Similar Place Avoidance: A statistical universal
KONSTANTIN POZDNIAKOV and GUILLAUME SEGERER
Linguistic Typology 11 (2007), 307–348 1430–0532/2007/011-0307
DOI 10.1515/LINGTY.2007.025 ©Walter de Gruyter
To †Sergei Starostin
Abstract
In recent years there has been an interest in the phenomenon of “Similar
Place Avoidance” (SPA), particularly as concerns Arabic CCC radicals. Although
little evidence has been considered outside Arabic, Hebrew, and perhaps
Semitic in general, where roots with successive consonants sharing the
same place of articulation are underrepresented, similarity avoidance has
sometimes been hypothesized as a universal tendency. Progressively extending
our scope from the Atlantic subgroup of Niger-Congo in its relation with other
Niger-Congo languages, which had been our original, diachronic concern, to
almost all of Africa and beyond, we undertook an extensive crosslinguistic investigation
of SPA and found impressive support for this notion.
Keywords: consonant co-occurrence, dissimilation, Obligatory Contour
Principle, phonology, place of articulation, root, Similar Place
Avoidance, word
1. Introduction
Based on what we know about phonological systems, it seems reasonable to
suppose that the major place features of two consonants which make up a
-C1VC2- root should be independent of each other. After all, no language has
been reported where, say, a C2 coronal consonant assimilates in labiality to a
C1 labial consonant, or vice versa. Thus, an input such as /bat/ is never realized
as *[bap] or *[dat]. Major place assimilation is not expected to apply across a
Brought to you by | University of Iowa Libraries
Authenticated
Download Date | 6/3/15 5:25 AM
308 Konstantin Pozdniakov and Guillaume Segerer
vowel.1 While opposite dissimilatory processes affecting place of articulation
are attested (Grammont 1895), they are rare and seldom regular. We therefore
do not expect an input such as /bap/ to be realized as *[bat] or *[dap] by regular
phonological rule. Given the relative independence of major features on
consonants separated by a vowel, it comes as somewhat of a surprise that there
are statistical biases in which transvocalic consonants can succeed each other
within roots. Specifically, in a number of languages which are discussed below,
we have noted that two consonants produced at the same place of articulation
are significantly underrepresented in lexical -CVC- sequences. Below we report
on statistical studies we have done on more than 30 genetically, typologically,
and geographically diverse languages. Our calculations reveal a striking
regularity in the underrepresentation of homorganic consonants in -CVC- se-
quences.
Such distributional irregularities involving consonant place have been long
noted in Semitic studies, particularly as concerns Arabic, whose triliteral√
CCC roots avoid consonants at the same place of articulation (Greenberg
1950, Fleisch 1961). The Arabic instantiation of Similar Place Avoidance
(henceforth SPA) has been studied in great detail by Frisch (1996) and Frisch
et al. (2004), who also demonstrate that speakers are aware of such statistical
biases, which they relate to the Obligatory Contour Principle (OCP): “Adjacent
identical elements are prohibited” (McCarthy 1986: 208). Some recent detailed
statistical studies have also confirmed SPA in Japanese (Kawahara et al. 2005),
in Muna (Coetzee & Pater 2006), and in Proto-Bantu (Teil-Dautrey to appear).
The purpose of the present article is to show that the SPA phenomenon is not
a specific property of Arabic, Japanese, or other isolated languages, but is in
fact observed in most, if not all languages of the world. It should be noted that
this result is not the confirmation of an a priori theoretical postulate. We approached
this issue with no bias – in fact, we stumbled on it quite by accident.
Surprised to discover SPA in a number of languages in the Atlantic sub-branch
of Niger-Congo,2 we believed we were dealing with an inherited genetic trait.
In order to verify the Atlantic hypothesis, we felt compelled to investigate possible
SPA effects in other languages. In the process, and based on extensive
crosslinguistic testing, we arrived at our current position that (statistical) SPA
is a likely universal property of human language.
1. As has been recently documented by Hansson (2001) and Rose & Walker (2004), nonadjacent
consonant harmony is typically limited to nasal, laryngeal, and secondary coronal
features such as anteriority and retroflection. Major place harmony is of course widely attested
in child language (see, for example, Pater & Werle 2001 and references cited therein).
2. For more information on Niger-Congo and its sub-branches, see Williamson & Blench (2000)
and the chapters in Bendor-Samuel (ed.) (1989).
Brought to you by | University of Iowa Libraries
Authenticated
Download Date | 6/3/15 5:25 AM
Similar Place Avoidance: A statistical universal 309
The remainder of this article is organized as follows. Section 2 introduces the
circumstances which initially led us to conduct our statistical counts as well as
the statistical techniques we adopted. Sections 3 to 5 respectively explore Atlantic
languages, other Niger-Congo languages, and non-Niger-Congo African
languages. In order to show that we are not dealing only with an African areal
phenomenon, we document SPA in a few non-African languages in Section 6.
Here we also address the question of how our findings might have been affected
in languages which we do not know well, particularly as concerns the morphological
structure of the lexical items used in our statistical analyses. In the final
discussion (Section 7) and conclusion (Section 8), we consider the implications
of our finding, presenting hypotheses and raising questions for future research.
2. Problem and methodology
The problem under discussion in this article attracted our attention during the
course of preparing a lexical corpus of a group of West African languages
which belong to the Atlantic sub-branch of Niger-Congo. Our goal was to do
lexical comparison for the purpose of subgrouping and ultimate reconstruction.
In fact, the Atlantic languages are lexically very heterogeneous, and although
most specialists continue to treat them as belonging to a single Niger-Congo
sub-branch, lexicostatistical counts often result in low cognate counts almost at
the level of chance (e.g., 5–7 % based on the Swadish 100-word list; cf. Sapir
1971). To take one example, an allegedly stable notion such as ‘big’ produces
several dozen different roots among the 40 Atlantic languages we examined.
Confronted by such variation, we developed a procedure which would permit
us to display lexical information in such a way as to reveal possible formal relationships
between the lexical items in question: for each notion, we constructed
a table where the languages are listed in the first column and the diverse consonant
combinations (C1-C2) head the additional columns. The cells of the table
are filled by the roots themselves, based on the place of articulation of their C1
and C2.
As a first step in preparing these tables, we assigned each phoneme of each
root to one of the major place classes: P, T, C, K. Thus, the phonemes /p, b, f, v,
w, m, mb, á/ are represented by the symbol P for the class of labial consonants.
Similarly, dental (and alveolar) consonants are symbolized as T, (alveo-)palatal
consonants as C, and velar consonants as K. By this procedure, the consonants
of all the examined roots will have been assigned to one of the symbols P, T,
C, K. Since the majority of the lexical roots examined in these languages have
the structure -CVC-, we therefore have a total of sixteen possible combinations
based on the four major classes P, T, C, and K. These sixteen combinations
are the sixteen columns which follow the language names in the table. Table 1
illustrates this procedure for the notion ‘hair’. As can be seen, the roots that are
Brought to you by | University of Iowa Libraries
Authenticated
Download Date | 6/3/15 5:25 AM
310 Konstantin Pozdniakov and Guillaume Segerer
Table 1. Words for ‘hair’ in the Atlantic languages
P-P P-T P-C P-K T-P T-T T-C T-K
Fula wa:r leá
Sereer wiil
Basari mban, fur
Bedik mbal mboy
Konyagi muul
Pen mban
Ndut fen
Noon fen
Safen fan
Lehar mul
Palor fen
Wolof *war
Buy bunk dung
Nyun
Biafada wey
Balanta
Joola 1 wal
Joola 2 wan
Manjaku wel, faal
Mankañ wel, fal
Pepel
Bijogo wen
Nalu fEl
Nalu Tob
Sua wiñ
Baga K. foon
Baga M. foon
Landuma foon
Temne fon
Bullom ring
Kisi
Sherbro
Gola dum
grouped together in the same column are not necessarily related. In addition,
depending on sound changes, historically related roots may appear in different
columns. Despite this, there is a greater probability that related roots will be
found in the same rather than a different column. This is the justification of the
procedure.
We applied this method and constructed this type of table for numerous basic
lexical notions. In the course of this we noticed that certain columns, on
average, seem “more empty” than others. For example, despite the numerous
reflexes seen in Table 1, the column K-K is completely devoid of words for
‘hair’, but also is strikingly empty for other basic lexical notions. For a given
gloss, it would be completely normal if certain columns were empty. The closer
Brought to you by | University of Iowa Libraries
Authenticated
Download Date | 6/3/15 5:25 AM
Similar Place Avoidance: A statistical universal 311
Table 1 (cont). Words for ‘hair’ in the Atlantic languages
C-P C-T C-C C-K K-P K-T K-C K-K
Fula gac
Sereer
Basari
Bedik
Konyagi
Pen
Ndut xoá
Noon
Safen
Lehar
Palor
Wolof *jaaw cok
Buy kum gen
Nyun gen
Biafada
Balanta yEEg hul
Joola 1 jab
Joola 2
Manjaku
Mankañ jab gaal
Pepel
Bijogo
Nalu
Nalu
Sua
Baga K.
Baga M.
Landuma
Temne
Bullom
Kisi yin
Sherbro zem
Gola
the attested reflexes are of an ultimate proto-root, the fewer filled columns there
should be. However, it was quite surprising to us that for dozens of basic notions,
we find almost the same columns empty.
What does this mean? Should one conclude that the Atlantic languages avoid
certain combinations of consonants? If so, could this be taken to be a property
of Proto-Atlantic which is preserved in the daughter languages? In this case, the
shunned combinations could provide a solid argument in favor of the existence
of the Atlantic group itself, for which no shared linguistic trait has been offered
as evidence of the alleged genetic sub-branch of Niger-Congo. In addition, if
valid, consonant distribution patterns of this sort might furnish very interesting
pathways for the reconstruction and comparision of Proto-Atlantic with
Brought to you by | University of Iowa Libraries
Authenticated
Download Date | 6/3/15 5:25 AM
312 Konstantin Pozdniakov and Guillaume Segerer
other subgroups of the Niger-Congo macro-family. In this case the appropriate
research strategy would be to establish the precise consonant combinations favored
or disfavored in languages from each of the other subgroups. To do so,
one would have to transform any intuitive impression into numeric values.3
Let us return to the “simplified” roots, where each consonant is represented
by the symbol of its class. For each biconsonantal root, there is an initial consonant
(C1) and a non-initial consonant (C2). For longer roots, for example of the
form CVCVC, the medial consonant is C2 with respect to the initial consonant,
but it is C1 with respect to the final consonant. In our statistical counts, the
lexicon is thus broken up into C1-V-C2 sequences, where the medial consonant
of a triconsonantal root is calculated both with respect to the consonant which
precedes and the consonant which follows it. For sequences including consonant
clusters, as for example CVCCVC, the cutting was CVC-CVC; when two
adjacent consonants are of the same class, as for example CVPPVC, then it
is equivalent to the CVPVC case, the P being final for the first sequence and
initial for the second one.
As an illustration, consider the consonant system of Balanta, displayed according
to place and manner of articulation in Table 2. There are eight labial
consonants /f, b, gb, w, m, mf, mb, Ngb/, nine dental consonants /t, th, d, l, r, n,
nt, nth, nd/, seven palatal consonants /c, s, j, y, ñ, ns, nj/, and five velar consonants
/k, h, g, N, Ng/, which we can symbolize by P, T, C, and K, respectively.4
Table 3 presents the measured frequencies of each place of articulation from
a Balanta lexicon of 766 entries containing 904 C1VC2 sequences (NdiayeCorréard
1970).5
3. For an application of statistical methods to comparative studies, see Pozdniakov (1991).
4. /th/ and /nth/ are interdentals. /s/ is treated as palatal, as it frequently occupies the “[S] slot”,
and labiovelars are treated as labials. These choices have been suggested by our knowledge
of the Atlantic languages, where for example /s/ is associated with /c/ in all the languages
showing consonant mutation, such as Fula, Sereer, the Tenda cluster, Wolof. We chose not to
change these while computing data from other languages. Other possibilities would have been
to treat /s/ as a dental, or labiovelars as velars, or both. This would have slightly changed the
figures, but not the tendencies. Moreover, as will be shown soon, dentals share some statistical
features with palatals while labials share some features with velars.
