions
spatial important) hang’) (‘attached by
spiking’)
Dutch 50Ⳮ, 15Ⳮ basic 3–10 (rel. none 4 (‘sit, stand, lie, fine distinctions
spatial important) hang’) (2 in’s, 2 on’s)
Lavukaleve ⬃12, 4 spatial 10Ⳮ one ‘general one ‘general ‘locative case’
(important) loc.’ case (or locative’ Ⳮ apparently
adpos.?) several optional reduced form of
positional verbs ‘generic
(clause chaining) adposition’
Basque 50Ⳮ, 15Ⳮ basic 50Ⳮ? one ‘general none (single adpositions
spatial? (important) loc.’ case locative verb) hard to
differentiate
from nouns
Ewe 7 prepositions (1 15Ⳮ? none? none one locative and postpositions
spatial); 15Ⳮ (maybe existential verb recent
postpostions postpostions ⳱ developments
(⬃10 spatial) nouns) from nouns
(e.g. often
homophonous
with body
parts)
Lao 2–10 (2 spatial; 5–10 none none nouns
maybe others) (important) apparently
grammaticalizing
into adpositions
(2 for sure;
others,
depending on
criteria)
Trumai ⬃5 adpositions 5–10 two cases one locative verb; spatial nouns
(none spatial) (important) (‘general’ vs. positional (body parts,
‘dispersed’ auxiliaries can be etc.) Ⳮ locative
locative) optionally added case mark
topological
relations
Yukatek 2–4 (1 generic many none many (a very rich a second ‘in’
spatial) (important) set, but not adposition
obligatory) seems to be
emerging
TABLE 2. The locative adpositional systems.
The formal unity of comparison. As can be seen, the subclass of locative adpositions
often overlaps both functionally and formally with the class of spatial nominals (‘top’,
‘bottom’, ‘side’, etc.). Such problems are present even in English, in which one could
arguably separate on top of as a complex locative adposition from on the top of, a
locative phrase headed by on. If locative cases also exist, they further complicate the
picture by introducing additional possibilities (for example, that nouns in the locative
case(s) may have developed into adpositions). Although the status of some expressions
is clear (Tiriyó adpositions are not hard to distinguish from nouns; see Meira 1999,
2003), most are not. In Basque, for instance, locative nouns (gaine-an ‘on (top of)’,
aurre-an ‘in front of’, and so on—the final -an is the locative case marker) usually
take their complement with the genitive marker -ren (e.g. eliza handia-ren aurre-an
‘in front of the big church’), but, under certain circumstances, caseless complements
are optionally possible (etxe aurrean ‘in front of the house’). Researchers are not agreed
on whether Basque has adpositions, or only locative nouns (cf. de Rijk 1999, Trask
1997); the same is apparently true for Ewe. Lavukaleve and Trumai are probably in a
similar situation, since they have many spatial nouns which may—or may not—have
become adpositions. There are at least some nouns in this situation in Lao, Yélı̂ Dnye,
and Yukatek. All this provides further justification for the decision not to make a clear-
cut separation between ‘true’ locative adpositions and spatial nominals in this study.6
A related problem is the presence, already mentioned, of LOCATIVE/POSITIONAL verbs
which also express significant information about the spatial (topological) relation be-
tween figure and ground. A study of the kinds of locative/positional systems that can
be found in the world is beyond the scope of this study. Ameka and Levinson (2003)
suggest that there are essentially two rather different systems of contrastive locative
verbs: one type has a small set of usually posture-derived verbs, the other a large set
of positional verbs making fine discriminations between figure-ground relations. The
small-set type is compatible with a rich set of adpositions (as in Dutch or Yélı̂ Dnye
in our sample), but the large-set type hardly ever occurs with a rich set of spatial
adpositions (compare Tzeltal with just one general adposition, or Likpe with one loca-
tive preposition). Such large-set systems are typical of, inter alia, the Mayan languages,
thus compensating for the relative absence of rich locative adpositional systems (see
Brown 1994, Levinson 1996). In our sample, we have just one language with the large-
set type available, namely Yukatek, but unlike in many other Mayan languages, a
general existential/locative verb now has completely general currency, motivating again
at least a handful of distinctions of an adpositional kind. Although these interactions
between adpositions and contrastive locative verbs are of considerable interest, we here
have to abstract away from them. Apart from Yukatek, all the languages in the sample
have either no contrastive locative verbs or only a small set thereof, the latter apparently
not inhibiting the development of rich spatial adposition systems. Thus we do not think
the overall picture will be distorted by ignoring the verbal systems for current purposes.
The semantic level of the comparison. We have illustrated the possibility that
adpositional systems can have substantial hierarchical structure. This raises the question
whether, for the purposes of semantic comparison, one ought to be comparing some
kind of basic-level oppositions rather than the full set of terminological distinctions.
We make crucial use below of an analogy to the basic color term work (Berlin & Kay
1969), so a reasonable question is: can one isolate basic adpositions, in such a way for
example, that we recognize in as a basic term in contrast to inside, which is both
formally marked and semantically a hyponym (green vis-à-vis chartreuse)? Unfortu-
nately, neither theory nor practice licenses this in the current state of knowledge about
the semantics of such systems. On the one hand, we can expect some opposition to the
6
Curiously, the adposition has little status as a form class in part-of-speech research, despite the considera-
ble work done on adpositional phrases in generative grammar (see Ayano 2001 for a recent review). There
is even fundamental disagreement over whether adpositions are functional or lexical categories (see Baker
2003:303–25 for discussion). The situation with spatial nominals is worse still: most languages have a
subclass of spatial nouns that behave in special ways, for example, serving directly as spatial adverbials (cf.
English home or north as in He went home/north), yet we know of no systematic study here.
treatment of adpositions as organized taxonomically, and on the other hand, the present
level of knowledge about the various systems found in the languages of our sample is
simply not deep enough for decisions on hierarchical status to be made in most cases. We
have therefore, as a first approach, treated any adpositional term displaying systematic
contrasts with other terms and occurring in the elicited data as a full-fledged member
of the system.
The semantic scope of the elements being compared. For several languages in our
sample (e.g. Lao, Yukatek), there were pictures in our elicitation tool for which re-
sponses did not include locative adpositions—they were scenes that were treated outside
the basic locative construction. Wilkins has shown that it is possible to scale spatial
scenes in such a way that there is a core of scenes (small unattached, manipulable
objects in canonical spatial relations) over which all languages will use their basic
locative constructions, and a periphery to which they may or may not extend them (see
Levinson & Wilkins 2003). Languages that avoid using a basic locative construction
for these peripheral scenes typically switch into a resultative or other descriptive mode.
This introduces another dimension of diversity, suggesting that languages perhaps differ
in what they consider a fundamentally spatial arrangement, or more specifically, in
how they extend the range or scope of the spatial topological domain. However, because
there is some underlying crosslinguistic systematicity here, we think it is legitimate to
proceed—all languages treated the great bulk of our scenarios within their basic locative
constructions. We should also point out possible extensions of topological spatial terms
in the other direction: some languages, but not all, extend the same (or derived) adposi-
tions to the motion domain, but others (like Yélı̂ Dnye) do not. In general, we approach
both horns of this dilemma (underextension of terms in some languages, overextension
with respect to our elicitation tool in others) heuristically: In order to have a starting
point, we assume that our elicitation tool is reasonably comprehensive and that the
elements being compared are reasonably systematic. We expect our results to be signifi-
cantly patterned; if it is the case that they are not, we should then revise our assumptions.
4. TESTING THE ORTHODOX ASSUMPTIONS AGAINST THE DATA. We concentrate first on
the prevailing orthodox assumption that languages basically agree on fundamental spa-
tial notions, so that in the topological domain notions like IN, ON, UNDER, AT, NEAR,
and so forth, are universal conceptual primitives that project directly into adpositional
meanings. We can formulate the presumption as follows:
Hypothesis 1: All languages agree on basic categories like IN, ON, UNDER, NEAR,
etc., in such a way that these notions form uniform, shared core-meanings for adposi-
tions across languages.
If this hypothesis holds in our small sample of half a dozen unrelated languages, it
would at least make the hypothesis plausible, although it could only be confirmed
on a much larger, truly representative sample. If, however, the hypothesis fails, the
universality of such notional content to adpositions is firmly ruled out. In order to test
the hypothesis, we make the assumption that similar intensions across languages will
share similar extensions. Nevertheless, we recognize that a category like IN might be
organized on prototype lines (Brugman 1981, Lindner 1981; see also Herskovits 1986:
36–41 on ‘ideal meanings’), so that a core concept of containment might be universal
while the boundaries might have variable extent across languages. Still, if that were
so, we would expect all languages to agree on grouping a core set of scenes together
as forming the heart of the category—say, the scenes that depict objects fully contained
within three-dimensional containers.
To test the hypothesis, it should suffice then to map all the languages’ adpositional
