Patterns of segmental modification in consonant inventories: a cross-linguistic study *.
It is well known that secondary articulation types, such as labialization and palatalization, as well as laryngeal modifications, such as aspiration and glottalization, can play a contrastive role in segment inventories. This article describes an investigation of such segmental modifications, whether laryngeal supralaryngeal or nasal, on the basis of detailed analyses of each of the 317 languages in the cross-linguistic database presented in Maddieson (1984). In the first part of this investigation, the hypothesis is tested that segmental modifications occur on natural classes of consonants rather than randomly or on isolated segments. A formalism is proposed to express this regularity. Recent theoretical proposals concerning the representation of complex segments, in particular affricates, lead to predictions regarding the patterning of segmental modification on such segments; in the second part of our investigation, these predictions are tested against the cross-linguistic data on segmental modification on affricates in the same sample of languages.
In terms of complexity, segments can be classified into three types: simple segments (consonants and vowels), complex segments such as affricates (e.g. /ts, tf/, i.e. "contour" segments, in the sense of Sagey 1986), and segments, whether simple or complex, that show modification like secondary articulation, such as /[p.sup.j]/ or /[t[integral].sup.w]/ respectively.
Secondary articulation is usually defined as the imposition of a vowel-like articulation on a consonant, including labialization, palatalization, velarization, etc. (see Layer 1994: [subsection] 11.4 6, Clements and Hume 1995: [section] 3.5, among others). The latter authors state that "the four most commonly-occurring types are labialization, typically realized as the addition of lip-rounding to the primary articulation; palatalization, typically involving the raising and fronting of the tongue body in the direction of the hard palate; velarization, typically realized as tongue backing; and pharyngealization, involving the retraction of the tongue root" (1995: 284-285, cf. 291-292). A similar definition is given by Ladefoged and Maddieson (1995), who observe that "the standard phonetic definition of a secondary articulation is that it is an articulation of a lesser degree of stricture accompanying a primary articulation of a higher degree", but who themselves "suggest that, from a phonetic point of view, secondary articulations are always approximant-like in nature" (1995: 354; cf. Laver 1994: 321).
Laryngeal modifications like aspiration, glottalization, etc., as well as (pre)nasalization can be studied in the same way, and we will refer to laryngeal, supralaryngeal (i.e. oral, secondary articulation) or nasal modifications together as segmental modifications (henceforth MODs). In section 3 we will show that the hypotheses to be presented directly below also seem to be relevant for, for instance, voice, retroflexion and length.
In this article we discuss two hypotheses regarding the patterning of types of modification. We investigate, first, to what extent segmental modification occurs on natural classes of segments, rather than on individual segments. This idea is not novel, it can also be found (for secondary articulation) in Trubetzkoy (1969: 138), for instance, where it is embedded in the Prague School notion of privative and equipollent oppositions. Obviously symmetry in segmental inventories leads to descriptive economy. In our study we explore the same idea on the conceptual basis of contemporary theories of phonological representation and on the empirical basis of an in-depth analysis of the consonant inventories of the languages in Maddieson's UPSID (the UCLA Phonological Segment Inventory Database) sample (Maddieson 1984). We tentatively explore a possible formalization of this insight and investigate some of the implications of this approach. We then investigate a second hypothesis, concerned with the position that affricates occupy in phonemic inventories, regarding the question to which natural class(es) these segments belong as far as their pattern of segment modification is concerned. Both questions will be discussed in more detail below.
The first hypothesis that we investigate is given in (1):
(1) Hypothesis 1 In segmental inventories types of segmental modification occur on a natural class of segments, rather than randomly or on isolated segments.
For instance, aspiration is often distinctive on all voiceless stops in a given system, and not just on a single stop. Labialization often occurs on all dorsal consonants. Both groups of segments, voiceless stops and dorsals, are straightforward examples of natural classes, which we will roughly define, following standard practice, as any group of segments that have a small number of distinctive features in common. The list of features that we adopt to describe both natural classes as well as segmental modification types is equivalent to that proposed by Clements and Hume (1995). For instance, the class of voiceless velar plosives can be represented by the set of features [-sonorant, -continuant, +dorsal, -voice] and the segmental modification of aspiration can be represented by [+spread glottis].
In sections 2-2.3 we test the hypothesis in (1) on the basis of consonant inventories of the 317 languages surveyed in Maddieson (1984), which is itself based on UPSID. We are fully aware of the fact that the descriptions provided by Maddieson (1984) are not always completely accurate (see the review in Basb[empty set]ll 1985, among others). This is partly a result of the fact that his survey represents an inevitable homogenization of descriptions, differing among themselves in phonetic detail and accuracy, of typologically very different languages. In addition, there may be misprints and copying errors, while inadequacies in Maddieson's sources may have been improved in later analyses. However, because of the large body of relevant observations, the results we report are probably statistically robust enough to test our hypotheses, even if there is a small margin of error in the input data. In a small sample of the languages mentioned in this article, namely, Nootka (Sapir and Swadesh 1955) and Yurak (Decsy 1966), we found no significant mismatches between Maddieson's account and the original grammars that he cites from (see also notes 5 and 6). Accordingly, we will not consider questions regarding Maddieson's representation of the segment inventories of the languages in his sample, such as, for example, phonemicity and the distribution of the segments and the segmental modifications in specific contexts, or their labelling, although in some cases there might be reasons to do so. For instance, in Japanese, palatalization may occur in the Mimetic stratum of the language, resulting in contrasts like /p/ vs. /[p.sup.j]/ and /k/ vs. /[k.sup.j]/. This case of palatalization is not mentioned by Maddieson. Similarly, the nonpalatalized sounds of Russian are velarized, but Maddieson lists only three velarized sounds for this language. In this article, we will take Maddieson's data at face value, although we realize the fact that the phonetic reality may in some cases be more complicated. We are fairly confident that an even more careful examination of the data than Maddieson's would not lead to dramatically different results than the ones found here. In section 2 we will first illustrate hypothesis 1, according to which segmental modification typically occurs on a natural range of segments, and test this hypothesis for the languages in Maddieson's sample, using quantitative and statistical techniques (sections 2.1-2.3).
In section 3 we will formalize the notion. To that end, we propose to capture regularities in the patterning of segmental modifications by way of "sympathetic" feature cooccurrence constraints (Pulleyblank 1998) in optimality theory (Prince and Smolensky 1993), such as the following, which would be suitable to describe the examples mentioned above:
(2) a. VOICELESSSTOPS/ASPIRATION Voiceless stops are or can be aspirated.
b. DORSALS/LABIALIZATION Dorsals are or can be labialized.
The first constraint expresses the fact that voiceless stops (may) have the segmental modification type stated behind the slash, that is, this natural class of segments has the potential of bearing aspiration. The second constraint expresses the fact that dorsals are or may be labialized. Constraints of this type offer a formal way of expressing regularities such as those noted above.
We will be concerned both with distinctive modification types, such as in a language in which aspirated and plain voiceless stops contrast, as well as "redundant" ones, such as in a language in which all voiceless stops are aspirated and nonaspirated ones do not occur. Both are equally relevant to the question whether segmental modification types appear on natural classes or not, that is, to our hypothesis 1. In sections 3 and 6 of this article we will briefly sketch out two possible approaches to this issue.(1)
Following Pulleyblank (1998), we assume that sympathetic feature cooccurrence constraints (whether contrastive or not; cf. section 3 below) are grounded in articulatory and acoustic phonetic facts. We believe that our study can contribute to the set of such constraints uncovered by Pulleyblank, but we will not attempt to give a fully fledged catalogue here. We will henceforth refer to sympathetic feature cooccurrence constraints that bear on segmental modification as MOD constraints. These constraints confine the ability of bearing distinctive MOD to particular segment types.
Such a constraint approach to segmental modification adds, we claim, to our understanding of the structure of segmental inventories. In addition, it may provide evidence about the internal structure of complex segments, since these are predicted to pattern with other segments in such an approach. Especially the representation of affricates has received some attention in the recent literature. It has been claimed that affricates behave as stops in phonological processes (e.g. Padgett 1991; Steriade 1993, among many others); at the same time it has been pointed out that affricates typically have the same place of articulation as fricatives in particular segmental inventories (van de Weijer 1996). Our investigation of patterns of segmental modification provides a novel perspective on this matter. Specifically, we will relate our approach to the structure of affricates. In sections 4-4.2, again on the basis of Maddieson's survey, we will test the following hypothesis:
(3) Hypothesis 2
Affricates are part of two natural classes with respect to the types of segmental modification they take: they may pattern with stops and/or they may pattern with fricatives.
The remainder of this article is organized as follows: in section 5 we will take a closer look at the relationship between the two hypotheses. In section 6 we draw some general conclusions, tentatively discuss how patterns of segmental modification and gaps in the patterns might come about, and briefly introduce several issues for further research.
2. Segmental modification in consonant inventories: hypothesis 1
We investigated the claim that types of MOD occur on a set of similar segments, rather than on individual, isolated segments. In (4) a list appears of the types of MOD that was adopted, supralaryngeal (4a) as well laryngeal (4b). We follow Maddieson (1984: 37) for the supralaryngeal types. There are no examples of consonants with uvularization or glottalization in his survey (though cf. note 3). We will also consider nasalization and prenasalization as MOD types (4c), in that we assume that the kind of patterns in which (pre)nasalization is involved can be studied in a way similar to the patterns of palatalization, etc. Iverson and Salmons (1996) analyze prenasalization in Mixtec as the phonetic expression of underlying voicing; to the extent that Mixtec does not have any plain voiced stops, it constitutes an example of a language in which a specific MOD type is not contrastive. Consonant nasalization did not occur in Maddieson's sample (but see section 2.1 below and note 4).