5. The number of sequences used for calculating the tables is indicated by “n=xx” in all tables
hereafter.
Brought to you by | University of Iowa Libraries
Authenticated
Download Date | 6/3/15 5:25 AM
Similar Place Avoidance: A statistical universal 313
Table 2. The Balanta consonant system
f t, th c, s k, h
b d j g
gb
w l, r y
m n ñ N
mf nt, nth ns
mb nd nj Ng
Ngb
Table 3. Balanta: Observed frequencies (O), in % (n=766)
P T C K
C1 26.8 31.7 23.7 17.8
C2 20.5 48.7 14.8 16.0
The next step in the procedure is to calculate the theoretical frequencies of
the different combinations. The theoretical frequency of any of the sixteen C1V-C2
combinations is obtained by multiplying the absolute frequency of the C1
consonant by the absolute frequency of the C2 consonant. The theoretical frequency
thereby obtained assumes an absence of correlation between the quality
of the C1 and that of the C2. Table 4 furnishes the theoretical frequencies of the
sixteen combinations for Balanta.
Table 4. Balanta: Expected frequencies (E), in %
C2
P T C K
C1
P 5.5 13.1 4.0 4.3
T 6.5 15.4 4.7 5.1
C 4.9 11.5 3.5 3.8
K 3.6 8.7 2.6 2.8
To illustrate, consider the example of K-P, that is, cases where any velar
is followed by any labial consonant. The frequency of C1 K is 17.8 %, and
the frequency of C2 P is 20.5 %. The frequency of the combination K-P is
therefore theoretically 17.8 % × 20.5 %, or 3.6 %. Based on the 904 sequences
examined, this means that one should find 904 × 3.6 %, or 33 sequences of
the form KVP. Among the 766 entries in the Balanta lexicon, one does in fact
Brought to you by | University of Iowa Libraries
Authenticated
Download Date | 6/3/15 5:25 AM
314 Konstantin Pozdniakov and Guillaume Segerer
find 34 sequences of this form. We can therefore consider this distribution as
“normal” (i.e., conforming to the anticipated theoretical frequency).
On the other hand, the combination K-K, for which we should find 17.8%×
16.0%×904 = 26 sequences, attests only nine such forms. For each language,
one thus compares the theoretical or expected (E) frequency with the actual or
observed (O) frequency of each combination. If a correlation exists between
the qualities of C1 and C2, this should be manifested by a significant discrepancy
between the E and O values. Thus, taking the example of the actual K-K
sequences in Balanta, the discrepancy is −65 % with respect to the theoretical
frequency. Table 5 presents the E/O discrepancies for C1-C2 in Balanta.
Table 5. Balanta: 100∗
(O−E)
E
C2
P T C K
C1
P −57.6 +30.7 +19.9 −38.2
T +22.6 −22.0 +1.1 +36.9
C +32.4 −20.3 −24.3 +42.8
K +3.2 +20.0 +0.6 −65.1
By convention and for purposes of readability, we adopt the following procedure
in presenting our results: (i) a discrepancy whose absolute value is less
than 15 % is considered to be non-significant and is not noted; (ii) a discrepancy
whose absolute value is between 15 % and 30 % is noted by a “+” or “–”
sign; (iii) a discrepancy whose absolute value is greater than 30 % is noted by
a double “+ +” or “− −” sign.6 The actual values, i.e., the observed number
of each combination, are given in the Appendix at the end of the article.
Table 6. Results of Table 5 in terms of +/− categories
C2
P T C K
C1
P − − + + + − −
T + − + +
C + + − − + +
K + − −
6. For ease of readability we do not use the X2 test. As another advantage, the method used here
preserves the direction of deviation with respect to the norm, which X2 does not.
Brought to you by | University of Iowa Libraries
Authenticated
Download Date | 6/3/15 5:25 AM
Similar Place Avoidance: A statistical universal 315
Following these conventions, the percentages in Table 5 produce the values
in Table 6. The presentation in Table 6 has the advantage of nicely exposing the
positive and negative tendencies. Thus, as indicated by the minuses along the
descending diagonal, Balanta systematically underrepresents combinations of
consonants at the same place of articulation. Not only is it the case that velars
rarely combine (as seen in the discussion of K-K above), but the same is true of
labials, palatals, and to a somewhat lesser extent dentals. The systematic nature
of the distribution observed in Table 6 (and elsewhere to follow) is a good
indicator of the validity of the general approach and of the specific method.
Such results in fact encourage us to seek other distributional regularities.
Before moving on to consider other languages, we should take note of two
other observations that can be made on the basis of Table 6. First, the presence
of negative discrepancies must be compensated for by positive discrepancies.
By definition, the total sum of the discrepancies with respect to the norm must
be zero. The negative discrepancies are almost exclusively due to the principle
of SPA: Balanta disfavors sequences where C1 and C2 are produced at the
same place of articulation. On the other hand, the positive discrepancies do
not seem to be principled. We thus do not see the relationship between the
fact that C1 palatals preferentially combine with C2 labials and velars and the
fact that C1 labials more frequently combine with C2 palatals. In other words,
while the distribution of the minuses is (relatively) regular and systematic, the
distribution of the pluses is not.
A second observation that can be made from Table 6 is that from a statistical
point of view, the four consonant classes P, T, C, and K can be grouped into
two “superclasses”: P and K vs. T and C. Within roots, not only do palatal
consonants show a statistical tendency to not combine with another palatal, but
also not with a dental. Similarly, labial consonants tend not to combine with
other labials, but also not with velars. We will refer to the two superclasses
as “peripheral” (P, K) and “medial” (T, C). The peripheral/medial opposition
corresponds exactly to the grave/diffuse distinction of Jakobson et al. (1952)
and inversely to the coronal/non-coronal opposition of generative phonology
(Clements & Hume 1995).
The two superclasses behave as basic classes in that consonants from within
each set rarely combine with each other. However, of equal significance is the
fact that the lack of combinations within a superclass corresponds to an excess
of combinations between the superclasses, as summarized in Table 7.
Brought to you by | University of Iowa Libraries
Authenticated
Download Date | 6/3/15 5:25 AM
316 Konstantin Pozdniakov and Guillaume Segerer
Table 7. Superclasses
Peripheral Medial
Peripheral − +
Medial + −
Table 8. Balanta: PKTC order
C2
P K T C
C1
P − − − + + +
K − − +
T + + + −
C + + + + − −
Because of these new groupings, we propose to modify the order of presentation
within the tables. Instead of the articulatory order PTCK, we shall
henceforth adopt the order PKTC, as shown for Balanta in Table 8. As seen by
the borders, we can now distinguish four quadrants in these tables: the upper
left and lower right quadrants enclose combinations where C1 and C2 belong
to the same superclass. As seen, all of the minus signs fall within these quadrants.
The lower left and upper right quadrants indicate combinations where
C1 and C2 belong to different superclasses. All of the plus signs fall within
these quadrants. Finally, the C1 and C2 which combine in the cells along the
descending diagonal belong to the same class (place of articulation). Here and
in the subsequent tables, these cells are shaded.
The Balanta facts have served as an illustration in presenting the methodology
and an initial set of results for comparison. In the following sections we
shall see that SPA by class and superclass is widespread in Africa and beyond.
3. Similar Place Avoidance in Atlantic
As mentioned, our initial interest was both diachronic and Atlantic-specific: in
all of the Atlantic languages we have examined, one finds the same distributional
tendencies concerning the classes P, T, K, C, as well as the peripheral and
medial superclasses. We present the results from twelve Atlantic languages in
Table 9, where the cells along the descending diagonal representing same place
combinations are shaded. The shaded cells along the descending diagonal are
almost all characterized by one or two minus signs. This corresponds to the tenBrought
to you by | University of Iowa Libraries
Authenticated
Download Date | 6/3/15 5:25 AM
Similar Place Avoidance: A statistical universal 317
dency for roots to avoid C1-C2 sequences made at the same (or similar) place
of articulation. The tables show that this tendency also affects consonants from
the same superclass: peripheral consonants tend not to combine, just as medial
consonants tend to avoid one another. The P/K and T/C groupings, which one
might have considered specific to our discussion of Balanta in Section 3, are
relevant in all of the languages examined, without exception.
Table 9. CVC combinations in twelve Atlantic languages
Fula (Labatut 1994; n=672)
P K T C
P − − − − +
K − − − − + +
T + + + + − −
C + + + − −
Palor (d’Alton 1987; n=2,116)
P K T C
P − − − + +
K − − − + +
T + + + + − −
C + + − −
Wolof (Fal et al. 1990; n=8,456)
P K T C
P − − − + +
K − − + +
T + + + −
C + + −
Nyun-Buy (Lespinay 1991; n=4,428)
P K T C
P − − + +
K − − − + + +
T + + + − − −
C + + −
Jaad (Ducos 1971; n=1,200)
P K T C
P − − − − + +
K − − +
T + + + + − − −
C + + + − −
Balanta (Ndiaye-Corréard 1970; n=904)
P K T C
P − − − − + + +
K − − +
T + + + −
C + + + + − −
Joola Kwaatay (Payne 1992; n=2,183)
P K T C
P − − − + +
K − − − +
T + + + −
C + + −
Manjaku (Buis 1990; n=3,145)
P K T C
P − − − − + +
K − − +
T + + + −
C + −
Brought to you by | University of Iowa Libraries
Authenticated
Download Date | 6/3/15 5:25 AM
318 Konstantin Pozdniakov and Guillaume Segerer
Bijogo (Segerer 2002; n=1,499)
P K T C
P − − − + + +
K − − + +
T + + + + − −
C + + − −
Sua (Segerer 1998; n=495)
P K T C
P − − − + +
K − − + n.s.a
T + + −
C +
Bullom (Nyländer 1814; n=827)
P K T C
P − + +
K − − + + n.s.
T + + + − − −
C
Kisi (Childs 2000; n=2,981)
P K T C
P − − + +
K − − + +
T + + + − − −
C + −
a The abbreviation “n.s.” (non-significant) indicates that the value of the norm (E) is too low to
have a statistical value. If the expected amount of a given combination is 2 and we find three
examples of this combination, it will make a 50 % positive deviation. We consider that in such
cases the deviation is not relevant, because the influence of chance is too great. We arbitrarily set
the minimal value for the norm at 10.
Let us separately examine the two superclass diagonals in Table 9, each one
consisting of two quadrants. One, the “grey super-diagonal”, represents the
combination of consonants of the same superclass (including the shaded descending
diagonal). The other, the “white super-diagonal”, represents the combination
of consonants of different superclasses. Given the twelve languages
examined, for each of the combinations peripheral-peripheral, peripheral-medial,
medial-peripheral, and medial-medial, there are 12 × 4 = 48 cells to fill
(since each combination of two superclasses corresponds to four combinations
of place). The result is striking: as seen in Table 10, all of the minus signs are
concentrated in the grey super-diagonal. In other words, the number of combinations
of consonants from the same superclass is always less than the norm,
and the number of combinations of consonants from different superclasses is
never less than the norm.