groupings onto a fixed arrangement of the scenes. This methodology has been pioneered
by Melissa Bowerman in a series of studies on adpositional and verbal coding of
topological spatial relations, where Venn diagrams are used to show how a language’s
spatial resources group scenes together (see Bowerman 1996, Bowerman & Choi 2001).
Thus, if seven of the seventy-one scenes in the TRPS elicitation booklet elicit the same
adposition from the majority of consultants for language L, these seven scenes can be
taken to represent the extensional category for that adposition; we can then go on and
see how speakers of another language apply their adpositions, and whether the same
scenes are grouped together or not. To make this visually inspectable, we lay the scenes
out on the plane surface in a fixed arrangement and map language after language onto
the same arrangement. We here use the best arrangement we have been able to find,
returning later to ask how that should be determined.
Figure 5 displays a mapping of the adpositional categories of one language, Tiriyó,
onto the fixed arrangement of scenes. Each picture is a stimulus that elicited a verbal
response, with the figure in the scene indicated by a small arrow. Lines of a single
color encircle those scenes grouped together by use of a shared adposition. Since ten
consultants’ answers have here been taken into account, the picture represents the usage
shared by the majority of consultants.
Inspection will show that Tiriyó uses sixteen distinct adpositions to cover this range
of scenes. Some of these adpositions are distinctly un-English, like the Tiriyó aquatic
adposition hkao ‘(be) in-water’ which groups scenes 32 and 11. Note too that the biggest
area is covered by an adposition pë(kë), which has no English equivalent but which has
the semantics of ‘attached to’. That attachment rather than, for example, nonhorizontal
support/contact is the decisive feature can be seen from the fact that several speakers
agreed that even normal horizontal contact can be described with pë(kë), as long as
the figure somehow ‘adheres’ to the ground, that is, offers some resistance to removal
(for instance, a piece of used chewing gum on a table). Adpositions specialized to
attachment scenarios turn out to be common enough crosslinguistically—this is an area
mostly covered by English on or, to a lesser extent, in. Bowerman and Pederson (1992;
see also Bowerman & Choi 2001:484–87) have argued that languages with adpositions
glossing ‘in’ and ‘on’ often invade the attachment area in an orderly manner, allowing
a scale of such scenes to be set up between prototypical containment at one extreme
and prototypical superadjacency (‘on’) at the other, a point to which we return.
Another exotic category in the Tiriyó map is awëe, glossed ‘astraddle’, which has
as its core meaning suspension involving a figure supported by a point such that the
figure hangs down on either side of the point. Tiriyó is not the only language with
exotic adpositional concepts—compare the groupings in Figure 6, contrasting Tiriyó
and Yélı̂ Dnye adpositions in the attachment/suspension area.
Still, the existence of exotic, language-specific spatial concepts like the Tiriyó aquatic
adposition (in fact shared across Cariban languages) is not perhaps the central issue at
stake. It could still be that all languages respect a central IN and ON area, for example,
while having in addition more language-specific categories. To test this, we map three
more languages in Figure 7 (about the maximum allowing graphical resolution) onto
the same arrangement used in Fig. 5. In the figure, different line types (dotted, dashed,
etc.) code the four languages, while the line colors code the distinct adpositions within
a language.
What is immediately evident from this superposition of adpositional categories is
that there is no crosslinguistic agreement on large IN, ON, or other categories even
‘NATURAL CONCEPTS’ IN THE SPATIAL TOPOLOGICAL DOMAIN
497
FIGURE 5. Tiriyó adpositions mapped onto fixed array of pictures.
FIGURE 6. Language-specific adpositional categories.
among just these four languages. There is not a single pairing of the scenes in the IN
and ON areas to the left of the diagram that each of the four languages agrees on. This
is—in the light of the orthodox assumptions sketched at the beginning—a very surpris-
ing finding. It is sufficient to refute hypothesis 1, for we see here no evidence at all
for prototype categories in these areas. The only grouping of three scenes that is agreed
upon across all four languages is an UNDER category at the bottom of the diagram,
and any such grouping may not survive when a larger number of languages is compared.
If hypothesis 1 is false, what other universal generalizations may be sustainable? We
have entertained the following hypothesis (Levinson & Wilkins 2003), which is perhaps
the next most restrictive position:
Hypothesis 2: Languages may disagree on the ‘cuts’ through this semantic space, but
agree on the underlying organization of the space—that is, the conceptual space formed
by topological notions is coherent, such that certain notions will have fixed neighbor-
hood relations.
To make this vivid, consider the findings from the comparative work on basic color
terms. It has been shown that the color space is treated by linguistic discriminations
as coherently organized, as a three-dimensional solid with a plane surface with the two
dimensions hue and intensity. No language has been found that, for example, collapses
yellow and purple in one category, or even green and blue unless it also encompasses
all intermediate hues. Further, there seem to be just six naturally salient foci, such that
languages will build their categories around one or more of these foci. Thus the domain
has an internal arrangement, independent of the categories and their boundaries, which
can be very variable. For example, a language with just three terms will have a so-
called composite category, which we might gloss ‘dark/cool’, that encompasses black,
blue, dark green, but these are all neighboring areas. Languages with more terms will
systematically fractionate out the three colors around the three subsumed foci (Kay &
McDaniel 1978).
If the topological domain has a similar internal coherence, it should be possible to
find a single fixed arrangement of the pictures such that those that are grouped together
in one language remain contiguous even if they are separated by a category boundary
in another language. We set out to test whether we can find such an arrangement,
starting, for simplicity, in two dimensions. Unfortunately, this is not a project that lends
itself to a computational solution, because with seventy-one pictures the number of
combinations on a plane surface is vast (71 factorial), beyond computation on a reasona-
ble time-scale. Instead, we have to proceed by hypothesis, assuming that, for example,
‘on’ scenes should be together and separated from ‘in’ scenes, and so forth. Using a
computer program that tested for neighborhood relations, we have found that the fixed
arrangement as already presented in Figs. 5 and 7 is the best fit we can find to the
adpositional groupings in our sample. This arrangement groups scenes in a number of
coherent notional categories, as outlined in the following diagram (Figure 8).
This arrangement, however, does not meet the criteria necessary to support hypothesis
2, for there are some language-discontinuous categories that are then dispersed across
the space, as illustrated in Figure 9, where two categories from the language Lavukaleve
fail to map contiguously, as does one from Yukatek. If we try to rearrange the pictures
to bring these categories together so that they map onto a single, contiguous area of
pictures, then the other language categories already displayed contiguously in Fig. 7
will now themselves be discontinuous in part.
From this endeavor we must conclude that it is not easy to find a fixed array of
scenarios that will yield contiguous categories in every language—we have not shown
that it is impossible (given the computational intractability), only that it is not obvious
that there is such an arrangement. If we had a much larger sample of languages, finding
such an arrangement would become even more difficult. But there are two important
caveats. First, we are looking for a fixed arrangement in only two dimensions—in three
dimensions many more possible arrangements exist, of course.7 The color work in fact
presumes a three-dimensional color solid, and has then found uniformity on one surface
of this solid (the surface of maximum saturation). Second, as compared with the color
work, both the semantic relations and the physical scenarios they describe are obviously
much more complex and offer different possible construals. For example, for a goldfish
in water in a bowl, what is the ground, water or bowl? Some of our Tiriyó consultants
went one way, using the aquatic adposition, while others went the other using an IN
adposition (had they all chosen the aquatic adposition, we would have had another
discontinuity).
There is further ground for optimism that hypothesis 2 may have some foundation.
Bowerman and Pederson (1992, 2003) have looked at all nonpredicative topological
relation markers—adpositions, cases, and spatial nominals—in some forty languages,
concentrating specifically on the extensions of terms coding the ON and IN areas. As
7
To say nothing of hyperspaces of higher dimensionality. In fact, in a higher-dimensional space, it would
be possible to construct an arrangement in which every one of the 71 pictures would be adjacent to all the
others. In such a configuration, any subset of pictures, even one composed randomly, would map without
discontinuities. This, however, is a trivial, uninteresting solution, since it would not distinguish true linguistic
categories from random ones. An interesting solution would be one in which not all pictures are adjacent,
and yet all adpositional categories from all languages would map continuously, as in the case of the color
solid. Only then could some insight be gained into how human cognition structures this particular domain.
FIGURE 7. Four languages’ adpositions mapped onto fixed array of pictures.