In (4) the different types of segmental modification are followed by the abbreviations we will use in the data matrix (see the Appendix for a specimen) and in the tables with results, as well as by some examples to illustrate the diacritics Maddieson uses to indicate them in his survey.
(4) a. secondary articulation (oral) labialization lab /[t.sup.w] [p.sup.w]/ palatalization pal /[t.sup.j] [p.sup.j]/ velarization vel /t p/ uvularization n.a. pharyngealization pha /[t.sup.[?]] [p.sup.[?]] b. phonation types (laryngeal) aspiration as[p.sup.2] /[t.sup.h] [p.sup.h]/ preaspiration pra /[.sup.h]p [.sup.h]t/ breathy voice brv /d b/ with breathy release brr /[t.sup.h] [p.sup.h] laryngealization lar /t p/ ejectivity eje /t p/ glottalization n.[a.sup.3] c. nasality nasalization n.[a.sup.4] prenasalization prn /[.sup.n]d [.sup.m]b/
Of course, the list in (4) is arbitrary to the extent that the selection of MOD types in the corpus on which Maddieson (1984) is based is arbitrary, so that a slightly different list might be found if another sample of languages were investigated. After looking into the cross-linguistic patterning of each single MOD type, we will aggregate the results for the three different MOD types in (4) in our survey below, but we will also point out interesting patterns of divergence between the three groups.
In this section, we will illustrate the idea underlying hypothesis 1 with data from four languages. In section 2.1 we will explain how this hypothesis was tested for every language in Maddieson's sample. The results will be analyzed statistically. On the basis of the outcomes of the statistical analyses we will first discuss some overall patterns in the data (section 2.2) and then we will test the hypothesis (section 2.3).
In (5) to (8) below, we present parts of the consonantal inventories of Nootka (language 730 in Maddieson's sample, a member of the Wakashan family, Northern Amerindian),(5) Kharia (301, Munda, Austro-Asiatic), Yurak (056, Samoyed, Ural-Altaic) and Nambakaengo (626, Central Melanesian, Indo-Pacific), taken from Maddieson (1984). These partial inventories serve to illustrate some of the ranges of segments with MOD; the numbers in parentheses refer to the numbers of the inventories in Maddieson. In each case an informal interpretation is provided that expresses the generalization in which, in our view, the particular classes of consonants are involved.
(5) Nootka (730) voiceless plosive p t k [k.sup.w] q [q.sup.w] vl. ejective stop p' t' k' [k.sup.w] q [q.sup.w] vl. nonsib, fric. x [x.sup.w] [chi] [[chip].sup.w]
(In Nootka voiceless stops may be distinctively ejective, and dorsals may bear distinctive secondary labialization.)
A language like Nootka thus confirms our hypothesis 1, for both types of MOD, ejectivity and labialization.
(6) Kharia (301) voiceless plosive p t t k vl. aspirated plosive [p.sup.h] [t.sup.h] [t.sup.h] [k.sup.h] voiced plosive b d d g breathy voiced plosive b d d g
(In Kharia, voiceless plosives may have distinctive secondary aspiration, and voiced stops may have distinctive breathy voicing.)
(7) Yurak (056) voiceless plosive p [p.sup.j] t [t.sup.j] voiced plosive b [b.sup.j] vl. sibilant affricate ts t[s.sup.j] vd. nonsibilant fricative [??] [[??].sup.j] vl. sibilant fricative s [s.sup.j] voiced nasal m [m.sup.j] n [n.sup.j] voiced trill r [r.sup.j] vd. lateral approximant l [l.sup.j]
(In Yurak, labials and coronals may bear distinctive palatalization.) (6)
(8) Nambakaengo (626) voices plos. p [p.sup.w] [p.sup.j] t [t.sup.w] [t.sup.j] k [k.sup.w] [k.sup.j] voices prenas. Plos. [sup.m]b [sup.m][b.sup.w] [sup.n]d [sup.n][d.sup.w] [sup.[eta]]g [sup.[eta]][g.sup.w] nasal m [m.sup.w] n [n.sup.w] [n.sup.j] [eta] [[eta].sup.w] (In Nambakaengo, voiceless stops, prenasalized stops, and nasals may be distinctively labialized, while all voiceless stops, and also the coronal nasal, may be distinctively palatalized.)
With the exception of Nambakaengo palatalization on the coronal nasal, these consonantal inventories illustrate the idea that MOD types occur on natural classes of segments, rather than on individual, isolated segments. The patterns of MOD in these four languages therefore confirm hypothesis 1. To make a general claim, however, this hypothesis should be tested in a large, well-balanced sample of languages. The results of our evaluation of this hypothesis for all relevant individual languages in the UPSID sample were aggregated and analyzed quantitatively. Before presenting the outcomes of these analyses, we will make our procedure explicit.
2.1. Testing hypothesis i in a large sample of languages: method
For each of the 317 languages described in Maddieson (1984), we determined whether it has consonants with MOD and, if so, which type(s) of MOD and whether it occurs (they occur) on a natural class (or natural classes) of consonants. MOD on vowels does occur (e.g. nasalization, retroflexion and pharyngealization), but is much less frequent than on consonants; hence it was not investigated in this study.
Let us illustrate how we proceeded to determine whether a MOD type occurred on a natural class, and, if so, which natural class was involved. Take a hypothetical language with the consonants /p t k/ and the corresponding ejectives /t' k'/. Such a pattern could be scored in different ways: (i) as ejectivity occurring on the class of nonlabial voiceless stops. This was rejected because "nonlabials" do not form a commonly recognized class, especially given the fact that [labial] is a privative feature in the Clements and Hume (1995) framework; and (ii) as ejectivity occurring on two natural classes: coronal and dorsal voiceless stops. We also rejected this because it would involve natural classes of only a single member, an approach which in our view would miss generalizations by abandoning all economy of description; (iii) as ejectivity occurring on voiceless stops but with a gap for the labial. The latter type of approach was adopted, following, in our view, standard phonological practice.
It is important to note that counterevidence against hypothesis 1 in our sample of languages comes in different degrees. Instances of MOD occurring on only a single consonant (to be referred to as the "only on consonant x" type of case) were scored as "genuine" counterexamples to hypothesis 1; in that case the relevant instance of MOD for the language at issue obtained a "_" score. This does not occur very frequently (30 out of 281 cases of modification in the entire sample); an example is Vietnamese (303; Vietmuong, Austro-Asiatic). As can be seen in (9), in Vietnamese the phonemic contrast between aspiration and nonaspiration occurs on a single consonant only. (7)
(9) Vietnamese (303) voiceless plosive t c k ? vl. aspirated plosive [t.sup.h] vd. implosive [??] d vl. nonsibilant fric. f x h vd. nonsibilant fric. v v etc.
Languages with MOD on two or more consonants showed a number of patterns. If the consonants concerned formed an obvious natural class (such as "sonorants," "dorsals," "voiceless stops"), and there were no other consonants belonging to this natural class, the instance of MOD obtained a "+" score. If a natural class consisting of more than two consonants had one segment without the expected MOD type (to be referred to as the "not on consonant x" type of case), this was marked as a mild counterexample to hypothesis 1, scored as "[+ or -]". Consider Tolowa in (10). Tolowa (Athabaskan, Northern Amerindian) has a contrast between plain and ejective stops in all oral noncontinuants; this contrast does not exist in the labials, however.
(10) Tolowa (704) vl. plosive p t k [k.sup.w] vl. ejective stop t' k' [k.sup.w']
The difference in the regularity of the distribution of ejectivity in the oral stops series between Tolowa and Nootka, in (10) and (5) respectively, is minimal. The patterning of ejectivity in Tolowa was therefore scored as a "mild counterexample" to the first hypothesis.
Hopi (Uto-Aztecan family, Northern Amerindian) constitutes another case of MOD forming mild counterevidence to hypothesis 1:
(11) Hopi(738) vl. plosive k [k.sup.w] voiced nasal [eta] [[eta].sup.w] voiceless nasal [eta]
In Hopi, there is no labialized counterpart of the voiceless velar nasal.
Cases in which a specific type of MOD occurs on all relevant members of a natural class as well as on a single additional consonant which cannot be considered as a member of that class (to be referred to as the "also on consonant x" type of case), were also analyzed as mild counterexamples to hypothesis 1, and scored as "[+ or -]".
If two (or more) consonants existed without a MOD type that seemed to range over a natural class of which these two segments would normally be considered to be part, this was marked as a genuine counterexample ("-"). The patterning of aspiration in Wintu (California Penutian, Northern Amerindian), is given in (12).
(12) Wintu (709) vl. plosive p t k q vl. aspir, plos. [p.sup.h] [t.sup.h]
It appears that in Wintu voiceless stops can be aspirated, but there are "gaps" for both the velar and the uvular series.
Another type of genuine counterevidence against hypothesis 1 is exemplified by Hausa laryngealization; see the partial consonant inventory in (13).