Table 10. Number of minuses in each quadrant
Peripheral Medial
Peripheral 41 / 48 0 / 48
Medial 0 / 48 30 / 48
Brought to you by | University of Iowa Libraries
Authenticated
Download Date | 6/3/15 5:25 AM
Similar Place Avoidance: A statistical universal 319
This very marked tendency found in all branches of the Atlantic group motivated
us to postulate a similar distribution in Proto-Atlantic. As Atlanticists,
the possibility of such a reconstruction was good news to us for two reasons.
First, it provided us with an element of proof in establishing the reality of the
Atlantic group. As already mentioned, no one up to this point had cited a single
linguistic trait found only in the Atlantic group. Second, the reconstruction of
SPA at the Proto-Atlantic level seemed to open new perspectives in seeking
regular correspondences with languages within other sub-branches of NigerCongo:
if SPA were the result of a Proto-Atlantic innovation involving place
dissimilation, then it might be that Atlantic labials in a certain context correspond
to Niger-Congo dentals, or that Atlantic velars correspond to NigerCongo
palatals. As seen in the following sections, what has instead turned out
to be “bad” news for Proto-Atlantic has wider consequences for the study of
language in general.
4. Similar Place Avoidance in Niger-Congo
If it is reasonable to postulate that Proto-Atlantic innovated SPA, it should
be the case that, statistically, other Niger-Congo subgroups do not exhibit the
same systematic restrictions in their consonant distributions. In other words,
if SPA is really an Atlantic innovation, it should be absent in other NigerCongo
subgroups. This was our belief, but we were wrong. In this section we
examine several other sub-branches of the Niger-Congo phylum for which we
have reconstructions.
Table 11 presents the results obtained by mapping out Moñino’s (1995) lexical
reconstructions of Proto-Gbaya. The Gbaya languages which are spoken
in Central Africa (Cameroun, Central African Republic) constitute a branch of
Niger-Congo, perhaps at the same level as Adamawa or Gur. As can be seen,
the situation is similar to what we saw in Atlantic: all of the same-place combinations
are avoided, and the remaining minus signs concern combinations
within the same superclass.
Table 11. Proto-Gbaya (Moñino 1995; n=761)
C2
P K T C
C1
P − − − − + + +
K − − + +
T + + + + − − −
C + + n.s.
Brought to you by | University of Iowa Libraries
Authenticated
Download Date | 6/3/15 5:25 AM
320 Konstantin Pozdniakov and Guillaume Segerer
Similar results are found in Proto-Ijo (Nigeria; reconstructed by Williamson
2004), which constitutes one of the highest branches of the Niger-Congo phylum.
Again, combinations of homorganic consonants are avoided.
Table 12. Proto-Ijo (Williamson 2004; n=509)
C2
P K T C
C1
P − − +
K − − − + n.s.
T + + + + − − n.s.
C + n.s. − n.s.
With distant Proto-Gbaya and Proto-Ijo joining Proto-Atlantic, the alleged
“Atlantic dissimilation” is obviously not an isolated fact. The probability of
systematic SPA occurring by inheritance in three Niger-Congo branches is extremely
low. At this point, we are conditioned to expect the same grey diagonal
minuses in other branches.
Table 13 shows that Proto-Mande, another early offshoot of Niger-Congo,
does not disappoint. (The calculations in Table 13 are based on the ProtoMande
reconstructions presented by Valentin Vydrine at the Workshop on
Proto-Niger-Congo, Paris 2004.) As in previously examined cases, all of the
minuses are concentrated within the grey super-diagonal, and the only double
minuses are in the grey cells within the narrow diagonal. Thus, as far back as
Proto-Mande we see a clear avoidance of consonant combinations at the same
place of articulation and, to a lesser extent, consonants belonging to the same
superclass. In addition, one observes that within the shaded cells along the narrow
diagonal, SPA is stronger among peripheral vs. medial consonants. In fact,
this observation is valid for all of the languages described up to now.
Table 13. Proto-Mande (Vydrine 2004; n=511)
C2
P K T C
C1
P − − − + +
K n.s. − − + n.s.
T n.s. + + n.s.
C n.s. + + − n.s.
Brought to you by | University of Iowa Libraries
Authenticated
Download Date | 6/3/15 5:25 AM
Similar Place Avoidance: A statistical universal 321
The only languages which show a certain deviation from the reported regularities
are the Bantu languages, whose combinatorial properties are documented
in Table 14. Here for the first time, a grey cell in the table has a “+ +”.
It appears that Proto-Bantu had access to a disporportionate number of palatal
consonant sequences. Since this has to do with medial consonants, another
tendency evoked above remains valid: peripheral consonants avoid each other
more. An examination of individual Bantu languages from different zones
(Guthrie 1967–1971) shows that the Proto-Bantu situation is well-reflected in
present-day daughter languages (see Table 15).
Table 14. Proto-Bantu (Bantu Lexical Reconstructions 1998; n=12,426)
C2
P K T C
C1
P − +
K − − +
T + + − −
C + + − − + +
Table 15. Four individual Bantu languages of four different zones
Swahili (zone G; Rugemalira 1993;
n=1,481)
P K T C
P − − +
K − − + + +
T + − −
C + − −
Mpongwe (zone B; Mouguiama 1994;
n=3,506)
P K T C
P + +
K − +
T + − −
C + − + +
Bemba (zone M; Mann 1995;
n=10,653)
P K T C
P − +
K − − + +
T + + −
C + − + +
Kiga-Nkore (zone J; Taylor 1959;
n=17,944)
P K T C
P −
K − + −
T + −
C + − +
One does not have to be a specialist in Bantu historical linguistics to assume
that the unusual statistical distribution of palatal consonants in the reconstructed
roots as well as in present-day languages reflects a Proto-Bantu
Brought to you by | University of Iowa Libraries
Authenticated
Download Date | 6/3/15 5:25 AM
322 Konstantin Pozdniakov and Guillaume Segerer
innovation with respect to Proto-Niger-Congo. In fact, the Bantu languages are
the only ones which show any tendency for C1 and C2 consonants to agree in
place of articulation – and, except for Swahili P-K, only among palatals.
5. Similar Place Avoidance as an African areal feature?
The Bantu deviation just discussed should not hide the fact that SPA represents
a formal characteristic of the entire Niger-Congo family. The question
addressed in this section is whether SPA is specific to Niger-Congo or whether
it is an African areal feature. If the former is true, then languages from other
African phyla should have behaviors that are significantly different from those
just seen in Niger-Congo.
We begin with Sara-Kaba-Na, a Nilo-Saharan language of the Sara-BongoBagirmi
branch. As seen in Table 16, the distributions are absolutely comparable
to what we have thus far observed in Niger-Congo: the grey diagonal
is entirely filled with minuses, and we find no minus outside the grey superdiagonal.
If these distributions could be shown to be a valid genetic marker,
they would have something to offer to those who favor a union of the NigerCongo
and Nilo-Saharan families into a even larger macro-family (see, however,
Section 6).
Table 16. Sara-Kaba-Na (Danay et al. 1986; n=3,300)
C2
P K T C
C1
P − − − − + +
K − − +
T + + + −
C + + − −
Let us therefore take a language of Africa which we know not to be involved
in this hypothesis. Based on its unique genetic source and relative isolation
from the continent, Malagasy, an Austronesian language, would not be
expected to share linguistic properties with Niger-Congo or Nilo-Saharan languages.
One nevertheless clearly sees in Table 17 the same discrepancies with
respect to the norm to which we have become accustomed. What better example
could we find of a typical Niger-Congo, even Atlantic distribution? All of
the minuses are in the grey super-diagonal, and all of the pluses are in the white
super-diagonal. At this point of the investigation, we have grounds for suspecting
that we are dealing with a more general phenomenon which goes beyond
the boundaries of genetic divisions. Could SPA be an African areal feature?
Brought to you by | University of Iowa Libraries
Authenticated
Download Date | 6/3/15 5:25 AM
Similar Place Avoidance: A statistical universal 323
Table 17. Malagasy, Sakalava (Lacroix 2001; n=1,944)
C2
P K T C
C1
P − − + +
K − − +
T + + − −
C + + − − −
The examination of African languages would not be complete without representation
from the Afro-Asiatic phylum, here represented by the Chadic subgroup.7
In this connection we have only tested two lexicons: the tentative ProtoChadic
reconstructions of Jungraithmayr & Ibrizsimow (1994) and the lexicon
of the Ader dialect of Hausa (Caron 1991). The results are presented in Table
18. In the Proto-Chadic corpus, the shaded diagonal is even more marked
than in Atlantic or other Niger-Congo languages. Recall that the “– –” sign
indicates that the number of combinations is at least 30 % below the norm calculated
according to the percentages observed independently in each position.
Here the situation is uniform for all four combinations of homorganic consonants.
In Hausa the tendency is a little less strong, but it is not contradictory,
since all of the minuses remain in the grey super-diagonal. As in Niger-Congo,
the tendency is greater for peripheral than for medial consonants.
Table 18. Chadic and Hausa (Ader dialect)
Proto-Chadic (n=1,306)
P K T C
P − − + + +
K − − + +
T + + + + − −
C + + + − − −
Hausa (n=3,880)
P K T C
P − − + +
K − − +
T + + −
C + + −
One might perhaps be less surprised that SPA is in full force within Chadic
than in the other language families that we have examined. Chadic is a branch
of the Afro-Asiatic macro-family to which Arabic, Hebrew, and the Semitic
subgroup also belong. As stated in Section 1, the avoidance of combinations of
7. Unfortunately we have not been able to study any language from Khoisan, the fourth African
macro-phylum.
Brought to you by | University of Iowa Libraries
Authenticated
Download Date | 6/3/15 5:25 AM
324 Konstantin Pozdniakov and Guillaume Segerer
similar consonants has been noted in these languages for some time, and with
Chadic we can extrapolate perhaps to the level of Afro-Asiatic itself.
Since Semitic takes us marginally outside of Africa, the possibility that SPA
is an African areal feature is somewhat weakened. The final blow comes in the
next section, where we demonstrate that SPA is a linguistic universal.
6. Similar Place Avoidance as a linguistic universal
Before presenting evidence for SPA from outside the African continent, where
we have less expertise, we wish to comment again on the steps that have been
involved in conducting this study. We reiterate that we first discovered the
SPA phenomenon in the Atlantic languages. As specialists of these languages,
we have developed tools for treating important corpora without making too
many methodological errors. However, we were surprised to see the SPA phenomenon
so distinctly manifested. It is clear that we do not have the same
expertise to treat the data from other families. Despite this, and despite any
possible biases, we have noted the same tendency working on lexicons which
have not been subject to the kind of phonological and morphological analyses
that should logically precede this kind of calculation.
To illustrate this, we conducted an experiment based on Fula. The whole
corpus has 1,153 items (1,651 CVC sequences). However, out of these 1,153
words, there are only 643 primary lexical stems (672 CVC sequences), the
others being derived. So we calculated the tables for both corpora. We can
see in Table 19 that the general tendency is the same in both cases, even if
the details are different: all the minuses are in the descending diagonal and
all the pluses are in the ascending one. In addition, we can see that SPA is in
every case more important for peripheral consonants than for medial ones. This
experiment shows that even with no expertise on a given language, we are able
to gather statistics on this language.
Table 19. Fula: difference between “clean” and “raw” corpora
“Clean” corpus, n=672
P K T C
P − − − − +
K − − − − + +
T + + + + − −
C + + + − −
“Raw” corpus, n=1,651
P K T C
P − − − +
K − − − + +
T −
C + + + + − −
In each of the languages examined up to now, the same tendencies have been
at work. The African languages presented in the preceding sections belong to
Brought to you by | University of Iowa Libraries
Authenticated
Download Date | 6/3/15 5:25 AM
Similar Place Avoidance: A statistical universal 325
four different families, but are still tied geographically. The question that now
arises is: What happens outside of Africa?