FIGURE 7. Continued.
FIGURE 8. Notional areas in fixed array.
mentioned, they in fact developed for that study the same stimulus set we have used
in this research. They have been able to establish a surprising fact: the relevant scenes
scale on a cline between a prototype IN (full containment) and prototype ON (superposi-
tion plus support); that is to say, if a language has a broad extension for, say, an ON
adposition, while another has a narrower one, it is pretty much predictable which scenes
will be included. Any language, for example, that codes ‘encirclement with contact’
(as by ring on finger) with the term used for ON, will also use the same term for ‘hang
with planar contact’ (as with picture on wall) or ‘sticky attachment’—in that way,
given the maximal extension of a term on the scale, all the other usages up to that
point on the scale are predictable. Using different methods from ours, namely Guttmann
scaling, they have shown that at least on one cut through our topological space (namely
the IN to ON dimension), there is a coherent structure to the space.
Our overall conclusion is that hypothesis 2 cannot be ruled out at this stage (and
indeed in §6 we further proceed on the assumption that there is indeed a coherence in
the domain, or at least in parts of it, even though this has yet to be established for the
whole space). Nevertheless, for the sake of argument here, let us admit that it is certainly
not obviously true. We now fall back on a still weaker hypothesis of a statistical kind:
Hypothesis 3: The domain of topological relations constitutes a coherent semantic
space with a number of strong ATTRACTORS, that is, categories that languages will
statistically tend to recognize even if some choose to ignore them.
To test this hypothesis, we proceed to analyze the data in a different way using multidi-
mensional scaling, as described in §5.
FIGURE 9. Discontinuous categories mapped onto fixed array.
5. MULTIDIMENSIONAL SCALING APPROACH. Hypothesis 3 is a statistical hypothesis,