(13) Hausa (266) vl. sibil, fric. s laryng, sibil, fric. s vd. central approx, j laryng, central approx. j
In Hausa (Chadic family, Afro-Asiatic) laryngealization is phonemically contrastive in two completely unrelated consonants, namely /s/ and/j/.(8) We consider this as a case in which a type of MOD occurs on an "unnatural" class of segments and consequently scored it as a real counterexample against hypothesis I. In the initial paragraphs of sections 3 and 4 as well as in the concluding section (section 6), we will develop the MOD constraint-based approach in such a way that it can deal with both the genuine and the mild types of counterevidence.
In the vast majority of cases, the procedure outlined above, though time-consuming, was simple enough. However, there is a small number of cases in which the patterning of MOD is not straightforward, at least not from the point of view of natural classes. We will very briefly discuss the most problematic cases here.
1. In Shilha (256, Berber, Afro-Asiatic) pharyngealization occurs on all nongeminate dental/alveolars--except for /n/--as well as on the nongeminate velar voiceless stop, but not on the other nongeminate voiceless stops. This combination of "not on" and "also on" types led us to consider the distribution of pharyngealization as a genuine counterexample to hypothesis 1;
2. In Washkuk (602, North New Guinea, Indo-Pacific), labialization occurs on both labial and velar consonants, except for the voiced labial fricative. Are we here dealing with two different natural classes? We decided to consider the relevant consonants as one natural class, namely the grave ones, encompassing all and only the acoustically similar labial and velar places of articulation (cf. Hyman 1975: 31-42 for discussion and references with regard to a feature [grave]);
3. In Otomi (716, Oto-Manguean, Northern Amerindian), aspiration occurs on all voiceless stops, but only on the plain ones, and not on the labialized voiceless velar stop (whether the latter segment is derived or not). Should we consider this as a "not on consonant x" case, or is the relevant natural class formed by the plain voiceless stops? We chose the former option, and scored a mild counterexample;
4. Also in Otomi, laryngealization occurs on voiced consonants--except for /1/ and /z/, as well as the voiced velar stop and the voiced alveolar flap. Consequently, we have scored this case of laryngealization as contradicting hypothesis 1. However, it does not seem wholly inconceivable to think of laryngealization in Otomi as taking place on two (or three) natural classes of consonants, namely the voiced central approximants (/j/ and /w/) as well as the voiced labial (/b/ and /m/) and coronal (/d/ and /n/) noncontinuants.
For each single case of MOD, our method forced us to decide whether or not the relevant group of consonants constitutes a natural class. For the largest part by far, the natural classes with MOD are constituted by stops, especially voiceless stops, and by stops at specific places of articulation. Obstruents also appeared as a class, and so did sonorants, and subsets thereof, such as nasals or laterals. Virtually none of these classes necessitate a deviation from generally accepted phonological models, except perhaps for the [grave] class mentioned above. A class which may need a few explanatory remarks is constituted by stops, nasals and laterals--in short by noncontinuants; they pattern in cases of MOD in a number of languages. Consider the examples in (14):
(14) Hindi-Urdu, Chuvash and Nyangumata (9) Hindi-Urdu (016) voiced plosive b d d g breathy vd. plosive b d d g voiced nasal m n breathy vd. nasal m n vd. lateral approx. l breathy vd. lateral appr. l Chuvash (060) voiceless plosive [t.sup.j] voiced nasal [n.sup.j] vd. lateral approx. [l.sup.j] yangumata (361) voiceless plosive [t.sup.j] voiced nasal [n.sup.j] vd. lateral approx. [l.sup.j]
These languages show that stops, nasals and laterals sometimes behave in identical ways, which should be reflected in feature systems.
A class which turns out to be particularly rare as far as MOD is concerned, is that of the "simple" continuants, that is, as a group per se, rather than as a subset of either the class of all obstruents or the class of all consonants. The rareness of this class has implications for testing the second hypothesis, which will be discussed in section 4.
To separate hypothesis 1 from hypothesis 2, MOD on affricates did not play a role in the question whether hypothesis 1 was upheld or not. Thus, for instance, in a language with aspiration on all voiceless stops, but not on affricates, the aspiration was considered to occur on a natural class (and scored as "+"). In such a case, the instance of aspiration in the language at issue would be assigned a "[+ or -]" score for the part of hypothesis 2 dealing with stops (see section 4.1).
Since glottal segments hardly participate at all in the patterning of suprasegmental segments,(10) these were systematically left out of consideration.
Two languages were excluded from the analyses because of ambiguities in their presentation.(11) This reduces the number of languages in our analyses to 315.
Maddieson's survey contains two click languages, Nama and !Xu (languages 913 and 918, respectively; both Khoisan, "other families"). Since it is not clear whether segments with ingressive air flow should be treated on a par with egressive segments, the clicks series of these two languages were not taken into consideration (see, however, section 4.1). The Nama clicks are the only segments represented in Maddieson's study in which nasalization occurs. The nonclick consonants of the two languages were included in our analyses.
All data resulting from our analyses were stored in a computer file, specifically in a data matrix legible for a (mainframe) version of SPSSx, a package of computational facilities for statistical analysis. A sample of the data matrix is presented in the Appendix.
2.2. Some general patterns
Before presenting those parts of the outcomes of the statistical analyses which are crucial to the evaluation of hypothesis 1, we should discuss some overall patterns in the data.
Of the 315 languages that were analyzed, 150 do not have any consonants with MOD. In the remaining 165 languages, there were 281 tokens of MOD. This means that over 52% of the languages studied have at least one type of MOD. There are languages with 2, 3, 4 and even 6 different types of MOD; there is none with 5, as can be seen in Table 1.
In general, in the subset of languages with MOD in consonants, the number of languages decreases as the number of different types of MOD increases. In 74 languages in the sample, two or more types of MOD cooccur; the question can be raised if MOD types randomly combine or, on the other hand, if there are any restrictions, that is, combinatorial preferences or prohibitions. An extreme degree of combination can be found in the language Twi (also known as Akan), where labialization and palatalization cooccur on obstruents, which are thus labio-palatalized (De Jong and Obeng 2000). Typological research might bring to light implicational relationships between MOD types (like, e.g., if a language has breathy voice in voiced stops, then it will also have aspiration on voiceless stops), especially if these can be formulated independently of natural classes (if a language has breathy voice, then it will also have aspiration). This, however, would probably require a larger sample of languages.
As shown in Table 2, the quantitative distribution of MOD over the 13 groups of language families represented in Maddieson's quota sample is not even. Whereas the Australian and Nilo-Saharan groups are relatively poor as far as MOD goes, the Austro-Asiatic and, and even more so, Northern Amerindian groups are particularly well-endowed proportionately. These, however, represent the extremes in the distribution; for the majority of groups the proportions of MOD are more homogeneous. Conclusions regarding genetic relationships between groups of language families cannot be drawn from these figures.
The frequency of occurrence of each single type of MOD in Maddieson's sample is given in the colunm totals in Table 4 (see section 2.3) and in Table 7 (see section 4.2). This distribution is quite uneven as well, ranging from a mere two occurrences (preaspiration and breathy release) to no less than 96 (aspiration, which is by far the most common type of MOD, accounting for over one third of the total number of cases). As for secondary articulation types, Ladefoged and Maddieson's (1995: 355, 361) claims that labialization is the most common type of MOD, whereas the number of languages with contrastive velarization is fairly small are both borne out by our findings.
The eleven different types of MOD occurring in our sample of languages can be grouped into the class nodes used in current models of feature geometry (see  above for the MOD types and Table 3). Compare the highly simplified feature tree in (15), after Clements and Hume (1995).
(15) root--laryngeal [nasal] oral
Our findings, presented in Table 3, reveal that the laryngeal types account for almost two thirds of the total number of occurrences, whereas the oral types account for less than one third--despite the fact that labialization occupies the third place in the rank order of cross-linguistic frequency of occurrence.
However, the fact should be taken into account that in the present sample of languages there are six different laryngeal types of MOD, against four different oral types of MOD and only one nasal type of MOD (in casu prenasalization); (4) in section 2 above. Weighing the figures in Table 3 thus renders the proportions (184/6):(79/4):(18/1) [approximatelty equal to] 3:2:2 roughly for laryngeal: oral: nasal types of MOD.
Cross-tabulating class nodes of MOD and language family grouping (not shown here) reveals that in all family groups but two, MOD follows the overall ranking 1. laryngeal, 2. oral, 3. nasal. The members of the Nilo-Saharan groups of language families deviate from this pattern in not having any oral types of MOD. The members of the Indo-Pacific group deviate from the general pattern in having 6 instances of oral, 4 of nasal and 3 of laryngeal MOD in total. The group of Northern Amerindian language families has by far the highest score for laryngeal and oral types of MOD: 64 out of the sample total of 184 and 25 out of the sample total of 79 cases, that is, slightly more than one third and slightly less than one third of the total number of occurrences of laryngeal and oral types of MOD in the entire sample, respectively. But as we have already seen, this group has the largest number of cases of MOD in the sample anyway, namely 90 out of 281. Only where nasal MOD is concerned, this group of language families scores far below its own average, having no more than one single instance of prenasalization.