It seems that even in Indo-European statistical skewings in the combination
of C1-C2 consonant sequences have not been very appreciated. Table 20
presents our findings for Proto-Indo-European (PIE), based on Starostin’s
(1998–2005) STARLING database.
Table 20. Proto-Indo-European (n=3,085)
C2
P K T C
C1
P − − +
K − − +
T + + + + − −
C + + −
Indo-European is without a doubt the most-studied language family, and that
for more than two centuries. As seen in Table 20, the same SPA tendency is
observed in Proto-Indo-European reconstructions. We note again that combinations
of consonants at the same place of articulation exist in most, if not all
languages, and that they are not necessarily rare. Some examples from IndoEuropean
taken from Starostin (1998–2005) include:
(1) a. labial-labial: *pib ‘to drink’, *paw ‘few, small’, *bhebhr-u-‘bear’
b. dental-dental: *tal-/-e- ‘earth, ground’, *del ‘long’, *nan-/*nen‘mother,
nurse’
c. palatal-palatal: *yes ‘to boil’
d. velar-velar: *(s)kek- ‘hair, beard’, *koks- ‘armpit’
It is likewise quite easy to find examples of consonant sequences at the same
place of articulation in African languages in general and in Atlantic in particular.
Often these are words of great frequency of usage, e.g., Wolof bopp ‘head’,
sàcc ‘steal’. From such examples one might easily conclude that there are no
constraints on transvocalic C1-C2 sequences. As this study has documented,
this would be an error. As we have shown, statistically, these combinations
are relatively rare.
Although the notion of a Proto-Nostratic existing at a considerably greater
time depth than Proto-Indo-European is quite controversial, Table 21 shows
that SPA is observed in the reconstructions proposed in the etymological dictionary
of Illiˇc-Svityˇc (1971–1984).
Brought to you by | University of Iowa Libraries
Authenticated
Download Date | 6/3/15 5:25 AM
326 Konstantin Pozdniakov and Guillaume Segerer
Table 21. Proto-Nostratic (n=318)
C2
P K T C
C1
P − − − + + −
K − − + +
T n.s. + + − − − −
C n.s. n.s. n.s.
Since Altaic is one of the branches of Nostratic, it is not surprising to find
SPA effects in the languages of that family. Table 22 presents the facts of “Classical”
Mongolian. In this table the tendencies are more marked than ever: not
only do the shaded cells contain seven minuses out of eight possible, but also
the cells of the inverse diagonal contain seven out of eight possible pluses.
(Only one minus and one plus occur outside these “narrow” diagonals, but
both are found within expected quadrants.) The distribution of these pluses
strikingly suggests that the heterorganic combinations P-C, C-P, T-K, and K-T
are strongly favored in Classical Mongolian. We have already remarked that
the two diagonals do not have the same status: while there is a tendency to find
the most minuses along the descending (shaded) diagonal in all of our tables,
there does not appear to be a corresponding tendency for the greatest number
of pluses to congregate along the inverse diagonal. Rather, these pluses appear
randomly distributed within the cells of the lower-left and upper-right quad-
rants.
Table 22. Classical Mongolian (n=66,407); the data were extracted from an online
dictionary of over 25,000 entries (Lainé 2004)
C2
P K T C
C1
P − − − + + +
K − − + +
T + + − −
C + −
While the other languages examined are not systematic in their preferences
for certain heterorganic sequences, the question naturally arises as to whether
the distribution of the pluses in Classical Mongolian is in fact principled. We
summarize the relevant facts as follows, where “>>” means “is preferred over”.
Brought to you by | University of Iowa Libraries
Authenticated
Download Date | 6/3/15 5:25 AM
Similar Place Avoidance: A statistical universal 327
(2) a. P-C >> P-T
b. T-K >> T-P
c. C-P >> C-K
d. K-T >> K-C
As seen, combinations of labials and palatals (P, C) are preferred over combinations
of labials and dentals (P, T), and combinations of dentals and velars (T,
K) are preferred over combinations of dentals and labials (T, P). Up until now
we only recognized peripheral (P, K) and medial (T, C) superclasses, which
group together the most similar places of articulation. The commonality of the
two places of articulation in each superclass can be defined either in terms of
their shared acoustic properties (grave vs. acute) or their shared articulatory
(non-)involvement of the front of the tongue (coronal vs. non-coronal). Mongolian
now suggests that anterior consonants (P, T) and posterior consonants
(C, K) also share a property, which corresponds roughly to [±high] (raising of
the body of tongue) in the Chomsky & Halle (1968) distinctive feature framework.
In this framework the four places of articulation would have the feature
values in Table 23.
Table 23. Shared features among P, T, C, K
P T C K
Coronal − + + −
High − − + +
Approached in these terms, we see that two groupings do not share either
feature: P, C and T, K. It could therefore be that Classical Mongolian has
the flip-side of SPA, namely the favoring of the most dissimilar consonant sequences.
We note that Classical Mongolian is the only language in our study
which has front-back vowel harmony, which may turn out to be a relevant factor,
hence worthy of further study.8
To summarize thus far, we have seen that SPA effects are widespread in the
world’s languages. Even if we have not tested all of the languages or language
families of the world, there is reason to believe that we are dealing with a universal
phenomenon. What would it take to be even more convincing, specifically
to rule out any possibility of an Afro-Eurasian genetic or contact phenomenon?
A genetically isolated language? A recently formed language, e.g.,
8. Many of the African languages cited have either ATR or height harmony which may be expected
to interact less with SPA than front-back vowel harmony. More languages having the
latter, as well as rounding harmony, need to be investigated (e.g., Turkish).
Brought to you by | University of Iowa Libraries
Authenticated
Download Date | 6/3/15 5:25 AM
328 Konstantin Pozdniakov and Guillaume Segerer
a pidgin? A language belonging to more distant language families (Australian,
American Indian)? Table 24 presents an example of each of these.9
Table 24. Other languages
Basque, Euskara
(http://weblandarbaso.miarroba.com;
n=3,140)
P K T C
P − −
K − − − − +
T + + + − +
C + + − −
Pidgin English, Port Moresby
(Barhorst & O’Dell-Barhorst
no date; n=2,215)
P K T C
P − − +
K − −
T + + + −
C +
Quechua (Ancey 1997;
n=5,254)
P K T C
P − − − + +
K − −
T + + + −
C + + + −
Kamilaroi, Australia
(Austin & Nathan 1998; n=980)
P K T C
P − − + +
K + + − −
T + +
C − + − −
In this arbitrary sample of four languages, we note no contradiction with the
tendencies previously seen. Of the sixteen shaded cells in Table 24, twelve contain
one or two minuses, and none contains any pluses. In contrast, the tendency
is less clear concerning the “wide” diagonal, i.e., combinations within the same
superclass. This is particularly striking in the case of Kamilaroi, where P-K,
K-P, T-C, and C-T are overrepresented. As seen, the minuses found along the
grey diagonal are unexpectedly compensated by pluses in the white cells of
the upper-left and lower-right quadrants. This case is unique among our sample.
Perhaps what we can say is that even if every language shows the effects of
SPA, each language preserves its own originality. Out of the 31 tables presented
above, no two are exactly identical. Evaluating the significance of the individual
differences is of course a task reserved for specialists of each language and
language family.
9. All the data for theses four languages were found on the internet. The links are given at the
end of the reference section.
Brought to you by | University of Iowa Libraries
Authenticated
Download Date | 6/3/15 5:25 AM
Similar Place Avoidance: A statistical universal 329
7. Discussion
7.1. Making sense of similarity avoidance
The preceding sections have clearly established that SPA is a likely universal
property of human language. We are aware that other studies have been concerned
with the tendency of like features or segments to resist repetition or be
kept at a distance from one another. Within non-linear phonology, the Obligatory
Contour Principle (OCP) is often cited as a universal tendency:
(3) Adjacent identical elements are prohibited (McCarthy 1986: 208)
Various authors have invoked the OCP or related principles under different
names to account for a variety of phenomena that minimize the same or similar
elements (consonants, vowels, tones, whole syllables, etc.). The following
quote from Tang (2000: 34) succinctly references much of this work:
In the literature, the principle in (1) [here (3)] has been called the “obligatory contour
principle”, the OCP (Leben 1973; Goldsmith 1976; McCarthy 1986), the “repeated
morph constraint” (Menn and MacWhinney 1984), ANTIHOMOPHONY
(Golston 1995), *REPEAT (Yip [...]), and IDAVOID (Brentari 1998). The effects
of this principle not only can be observed in autosegmental phonology and
feature geometry (Leben 1973; Goldsmith 1976; McCarthy 1986; Myers 1987;
Yip [1988]; Pierrehumbert 1993, among many others) but also can be found in
morphology (Stemberger 1981; Menn and MacWhinney 1984; Mohanan 1994;
Golston 1995; Yip 1995, 1998; Brentari 1998, among many others).
Within phonology, SPA effects of the type described in this article have long
been observed within Semitic languages and continue to be the subject of study,
especially as concerns Arabic (Frisch et al. 2004) and Hebrew (Berent & Shimron
2003). While SPA effects have been presented as static distributional tendencies,
the diachronic process of consonant dissimilation has received considerable
attention for over a century, mostly notably in Grammont (1895). The
question we would like to raise in this section is whether the SPA effects we
have reported in tabular form are one of the manifestations of the OCP. As attractive
as this may seem, our approach and findings differ from some of the
above work in at least three ways.
First, in spite of a few exceptions (Yip 1995, Frisch et al. 1996, MacEachern
1999, Berent & Shimron 2003, Coetzee & Pater 2006), the universal OCP
mostly concerns identical elements. In our case, we have dealt not only with
restrictions on combinations of identical, but also similar elements: we have
been concerned with restricted sequences of consonants made in a same or similar
place of articulation. We have also shown that the restrictions hold not only
of exact homorganic consonants, but also of consonants that belong to the same
“superclass”. In this connection, we have seen the SPA at work within both the
Brought to you by | University of Iowa Libraries
Authenticated
Download Date | 6/3/15 5:25 AM
330 Konstantin Pozdniakov and Guillaume Segerer
medial and peripheral superclasses, although the tendency has a greater effect
among peripheral consonants.
Second, the results obtained for Semitic exploit the fact that in these languages
consonantal roots have a concrete reality which can be observed in their
templatic morphology. Their isolability makes the calculations relatively easy.
Our measurements have involved languages where the notion of “consonantal
root” is rarely justified. We have even evaluated corpora with no preliminary
knowledge of the morphological structures of the languages in question. Despite
this, the SPA effects were still evident.
Third, whereas recent discussions of SPA, especially those based on Semitic,
seek above all to discover the nature of this phenomenon in synchrony,10 for
us the phenomenon has great diachronic consequences for the comparative
method. If consonant combinations have to satisfy a principle of equilibrium,
the phonetic changes that affect consonants will presumably be in part conditioned
by SPA. Thus, alongside the two major sources of language change,
regular sound change and analogical change, a third factor must be at play
which we can call “dissimilation consonantique”. Within this category, we can
classify the examples given by Greenberg (1968: 107–108) such as Latin arbor
> Spanish arbol or Latin anima > *anma > Spanish alma. Here we have
neither a regular sound change (*r > l or *n > l) nor analogical change, but
rather “sporadic” dissimilations: of two r’s in the first case, of two nasals in
the second. Greenberg formulates this tendency only for sonorants (liquids and
nasals) and s(h)ibilants, thus for specific manners of articulation. We have tried
to show the importance of similarity avoidance for place of articulation.