requiring statistical testing. We are interested less in normal tests of statistical signifi-
cance than in a broad heuristic way of seeing whatever main trends there are in the
data. Since we are interested in how languages group the scenarios, we need a form
of cluster analysis, and we adopt here a multidimensional scaling technique.
The procedure was as follows. First, each language was treated on its own. An
average of the consultants’ responses was calculated: for the languages with many
consultants (Tiriyó, Basque, Dutch), a picture was ascribed to a certain adposition when
more than 50% of the consultants used it; for languages with four or five consultants,
a picture was ascribed to a certain adposition if at least two of them used it; for the
languages with three or fewer consultants, a picture was ascribed to a certain adposition
if any of the consultants used it. Notice that a picture can be ascribed to more than one
adposition (since consultants often gave multiple answers, there can in principle be
50% agreement on several adpositions for the same picture). After that, a list of adposi-
tions with the pictures ascribed to them was constructed (e.g. ADP-1: pictures a, b, c
. . . ; ADP-2: pictures x, y, z . . . ) for every language.
Since there were 71 pictures, a 71 ⳯ 71 matrix was constructed to represent the
relation—similar or different—between the linguistic treatment of each pair of pictures.
For each pair (in ith column and jth row), we need a measure of similarity or, as is
conventional in these treatments, of dissimilarity. We considered that pictures which
were ascribed to the same adposition are more similar to each other than pictures which
were ascribed to different adpositions. Based on this, the dissimilarity values (ranging
from 1, totally dissimilar, to 0, totally similar) for each cell of the matrix were calculated
as follows:
(a) a picture was considered perfectly similar to itself (i.e. 0 dissimilarity); therefore,
the diagonal of the dissimilarity matrix consists only of zeroes;
(b) for any other pair of pictures pi and pj , the SIMILARITY between them is roughly
proportional to the number of adpositions that were used to describe both of them (i.e.
the number of adpositional sets in which both pi and pj occur). Thus, if both in and inside
were used for pi and pj , they should be considered more similar (lower dissimilarity) than
if only in, but not inside, could be used. To make this a measure of dissimilarity, we
subtracted the number of adpositions that treat pi and pj alike from the total number
of adpositions in a given language. Thus, in a language with ten adpositions, if two
(e.g. in and inside) occurred with both pi and pj , the dissimilarity would be 10 ⳮ 2
⳱ 8, that is, the number of adpositions which did NOT treat them alike. In order to
keep dissimilarity values between 0 and 1, we divided the result by the total number
of adpositions (8/10 ⳱ 0.8). The procedure is summarized in the following formula.
(total adpositions) ⳮ (adpositions that treat pi and pj alike)
D⳱
(total adpositions)
For example, Tiriyó consultants used both the postpositions awë and tao for the two
pictures p2 and p54 ; since there were sixteen adpositions used altogether by consultants
for the stimuli set, the dissimilarity value for these two pictures is:
16 ⳮ 2
D⳱ ⳱ 0.87
16
Incidentally, this formula considers pictures maximally dissimilar unless they occur
together in the extension set of at least one adposition. Thus, for a given language,
pictures that did not elicit adpositional responses (that is, they were described with
constructions other than simple locatives) are considered maximally dissimilar to any
picture which had adpositional responses, and also to each other.
The resulting coefficients were calculated for every pair of pictures and tabulated in
the dissimilarity matrix for each language. Then the set of dissimilarity matrices for
all the languages were simply summed (each cell was the sum of the corresponding
cells in each language matrix), producing a composite matrix, which represented for
all the languages the extent to which any pair of pictures was or was not treated as
representing the same kind of spatial relation. For instance, p2 and p54 had dissimilarity
value of 0.87 in Tiriyó, 0.93 in Dutch, 0.88 in Basque, and so on; the final dissimilarity
was 0.87 Ⳮ 0.93 Ⳮ 0.88 Ⳮ . . . ⳱ 8.15. The composite dissimilarity matrix was then
input into the multidimensional scaling algorithm ALSCAL (SPSS v. 7.5), and a plot
was obtained using an Euclidean model (keeping ties untied). This plot for all the
selected languages is shown in Figure 10. As mentioned in §3, we have restricted the
plot to all the languages we have in the sample that are unrelated, to avoid a statistical
biasing towards, for example, Indo-European patterns.
It is immediately evident from the plot that pictures do tend to cluster, and that they
are not randomly distributed across the surface of the plot, as they would be if there
were no crosslinguistic generalizations at all. Of course, a crucial consideration is
whether this particular pattern is an artifact of the particular languages we happened
to have selected. That is a question we cannot answer definitively—we can say only
that the patterns now showing seem quite stable when further languages are added,8
8
In a larger sample, we included different languages from some of the same language families (e.g. Indo-
European, Mayan), but have excluded them here in order not to bias the sample.
FIGURE 10. ALSCAL plot for Tiriyó, Yélı̂ Dnye, Ewe, Lavukaleve, Trumai, Yukatek, Lao, Dutch, and Basque.
and that these patterns are at least suggestive of hypotheses about universal tendencies,
remarks we develop in the next section.
Some quite interesting observations emerge from the multidimensional scaling plot.
The annotations show the immediate generalizations under conventional labels (not to
be taken too literally): there are some dense clusters of scenes, showing that the majority
of languages treat these scenes as related by their spatial adpositions.
(1) In the middle, there is a large, relatively loose, cluster of ‘attachment’ scenes,
reinforcing the observation made above that many languages find ‘attachment’
a central topological notion, although it is not a concept predicted by the orthodox
assumptions, based as they are on European languages.9 Figure 11 shows a
blowup of this area of the chart near the center of the plot in Fig. 10. We have
superimposed thumbnails of the pictures, illustrating the relevant topological
relations, for example, letters on T-shirt (p68 ), papers on spike (p22 ), ring in ear
(p69 ), apples on tree (p45 ).
(2) To the right of the plot in Fig. 10 there is a tight cluster of scenes that collapse
together all kinds of superposition, with or without contact—hence the ON and
OVER relations presumed by the orthodox assumptions are collapsed rather than
9
A reviewer questions to what extent ‘attachment’ (and indeed other notions like ‘containment’ and
‘support’) are really spatial as opposed to mechanical in conception. Such doubts are certainly in order. Note
that we began from the orthodox assumptions, wherein spatial topological concepts are universally given,
forming the mold for crosslinguistic categories, and have found that position untenable. In this section we
are trying to extract what tendencies are in fact empirically found in the semantics of adpositions used in
answers to where questions, and ‘attachment’ (which has both topological spatial and mechanical conceptual
elements) emerges as a significant generalization. That neo-Kantian categories do not exactly match the
empirical generalizations should be part of the interest of the exercise, although perhaps no more surprising
than the finding that grammatical tense distinctions cannot always be captured purely in terms of Newtonian
time concepts (see e.g. Comrie 1985:43–48).
FIGURE 11. Blowup of ATTACHMENT area.
FIGURE 12. Blowup of area including ON/OVER cluster.