For the sake of the testing of hypothesis 1, the analysis of the distribution of segmental modification types in each single language in the sample was crucially based on the concept of natural class. The data matrix, a sample of which is shown in the Appendix, contains among other things an informal description of the natural classes relevant to each instance of MOD in the relevant 165 languages in Maddieson's sample. However, the overall set of natural classes which emerged has not as such been included as an independent variable in our statistical analyses; there are two reasons for this decision. First, the absolute or relative quantitative representation of all the (more and less common) natural classes relevant to MOD may not be of great interest. It might be interesting to establish whether there are any clusterings between natural classes of consonants on the one hand and types of MOD or groups of MOD, such as the ones under the three class nodes, on the other. (12) This, however, is hardly feasible because of the second reason, which is of a more practical nature: the total number of different natural classes in connection with MOD in our sample of languages is too large to allow for any revealing statistical analyses. If there was an obvious way to transform the variable "natural class" into a variable with a limited number of values (i.e. a way to reduce the amount of data without losing the information), this exercise would probably be worth it. This is one of the main questions for further research.
2.3. Evaluating hypothesis 1
As we pointed out in section 2.1 above, each single instance of MOD was interpreted as positive evidence for hypothesis 1 ("+" score), as a mild counterexample (scored as "[+ or -]"), or as a genuine counterexample ("--") to hypothesis 1. Apart from these three categories, and for the sake of an unambiguous test of the two hypotheses, (13) a fourth one was also necessary, as there appeared to be three languages (14) which have a type of MOD which occurs in their affricates series only--aspiration in all three cases.
In Table 4 the numbers of instances of each type of MOD that support, mildly contradict or seriously contradict (15) hypothesis 1 are presented. In this table, as well as in Tables 7 and 9, in the column under "noMOD" the number of languages lacking MOD is mentioned; the row headed by "noMODh1" contains the number of cases which are not relevant to the testing of hypothesis 1. The latter category consists of the three languages just mentioned, which have a type of MOD occurring in their affricates series exclusively, and, of course, the 150 languages which do not have any MOD at all.
Among the 165 languages which have consonants with MOD, there is one language--one of the three just mentioned--in which MOD is exploited phonemically in affricates only. Testing hypothesis 1 is hence based on the analyses of data for 278 cases of MOD in 164 languages.
There appears to be overwhelming evidence in favour of hypothesis 1: in 182 out of 278 instances, MOD occurs on a natural class of consonants. Of these 182 instances, aspiration accounts for 76 cases. In the category "mild counterexample" (n = 45), ejectivity has the highest score, whereas laryngealization scores highest in the category "genuine counterexample" to hypothesis 1.
The row totals for the three categories allow us to test hypothesis 1 statistically. (16) Testing the observed frequencies for each of the three categories for "goodness of fit" against the null hypothesis ("MOD does not occur on natural classes of segments") indicates that the pattern is highly significant ([chi square] = 129.37, df = 2, p = 0.000). If we test more conservatively, by contrasting positive evidence against mild and genuine counterevidence taken together (+ versus [+ or -] plus -), the observed pattern still turns out to be highly significant ([chi square] = 26.60, df = 1, p < 0.001). We can therefore safely conclude that hypothesis 1 is borne out for the data for the UPSID sample.
In the remainder of this section we will discuss two related questions. The first is whether there is a systematic relationship between the types of MOD and the extent to which hypothesis 1 is confirmed. In particular: does the cross-linguistic frequency of occurrence of the several types of MOD or their grouping in terms of class nodes, mentioned in the impoverished feature tree in (15), play a role? Consider Table 5.
The average frequency of occurrence in the sample of languages of the types of MOD which confirm hypothesis 1 in most cases is 31.7, whereas the average frequency of occurrence of the types of MOD which contradict the hypothesis in most cases is 11. (17) So there seems to be a tendency for types of MOD to comply with the hypothesis according to their relative cross-linguistic frequency of occurrence. In case naturalness is defined in terms of relative typological frequency, the parameter underlying the degree to which different types of MOD support the first hypothesis may be their naturalness. However, naturalness seems a fairly poorly defined notion, also in phonology. It therefore seems wise to confront the present operationalization in terms of cross-linguistic frequency with other language independent operationalizations and then check whether the above ordering would be obtained, at least approximately.
Remarkably, five out of the seven types of MOD which usually comply with the first hypothesis fall under the laryngeal node, whereas all three types of MOD whose patterning contradicts the hypothesis for the most part fall under the oral node; as (4a) shows, the UPSID sample contains no more than four oral MOD types. This pattern also shows up if we aggregate the data for the three categories relevant to the evaluation of hypothesis 1 (namely +, [+ or -] and -) under the three class nodes. As the figures in Table 4 show, most support for hypothesis 1 by far is provided by the laryngeal types of MOD: 130 out of 182 + scores. The numbers of counterexamples (the category labelled -) are almost equal for laryngeal and oral types of MOD (24 and 25, respectively), but just as in our sample of languages oral types of MOD occur much less frequently than laryngeal ones, so we find most falsifications of hypothesis 1 among the oral types, that is, among what, for example, Sagey (1986), Halle (1992: 209) and Clements and Hume (1995) label "minor" articulation. Principled explanations of these remarkable tendencies await further investigation of the natural classes typically associated with these segmental modification types, and might come from either the phonetics or the phonology. With respect to the phonetics, the acoustic signal might be altered significantly differently by laryngeal, as opposed to supralaryngeal MOD types. Note that oral MOD types would appear to be able to change the primary place of articulation more easily than laryngeal ones. For instance, palatalization on coronals or velars easily gives rise to affrication, with concomitant shift to palato-alveolar place of articulation. Consonants with laryngeal MOD types appear to be more "stable", both in the course of diachrony as well as in synchronic segmental inventories. With respect to the phonology, the explanation of both tendencies might be related to the question whether the MOD types are contrastive or redundant. Both paths of explanation, which might be interrelated, await further investigation (cf. also section 6).
The last question to be briefly addressed here is whether there is a systematic relationship between the groups of language families on the one hand and the extent to which hypothesis 1 is confirmed on the other. In other words, is there a genetic bias where hypothesis 1 is concerned? We will briefly summarize the quantitative patterns. First, the overall distribution is fairly even. Second, the group of Northern Amerindian language families shows the largest numbers of cases of MOD supporting the hypothesis, but it also shows the largest numbers of cases of mild and serious counterexamples. This is, of course, not surprising since, as we saw in Table 2, this group accounts for a disproportionately large number of cases of MOD. Third, all groups of language families except for two predominantly support the hypothesis. The Nilo-Saharan group has equal numbers of cases constituting positive and mild counterevidence, namely three; but overall there are only eight relevant cases (18) in this group, and because of these small numbers generalizations are not warranted. The group labelled "other families" (group 13) has almost equal numbers of cases supporting and seriously contradicting the hypothesis, namely nine and eight, respectively, with another five cases that constitute mild counterevidence against the hypothesis. So it seems that only this latter group does not support our first hypothesis.
The results for the 281 cases of MOD in 165 different languages, discussed in the preceding sections, mostly confirm the first hypothesis. We now turn to a way of formalizing this finding with the aid of MOD constraints (introduced in section 1 above), which will pave the way for a proper introduction of hypothesis 2.
MOD constraints interact with other constraints: both faithfulness constraints, demanding that underlying aspiration is retained, and markedness constraints, barring segmental modification altogether, resulting in a typology of languages that either have no aspiration at all, or aspiration on only natural classes of segments, such as the voiceless stops, either contrastively or not. To illustrate this, consider a language which has only aspirated voiceless stops, but no plain, nonaspirated voiceless stops. In such a language, regardless of whether stops in underlying forms are aspirated or not, all such stops will be aspirated as a result of VOICELESSSTOPS/ASPIRATION being ranked higher than the faithfulness constraint IDENT(IO)(ASPIRATION) and the markedness constraint NoASPIRATION, hence VOICELESSSTOPS/ASPIRATION >> IDENT(IO)(ASPIRATION), NoASPIRATION. In cases like this, the operation of the relevant constraints resembles that of redundancy rules in earlier generative models. In languages without aspiration, the general markedness constraint against aspiration will forbid any aspiration being introduced (NoASPIRATION >> VOICELESSSTOPS/ASPIRATION, IDENT(IO) (ASPIRATION). In a third type of languages, namely those where aspiration is contrastive in the same set of segments, the ranking IDENT(IO)(ASPIRATION >> VOICELESSSTOPS/ASPIRATION, NoASPIRATION holds:
(6) IDENT (IO) VOICELESSSTOPS/ NOASPIRATION /[t.sup.h]an]/ (ASPIRATION) ASPIRATION [A] [t.sup.h]an] * [tan] *! * /tan/ [t.sup.h]an] *! * [A] [tan] * Legend: [A] Index finger
In such a language, the underlying pattern of aspiration contrasts will be preserved. Such a situation is also encountered in cases which are counterexamples to the generalization that modifications typically occur on natural classes of segments, that is in those cases that score "-" for hypothesis 1. We assume that if a language has a natural MOD pattern, lexicon optimization (Prince and Smolensky 1993: 172) will ensure that this pattern is also preferably found in the input forms.