Since SPA is general in language, we should expect to find processes that affect
C1VC2 sequences where C1 and C2 belong to the same class or superclass
with respect to place. Logical possibilities include one consonant dissimilation
from the other in place, or dropping out under identity. Another possibility is
that lexical items that repeat the same place of articulation may be disfavored
and drop out – or may not have been formed in the first place. On the other
side of the equation, when we find a language which violates SPA in an unexpected
way, we might conclude that the language in question has undergone
a specific diachronic change to produce the unusual situation. For example,
in Bantu we observed an excess of C-C (palatal-palatal) combinations. Since
Bantu diverges significantly from the norm in this respect, it behooves the Bantuist
scholar to seek a diachronic explanation. From a wider comparative point
of view, this unique distribution leads one to hypothesize that the homorganic
C-C sequences of Proto-Bantu must correspond to other consonant combinations
in other groups of Niger-Congo.
10. Cf. Frisch et al. (2004), especially Section 4.1 “The Psychological Reality of OCP-Place”.
Brought to you by | University of Iowa Libraries
Authenticated
Download Date | 6/3/15 5:25 AM
Similar Place Avoidance: A statistical universal 331
The Bantu example might give the impression that only such an “anomaly”
can be exploited for comparative purposes. However, although all of the languages
presented here show the same general SPA tendency, they all differ in
their statistical details, which may therefore furnish precious indices for com-
parison.
As we indicated in Section 6, there may also be an important interest in
closely studying the possible relationship between SPA and the inverse tendency
of vowel harmony. Vowel harmony is of course a quite different phenomenon
operating synchronically and productively, contrary to consonantal
incompatibility. As we hypothesized in the case of Classical Mongolian (Table
22), the tendency for consonant place to be dissimilar may be more accentuated
in languages with front-back vowel harmony. If the distribution of pluses
in the inverse diagonal of Table 22 is not fortuitous or isolated, this could mean
that there is a tendency for an equilibrium to be established between the two
subsystems (vowels and consonants) within the phonological structure of the
word.
There is one limitation on SPA that we have not yet addressed. Contrasting
with the very clear tendency to avoid successive consonants of similar place,
the Atlantic languages shows an inverse tendency which at first appears contradictory:
in Atlantic, as in other languages of the world, identical consonants
combine easily, especially in certain lexical subclasses, e.g., ideophones, intensifier
adverbs, and other iconic words. One of the possible sources of combinations
of identical consonants is reduplication, which is often associated with
the expression of intensity. However, the statistical results presented here in
support of SPA are robust despite these well-known cases of identical consonant
combination. It is not impossible that these two opposing tendencies are
interrelated.
To test the effect of identical C1-C2 consonants on our results, we did a more
detailed statistical count on Wolof which distinguished between sequences of
identical vs. non-identical homorganic consonants. This further study yielded
the following results:
(i) If one takes into account the relative frequencies of the different consonants
in the dictionary, there is no general tendency to combine identical
consonants, with the exception of two special cases: (a) combinations
of identical nasal or prenasalized consonants, especially mb-mb, nd-nd,
m-m, and n-n, are considerably more frequent than expected; (b) the sequences
f-f and c-c are particularly frequent.
(ii) Besides those combinations whose frequency exceeds or corresponds to
the norm, two combinations of identical consonants have a frequency below
the norm: r-r and t-t.
(iii) An analysis of lists of words which present sequences of identical consonants
shows that these words are not generally formed by lexical reduBrought
to you by | University of Iowa Libraries
Authenticated
Download Date | 6/3/15 5:25 AM
332 Konstantin Pozdniakov and Guillaume Segerer
plication. They are, however, often formed by a sort of “grammatical”
reduplication. In Wolof, for each noun class there is a series of determiners
of the form CooCV where C is the consonant of the noun class and
V is a “deictic” vowel (i for ‘near’, a for ‘far’, u for ‘unmarked’). Other
forms exist which are accompanied by a particle with an emphatic value
(i or le), which increases the number of words containing a sequence of
two identical consonants. For example, for the “M class”, we find:
(4) a. muus mi ‘this cat’
b. muus moomu ‘this cat (in question)’
c. muus moomule ‘idem (emphatic)’
d. muus mooma ‘that cat (of which you had spoken and which is
not present)’
e. muus moomale ‘idem (emphatic)’
f. muus moomee ‘idem (emphatic)’ (< *mooma + i)
g. muus moomii ‘this cat here (of which you had spoken,
emphatic)’
Besides the above, the consonant n is repeated in the lexical base meaning
‘other’, e.g., m-eneen in the M class. The presence of these multiple series
increases the frequency of combinations of identical consonants.
In the above examples, the repetition of identical consonants is associated
with the morphology of the language. They are thus exempt from any possible
phonetic or phonological motivation for SPA. If we remove all of the words
that have a sequence of identical consonants from the wordlists, the tendency
for successive consonants to be of different place is of course enhanced. This
is seen in Table 25, which compares the percentages of each combination with
vs. without identical C1=C2 sequences.
Table 25. Wolof percentages with vs. without C1=C2 sequences
With P K T C
P −54 −17 +23 +17
K −9 −69 +27 +15
T +27 +31 −22 −4
C +26 +30 −14 −26
Without P K T C
P −74 −15 +28 +23
K −6 −83 +30 +20
T +35 +37 −31 −4
C +33 +33 −10 −49
Instead of considering the combinations of identical consonants with respect
to the whole dictionary, we shall limit ourselves to the expected number of
C1=C2 words compared to the total set of consonants made at the same place
of articulation. For example, the list of Wolof words which contain a TVT sequence
(where T=any dental/alveolar) provides 988 sequences. The frequency
Brought to you by | University of Iowa Libraries
Authenticated
Download Date | 6/3/15 5:25 AM
Similar Place Avoidance: A statistical universal 333
of C1 l is 16 % among the TVT sequences. Its frequency in C2 position is 28 %.
In the absence of mutual influence, we should find 988×16%×28% = 45 occurrences
of lVl sequences in the Wolof dictionary. It turns out that we find 47.
We can therefore conclude that lVl either escapes the effects of SPA, or that the
effect of SPA is canceled out by the inverse tendency to favor identical consonants,
in this case l’s. Let us now compare nVn sequences. The frequence of C1
n is 14 % (among TVT sequences). Its frequency in C2 position is 18 %. We
should therefore find 988×14%×18%= 25 occurrences of the sequence nVn.
In this case we find 41, which represents a deviation of 65 % with respect to
the norm. This discrepancy is sufficiently large enough to suggest that the sequence
nVn is relatively privileged among the possible combinations of dental
consonants (TVT).
We have chosen these specific illustrations because it is among dental consonants
that we find the only negative discrepancies for identical consonants.
Within the labial, palatal, and velar series combinations of identical consonants
are systematically favored. For the dental place, the results are mixed, as seen
in Table 26.
Table 26. Wolof percentages for dentals, where C1=C2
ndVnd nVn dVd lVl tVt rVr
+132 % +65 % +41 % +4 % −35 % −64 %
To summarize, two inverse tendencies are at work in Wolof. In a general way,
combinations of homorganic consonants are avoided (SPA). However, in the
midst of such homorganic consonants, combinations of identical consonants
are statistically favored, sometimes reflecting the fact that these sequences are
charged with a grammatical function. This second tendency confirms that it is
the place of articulation which constitutes the relevant context for statistical
biases in the combination of consonants.
But there is something more: we have pointed out several times that our
statistical counts show strong tendencies, even with a poor knowledge of the
phonology of the language studied. Languages often have additional features
which, if taken into consideration, might make the counts more accurate or
revealing. Thus, for every language, good knowledge of the phonology could
help find some otherwise hidden tendencies. Let us illustrate this with Wolof
again. In Wolof, there is a vowel length contrast that has not been taken into
account for the tables presented here (cf. Tables 9 and 25). It is interesting,
however, to calculate separate tables for short and long vowels respectively.
The result is shown in Table 27.
Brought to you by | University of Iowa Libraries
Authenticated
Download Date | 6/3/15 5:25 AM
334 Konstantin Pozdniakov and Guillaume Segerer
Table 27. Wolof percentages with C1VC2 vs. C1VVC2 sequences; the -CVC- and
-CVVC- counts are based on 5,991 CVC sequences and 2,104 CVVC sequences, re-
spectively.
C1VC2 P K T C
P − − − + + +
K − − + +
T + + + −
C + + + − − −
C1VVC2 P K T C
P
K − +
T +
C +
There is an enormous difference between these two tables. For CVVC sequences,
there is no trace of SPA! Only four cells show a slight deviation from
the expected frequency, but there is nothing systematic here. So, it seems that
the “phonological distance” between two consonants is too big for the SPA
effect to appear. This kind of analysis has not been made for other languages
with a vowel length contrast. Therefore, we cannot consider this phenomenon
as a universal. But it might well be, for the theoretical explanation involving
“phonological distance” seems reasonable enough. In addition to the lack of
SPA in the CVVC table above, we observe a reinforcement of SPA in the CVC
table, which is quite logical, given that SPA was sensible even on the global
corpus.
7.2. The problem of reconstructed languages
As pointed out by an anonymous reviewer, the use of reconstructed lexical
forms could introduce some bias in the tables. There are several reasons that
make us think that we still can use these. First, the nature of the lexicon is different:
in a reconstructed one, there are usually no borrowings or ideophones,
and it is precisely those items that can blur the observed tendencies by having
irregular phonological shapes. So the observed tendencies can only be stronger.
However, this is not always the case. For example, the Proto-Bantu lexicon as
elaborated by the Tervuren group contains all the reconstructed dialectal variants
of all zones of Bantu. Thus, the proto-lexicon has far more items than any
of the present-day Bantu languages surveyed for this study. Here we can expect
the statistical tendencies to be slightly different, and that is the reason why
we included not only the Proto-Bantu table, but also four present-day Bantu
languages.
Second, concerning diachronic aspects of the problem, it is important to determine
if the SPA phenomenon results from historical processes of dissimilation
or if its effects are purely synchronic. If we compare the measurements
made of the lexicons of proto-languages and their descendent living languages,
we find that the differences are of exactly the same order as those observed
Brought to you by | University of Iowa Libraries
Authenticated
Download Date | 6/3/15 5:25 AM
Similar Place Avoidance: A statistical universal 335
between the individual living lexicons and their average. (Compare, for example,
Proto-Bantu in Table 14 with “Average Bantu” in Table 29 below.) This
means that the data from the proto-language better represent the family of the
descendent languages than any one of these languages taken by chance. Thus,
the Proto-Indo-European lexicon is more representative of Indo-European in
general than is, say, the Albanian lexicon. Consequently, in the absence of rigorously
established reconstructions, it is justified to use the average calculated
on the basis of the living languages of a family, as we do, for example, with the
Altantic languages in Section 8.
8. Similar Place Avoidance and the hierarchy of combinations
So far, we have shown that SPA is a reality for every individual language. But
a given language may well show some deviations with respect to SPA, especially
as far as superclasses are concerned. For example, a K-K combination
is avoided in all languages, but K-P is overrepresented in Quechua and Kamilaroi
and underrepresented in Basque (see Table 24). This might raise some
doubts about the existence of superclasses. A more general question is whether
there would be a kind of hierarchy with respect to the respective “rate of avoidance/affinity”
of each of the sixteen possible combinations. To address these
issues, we need a general overview of the values disseminated in the languageindividual
tables. This table can be obtained in the following way: in Table 28,
for the 31 languages examined, we have put the total number of the six possible
E/O values (“+”, “+ +”, etc.) for each of the sixteen combinations of P, K, T,
and C. For example, among the 31 tables presented above, the T-C combination
has never shown up as “+ +”, but is attested twice as “+”, ten times as “−”,
eigth times as “− −”, nine times empty (i.e., with an E/O discrepancy of less
than 15 %), and twice as non-significant (“n.s.”) for lack of sufficient T-C occurrences.