distinguished.10 Figure 12 shows a blowup of the ON/OVER cluster found to

the far right of the plot in Fig. 10. What cluster together are relatively small,
moveable objects, mostly inanimate (certainly nonhuman), which are in a super-
adjacent relation—with or without contact—to a relatively immovable ground
which has its base at ground level. It is immediately clear that all sorts of scenes
that in English would seem like good on relations, like the scenes in our stimuli
representing a book on a shelf, or a tablecloth on a table, or a tree on a mountain,
or a man on a roof, are not part of the cluster. They are in fact rather distant on
the plot, forming another cluster near the middle of Fig. 10, which we can label
the ON-TOP cluster, since they mostly involve figure objects on elevated ground
objects, or figures covering the ground object—see Figure 13 for a graphical
depiction of the relation between the ON/OVER cluster and the ON-TOP cluster.
FIGURE 13. Graphical representation of larger area including ON/OVER cluster and ON-TOP cluster.
(3) We look now for the presumed universal ‘IN’ relations. These do indeed form
a recognizable cluster, as depicted in Figure 14. Note the subcluster in the top
left corner of the figure, where relatively small, moveable objects are more or
less wholly contained within the ground object. What is notably missing both
from this subcluster and the wider cluster in Fig. 14 are the cases of two-dimen-
sional enclosure in a plane surface. For example, the pictures depicting an apple
inside a ring (p19 ) or a house inside a fence (p60 ) are very distant on the plot,
as is partial enclosure in a three-dimensional solid, as with an arrow through an
apple (p30 ) (these are all to be found distributed around the far top-left edges of
the whole plot in Fig. 10). The evidence then is that there is a strong crosslinguis-
tic tendency to have dedicated TRMs to encode the notion of containment within
artifacts made as containers (the cultural rather than nativist nature of this general-
ization is made clear by the exceptions, for example Australian languages, men-
tioned immediately below and in the conclusion; cf. Vandeloise 1986).
10
Although only p36 (cloud over mountain) belongs closely to this cluster, if we subtract the effect of
Basque, p13 (lamp over table) also migrates to this cluster. We feel the ON/OVER conflation is a firm
crosslinguistic tendency, reflected directly in languages like Japanese and Arrernte, not in our sample here.
FIGURE 14. Blowup of area containing IN cluster.
(4) Among other clusters that are noteworthy in Fig. 10, despite playing no large
role in the orthodox assumptions, are UNDER relations which, we already noted,
showed through as the sole cluster surviving our four-language map. In our
plot they form a cluster with NEAR relations, from which they are not clearly
separated—see Figure 15. If Basque is subtracted from our nine-language plot,
this cluster of UNDER/NEAR relations merges with the IN cluster, which is
interesting given conflations in Australian languages between IN and UNDER.
Summing up so far: we have found that crosslinguistically certain extensional classes
tend to be shared—inspection suggests these cluster around the notions of attachment,
superadjacency, full containment, subadjacency, and proximity. Note that the ‘confla-
tion’ of ON/OVER suggests that ON simpliciter is not a primitive (as on the orthodox
view) but is composed of superposition plus or minus contact—the alternative view that
it represents a true conflation of primitives we return to below. The general treatment of
certain scenes as ISOLATES, separated from any tight cluster, is also interesting. These
are scenes with negative figures (cracks, holes), part-whole relations (straps on bag,
handle on door), and other scenes which we have found to often evoke a different
construction than the basic locative construction (see Levinson & Wilkins 2003).
These findings suggest that hypothesis 3, if not positively confirmed (for this would
require a much larger sample), is at least compatible with the data in hand. We now
go on to ask, assuming that the pattern we have is stable when more languages are
processed, what further typological generalizations are possible.
6. EMERGING GENERALIZATIONS ABOUT THE SEMANTIC TYPOLOGY OF ADPOSITIONS. Se-
mantic typology is a nascent field, where serious collective work has hardly begun (the
FIGURE 15. Blowup of NEAR/UNDER cluster.
color work stands out as an exception), although very interesting observations have
been made by individual scholars such as Talmy (2000). The basic principles of this
field have yet to be established. Clearly it has kinship with the morphosyntactic typology
started by Greenberg (1966), which now passes under the simple rubric of linguistic
typology. The difference in slogan form between the two approaches is that while in
Greenbergian typology meaning is used only to get at the forms, in semantic typology
the forms are used only to get at the meaning. There is one area where both kinds of
typology overlap—namely in the importance of markedness, which has both a formal
and semantic dimension. Semantic markedness involves privative oppositions, already
discussed above under the rubric of taxonomic structure and semantic versus pragmatic
content. In many other respects, though, we can expect the principles of the two fields
to diverge. Adpositional meanings illustrate this nicely.
We have entertained the following kind of Greenbergian implicational scale over
adpositional notions (for a related scale supported by studies of crosslinguistic acquisi-
tion, see Johnston & Slobin 1979).11
The prediction here is that any language that, for example, has an INSIDE adposition
also has an ON-TOP and an OVER, as well as an IN. Some adjustment needs to be
made for English, which has no ATTACHED adposition (see below); otherwise, our
sample supports such a generalization. The evidence comes not just from the overall
inventories of our languages, but also from language-internal evidence of adpositions
11
Johnston and Slobin (1979) predicted the following scale (order of acquisition) on the grounds of
conceptual difficulty and salience in acquisition: IN/ON/UNDER ⬍ BESIDE ⬍ BACKintrinsic ⬍ FRONTintrin-
sic ⬍ BETWEEN ⬍ BACKprojective ⬍ FRONTprojective . Acquisition in four languages (Turkish, Italian, English,
Serbo-Croatian) substantiated that the IN/ON/UNDER cluster of adpositions are learned first, and the projec-
tive ones last, but the middle adpositional concepts were learned in variable order.
FIGURE 16. Implicational scale over adpositional notions.
having more versus less grammatical status. The IN adposition of Yukatek, for example,
is clearly more grammatical in nature (more prepositional, less nominal) than the ON
and UNDER ones, and the INSIDE in Yélı̂ Dnye is clearly less grammatical (more
nominal, still carrying a possessive) than the IN and ATTACHED adpositions.
Still, from a semantic point of view this generalization is unsatisfactory. For example,
a language with only AT, IN, and ON has a quite different concept for ON than one
that has both an ON-TOP and an OVER—the former concept of ON effectively encom-
passes OVER and ON-TOP, which are differentiated in the latter. In fact, this is a
typical semantic correlate of Greenbergian implicational scales. Compare the well-
established scale (Croft 1990:66):
SINGULAR ⬍ PLURAL ⬍ DUAL ⬍ TRIAL/PAUCAL
implying that any language with a dual has a plural. But of course a language with
only a singular and plural has a different SENSE of plural than a language with a dual—in
the former plural means ‘two or more’, while in the latter it means ‘three or more’.
Typological hierarchies of this sort are thus not meaning-preserving.
From the point of view of semantic typology we need another model that is explicit
about these semantic changes. Here we turn to the only well-developed model of seman-
tic typology, the work on color terminologies. The original theory in Berlin & Kay
1969 is the textbook version, now long superseded. In the original theory, a selection
of terms, each focused around a single salient ‘best color’, gave us an ‘evolutionary
sequence’, as in Figure 17.
FIGURE 17. Berlin and Kay ‘evolutionary sequence’ of basic color terms.
The foci of the basic color terms were prototypes, and the boundaries of color categor-
ies ill defined. Subsequently, this theory was replaced by the new COMPOSITE CATEGORY
theory, developed by Kay and McDaniel (1978). Building on Heider 1972, they intro-
duced a different idea about the foci of basic color terms: early color terms in this
sequence are COMPOSITE TERMS with more than one focus (see Levinson 2000c for a
review). Thus the Dani two-term system should be seen as a warm/cool system rather
than a black/white system, since the warm term may be focused in white, yellow, or
red, and the cool term in black, green, or blue. A three-term system fractionates out
the white, leaving a warm with two foci (red or yellow), and the cool with three foci
as before. These composite terms are always formed from adjacent foci. The next stages
break these composite terms down further until each of the first six categories has just
the one focus, black, white, red, yellow, green, or blue. Thereafter new color terms
(such as brown) are formed as intersections, or perceptual blends, between the six core
terms. There are thus three distinct kinds of color terms: composite categories, six
primary categories, and derived blends. Together they exhaust the color space in any
one language. This is now the standard theory (which we will refer to as BCT theory),
although it undergoes interesting and important revisions to this day (see Kay & Maffi
1999). Levinson (2000c) introduced a further wrinkle: not all languages treat the domain
as one that must be exhaustively covered by color terms—some permit large gaps in
coverage (a feature relevant for the topological domain). This model may offer an
important analogy for the adpositional domain. As we have seen, the standard typologi-
cal implicational scale equivocates on meanings, just as the original Berlin and Kay
theory did (an early WHITE term in fact has a huge extension beyond white, and encom-
passes other foci too). The new composite-category theory precisely allows for meaning
change of terms as others are added. So, a language with just one adposition in the
topological domain (like Tzeltal) is a one-term composite category, covering all the
potential foci in this domain. When Yukatek innovates more terms, it first (to judge from
grammaticalization evidence) introduces an IN term focused on container-inclusion but
including plane-inclusion, then a general superposition term (covering OVER and ON)
and an UNDER. It retains the general term with full coverage, as in Tzeltal, thus
introducing hierarchy or privative opposition into the system (in basic color term theory,
such hierarchical relations are excluded from consideration, but for us they are impor-
tant). Like the color systems that do not exhaust the color domain, Yukatek also excludes
about a third of the stimulus scenarios, finding other nonspatial means to describe them.
Yukatek thus has, like early color systems, a number of composite categories and one
primary category with just the one focus (UNDER).
In this sort of way, we think the color model can be usefully applied to the topological
domain. There remain though a number of difficulties. One problem (that should proba-
bly be set aside pending further research), is that it is not obvious that the domain is
a single, coherent, fixed space like the psychophysical color solid. This lack of strict
coherence may be the case, and yet the model may still be good enough to offer a
useful approximation. A second obvious problem is exactly what would constitute the
foci in this domain. When talking about the clusters in our multidimensional scaling
plot, we talked of attractors in the conceptual space of spatial topology. But we cannot
directly equate the clusters we have obtained with the underlying foci or attractors—for
these are the category tendencies, and such categories we are now supposing may be
composite in nature. Let us take the ON/OVER cluster as an example. We can conceive
the corresponding intension in two different ways:
(a) as an underspecified meaning, say [ⳭSuperposition, ⳲContact], general over
presence or absence of contact between figure and ground; or
(b) as a composite category with two foci, [ⳭSuperposition, ⳮContact], [ⳭSuper-
position, ⳭContact].
Position (a) presumes categories without prototypes, but with necessary conditions and
thus relatively sharp boundaries, while position (b) presumes prototype categories with
elastic, ill-defined boundaries. Note that on either analysis the foci are not semantically
primitive, but analyzable into constitutive concepts.
If we follow the model of BCT theory, we would take position (b) (some evidence
for elastic boundaries of spatial categories around prototypes can be found in the psycho-
linguistic literature; see for example Hayward & Tarr 1995, Carslon-Radvansky &
Irwin 1993). This allows categories to have a disjunctive character, just as languages
that lack a ‘blue’ and ‘green’ term often have a ‘grue’ focused in both blue and green.
As an example, consider the category that subsumes OVER and ON as in Yukatek or
Ewe. In the same way, English in would have two foci distinct in topological space,
one specifying containment in a three-dimensional container, the other inclusion in a
notional two-dimensional plane. English on is a larger composite category excluding
OVER but including ATTACHMENT, location above eye-level (ON-TOP), and yet
other areas. These large categories will tend, on this theory, to be split into primary
(single-focus) categories over time under particular functional pressures. We sketch
the sort of sequence we imagine in Figure 18, which specifies the possible routes of
development implied by the first four stages in the implicational scale introduced above.
FIGURE 18. Successive fractionation of composite concepts in the topological domain. (Bifurcating arrows
indicate alternate routes; braces indicate categories that resolve together; AT is a residual category with
successive reductions marked by subscript.)
In the figure, we have to allow for distinct routes through developmental space, just
as in BCT theory—these are indicated by bifurcating arrows. The unique term at the
left is a general locative adposition denoted AT. This is a residual category, which is
successively reduced (compare the COOL term in BCT theory), but perhaps has spatial
coincidence as a primary focus (like English at). The capital letter meta-terms in this
diagram must now be taken to indicate specific prototypes; for example, ON is superpo-
sition plus contact, and IN is full containment in three-dimensional container under
convex closure (see Herskovits 1986). English on is represented by the box to the right
including ON, ON-TOP, and ATTACHMENT—it is a composite category. The Tiriyó
counterpart has extracted out both ATTACHMENT and ON-TOP as separate categor-
ies, leaving ON as a primary category with single focus, and so forth.
Finally, we should ask how we should conceive of the language-specific categories
like the Tiriyó aquatic or the Yélı̂ Dnye spiked adpositions within such a framework.
Note that in BCT theory there are blends between neighboring focal categories (fuzzy-
set intersections), giving us colors like ‘brown’, ‘pink’, and ‘orange’. But these are
interstitial categories, while the Tiriyó aquatic crosscuts a number of categories, and
the Yélı̂ spiked is a clear subcase of attachment, and thus are not interstitial in the same
way. One can perhaps find such interstitial categories in the topological domain—En-
glish against may constitute a case, being arguably interstitial between the large on
category and a near category. We think we simply have to recognize that such language-
specific categories just go beyond the theory, in the same way that an English nonbasic
color term like crimson or vermillion goes beyond basic color term theory.
An account of this kind—if it can be made to stick—finds universal structure in
diversity: first, there are universal prototypes; second, there are universal constraints
on category formation, requiring only neighboring prototypes to coalesce into composite
categories; third, there are constraints on synchronic sets of categories, as represented
by the routes through the developmental sequence.
How seriously should we take the diachronic implications of the model? Such an
‘evolutionary’ perspective is partly fictional of course (given the many languages in
our sample without a written history), as it is in the color domain, but the added temporal
dimension allows us to conceive of different unfolding patterns, each with its own
implicational relations. It also helps to account for the hierarchical patterns we earlier
adduced for Tiriyó, for example, where a more specific adposition can develop, leading
to a privative relation with the more general superordinate category which it entails.
Whether the distinct routes through the developmental sequence can be substantiated
in languages with extensive historical records is an empirical matter, and an advantage
of the model is that it at least suggests such hypotheses.
7. CONCLUSIONS. This article should be taken as a pilot study in a complex domain.