The MOD constraints express valid generalizations with respect to segment inventories. Taking the inventories in (5) to (8) in section 2 as examples, we will demonstrate that it is not necessary to state on the level of individual segments that they can be labialized or aspirated, etc, if the proper MOD constraints are adopted. The part of the phoneme inventory of Nootka that we considered in (5) in section 2, for instance, can be reduced to the system in (5'):
(5') Nootka (730) restated voiceless plosive p t k q vl. nonsib, fric. x [CHI] Nootka MOD constraints: VOICELESSSTOP/EJECTIVE Voiceless stops are ejective DORSALS/LABIALIZATION Dorsals are labialized
Here the MOD constraints express the fact that voiceless stops can be distinctively ejective. It is possible, as it were, to "unfold" the partial inventory of (5') yielding the surface inventory in (5) by means of a constraint hierarchy in which the MOD constraints play a role. In this way, a generalization about the language is expressed. As we saw in the previous section, the possibility to carry MOD is regular and hence predictable in 182 of the 278 cases we considered; among these 182 cases are Nootka ejectivization and labialization. If predictable information should not be included in the underlying, supposedly nonredundant, representations of the individual segments themselves, the MOD constraints offer a way of doing so; however, in a language in which absence vs. presence of these modifications is contrastive, the lexical representation of the relevant morphemes should of course contain MOD specifications, and the language should faithfully retain these underlying contrasts onto the surface, through the appropriate ordering of MOD constraints relative to both faithfulness and markedness constraints.
The other three (parts of) consonantal inventories that served as illustrations in section 2 may be restated as follows:
(6') Kharia (301) restated voiceless plosive p t t k voiced plosive b d d g Kharia MOD constraints: VOICELESSSTOPS/ASPIRATION VOICEDSTOPS/BREATHYVOICE
Note that it would be easy to push this formalization further, and that it would be possible also to reduce the voiced-voiceless contrast in (6'), by making use of another MOD constraint saying that stops typically display a voicing contrast. We do not wish to pursue this line of inquiry here, if only because then different constraints would have to be crucially ordered with respect to one another.
The ordering of different types of MOD also seems to be relevant in, for example, Zuni (748, Zuni family, Northern Amerindian); in this language, labialization only occurs on ejectivized and aspirated consonants. However, the ordering of actual and possible types of MOD is an issue for further research.
Now reconsider Yurak:
(7') Yurak (056) restated voiceless plosive p t voiced plosive b vl. sibilant affricate ts vd. nonsibilant fricative [??] vl. sibilant fricative s voiced nasal m n voiced trill r vd. lateral approximant l Yurak MOD constraints: LABIALS/PALATALIZATION CORONALS/PALATALIZATION
In (7'), labials and coronals do not form one natural class in any generally accepted feature systems. (19) We therefore treat cases like these as if they concerned two separate cases of MOD, each in its own (natural) class.
(8') Nambakaengo (626) restated voiceless plosive p t k voiced plosive [sup.m]b [sup.n]d [sup.n]g voiced nasal m n n Nambakaengo MOD constraints: NONCONTINUANTS/LABIALIZATION VOICELESSSTOPS/PALATALIZATION (20)
It is possible that other aspects of phonological representation are also susceptible to an expression as a MOD constraint. In a language like Pashto (014, Iranian, Indo-European), for instance, all noncomplex dental/alveolar consonants have retroflex counterparts (the dental alveolar affricates do not take part in the generalization). This is shown in (17): (21)
(17) Pashto (014) voiceless plosive "t" t voiced plosive "d" d vl. sibilant fricative "s" s vd. sibilant fricative "z" z voiced nasal "n" n voiced trill "r" r
We have not treated retroflexion as a segmental modification, following Maddieson (1984) and the other authors mentioned in the introduction, section 1 above. However, again, it might be argued that in such an inventory the occurrence of retroflexion is also predictable for a natural class of segments, so that a feature cooccurrence restriction stipulating the possibility for coronals to be retroflexed might be drawn up. The same applies to the gemination of consonants. Consider part of the inventory of Shilha (256):
(18) Shilha (256) voiceless plosive "t" k long vl. plosive "t:" k: voiced plosive b "d" g long vl. plosive b: "d:" g: vl. nonsibilant fric. f x long vl. nonsib, fric. f: x: vd. nonsibilant fric. V long vd. nonsibilant fric. V: vl. sibilant fric. "s" S long vl. sibilant fric. "s:" [integral] vd. sibilant fric. "z" 3 long vd. sibilant fric. "z:" 3: voiced nasal m "n" long vd. nasal m: "n:" voiced trill "r" long voiced trill "r:" vd. lateral approx. "l" long vd. lat. approx. "l:"
In this language, all consonants except the ones with pharyngealization and the central approximants /j/ and /w/ (neither of which classes are shown in ), can occur either as a singleton or as a geminate. This fact, too, is redundant if it is stated for every individual consonant. We leave the question of which aspects of phonological representation can be brought into the register of the format of a MOD constraint for further discussion.
4. The structure of affricates: hypothesis 2
We will now consider the question how affricates pattern with respect to the types of MOD they have. Consider a "standard" view of the structure of affricates (namely that advanced in Sagey 1986), where stops have a feature [-cont], fricatives have [+cont], and affricates have both. (22) Schematic representations are given in (19):
(19) C--[cont] stops C--[+cont] fricatives C--[-cont] affricates [+cont]
Since affricates have characteristics in common with stops as well as fricatives, they could behave either like stops, or like fricatives, or pattern with both.
A MOD constraint approach predicts that if a modification occurs on the natural class of the [-cont] segments, it will also occur on affricates, since these also have a [-cont] feature. For instance, in a segmental inventory in which all stops can be distinctively aspirated, it is predicted that affricates can also be distinctively aspirated. If affricates are not distinctively aspirated in such an inventory, this must be marked as an exception to the constraint, for instance, by way of a highly ranked markedness constraint pertaining to aspiration on affricates; or, more generally, a constraint against the combination of two types of complexity within one segment.
An example in which affricates pattern with voiceless stops, in terms of their being potentially ejective, is Nootka, as shown in (20). For this language the relevant MOD constraint can therefore be formulated very generically.
(20) Nootka (730) affricates; stops see (5) and (5') vl. sibilant affric, ts t[integral] vl. sib. eject, affr. ts' t[integral]' vl. lateral affricate t[??] vl. lat. eject, affr. t[??]
We have seen that Nootka allows distinctive ejectivity on voiceless stops (expressed by the first MOD constraint in [5']). Since the Nootka affricates also have the features [-cont] and [-voice], these segments are predicted to have ejective counterparts. This is indeed the case in Nootka.
A language in which affricates do not behave identically to (voiceless) plosives is Kirghiz (Turkic, Ural-Altaic). In this language, voiceless stops can be distinctively aspirated, but affricates cannot (although, in general, aspirated affricates do exist, see, for example, languages like Thai, Mandarin, and Chipewyan in Maddieson 1984).
(21) Kirghiz (062) stops and affricates vl. plosive p t q vl. aspirated plos. [p.sup.h] [t.sup.h] [q.sup.h] vl. sibilant affr. ts t[integral]
These two cases show that as far as MOD is concerned affricates can, but need not, pattern with stops.
After having illustrated the basic idea, let us now verify the issue from a general viewpoint, again on the basis of a quantitative approach to the data obtained from Maddieson (1984). Hypothesis 2 is repeated in (22), from (3) above:
(22) Hypothesis 2 Affricates are part of two natural classes with respect to the types of segmental modification they take: they may pattern with stops and/or they may pattern with fricatives.
We will approach this hypothesis by testing two subhypotheses, which we will refer to as subhypotheses 2a and 2b, given in (23):
(23) Subhypothesis 2a Affricates pattern with stops with respect to their segmental modification types. Subhypothesis 2b Affricates pattern with fricatives with respect to their segmental modification types.
As will be obvious, to test subhypothesis 2a in a particular language, the language must have both stops and affricates. Moreover, the language needs to have one and the same type of MOD on a series of stops and/or on the affricates. In such a language, subhypothesis 2a is only assumed to be confirmed if the patterning of the MOD in both the stops series and all affricates can be captured in one generalization. Only in that case does the MOD token obtain a "+" score for subhypothesis 2a.
In the case of Nootka (20), which has phonemic ejectivity on stops as well as affricates, subhypothesis 2a is confirmed; hence we assign it a plus score. For Kirghiz (21), subhypothesis 2a is assigned a minus score. Note that both Nootka and Kirghiz are neutral with respect to subhypothesis 2b. This does not mean that subhypothesis 2b is disconfirmed, of course: for subhypothesis 2b to be disconfirmed a language has to be found in which distinctive segmental modification on fricatives occurs, but not on affricates, or vice versa. In general, for a language to allow testing subhypothesis 2b (or 2a or both), it needs to have MOD on fricatives (or stops or both, respectively) and/or affricates. As we pointed out already, there are relatively few languages with MOD on fricatives: the UPSID sample contains only 23 such languages. In all, these 23 languages have 26 cases of MOD on fricatives.