In the last column we have put the numeric value corresponding to
the sum of the plus and minus values in the preceding columns. This total represents
the E/O discrepancy proper to each combination: a positive value represents
an overrepresented combination of consonants, while a negative value
signals an underrepresented combination. To highlight the wide range obtained
in these values, the rows have been arranged with the greatest negative value
(−31) at the top and the greatest positive value (28) at the bottom.
Brought to you by | University of Iowa Libraries
Authenticated
Download Date | 6/3/15 5:25 AM
336 Konstantin Pozdniakov and Guillaume Segerer
Table 28. Summary of the preceding tables
+ + + − − − norm n.s. Total
K-K 5 26 -31
P-P 6 24 1 -30
T-T 17 6 8 -23
C-T 1 14 2 14 -15
T-C 2 9 7 11 2 -14
P-K 1 1 9 6 14 -13
C-C 2 1 6 8 10 4 -11
K-P 1 8 2 19 1 -9
Total 9 155 77 6
K-C 4 10 1 1 13 2 12
P-C 10 8 1 12 17
C-P 7 10 12 2 17
C-K 9 11 1 8 2 19
P-T 4 17 10 21
K-T 7 14 10 21
T-K 13 14 2 2 27
T-P 18 10 3 28
Total 166 4 70 8
Two facts are immediately visible:
First, the table is divided into two equal parts of eight positive and eight
negative rows each. This means that there are as many overrepresented combinations
as there are underrepresented ones. In addition, the zeros and “n.s.” are
also equally distributed within the upper and lower halves of the table.
Second, all of the combinations of consonants from the same class (represented
by the gray cells) are in the upper part of the table, indicating that these
combinations are globally underrepresented. This is the concrete trace of the
SPA phenomenon.
Another important phenomenon can also be noted: all of the combinations
within the same superclass (i.e., peripheral K/P vs. medial T/C) are also underrepresented.
Thus, the eight combinations which show a negative total are
exactly the combinations of consonants within these superclasses,
whether the consonants are homorganic or not.
Recall that this result is a global one. In an individual language, one or another
of these combinations can be overrepresented, as seen in five of the eight
rows in the upper half of Table 28. We also observe that the three remaining
rows which lack a positive value are all combinations of homorganic consonants
(K-K, P-P, and T-T).
Brought to you by | University of Iowa Libraries
Authenticated
Download Date | 6/3/15 5:25 AM
Similar Place Avoidance: A statistical universal 337
The consequence of the preceding observations is that the combinations with
a positive total, i.e., those which are globally overrepresented, are all combinations
of consonants belonging to different superclasses.
Among the 31 languages examined, seventeen belong to Niger-Congo:
twelve Atlantic languages and five Bantu languages, among which Proto-Bantu.
Afro-Asiatic is represented by two Chadic lexicons. All of the other languages
are the only representatives of their group. Thus, in order to avoid any bias
which might be due to the consideration of related languages, we have recalculated
the general table in such a way that each group of languages is represented
by a single language. For the Atlantic languages and Bantu, we have taken the
average of the observed individual values shown in the following tables.11
Table 29. Atlantic and Bantu average values
Average Atlantic
P K T C
P − − − + +
K − − − + +
T + + + + − −
C + + −
Average Bantu
P K T C
P − +
K − − +
T + + −
C + + − +
For the Chadic group, we have eliminated the relatively small Proto-Chadic
lexicon which shows too much internal variability.
The fifteen groups of languages now each having one set of values are
the following: Atlantic, Bantu, Mande, Kwa, Ubangi (Niger-Congo); SaraBongo-Bagirmi
(Nilo-Saharan); Chadic (Afro-Asiatic); Malagasy (Austronesian);
Indo-European;Nostratic; Mongolian (Altaic); Basque; Quechua; Kamilaroi
(Australian); and Port Moresby Pidgin English. With these changes, Table
30 represents the recomputed values.
11. The average is calculated by dividing the difference of the number of “+” and the number
of “−” by the number of languages. For example, the combination C-C in the five Bantu
languages presents the value “++” three times, the value “+” once, and the value “−” once.
The average is therefore: ((3×2)+1−1)
5 = 1.2, which we round to 1. For the combination C-C,
“Average Bantu” will thus have the value “+”.
Brought to you by | University of Iowa Libraries
Authenticated
Download Date | 6/3/15 5:25 AM
338 Konstantin Pozdniakov and Guillaume Segerer
Table 30. Summary table, recomputed
+ + + − − − norm n.s. Total
K-K 1 14 −15
P-P 1 14 −15
T-T 6 4 5 −10
C-C 1 3 4 3 4 −6
P-K 1 4 2 8 −5
T-C 2 5 2 4 2 −5
C-T 1 6 8 −5
K-P 1 3 1 9 1 −3
Total 6 70 37 6
P-C 4 2 1 8 5
K-C 2 3 8 2 5
C-K 5 4 1 3 2 8
C-P 2 6 5 2 8
K-T 1 7 7 8
P-T 3 8 4 11
T-P 6 6 1 2 12
T-K 10 3 2 13
Total 72 2 38 8
While Table 30 is comparable overall to Table 28, the tendencies are now
even more evident:
(i) This time, the four combinations of homorganic consonants are at the top
of the table, which signifies that they are the most underrepresented (the
totals go from −15 to −6). Among these four combinations, only one
positive value occurs, viz. C-C in “Average Bantu”.
(ii) Among the combinations forming the second part of the upper half of
the table (i.e., non-homorganic combinations from the same superclass),
there does not appear to be any neat hierarchy. One can just point out that
the combination K-P is the most “normal” of the four, the totals ranging
from −5 to −3. The peripheral combinations are favored twice with
“+ +” but only in Kamilaroi, with both K-P and P-K. The medial combinations
are also favored three times, but only with “+” (T-C in Basque
and both T-C and C-T in Kamilaroi).
(iii) The lower half of the table, which contains the eight combinations of
medial + peripheral consonants reveals a surprising internal structure: the
first four combinations are exactly and only those which contain C (totals
of +5 to +8). The lowest four combinations are exactly and only those
which contain T (totals from +8 to +13). Among all these combinations,
Brought to you by | University of Iowa Libraries
Authenticated
Download Date | 6/3/15 5:25 AM
Similar Place Avoidance: A statistical universal 339
only P-C (in Nostratic) and C-K (in Kamilaroi) present a single case each
of a negative value.
(iv) The two subgroups in the lower half of the table in turn present an identical
internal structure: combinations in which the medial consonant precedes
the peripheral consonant are preferred to the reverse. T-P and T-K
have a higher global value than P-T and P-K. Similarly, C-P and C-K
have a higher value than P-C and K-C. This preference is graphically
symbolized as follows:
T
K P
C
K P
Combinations of peripheral consonants (P, K) with dentals (T) are never underrepresented.
Even more striking, the combinations T-K and T-P are nearly
always overrepresented (thirteen of the fifteen linguistic groups examined).
Contrary to what we proposed earlier, this means that more is going on than a
simple compensation for the underrepresented restricted sequences. If this were
the case, we would not expect the “+” and “+ +” values to represent such a hierarchical
structure, rather that each of the “compensatory” combinations would
have approximately the same values in Table 30. We are therefore forced to
conclude that besides the rather spectacular restrictions concerning consonants
of the same (super-)class, certain cross-superclass combinations are “favored”
by languages. In other terms, beside “bad” words such as toad and bug, one
finds “good” words such as dog and cat. This conclusion, which goes beyond
the objectives of this article, merits a detailed study of its own. It does, however,
raise a further possibility: If there are good words and bad words, are
there also good and bad languages? In fact, the very nature of Table 30 allows
us to calculate the average values for each combination. By doing so, we obtain
what could be an “average” language in Table 31, one which conforms exactly
to the tendencies shown by all the languages as a whole. Obviously Table 31 is
not that of any of the languages studied.
Table 31. An average language
P K T K
P − − + +
K − − +
T + + + −
C + + −
Brought to you by | University of Iowa Libraries
Authenticated
Download Date | 6/3/15 5:25 AM
340 Konstantin Pozdniakov and Guillaume Segerer
Finally, Table 30 allows us to establish a hierarchized list of the constraints:
(i) pure SPA: adjacent identical classes are disfavored;
(ii) extended SPA: adjacent identical superclasses are disfavored;
(iii) combinations involving dentals (T) are preferred over combinations involving
palatals (C);
(iv) the order “medial > peripheral” is preferred over the order “peripheral >
medial”.
While it must be repeated that these constraints do not describe dynamic processes,
it is worth noting that they might be responsible for dynamic effects,
as when the Bantu language, Tiene, metathesizes CVP-VT and CVK-VT to
CVT-VP and CVT-VK (Hyman 2006).
9. Conclusions and hypotheses
In the course of this study we have reached the following conclusions.
First, the phenomenon of Similar Place Avoidance (SPA), previously described
for Semitic languages, seems to be a linguistic universal, being observed
in languages which are both genetically and geographically unrelated.
Second, since the effects of SPA are non-categorical and vary slightly from
language to language, SPA is best seen as a statistical tendency. This tendency
can be observed in spite of the following factors that may lower its
effects:
(i) reduplication. Reduplication processes are well documented and often
operate on morphological grounds. This leads to numerous sequences of identical
consonants separated by a vowel. Not only are these sequences not forbidden,
but they are rather favored in some grammatical categories (cf. Wolof
demonstratives above, but also ideophones, baby talk, and other words with
expressive or intensive meaning).
(ii) preference for identical vs. similar consonants. At the phonological
level, a close examination of similar-place consonant sequences shows that
SPA may not operate equally on all types of same-place consonant sequences.
For example, in Russian, the number of initial CVC sequences involving labial
consonants is inferior to what is expected, but there are important discrepancies
as for the possibilities of combinations within this subset: pVp- or bVbsequences
are relatively frequent (while their number is still inferior to what
is expected), but bVp- sequences, for which the two consonants are “dangerously”
similar, are strictly limited to a very small number of borrowings: baptist
(and a few derived forms), biplan, bipol’arnyj (‘baptist’, ‘biplane’, ‘bipolar’).
Furthermore, two of these show a visible prefix bi-.
(iii) Consequently, sequences containing a morphological boundary may be
more tolerant with respect to SPA, as illustrated with two of the three bVp- exBrought
to you by | University of Iowa Libraries
Authenticated
Download Date | 6/3/15 5:25 AM
Similar Place Avoidance: A statistical universal 341
amples in Russian above. This is even more evident for pVb- sequences where,
aside from two borrowings publika (and a few related forms) and pubertal’ny
(‘public’, ‘pubescent’), all the 147 forms among the 97,328 items of Zalizniak’s
(1977) dictionary involve the prefix po-, so that the sequences are actually
po-b-, b- being the real initial of the stem. As another example, Larry
Hyman (2006, personal communication) points out that in Chichewa, Ciyao,
and many other Bantu languages, the unproductive verb extensions -am- and
-at- almost fail to occur following CVP and CVT, respectively. On the other
hand, productive extensions such as -i(t)s- ‘causative’, -il- ‘applicative’, and
-an- ‘reciprocal’ show no such effects.