But we believe that we have already established some basic foundations for future
research. First, the orthodox assumptions mentioned at the outset do not stand up to
close scrutiny if taken in any literal way. Generalizations about universal patterns must
take into account that we are dealing with much more diversity than the orthodox view
suggests. Although this diversity could be adduced simply to disprove the orthodox
assumptions, we believe that the diversity is itself organized in ways that may allow
us to discern universal tendencies. We believe that a BCT-type model, which brings
the extra temporal dimension into the picture, allows us to see a pattern where otherwise
it would be obscure. Such a model may prove tenable when a much larger sample of
languages is processed, and it is not a set of simple generalizations. Nevertheless, the
idea of a successive fractionation of categories formed over primary foci in a fixed
topological conceptual space may prove a very helpful heuristic for understanding the
patterns in the diversity.
In addition, we hope that the methods employed here will raise the standards of work
in semantic typology. The use of fixed stimuli across languages and the different ways
of processing the results will, we hope, inspire further efforts in this and related domains.
Among the more important substantial findings from this project has been the crosslin-
guistic importance of the ATTACHMENT category. In the earlier work by Bowerman
and Pederson (1992; see also Bowerman & Choi 2001) the ATTACHMENT area was
seen as a category invaded by the encroachment of large IN and ON categories. Here
we reverse the implicit developmental sequence and suggest that the ATTACHMENT
area starts out conflated with ON/OVER/ON-TOP notions, and on one developmental
sequence is fractionated out first (see bottom sequence in Fig. 18). In any case, AT-
TACHMENT has at least one clear focus of its own and is an important category that
tends to be recognized in language after language.12
Finally, we should attend to the conceptual status of the foci—where do they come
from? Are they universal foci, innate natural categories given by our biological endow-
ment? Not necessarily, of course. We think they should be seen in a functional perspec-
tive, given universals and tendencies in human organization of the environment.
Consider, for example, the IN category, or more exactly the IN-CONTAINER focus:
nearly all contemporary cultures have large sets of containers for different purposes.
The ethnographic record makes clear, however, that hunter-gatherers like the Australian
Aboriginals had little traditional use for containers, using for the most part only flattish
trays or coolabahs. Interestingly, across Australian languages there is a conflation of
IN/UNDER notions in a single spatial nominal (Australian languages generally lack
adpositions, using case and spatial nominals instead).13 In our plot in Fig. 10, the IN
cluster and UNDER/NEAR clusters are neighbors; in fact, as mentioned above, if we
take out Basque, the two clusters coalesce. The Australian facts suggest that cultural
factors play a role in making certain categories worth isolating—whether they actually
create those categories and their foci or merely resolve potential, incipient categories
is up for grabs. Similarly, for the ON relation, cultures that elevate working surfaces
and storage off the ground level clearly have a special interest in distinguishing such
relations from, for example, OVER ones. That UNDER is more clearly universal as a
simplex category than either fully resolved ON or IN relations perhaps speaks to univer-
sal features of the natural environment as opposed to features of the human, built,
environment. Further, in addition to these cultural pressures for the distinction between
special spatial relations, the shared nature of our human stance and preoccupations in
a terrestrial environment with its uniform gravitational field offer additional functional
sources for universal tendencies (see H. Clark 1973). The ATTACHMENT focus is,
in this respect, somewhat ambiguous: it may depend on cultural factors (like the pres-
ence and importance for a given culture of means of attachment, such as ropes, spikes,
and so on), but also on natural factors (e.g. part-whole relationships, from which adhe-
sion might develop: something that adheres or is attached to a ground is not unlike a
part of the ground, since it would resist removal like all other parts). We may need
little recourse to innate ideas or nativist categories, which is just as well, since the
neurocognitive or genetic evidence for any extensive body of such beasts is equivocal
at best (Elman et al. 1996).
REFERENCES
AMEKA, FELIX, and STEPHEN C. LEVINSON. 2003. Positional and postural verbs. Special issue
of Linguistics, to appear.
AYANO, SEIKI. 2001. The layered internal structure and the external syntax of PP. Durham:
University of Durham dissertation.
BAKER, MARK C. 2003. Lexical categories: Verbs, nouns, and adjectives. Cambridge: Cam-
bridge University Press.
BERLIN, BRENT, and PAUL KAY. 1969. Basic color terms: Their universality and evolution.
Berkeley: University of California Press.
BOWERMAN, MELISSA. 1996. Learning how to structure space for language: A cross-linguistic
perspective. Language and space, ed. by Paul Bloom, Mary Peterson, Lyn Nadel, and
Merrill Garrett, 385–436. Cambridge, MA: MIT Press.
12
Some studies suggest that ATTACHMENT may make itself felt even in European languages (see Nüse
1999 on some of the uses of German auf and an).
13
We are indebted to forthcoming work by David Wilkins on this issue.
BOWERMAN, MELISSA. 2003. Space under construction: Language-specific spatial categoriza-