It is noteworthy that subhypothesis 2b is supported by a number of cases in which there is no range of fricative segments that bear a type of segmental modification. Consider the partial consonant inventories of Kashmiri (018, Indic, Indo-European), Akan (115, Kwa, Niger-Kordofanian--this language is also known as Twi, see section 2.2) and Wichita (755, Macro-Siouan, Northern Amerindian) in (24):
(24) Kashmiri (018) vl. sibilant affricate [t[integral].sup.j] vd. sibilant affricate d[z.sup.j] vl. sibilant fricative [[integral].sup.j] Akan (115) vl. nonsibilant affricate [cc.sup.w] vd. nonsibilant affricate [[??]j.sup.w] vl. nonsibilant fricative [c.sup.w] Wichita (755) vl. sibilant eject, affricate ts' vl. sibilant eject, fricative s'
In these languages there is only a single continuant with segmental modification if affricates are left out of consideration; (23) hence, they do not support hypothesis 1. However, the segmental modification type does range over a number of continuants if affricates are taken into account, so in these cases subhypothesis 2b is supported. No cases in which subhypothesis 2a would be supported in a similar way were found.
4.1. Testing hypothesis 2: method
In testing hypothesis 2, we obviously selected only those languages in the UPSID sample which have affricates. In each single case we examined whether or not the patterning of a modification in the stops and/or fricatives recurs in the affricates. For each single case, there are in principle [2.sup.3] = 8 different possible scenarios: 1. no MOD in affricates, stops, and fricatives; 2. MOD in affricates, but not in stops and fricatives; 3. no MOD in affricates, MOD in stops, no MOD in fricatives; 4. the same pattern of MOD in affricates and stops, no MOD in fricatives; 5. no MOD in affricates and stops, MOD in fricatives; 6. MOD in affricates, no MOD in stops, MOD in fricatives, along the same pattern as in the affricates; 7. no MOD in affricates, MOD in stops and fricatives; 8. the same pattern of MOD in affricates, stops and fricatives.
Apart from these eight, there are two other possible scenarios: one occurs in those languages which do not have any affricates (scenario 0), the other in languages which only have affricates which do not fall in the range of the generalization regarding the natural class of the other obstruents in which MOD occurs, that is, in the ones that were included in the testing of hypothesis 1 (scenario 9). In other words, this last scenario is formed by cases of MOD on plosives and/or fricatives of which there are no affricated counterparts in the language at issue; the affricates that do exist do not allow testing the second hypothesis. An example is velarization in Yakut (language 611 in Maddieson 1984). In Yakut, laterals are velarized, but lateral affricates do not exist in this language (unlike in 10 other languages in Maddieson's sample--cf. 1984: 225).
An overview of the possibilities with respect to the evaluation of hypotheses 2a and 2b is given in Table 6. In this table and in the rest of this text we will refer to "score type" instead of "scenario".
Only the score types 2 to 8 above bear on subhypotheses 2a and 2b. Specifically, every case of MOD of the score types 2, 3, 6 and 7 is counterevidence (symbolized as "-") to subhypothesis 2a, whereas score types 4 and 8 constitute positive evidence ("+") for this hypothesis. Similarly, each case of MOD of the score types 2, 4, 5 and 7 is counter evidence to subhypothesis 2b, whereas score types 6 and 8 constitute positive evidence ("+") for this hypothesis. Only the score types 2, 7 and 8 are logically relevant to both hypotheses simultaneously; only in the case of the latter score type are both hypotheses borne out at the same time.
As we pointed out in section 2.1, the UPSID sample includes two click languages. The clicks series of these two languages were excluded from the testing of hypothesis 2. However, if there is independent reason to consider the nonaffricated clicks as obstruents, at least in the case of !Xu (language 918), the five types of MOD in the clicks series would all support the general version of hypothesis 2, since the patterning of MOD in the affricated clicks turns out to be identical to that in the nonaffricated clicks. As we said in section 2.1, only the nonclick consonants of both languages were taken into consideration in connection with the testing of both hypothesis 1 and hypothesis 2.
4.2. Evaluating hypothesis 2
Table 7 contains the numbers of instances of each type of MOD for each of the ten score types in connection with hypothesis 2, described in section 4.1. Just as in Table 4, in the column under "noMOD" the number of languages lacking MOD is mentioned.
Among the 150 languages lacking MOD, 65 do not have any affricates either (cf. score type 0). In all, about one third of the languages in the UPSID sample lack affricates. The three instances of MOD of the score type 2 (i.e. MOD occurring in affricates, but not in stops and fricatives) are the ones mentioned at the beginning of section 2.3 and in note 14. The score type 4, that is, the type of cases where we find identical patterns of MOD in affricates and stops, yet no MOD in fricatives, in other words the type of cases which support subhypothesis 2a, shows very high numbers for aspiration and ejectivity. Inspecting the columns in Table 7, it appears that for the latter two types of MOD as well as for breathy voice, the score type supporting subhypothesis 2a alone is the one attracting by far the largest numbers in this sample of languages. Remarkably, in the overwhelming majority of instances of labialization, the relevant languages do have affricates, but these affricates are not in the range of the generalization, that is, they are not members of the natural class formed by the other obstruents on which the MOD occurs (score type 9), so they do not allow testing the second hypothesis.
The row totals for the relevant score types permit the statistical testing of subhypothesis 2a, subhypothesis 2b as well as hypothesis 2 in general (affricates pattern with both stops AND fricatives, that is, the conjunction of subhypotheses 2a AND b). Testing the observed frequencies pro and contra hypothesis 2 and the two subhypotheses for all eleven MOD types taken together yields the following results:
Subhypothesis 2a pro: score types 4, 8 = 103 + 13 = 116 contra: score types 2, 3, 6, 7 = 3+37+4+7 = 51 [chi square] = 25.30 df = 1 p < 0.001--hypothesis accepted Subhypothesis 2b pro: score types 6, 8 = 4 + 13 = 17 contra: score types 2, 4, 5, 7 = 3+103+2+7 = 115 [chi square] = 72.76 df = 1 p < 0.000--highly significant, but hypothesis rejected; opposite alternative hypothesis ("affricates do NOT pattern with fricatives with respect to their segmental modification types") accepted Hypothesis 2 pro: score type 8 = 13 contra: score types 2, 7 = 3+7 = 10 [chi square] = 0.39 df = 1 not significant--hypothesis rejected
We conclude that only subhypothesis 2a is borne out, which means that in the data from the UPSID sample affricates typically pattern with stops with respect to their segmental modification types. There is overwhelming evidence against subhypothesis 2b, which says that as far as segmental modification is concerned affricates pattern with fricatives. (24) There is even evidence for the reverse of subhypothesis 2b, that is, there is evidence that affricates do NOT pattern with fricatives with respect to their MOD types. This ties in well with recent proposals by Steriade (1993) and Clements (1999) in which affricates are basically stop segments with certain special properties. Steriade (1993: 402) assumes that in affricates, while they have both a closure specification ([A.sub.o]--total absence of airflow) and a fricative specification ([A.sub.f]--degree of total aperture sufficient to produce a turbulent airstream) "the fricative release of affricates is not a distinctive property" (1993: 404). In the light of the devastation of subhypothesis 2b it comes as no surprise that there is not sufficient support for the demanding hypothesis 2, that affricates pattern with both stops and fricatives with respect to their segmental modification types.
In the rest of this section we will discuss two more questions. The first one is whether the three classes of MOD types behave differently vis-a-vis hypothesis 2. Breaking down the results for the three classes of MOD types, it becomes clear that the overall picture is fully replicated for the laryngeal types, whereas subhypothesis 2a, subhypothesis 2b and hypothesis 2 in general need to be rejected for both the oral and the nasal types, to the extent that they can at all be tested--see Table 8.
This finding means that in the UPSID sample (a) the laryngeal MOD types are prevalent, and (b) as far as the patterning of MOD types in the three classes of obstruents is concerned, phonation, secondary articulation and (pre)nasalization are essentially different. In 97 out of the 184 cases in which they occur, the laryngeal types of MOD support subhypothesis 2a and subhypothesis 2a alone (score type 4). This, however, does not come as a big surprise in the light of the very high number of cases of aspiration and ejectivity which constitute positive evidence for subhypothesis 2a, as we already saw in our discussion of the findings presented in Table 7. The laryngeal types of MOD are also the ones which contribute by far most to those cases which support hypothesis 2, even though this hypothesis is not supported on any level.
Again, we should of course bear in mind that in the UPSID sample of languages there are six different laryngeal types of MOD, against four different oral types of MOD and only one nasal type (prenasalization). On the other hand, these proportions can still not account for the fact that of the 103 cases of MOD supporting only subhypothesis 2a, 97 are laryngeal, 2 oral and 4 nasal and for the fact that of the 13 cases of MOD supporting hypotheses 2, 9 are laryngeal, 4 oral and, consequently, 0 nasal.
Throughout the sample, there are only four cases in which only subhypothesis 2b is supported (score type 6); three of these concern oral types of MOD.
The second and, at the same time, final remaining question to consider is whether there is any typological bias in the sample as far as the second hypothesis is concerned. As far as the figures are concerned, (25) this distribution does not show any spectacular skewing either. Three things deserve to be pointed out, however. First, although they all support hypothesis 1, the few instances of MOD in the group of Australian language families have no bearing on our subhypotheses 2a and 2b, as they occur in languages which have no affricates (score type 0). (26) Second, the group of the 26 Indo-Pacific languages in Maddieson's sample is the one with the largest number of languages lacking affricates (n = 21), which thus do not allow testing hypothesis 2. Third and last, the group of Northern Amerindian language families have disproportionately large numbers of cases of MOD supporting subhypothesis 2a as well as hypothesis 2 in general. So whereas the Indo-Pacific and the Australian language families have little or nothing to contribute to the testing of hypothesis 2, the Northern Amerindian language families overwhelmingly support the hypothesis, especially subhypothesis 2a, even if one takes the fact into account that overall they have a relatively high incidence of MOD.