(iv) As illustrated with the bi-p- examples in Russian above, borrowings may
be less subject to SPA. This may be due to differences in morphological analysis
in source and target languages: if a word contains a morphologicalboundary
in its source language, the speakers of the language that has borrowed it have
no awareness of this boundary, and thus treat the word as a monomorphemic
one. Or, to put it in a different way, a word treated as monomorphemic in a
language may have a “disfavored” shape because its origin was bimorphemic
and therefore less sensitive to SPA.
(v) We are responsible for two additional sources of discrepancies. The first
one inevitably arises from our lack of competence for morphological segmentation.
For a number of languages, we simply took the data without doing any
segmentation at all. The second one is our arbitrary classification of phonemic
features into four classes, whereas some languages could require more contrasts,
e.g., labiovelar consonants, which have been included in the labial class
(Bijogo, etc.), and postvelar consonants, which have been placed in the velar
class (Wolof, etc.). Moreover, the particular status of some elements may be
different from one language to another. The most problematic case is undoubtedly
that of s (resp. z), which we have always included in the palatal class. In
languages where there is a contrast between s and sh (French, English, etc.), /s/
would better fit in the dental class. In some cases, we have computed wordlists
that we found on the internet. These files generally came with no information
about the orthographical conventions. For Quechua and Basque, we could
assume that the spelling was influenced by Spanish, but we had no such information
for Malagasy, Pidgin English, or Kamilaroi.
In spite of all these factors blurring the tendency, it is still present not only
in each individual language, but also as an average for all the languages.
Third, given the universality of SPA, it follows that any counter-tendency
in a language must be regarded as an anomaly, e.g., the overrepresentation of
sequences of palatal consonants in Bantu (Tables 14 and 15).
Fourth, while Frisch (1996) has hypothesized that constraints on consonant
sequences should be proportional to the number of shared phonological features
(with C1=C2 being a special case), additional counts not presented here
Brought to you by | University of Iowa Libraries
Authenticated
Download Date | 6/3/15 5:25 AM
342 Konstantin Pozdniakov and Guillaume Segerer
reveal that SPA is more sensitive to some feature classes than others. Our measurements
show that the dominant effect concerns place of articulation, not
manner, nasality, or state of the glottis – which may in fact tend to harmonize
(Hansson 2001).
Fifth, we have shown that SPA effects justify grouping the four places of
articulation into two superclasses: peripheral P, K (grave, non-coronal) vs.
medial T, C (acute, coronal). While affecting both superclasses, it appears that
SPA has a stronger effect on peripheral than on medial consonants.
Sixth, as suggested by the Classical Mongolian data, it is possible that an
elevated level of SPA effects may be compensated by processes of vowel assimilation,
especially by back and round vowel harmony. More such languages
need to be investigated, however, to test this potential interaction.
Seventh, in the course of our investigations we have noted that the statistical
biases attributable to SPA are even more robust if the counts are limited to
basic lexical items, i.e., a part of the lexicon that includes fewer derived words,
borrowings, and elements with an “expressive” value (e.g., ideophones). We
have used dictionaries containing as many as 25,000 entries, but also as few as
318 CVC sequences (in the Nostratic wordlist). Although one might a priori
tend to doubt results based on such a small number of items, SPA effects were
found in corpora of all sizes.
Eighth, when examined in detail, restrictions due to SPA reveal internal
hierarchies. By compiling all measurements of consonant co-occurrence restrictions
for our sample of fifteen genetic units (that is, languages or protolanguages
representative of their genetic family), we have found that this hierarchy
involves not only restriction constraints, but also preference ones. The
formers concern classes and superclasses and are ordered as follows:
(i) pure SPA: adjacent identical classes are prohibited;
(ii) extended SPA: adjacent identical superclasses are disfavored.
Within the four classes P, K, T, and C, there exists another hierarchy: peripheral
classes (P and K) tend to combine less than medial ones.
The preferences may be attributed a different status. In fact, the more we find
restrictions, the more we can expect “preferences” to be compensatory. Thus,
their distribution is expected to be arbitrary. While this is often the case for
individual languages, the distribution of preferences shows more consistence
when we consider the summary of all the data, as presented in Table 30. The
preferences concern combinations of different superclasses, as expected, and
are ordered as follows:
(iii) combinations involving dentals (T) are preferred over combinations involving
palatals (C);
(iv) the centrifugal order (medial > peripheral) is preferred over the centripetal
one (peripheral > medial).
So, not only is SPA worth studying, but we are convinced that the study of
Brought to you by | University of Iowa Libraries
Authenticated
Download Date | 6/3/15 5:25 AM
Similar Place Avoidance: A statistical universal 343
preferences, which we can label CPA (for “Centrifugal Place Asymmetry”),
will lead to many important discoveries.
These conclusions have relevance not only to synchronic phonology, but
also to comparative and historical linguistics. Grammont (1895) and Greenberg
(1968) have recognized consonant dissimilation as one of the three important
factors playing a role in phonetic change, alongside regular sound change
and analogical change. If the phenomenon of SPA is universal, and if language
imposes a certain phonetic contour within the limits of the word, questions
naturally arise as to how SPA effects come into being and are maintained
in the face of the different diachronic pressures to which the shapes
of words are subjected. Is SPA a statistical property of proto-language that has
survived with different nuances in all of the world’s languages? While cases
of palatalization and labialization are well-known, most processes of sound
change affect features other than place: spirantization/affrication, nasalization,
voicing/devoicing, aspiration/deaspiration, etc. However, there are changes
which affect place of articulation, sometimes limited to C2, as when final *m
and *p become n and t in the history of Chinese (Chen 1973). It is not hard
to imagine possible, but as far as we know unattested, SPA effects such as the
following:
(5) a. C2 *m > n, unless C1=dental
b. C2 *m > n only if C1=labial
c. C2 > Ø, if it is identical in place to C1
Such hypothetical changes, however interesting, appear to be a misapplication
of the statistics presented here, which should not be taken for what they are
not. SPA is not a law, but rather a universal tendency. There is no categorical
prohibition against words containing sequences of homorganicconsonants, and
hence no expectation that sound changes such as the above will ever take place.
More reasonable to us might be cases where certain words or combinations
of morphemes within words are avoided if they produce violations of SPA.
Nevertheless, the highly predictable nature of the tendency suggests that words
that violate SPA may be more susceptible to change than those which do not.
All such speculations can and, of course, should be tested against further data.
Received: 31 January 2006 LLACAN (CNRS, INALCO)
Revised: 19 December 2006
Correspondence address: LLACAN, 7, Rue Guy Môquet, Bât. C, 94801 Villejuif Cedex, France;
e-mail: (Pozdniakov) pozdniakov@vjf.cnrs.fr, (Segerer) segerer@vjf.cnrs.fr
Brought to you by | University of Iowa Libraries
Authenticated
Download Date | 6/3/15 5:25 AM
344 Konstantin Pozdniakov and Guillaume Segerer
Acknowledgements: The ideas developed here were first presented at the Workshop on ProtoNiger-Congo
held in Paris, October 11–16, 2004, which was organized by the Santa Fe Institute
and the CNRS-LLACAN in the context of the Evolution of Human Languages Project directed by
†Sergei Starostin. We would like to express our grateful feelings to Larry Hyman. Not only has
he translated this paper from French, but his suggestions, comments, and encouragement at every
stage of this paper have been especially helpful to us. Any possible errors or misinterpretations
remain entirely our responsiblity.
Appendix: Observed number of every consonantal combination for each
language examined
PP PK PT PC KP KK KT KC TP
*Bantu 557 812 1,635 205 614 470 1,543 215 999
*Chadic 23 65 205 66 69 16 183 52 133
*Gbaya 24 33 163 20 33 23 129 18 43
*Indo-European 58 145 586 117 159 71 455 83 339
*Ijo 25 38 131 13 19 14 71 7 62
*Mande 2 23 132 22 8 9 101 6 14
*Nostratic 2 14 47 9 25 12 78 34 18
Balanta 21 24 154 43 34 9 94 24 72
Basque 20 69 344 131 39 47 542 163 106
Bemba 655 664 1,638 392 424 247 1,004 320 906
Bijogo 24 78 249 64 48 23 167 31 133
Bullom 39 70 157 40 20 20 100 8 46
Hausa 92 145 279 226 249 166 563 287 293
Jaad 44 48 226 95 50 11 127 51 84
Joola Kwaatay 59 82 299 140 68 50 222 75 170
Kamilaroi 36 79 181 47 68 22 161 47 48
Kiga-Nkore 391 742 1,434 672 747 840 2,423 697 1,059
Kisi 164 213 436 140 98 119 268 59 252
Malagasy 82 133 349 88 69 48 182 44 175
Manjaku 119 111 515 156 73 73 238 85 300
Mongolian 153 2,108 4 1,332 1,233 3,399 11,245 2,525 1,92
Mpongwe 155 223 411 135 198 155 454 142 257
Nyun-Buy 94 220 478 183 179 68 703 141 349
Palor 47 97 222 100 84 57 311 106 169
Fula (clean) 33 7 109 30 23 9 73 34 65
Fula (raw) 58 24 267 48 34 20 186 49 132
Pidgin English 126 94 406 111 108 32 184 49 307
Quechua 74 181 402 631 160 118 353 737 141
Sara Kaba Na 71 105 384 139 208 172 491 262 188
Sua 21 18 74 33 9 9 37 7 34
Swahili 75 131 151 86 76 49 134 74 120
Wolof 178 326 1,171 397 283 94 938 306 716
Brought to you by | University of Iowa Libraries
Authenticated
Download Date | 6/3/15 5:25 AM
Similar Place Avoidance: A statistical universal 345
TK TT TC CP CK CT CC Total Table
*Bantu 1,164 1,457 69 724 821 803 338 12,426 14
*Chadic 99 123 48 85 47 82 10 1,306 18
*Gbaya 63 90 5 18 32 65 2 761 11
*Indo-European 308 414 145 50 45 92 18 3,085 20
*Ijo 39 35 6 14 15 17 3 509 12
*Mande 24 61 5 14 27 57 6 511 13
*Nostratic 23 13 7 7 7 19 3 318 21
Balanta 63 109 43 58 49 83 24 904 8; 9
Basque 237 550 324 69 84 335 80 3,14 24
Bemba 857 1,229 309 513 508 596 391 10,653 15
Bijogo 207 196 53 53 64 93 16 1,499 9
Bullom 101 86 12 18 37 61 12 827 9
Hausa 229 352 167 278 148 247 159 3,88 18
Jaad 56 112 35 77 45 109 30 1,2 9
Joola Kwaatay 164 193 113 132 139 185 92 2,183 9
Kamilaroi 38 110 40 15 11 71 6 980 24
Kiga-Nkore 1,657 2,189 103 899 922 1,348 894 17,944 15
Kisi 260 226 64 179 223 192 88 2,981 9
Malagasy 178 261 59 61 94 106 15 1,944 17
Manjaku 226 393 130 145 174 325 82 3,145 9
Mongolian 12,184 6,654 2,715 1,586 6,082 7,567 1,704 66,407 22
Mpongwe 306 262 79 176 258 215 80 3,506 15
Nyun-Buy 566 472 60 271 231 326 87 4,428 9
Palor 231 138 93 92 159 147 63 2,116 9
Fula (clean) 23 66 19 59 29 76 17 672 9; 19
Fula (raw) 54 364 48 108 46 190 23 1,651 19
Pidgin English 137 266 117 88 47 104 39 2,215 24
Quechua 283 236 417 214 379 259 669 5,254 24
Sara Kaba Na 278 296 177 101 177 186 65 3,3 16
Sua 37 71 18 25 32 52 18 495 9
Swahili 97 90 50 115 98 70 65 1,481 15
Wolof 722 1,025 454 445 445 734 222 8,456 9
References
Ancey, Jean-Luc (1997). Dictionnaire quechua–français. http://members.tripod.com/~jlancey/Peda/
Quecfran.htm
Austin, Peter & David Nathan (1998). Kamilaroi/Gamilaraay dictionary. http://coombs.anu.edu.
au/WWWVLPages/AborigPages/LANG/GAMDICT/GAMDICTF.HTM
Barhorst, Terry D. & Sylvia O’Dell-Barhorst (no date). Pidgin/English dictionary. http://www.
june29.com/HLP/lang/pidgin.html
Bendor-Samuel, John (ed.) (1989). The Niger-Congo Languages. London: University Press of
America.