tion in first language acquisition. Language in mind, ed. by Dedre Gentner and Susan
Goldin-Meadow, 387–427. Cambridge, MA: MIT Press.
BOWERMAN, MELISSA, and SOONJA CHOI. 2001. Shaping meanings for language: Universal
and language-specific in the acquisition of spatial semantic categories. Language acqui-
sition and conceptual development, ed. by Melissa Bowerman and Stephen C. Levinson,
475–511. Cambridge: Cambridge University Press.
BOWERMAN, MELISSA, and ERIC PEDERSON. 1992. Cross-linguistic studies of spatial semantic
organization. Annual Report of the Max Planck Institute for Psycholinguistics 1992,
53–56.
BOWERMAN, MELISSA, and ERIC PEDERSON. 2003. Cross-linguistic perspectives on topologi-
cal spatial relationships. Eugene: University of Oregon, and Nijmegen: Max Planck
Institute for Psycholinguistics, MS.
BROWN, PENELOPE. 1994. The INs and ONs of Tzeltal locative expressions: The semantics
of stative descriptions of location. Linguistics 32.4/5.743–90.
BRUGMAN, CLAUDIA. 1981. The story of ‘over’. Berkeley: University of California master’s
thesis.
CARLSON-RADVANSKY, LAURA A., and DAVID A. IRWIN. 1993. Frames of reference in vision
and language: Where is above? Cognition 46.223–44.
CLARK, EVE V. 1974. How children describe events in time. Advances in psycholinguistics,
ed. by Giovanni B. Flores d’Arcais and Willem J. M. Levelt, 275–84. Amsterdam:
North-Holland.
CLARK, HERBERT H. 1973. Space, time, semantics, and the child. Cognitive development
and the acquisition of language, ed. by Terrence E. Moore, 28–64. New York: Academic
Press.
COMRIE, BERNARD. 1985. Tense. Cambridge: Cambridge University Press.
CROFT, WILLIAM. 1990. Typology and universals. Cambridge: Cambridge University Press.
DE RIJK, R. P. G. 1999. Location nouns in Standard Basque. ASJU (Anuario del Seminario
de Filologı́a Vasca) XXIV-1.3–20.
DRYER, MARTIN S. 1989. Large linguistic areas and language sampling. Studies in Language
13.2.257–92.
DRYER, MATTHEW S. 1992. The Greenbergian word order correlations. Language 68.81–138.
ELMAN, JEFFREY; ELIZABETH BATES; MARK JOHNSON; ANNETTE KARMILOFF-SMITH; DOMEN-
ICO PARISI; and KIM PLUNKETT. 1996. Rethinking innateness: A connectionist perspec-
tive on development. Cambridge, MA: MIT Press.
GREENBERG, JOSEPH. 1966. Language universals, with special reference to feature hierar-
chies. The Hague: Mouton.
HAYWARD, WILLIAM G., and MICHAEL J. TARR. 1995. Spatial language and spatial representa-
tion. Cognition 55.1.39–84.
HEIDER, ELEANOR. 1972. Universals in color naming and memory. Journal of Experimental
Psychology 93.10–20.
HERSKOVITS, ANNETTE. 1986. Language and spatial cognition: An interdisciplinary study
of the prepositions in English. (Studies in natural language processing.) Cambridge:
Cambridge University Press.
JACKENDOFF, RAY. 1983. Semantics and cognition. Cambridge, MA: MIT Press.
JOHNSTON, JUDITH R., and DAN I. SLOBIN. 1979. The development of locative expressions
in English, Italian, Serbo-Croatian and Turkish. Journal of Child Language 6.529–45.
KAY, PAUL, and CHAD K. MCDANIEL. 1978. The linguistic significance of the meanings of
basic color terms. Language 54.610–46.
KAY, PAUL, and LUISA MAFFI. 1999. Color appearance and the emergence and evolution of
basic color lexicons. American Anthropologist 101.4.743–60.
LANDAU, BARBARA, and RAY JACKENDOFF. 1993. ‘What’ and ‘where’ in spatial language
and spatial cognition. Behavioral and Brain Sciences 16.217–38.
LEVINSON, STEPHEN C. 1994. Vision, shape and linguistic description. Tzeltal body-part
terminology and object description. Space in Mayan languages (Special issue of Lin-
guistics 32.4), ed. by John B. Haviland and Stephen C. Levinson, 791–855. Berlin:
Mouton de Gruyter.
LEVINSON, STEPHEN C. 1996. Relativity in spatial conception and description. Rethinking
linguistic relativity, ed. by John J. Gumperz and Stephen C. Levinson, 177–202. Cam-
bridge: Cambridge University Press.
LEVINSON, STEPHEN C. 2000a. H. P. Grice on location on Rossel Island. Berkeley Linguistics