5. The relationship between the two hypotheses
Put generally, the two main ideas in this study are that MOD in consonants will prefer to occur in natural classes (hypothesis 1) and, in particular, that MOD in affricates will pattern with that in the natural class of stops and/or fricatives (hypothesis 2). The data from the languages in the UPSID sample made it possible to test both ideas in a refined way. Hypothesis 1 is convincingly supported by the data. As far as hypothesis 2 is concerned, our findings indicate that as far as segmental modification goes, affricates pattern with stops rather than fricatives; for further discussion, see section 6 below.
In the present section we will briefly discuss some aspects of the logically possible ways for evidence pro and contra both hypotheses to cluster. We will do so on the basis of several "numerical" trends in the 315 languages in the sample. In Table 9, the total numbers of instances of MOD and the absence of MOD have been cross-tabulated for the ranges of outcomes in connection with both hypotheses.
Four things may be touched upon briefly in connection to Table 9. First, out of the 281 instances of MOD there appear to be 8 instances in which MOD is completely systematic in terms of natural classes (pro hypothesis 1), although no obstruents are involved in these natural classes; these are among the cases in which it was not possible to test hypothesis 2. Second, out of the 37 cases of MOD which contradict only subhypothesis 2a (no MOD in affricates, MOD in stops, no MOD in fricatives), 26 support hypothesis 1. In other words, counterexamples to subhypothesis 2a are for a large majority well-behaved vis-a-vis hypothesis 1. Third, in 82 cases the pattern of MOD is completely systematic in terms of natural classes (pro hypothesis 1) and the MOD in the affricates patterns with that in the stops, while there is no MOD in the fricatives (contra hypothesis 2b). Fourth, of the 13 cases in which the distribution of MOD in affricates supports both hypothesis 2a and hypothesis 2b (MOD in affricates shows the same pattern as in stops and fricatives), 6 constitute positive evidence for hypothesis 1, whereas 5 cases mildly contradict and 2 seriously contradict the first hypothesis. These differences are not statistically significant, however.
The main conclusion to be drawn from this part of our analyses seems to be that in the languages in Maddieson's (1984) sample positive evidence for hypothesis 1 and negative evidence for subhypothesis 2a coagulate (26 cases), although this tendency is clearly overshadowed by the heavy clustering of positive evidence for both hypothesis 1 and subhypothesis 2a (82 out of the 278 relevant cases of MOD in the entire language sample). The question can be raised which MODs and which natural classes are involved in these two tendencies. This is a more specific variant of the question whether there is any cross-linguistic clustering between natural classes of consonants and types of MOD, raised in section 2.2.
6. Conclusion, discussion, and issues for further research
In this article we have examined one aspect of consonant systems, namely the way in which types of "phonemic" MOD typically pattern. We have shown that in the vast majority of cases these occur on natural classes of segments, rather than on individual, isolated ones or randomly. This finding was formalized by introducing the notion of MOD constraint, which expresses general statements about MOD in segmental inventories and shows that segments are not primitives. We suspected that such constraints are grounded in the phonetics (following Pulleyblank 1988); for example, voiceless stops are perceptually more salient if they are realized with aspiration (cf. Stevens and Keyser's  enhancement theory). We have not, however, attempted to draw up a comprehensive set of MOD constraints.
Our approach also made it possible to investigate the question what position affricates take in consonantal inventories: it turned out that these usually pattern with stops but not with fricatives, with respect to their MOD pattern. This finding seems to imply that affricates might be looked upon as stops with special characteristics (as also in Jakobson, Fant and Halle's  proposal to regard affricates as "strident stops"). It may also mean that diachronically affricates tend to develop out of stops, rather than out of fricatives.
Affricates pattern with fricatives with respect to their place of articulation (van de Weijer 1996). With respect to MOD, on the other hand, they pattern with stops. In this latter connection, we might conclude that our results accord well with recent insights of Steriade (1993) and Clements (1999), who propose that affricates are stops with certain special properties. Under such a conception, it comes as no surprise that affricates usually pattern with stops rather than fricatives with regard to their type of MOD.
We observed in the beginning of this article that the notion of MOD is not uniformly defined. Different aspects of segmental structure may behave in the regular way that we have uncovered, such as the imposition of certain vowel qualities (palatalization, labialization), voice qualities (aspiration, glottalization), prenasalization, as well as retroflexion, and even length, although the latter two (which were not involved in our analyses) are usually not considered as segmental modification types. These aspects are expressed in different ways in various current feature geometries (see Sagey 1986, Clements and Hume 1995), but never uniformly. This led us to raise the question how far one can take the concept of MOD constraints: can it be extended to gemination, retroflexion and voicing? Obviously, further research is needed here.
Before we list a number of additional questions for further research, we will briefly discuss (a) the diachronic relevance of patterns of segmental modification as well as (b) possible explanations of gaps in patterns of segmental modification.
Segmental inventories are not necessarily stable in the course of time; see, for example, Lass (1984: 164), who distinguishes between "static" and "dynamic" palatalization, pointing out that the former can be the diachronic result of the latter. In our view, the probability should not be excluded that MOD plays a role in the restructuring of phoneme inventories. The consonant inventory of Muinane (806, Witotoan, Southern Amerindian) might be illustrative in this respect.
(25) Muinane (806) voiceless plosive p t c k ? voiced plosive b d J g vl. sibilant affr. t[integral] vd. sibilant affr. [d.sub.3] vl. nonsib, fric. [phi] x vd. nonsib. fric. [beta] vl. sibilant fric. s [integral] vd. sibilant fric. 3 voiced nasal m n [eta] voiced r-sound rr [rr.sup.j]
As can be seen in (25), in Muinane contrastive secondary palatalization only occurs on /rr/, resulting in /[rr.sup.j]/. This contrast does not appear to play a role in the rest of the Muinane consonantism. (27) However, it is not inconceivable that Muinane's five palatals result from a past "fusion" of previously secondarily palatalized dentals with their segmental modification feature, in that palatalization developed into a primary articulation in these consonants. The /rr/ ~ /[rr.sup.j]/contrast thus probably constitutes a relic of this hypothetical former stage of the language. Ladefoged and Maddieson (1995: 368) expressed the role secondary articulation can play in diachrony as "Today's secondary articulations may be the primary articulations of the future." This results from the sort of process Clements (1991) calls "promotion." (28)
Another type of promotion that may be relevant here is "debuccalization", the loss of the primary oral articulation (cf. Fallon 1998). In some cases, such as Irish /[t.sup.j], [s.sup.j]/>/hj/, the secondary articulation is preserved. In others, even the secondary articulation is given up, in favour of a phonation type, that is, debuccalization of /[k.sup.w]/to result in a glottal stop, as in Takelma (Sapir 1912: 44), in which a labialized velar ejective before /xC/ becomes a glottal stop. Before /xV/, the labialization is transferred.
The study of the patterning of MOD, both from phonological and typological points of view, might cast new, if only indirect, light on some of the more recent divisive elements of historical linguistics, such as the supposed origin of the Modern German affricates in former aspirated voiceless stops (Hock 1991: 436, 666-667) (29) and "glottalic theory." Glottalic theory concerns the reconstruction of the stop series of Proto-Indo-European. In referring to this theory, we deliberately avoid using the definite article, since it has several variants; we will briefly discuss the aspects that most variants seem to have in common.
According to glottalic theory, Proto-Indo-European, rather than having voiceless stops, voiced stops (lacking /b/, however) and aspirated voiced stops, distinguished between a voiceless, a voiceless "glottalized", that is, ejectivized, (30) and a voiced series. According to some, both the voiceless and the voiced series were aspirated, while according to others, both the voiceless and the voiced series had aspirated and unaspirated allophones. Yet others think that both the voiceless and the voiced series were aspirated, with the voiceless stops having nonaspirated allophones (Watkins 1992). All glottalic models posit (a) ejectives instead of the voiced stops traditionally postulated, and (b) that the labial stop was absent from the ejectivized voiceless series. (31) According to its proponents, this is a much less marked state of affairs (Gamkrelidze and Ivanov 1973; cf. the critical discussion in Hock 1991: 621-626).
A basis for an investigation of glottalic theory might be a survey of language families of which the members differ as to the relationship between secondary ejectivity and (primary) voice quality in the obstruent series. More in particular, language families would have to be found of which certain members appear to be more conservative in retaining the secondary ejectivity, perhaps in conjunction with its absence in certain (types of) consonants. For those instances of secondary ejectivity which constitute "mild" counterexamples to our hypothesis 1, the question could be considered whether the gaps tend to occur typically in the voiceless rather than voiced labials, or somewhere else.