Berent, Iris & Joseph Shimron (2003). Co-occurrence restrictions on identical consonants in the
Hebrew lexicon: Are they due to similarity? Journal of Linguistics 39: 31–55.
Brentari, Diane (1998). Comment on the paper by Yip. In Lapointe et al. (eds.) 1998, 247–258.
Buis, Pierre (1990). Essai sur la langue manjako de la zone de Bassarel. Bissau: Instituto Nacional
de Estudos e Pesquisas.
Caron, Bernard (1991). Le haoussa de l’Ader. Berlin: Reimer.
Brought to you by | University of Iowa Libraries
Authenticated
Download Date | 6/3/15 5:25 AM
346 Konstantin Pozdniakov and Guillaume Segerer
Chen, Matthew (1973). Cross-dialectal comparison: A case study and some theoretical considerations.
Journal of Chinese Linguistics 1: 38–63.
Childs, G. Tucker (2000). A Dictionary of the Kisi Language with an English-Kisi Index. Köln:
Köppe.
Chomsky, Noam & Morris Halle (1968). The Sound Pattern of English. New York: Harper & Row.
Clements, George N. & Elizabeth V. Hume (1995). The internal organization of speech sounds. In
John A. Goldsmith (ed.), The Handbook of Phonological Theory, 245–306. Oxford: Black-
well.
Coetzee, Andries & Joe Pater (2006). Lexically ranked OCP place constraints in Muna. Unpublished
manuscript, University of Michigan and University of Massachusetts, Amherst. Available
at http://roa.rutgers.edu/view.php3?id=1219
d’Alton, Paula (1987). Le palor: Esquisse phonologique et grammaticale d’une langue cangin du
Sénégal. Paris: Editions du CNRS.
Danay Kamis, Mando Makode, Ganda Nikubu, Maurice Tambyo, Namala Ngarassim, & Augustin
Goytisolo (1986). Dictionnaire sara-kaba-na–français, Kyabe (Tchad). Sarh: Centre d’Etudes
Linguistiques, Collège Charles-Lwanga.
Ducos, Gisèle (1971). Structure du badiaranké de Guinée et du Sénégal (phonologie, syntaxe).
Paris: SELAF.
Fal, Arame, Rosine Santos, & Jean Léonce Doneux (1990). Dictionnaire wolof-francais suivi d’un
index francais-wolof. Paris: Karthala.
Fleisch, Henri (1961). Traité de philologie arabe (vol. 1). Beyrouth: Imprimerie catholique.
Frisch, Stefan A. (1996). Similarity and frequency in phonology. Doctoral dissertation, Northwestern
University, Evanston, Ill.
Frisch, Stefan A., Janet Pierrehumbert, & Michael Broe (2004). Similarity avoidance and the OCP.
Natural Language and Linguistic Theory 22: 179–228.
Goldsmith, John (1976). Autosegmental phonology. Doctoral dissertation, Massachusetts Institute
of Technology.
Golston, Chris (1995). Syntax outranks phonology: Evidence from Ancient Greek. Phonology 12:
343–368.
Grammont, Maurice (1895). La dissimilation consonantique dans les langues indo-européennes et
dans les langues romanes. Dijon: Darantière.
Greenberg, Joseph H. (1950). The patterning of root morphemes in Semitic. Word 6: 161–182.
— (1968). Anthropological Linguistics: An Introduction. New York: Random House.
Guthrie, Malcolm (1967–1971). Comparative Bantu: An Introduction to the Comparative Linguistics
and Prehistory of the Bantu Languages. 4 volumes. London: Greggs.
Hansson, Gunnar (2001). Theoretical and typological issues in consonant harmony. Doctoral dissertation,
University of California at Berkeley.
Hyman, Larry M. (2006). Affixation by place of articulation: Rare and mysterious. Paper for
the proceedings of the Rara and Rarissima conference, Max-Planck-Institut für evolutionäre
Anthropologie, Leipzig, March 29–April 1, 2006.
Illiˇc-Svityˇc, Vladislav (1971–1984). Opyt sravnenija nostratiˇceskix jazykov. 3 volumes. Moskva:
Nauka.
Jakobson, Roman, C. Gunnar M. Fant, & Morris Halle (1952). Preliminaries to speech analysis:
The distinctive features and their correlates. (Technical Report, 13.) Cambridge, Mass.:
Acoustics Laboratory, Massachusetts Institute of Technology.
Jungraithmayr, Herrmann & Dymitr Ibriszimow (1994). Chadic Lexical Roots. 2 volumes. Berlin:
Reimer.
Kawahara, Shigeto, Hajime Ono, & Kiyoshi Sudo (2005). Consonant co-occurrence restrictions in
Yamato Japanese. In Timothy Vance (ed.), Japanese/Korean Linguistics, Volume 14, 27–38.
Stanford, Cal.: CSLI.
Labatut, Roger (1994). Initiation au peul. Paris: INALCO.
Lacroix, Jean-Michel (2001). Lexique des termes sakalava. http://www.zomare.com/lts_ab.html
Brought to you by | University of Iowa Libraries
Authenticated
Download Date | 6/3/15 5:25 AM
Similar Place Avoidance: A statistical universal 347
Lainé, Bruno (2004). Dictionnaire mongol-français. http://membres.lycos.fr/brunogml/sub/corps.
htm
Lapointe, Steven, Diane K. Brentari, & Patrick M. Farrell (eds.) (1998). Morphology and its Relation
to Syntax and Phonology. Stanford, Cal.: CSLI.
Leben, William R. (1973). Suprasegmental phonology. Doctoral dissertation, Massachusetts Institute
of Technology.
Lespinay, Charles de (1991). Langue et parlers baynunk: Lexique comparatif. Compte-rendu
d’enquêtes et synthèse de lexiques anciens (17e/18e s. – 1988). 2nd edition. Paris: Centre
de Recherches Africaines.
MacEachern, Margaret R. (1999) Laryngeal Cooccurrence Restrictions. New York: Garland.
Mann, Michael (1995). Bemba–English dictionary. http://www.cbold.ddl.ish-lyon.cnrs.fr/
McCarthy, John J. (1986). OCP effects: Gemination and antigemination. Linguistic Inquiry 17:
207–263.
Menn, Lise & Brian MacWhinney (1984). The repeated morph constraint: Toward an explanation.
Language 60: 519–541.
Mohanan, Tara W. (1994). Case OCP: A constraint on word order in Hindi. In Miriam Butt, Tracy
Holloway King, & Gillian Ramchand (eds.), Theoretical Perspectives on Word Order in South
Asian Languages, 185–216. Stanford, Cal.: CSLI.
Moñino, Yves (1995). Le Proto-Gbaya: Essai de linguistique comparative historique sur vingt-etune
langues d’Afrique centrale. Leuven: Peeters.
Mouguiama, Patrick Daouda (1994). Mpongwe word list. http://www.cbold.ddl.ish-lyon.cnrs.fr/
Myers, Scott (1987). Tone and the structure of words in Shona. Doctoral dissertation, University
of Massachusetts at Amherst.
N’Diaye-Corréard, Geneviève (1970). Etudes fca ou Balanta (dialecte ganja). Paris: SELAF.
Nyländer, Gustav Reinhardt (1814). Grammar and Vocabulary of the Bullom Language. London:
Church Missionary Society by Ellerton and Henderson.
Pater, Joe & Adam Werle (2001). Typology and variation in child consonant harmony. In Caroline
Féry, Antony Dubach Green, & Ruben van deVijver (eds.), Proceedings of HILP5, 119–139.
Potsdam: Universität Potsdam.
Payne, Stephen (1992). Une grammaire pratique avec phonologie et dictionnaire de kwatay (parler
du village de Diémbéring, Basse Casamance, Sénégal). (Cahiers de Recherche Linguistique,
1.) Dakar: Société Internationale de Linguistique.
Pierrehumbert, Janet (1993) Dissimilarity in the Arabic verbal roots. North Eastern Linguistics
Society 23: 367–381.
Pozdniakov, Konstantin (1991). Perspectives of comparative studies on the Mande and West Atlantic
language groups: An approach to the quantitative comparative linguistics. Mandenkan
22: 39–69.
Rose, Sharon & Rachel Walker (2004). A typology of consonant agreement as correspondence.
Language 80: 475–531.
Rugemalira, Josephat (1993). Nyambo word list. http://www.cbold.ddl.ish-lyon.cnrs.fr/
Sapir, J. David (1971). West Atlantic: An inventory of the languages, their noun class systems and
consonant alternation. In Thomas A. Sebeok (ed), Current Trends in Linguistics, Volume 7:
Linguistics in Sub-Saharan Africa, 45–98. The Hague: Mouton.
Segerer, Guillaume (1998). Lexique sua. Unpublished manuscript.
— (2002). La langue bijogo de Bubaque. (Afrique et Langage, 3.) Leuven: Peeters.
Starostin, Sergei (1998–2005). STARLING database. http://starling.rinet.ru/
Stemberger, Joseph P. (1981). Morphological haplology. Language 57: 791–817.
Tang, Sze-Wing (2000). Identity avoidance and constraint interaction: The case of Cantonese. Linguistics
38: 33–61.
Taylor, Charles (1959). Kiga–Nkore dictionary. http://www.cbold.ddl.ish-lyon.cnrs.fr/
Teil-Dautrey, Gisèle (to appear). Et si le proto-bantou était aussi une langue ... avec ses contraintes
et ses déséquilibres. Diachronica.
Brought to you by | University of Iowa Libraries
Authenticated
Download Date | 6/3/15 5:25 AM
348 Konstantin Pozdniakov and Guillaume Segerer
Tervuren Bantu Group (1998). Bantu lexical reconstruction 2. http://www.cbold.ddl.ish-lyon.cnrs.
fr/
Vydrine, Valentin (2004). Mande reconstructions. Unpublished manuscript.
Williamson, Kay (2004). Ijo reconstructions. Unpublished manuscript.
Williamson, Kay & Roger Blench (2000). Niger-Congo. In Bernd Heine & Derek Nurse (eds.),
African Languages: An Introduction, 1–42. Cambridge: Cambridge University Press.
Yip, Moira (1988) The Obligatory Contour Principle and phonological rules: A loss of identity.
Linguistic Inquiry 19: 65–100.
— (1995) Repetition and its avoidance: The case of Javanese. In Keiichiro Suzuki & Dirk Elzinga
(eds.), Proceedings of 1995 South Western Workshop on Optimality Theory, 238–262. Tucson:
Coyote Papers, University of Arizona.
— (1998). Identity avoidance in phonology and morphology. In Lapointe et al. (eds.) 1998, 216–
246.
Zalizniak, Andrei A. (1977). Grammatiˇceskij slovar’ russkogo jazyka. [Grammatical Dictionary of
the Russian Language.] Moskva: Russkij Jazyk.
Brought to you by | University of Iowa Libraries
Authenticated
Download Date | 6/3/15 5:25 AM