Society 25.210–24.
LEVINSON, STEPHEN C. 2000b. Presumptive meanings: The theory of generalized conversa-
tional implicature. Cambridge, MA: MIT Press.
LEVINSON, STEPHEN C. 2000c. Yélı̂ Dnye and the theory of basic color terms. Journal of
Linguistic Anthropology 10.1.3–55.
LEVINSON, STEPHEN C. 2001. Space, linguistic expression of. International encyclopedia of
social and behavioral sciences, ed. by Neil J. Smelser and Paul Baltes, vol. 22,
14749–52. Amsterdam/Oxford: Elsevier Science.
LEVINSON, STEPHEN C. 2003. Space in language and cognition: Explorations in cognitive
diversity. Cambridge: Cambridge University Press.
LEVINSON, STEPHEN C.; SOTARO KITA; DANIEL HAUN; and BJOERN RASCH. 2002. Re-turning
the tables: Language affects spatial reasoning. Cognition 84.155–88.
LEVINSON, STEPHEN C., and DAVID WILKINS (eds.) 2003. Grammars of space. Cambridge:
Cambridge University Press, to appear.
LI, PENNY, and LILA GLEITMAN. 2002. Turning the tables: Language and spatial reasoning.
Cognition 83.265–94.
LINDNER, SUE. 1981. A lexico-semantic analysis of verb-particle constructions with ‘up’
and ‘out’. San Diego: University of California dissertation.
LUCY, JOHN. 1997. The linguistics of ‘color’. Color categories in thought and language, ed.
by Larry Hardin and Luisa Maffi, 320–46. Cambridge: Cambridge University Press.
MEIRA, SÉRGIO. 1999. A grammar of Tiriyó. Houston, TX: Rice University dissertation.
MEIRA, SÉRGIO. 2003. The grammar of space in Tiriyó. Grammars of space, ed. by Stephen
C. Levinson and David P. Wilkins. Cambridge: Cambridge University Press, to appear.
MILLER, GEORGE, and PHILIP N. JOHNSON-LAIRD. 1976. Language and perception. Cam-
bridge, MA: Harvard University Press.
NÜSE, RALF. 1999. General meanings for German an, auf, in and unter towards a (neo)classi-
cal semantics of topological prepositions. Berlin: Humboldt University dissertation.
PIAGET, JEAN, and BÄRBEL INHELDER. 1956 [1948]. The child’s conception of space. London:
Routledge and Kegan Paul.
PINKER, STEVEN. 1994. The language instinct. New York: William Morrow.
SINHA, CHRIS, and TANYA KUTEVA. 1994. Spatial and non-spatial uses of prepositions. Kogni-
tive Semantik, ed. by Monika Schwarz, 215–38. Tübingen: Narr.
TALMY, LEN. 1983. How language structures space. Spatial orientation: Theory, research
and application, ed. by Herbert Pick and Linda Acredolo, 225–82. New York: Plenum
Press.
TALMY, LEN. 2000. Toward a cognitive semantics. Volumes 1 and 2. Cambridge, MA: MIT
Press.
TRASK, ROBERT L. 1997. The history of Basque. London: Routledge.
VANDELOISE, CLAUDE. 1986. L’espace en français. Travaux Linguistiques. Paris: Seuil.
WHALEY, LINDSAY J. 1997. Introduction to typology: The unity and diversity of language.
Thousand Oaks, CA: Sage Publications.
WIERZBICKA, ANNA. 1980. Lingua mentalis: The semantics of natural language. Sydney:
Academic Press.
Levinson [Received 23 July 2002;
Max Planck Institute for Psycholinguistics accepted 8 March 2003]
PO Box 310
NL-6500 AH Nijmegen
The Netherlands
[levinson@mpi.nl]
Meira
Leiden University
TCIA, Department of Linguistics
PO Box 9515
NL-2300 RA Leiden
The Netherlands
[S.Meira@let.leidenuniv.nl]

Levinson 2003 Natural

Uploaded by

Copyright:

Available Formats

Levinson 2003 Natural

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Levinson 2003 Natural

Uploaded by

Copyright:

Available Formats

‘NATURAL CONCEPTS’ IN THE SPATIAL TOPOLOGICAL

DOMAIN—ADPOSITIONAL MEANINGS IN CROSSLINGUISTIC

1. ADPOSITIONS AS A DISTILLATION OF HUMAN SPATIAL COGNITION. In studies of spatial

2. METHODS FOR COLLECTING DATA. First, we need an operational definition of the

FIGURE 1. Standard comparative stimulus material.

FIGURE 2. Hierarchy in Tiriyó adpositions.

FIGURE 3. Choices between adpositions (Yélı̂ Dnye); 4 consultants, 7 scenes.

SPATIAL LOCATIVE POSITIONAL/

FIGURE 6. Language-specific adpositional categories.

FIGURE 7. Four languages’ adpositions mapped onto fixed array of pictures.

FIGURE 8. Notional areas in fixed array.

FIGURE 9. Discontinuous categories mapped onto fixed array.

5. MULTIDIMENSIONAL SCALING APPROACH. Hypothesis 3 is a statistical hypothesis,

FIGURE 11. Blowup of ATTACHMENT area.

FIGURE 12. Blowup of area including ON/OVER cluster.

distinguished.10 Figure 12 shows a blowup of the ON/OVER cluster found to

FIGURE 14. Blowup of area containing IN cluster.

FIGURE 15. Blowup of NEAR/UNDER cluster.

FIGURE 16. Implicational scale over adpositional notions.

7. CONCLUSIONS. This article should be taken as a pilot study in a complex domain.

BOWERMAN, MELISSA. 2003. Space under construction: Language-specific spatial categoriza-

LEVINSON, STEPHEN C. 2000a. H. P. Grice on location on Rossel Island. Berkeley Linguistics

You might also like