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.
Article Phonetic Symbol Filter
Variation of Semantic Map Display is Necessary
Author’s Reply to “What Should Be on a Map?”
The comment by Narrog (2010) raises various issues
concerning my approach to semantic maps as presented in Cysouw (2010a). I thank
Narrog for pointing out these issues because his comment gives me the
opportunity to clarify a few conceptual assumptions that might not have been
made explicit in the original paper (some of those assumptions are discussed in
Cysouw (2008), which can be seen as the companion paper to Cysouw (2010a)). As a
general conclusion, though, I wholeheartedly agree with Narrog that the endeavor
of creating semantic maps is time well spent. My plea to be careful with their
interpretation—because semantic maps are always to be interpreted relative
to the many analytical decisions made by a researcher (thus, more than one map
is possible with the same data)—should surely not be interpreted as
claiming that the creation of semantic maps is a futile and wholly relativistic
research program. Far from that, I believe that the basic insight leading to
semantic maps, namely that linguistic diversity can be used as a tool to
investigate meaning, allows for breathtakingly new theories of language. My main
problem with the “classical” approach to semantic maps is that it is
not using the full power of this insight and unnecessarily restricts itself to a
single form of display. More variation in the graphical display of semantic maps
is possible, and it is sorely needed in order to deal with the many different
kinds of data amenable to the semantic map approach.
As for the issues raised by Narrog, I will reply here to four of them.
First he claims that a semantic map based on measurable distances
“practically excludes the so-called ‘classical’ or
‘traditional’ maps”. This is a clear misunderstanding because
classical maps can be analyzed as being based on a very simplistic measurable
distance. The classical maps are thus, strictly speaking, a subset of all maps
based on measurable distances. In classical maps, only two possible distances
are used, namely “never encoded together by a form in any known
language” and “encoded identically in at least some language”
(see Cysouw 2007 for full argumentation). So, it is trivial to turn more
detailed measured distances into classical semantic maps by, for example,
rounding off the distances. Furthermore, it is also unproblematic to turn more
detailed measured distances into a graph similar to classical semantic maps.
There is actually a large field of graph theory explicitly dealing with the
question of how to draw sensible graphs on the basis of complex data. An example
of such a method is the Fruchterman & Reingold (1991) graph-layout algorithm
which I used in Cysouw (2010b) commenting on Perrin (2010).
Second, Narrog claims that the
causative/inchoative alternations used in my paper are not cases of polysemy or
polyfunctionality as traditionally used to establish semantic maps. Indeed,
there is no direct polysemy between “wake up” and
“teach”. However, there is polysemy in the distribution of
causative/anticausative morphology for those meanings (as is set out in detail
in the appendix to my paper). This is really nothing intrinsically different
from the more classical kinds of polysemy used to draw semantic maps.
Third, in discussing Figure 2 in my paper, Narrog says “the first
thing that I would have expected was a map […] presenting the two (or
more) dimensions [of the MDS].” Figure 2 in my paper only presents the
first dimension of the MDS on the y-axis and something else on the x-axis,
arguing that there is a correlation between the two factors. I think the reason
behind Narrog’s expectation is that very many presentations of MDS-results
are given in the form of two-dimensional graphics. However, the reason for this
predominance is only a matter of graphical convenience: two dimensions are
easily printed on paper—it is therefore often the first two dimensions of
an MDS that are presented. However, in my paper then I would have had to add a
third dimension to argue for the correlation with the additional factor (now
simply shown on the x-axis). More importantly, nothing intrinsically in MDS
favors the first two dimensions over the others. There are always as many
dimensions in an MDS as the number of entities depicted minus one. So, in the
case of the 31 lexical concepts in my original paper, there are up to
31–1=30 dimension calculated by the MDS. Yet, the crucial principle of MDS
is that the importance of the dimensions diminishes for higher dimensions. The
first dimension captures the largest chunk of variation, the second dimension a
slightly smaller (and independent) chunk, and so on. So, only showing the first
dimension of an MDS is just as well a possibility as showing the first two
dimensions, or the first three. The reason for only showing the first dimension
in my paper is motivated by the wish to show yet some other measure in relation
to the MDS results.
Finally, Narrog is struck by the position of “die/kill” in
Figure 2 in Cysouw (2010a). The closeness to “freeze” and
“boil” does not make sense semantically to him. However, he does not
make explicit what semantic association would be more suitable; he even asks
“what should be represented on a map”. Now, Haspelmath’s
(1993) original paper gives a concrete semantic interpretation of the
causative/anticausative fraction, namely in the form of the spontaneity scale.
The more spontaneously an inchoative action can happen, the more linguistically
marked the causative is expected to be (i.e. the less anticausative
constructions are expected cross-linguistically). It is indeed problematic to
assess the position of “die/kill” on the spontaneity scale solely on
the basis of causative/inchoative patterns, because the “die/kill”
opposition is suppletive in so many languages (so the resulting fraction of
anticausatives is statistically not robust for this alternation). However, as
argued in Cysouw (2008), the fraction of anticausative encodings is highly
correlated with the average difference in length between the inchoative and
causative forms across a wide variety of languages. The main outlier on this
correlation is “die/kill”. I interpret this result in that paper as
showing that the position of “die/kill” on the spontaneity scale can
more truthfully be established on the basis of average length difference than on
the basis of the fraction of anticausative constructions. This interpretation
results in “die/kill” being roughly on the same level of spontaneity
as “learn/teach” and “wake up”. I hope that that makes
more sense to Narrog’s intuition.
Cysouw, Michael. 2007. Building semantic maps: the case of person
New Challenges in Typology, ed. by Bernhard Wälchli and
Matti Miestamo, 225-248. Berlin: Mouton de Gruyter.
-----. 2008. Generalizing scales,
Scales, ed. by Marc
Richards and Andrej Malchukov, 379-96. Leipzig: Institut für
-----. 2010a. Semantic maps as metrics on meaning,
, this volume.
-----. 2010b. Drawing networks from recurrent polysemies,
Linguistic Discovery, this volume.
Fruchterman, Thomas M. J. and Edward M. Reingold. 1991. Graph
drawing by force-directed placement, Software: Practice and Experience,
Haspelmath, Martin. 1993. More on the typology of
inchoative/causative verb alternations,
Causatives and Transitivity, ed.
by Bernard Comrie and Maria Polinsky, 87-120. Amsterdam:
Perrin, Loïc-Michel. 2010. Polysemous qualities and universal
Linguistic Discovery, this volume.
Narrog, Heiko. 2010. What should be on a map?
, this volume.