We suspect that such gaps in the patterning of MOD types can be phonetically motivated, just like gaps of nonmodified consonants. Examples of the latter are the /p/ gap in Hausa (in [p] there is hardly any resonance in the oral cavity so that the stop burst is difficult to detect acoustically) and the /g/ gap in Dutch (as voicing and velarity are articulatorily hard to maintain simultaneously). As we saw, in Shilha pharyngealization occurs on all nongeminate dental/alveolars except for /n/ as well as on nongeminate velar voiceless stops. This gap may be grounded in physiology; pharyngeal nasals are excluded, in principle, in the IPA. For examples of the phonetic plausibility of certain gaps, see also Ohala (1983). Such gaps can be formalized as highly specific feature cooccurrence constraints, which would be high-ranked in the languages concerned. Mutatis mutandis, a comparable approach might be useful in explaining the opposite case: instances of MOD occurring on a single consonant, that is, the genuine counterexamples to hypothesis 1. In our discussion of aspiration in Vietnamese ( in section 2.1 and note 7) we already hinted at a possible diachronic explanation for the fact that this MOD type occurs on the voiceless coronal stop alone.
In the first part of section 2.2 we asked if there are any combinatorial preferences of different MOD types within individual languages. At the end of the same section we pointed out that further research could focus on the question whether there are any systematic relationships between natural classes of consonants on the one hand and types of MOD on the other. Are there any cross-linguistic regularities in the distribution of MOD types over natural classes? At the end of section 5 we presented a more specific version of this question. One way of tackling this problem would be to incorporate the natural class of the consonants on which MOD has been found to occur in our sample of languages into the statistical analyses; to this end, however, a procedure needs to be developed to reduce the well over 30 different natural classes into a variable with a much smaller number of values. Universal parameters of phonological organization (such as e.g. manner, place and "voice"/laryngeal) might well turn out to be useful in this respect.
Another issue on which future research on MOD could focus (addressed in the first paragraphs of section 3) is the question whether there are any systematic differences in the typological patterning of MOD types between those cases in which there are no "plain" counterparts of the consonants with segmental modification and those cases where "plain" counterparts do exist. We dealt with this question in section 3 by using different orderings of MOD constraints, markedness constraints and faithfulness constraints. Another way of dealing with this issue would be by using "antagonistic" feature cooccurrence constraints (see again Pulleyblank 1998), which forbid the combination of specific features or feature sets, whereas sympathetic feature cooccurrence constraints demand cooccurrence of specific features or feature sets. All MOD constraints that we saw so far were of the sympathetic type. It is not clear to us at this point whether there are any principled differences or empirically different predictions associated with either approach. We investigate this issue in more detail in van de Weijer and Hinskens (forthcoming).
A final issue for further research we would like to mention here is the question whether there is a systematic relationship between hetero- versus homorganicity of MOD and the consonants it occurs on.
Received 23 July 2001 Revised version received 23 July 2002
Meertens Instituut, Amsterdam University of Leiden
In this appendix we list a part of the languages (in Maddieson 1984) that we checked for the presence of one or several of the types of segmental modification (MOD) listed in (4) in the paper. Apart from the type of MOD, we also indicate whether the pattern of secondary articulation confirms or contradicts hypothesis 1, by way of "+", "-" or "[+ or -]" marks (see section 2.1) as well as whether hypotheses 2, 2a and 2b are relevant and, if so, if they are confirmed or rejected.
The following abbreviations are used:
LAB = labialization PAL = palatalization VEL = velarization PHA = pharyngealization ASP = aspiration PRN = prenasalization NAS = nasalization PRA = preaspiration BRV = breathy voicing BRR = breathy release LAR = laryngealization EJE = ejectivity GLO = glottalization
Column 1: first three digits: language number in Maddieson (1984) fourth digit: nth case of MOD in this language fifth digit: total number of different MOD types in this language
Column 2: type of MOD (000 = none)--cf. section 2
Column 3: evaluation of hypothesis 1--cf. section 2.1
Column 4: natural class of consonants on which MOD occurs
Column 5: reevaluation of hypothesis 1:
0 0 0 0 = hypothesis n.a.
+ + + + = hypothesis borne out
NOTO = MOD occurs on all available members of the natural class but one
ALSO = MOD occurs not only on all members of the natural class, but also on an isolated other consonant
ONLY = MOD occurs on either one or two unrelated consonants--cf. section 2.1
Column 6: evaluation of hypothesis 2--cf. section 4.1
Table 1. The distribution of the number of different types of MOD Types of MOD Cases Languages n n % n % 0 150 34.80 150 47.62 1 91 21.11 91 28.89 2 94 21.81 47 14.92 3 42 9.74 14 4.44 4 48 11.14 12 3.81 5 0 0.00 0 0.00 6 6 1.39 1 0.32 Total 431 100.00 315 100.00 Table 2. The number of languages studied and the number of cases of MOD per group of language families Group of language families Languages in sample Cases of MOD n n l. Indo-European 21 22 2. Ural-Altaic 22 10 3. Niger-Kordofanian 31 17 4. Nilo-Saharan 21 8 5. Afro-Asiatic 21 23 6. Austro-Asiatic 6 10 7. Australian 19 4 8. Austro-Tai 25 13 9. Sino-Tibetan 18 18 10. Indo-Pacific 26 13 11. Northern Amerindian 50 90 12. Southern Amerindian 37 30 13. Other families 18 23 Total 315 281 Table 3. The number of cases of MOD per class node Class node Cases of MOD n Laryngeal 184 Nasal 18 Supralaryngeal/oral 79 Total 281 Table 4. The relationship between the 11 types of MOD (as well as the absence of MOD) and the evaluation of hypothesis 1 Hyp 1 LAR asp brr brv eje lar pra + 76 2 6 31 13 2 [+ or -] 10 0 0 15 2 0 - 7 0 1 5 11 0 noMODhl 3 0 0 0 0 0 Total 96 2 7 51 26 2 Hyp 1 ORA NAS noMOD total lab pal pha vel prn + 30 5 1 2 14 0 182 [+ or -] 8 6 1 1 2 0 45 - 8 9 4 4 2 0 51 noMODhl 0 0 0 0 0 150 153 Total 46 20 6 7 18 150 431 Table 5. The evaluation of hypothesis 1 in connection with the sample frequency of the several types of MOD Evaluation of hyp 1 Type of MOD Cases n mainly + asp 96 eje 51 lab 46 prn 18 brv 7 brr 2 pra 2 dubious lar 26 mainly - pal 20 vel 7 pha 6 Table 6. The procedure in connection with the evaluation of hypothesis 2 Natural class MOD? affricates - + - + - + - + stops - - + + - - + + fricatives - - - - + + + + Score type 1 2 3 4 5 6 7 8 (a) evaluation hyp 2a - - + - - + evaluation hyp 2b - - - + - + evaluation hyp 2a and b - - + (a) Additional score types: 0 = language lacks affricates. 9 = language lacks affricates in the natural class of the plosives and/or fricatives in which MOD occurs. Table 7. The relationship between the 11 types of MOD (as well as the absence of MOD) and the evaluation of hypothesis 2 Hyp 2 LAR ORA asp brr brv eje lar pra lab 0 13 0 0 5 4 0 8 1 0 0 0 0 11 0 0 2 3 0 0 0 0 0 0 3 25 0 1 1 1 0 1 4 52 2 6 34 2 1 0 5 0 0 0 0 0 0 1 6 0 0 0 1 0 0 2 7 0 0 0 0 2 0 1 8 1 0 0 7 0 1 2 9 2 0 0 3 6 0 31 Total 96 2 7 51 26 2 46 Hyp 2 NAS noMOD total pal pha vel prn 0 3 2 0 7 65 107 1 1 1 2 0 85 100 2 0 0 0 0 0 3 3 2 0 1 5 0 37 4 1 0 1 4 0 103 5 0 0 1 0 0 2 6 1 0 0 0 0 4 7 4 0 0 0 0 7 8 2 0 0 0 0 13 9 6 3 2 2 0 55 Total 20 6 7 18 150 431 Table 8. Testing hypotheses 2a, 2b and 2 for the three classes of MOD types (for every cell in the matrix df = 1) Hypothesis 2a Laryng [chi square] = 30.96, p < 0.001 Oral [chi square] = 0.66, NS Nasal [chi square] = 0.05, NS Hypothesis 2b Laryng [chi square] = 75.58, p < 0.001 opposite alternative hyp. accepted Oral [chi square] = 0.25, NS Nasal data do not allow testing Hypothesis 2 Laryng [chi square] = l.14, NS Oral [chi square] = 0.06, NS Nasal data do not allow testing NS = not significant. Table 9. The relationship between the evaluation of hypthesis l and the evaluation of hypothesis 2 Hyp 1 Hyp 2 0 1 2 3 4 5 + 28 8 0 26 82 1 [+ or -] 5 0 0 5 17 1 - 9 7 0 6 4 0 noMODh1 65 85 3 0 0 0 Total 107 100 3 37 103 2 Hyp 1 6 7 8 9 total + 0 1 6 30 182 [+ or -] 0 2 5 10 45 - 4 4 2 15 51 noMODh1 0 0 0 0 153 Total 4 7 13 55 431 Column 6: evaluation of hypothesis 2--cf. section 4.1 00000 000 0 0 0 0 0 0 0 1 00113 PAL - - - all consonants NOTO 7 00123 VEL - - - bilabials ALSO 9 00133 ASP + + + vl. plosives + + + + 3 00211 ASP + + + vl. plosives + + + + 0 00411 ASP + + + vl. plosives + + + + 3 00611 ASP + + + vl. plosives + + + + 0 00711 PAL - - - all consonants NOTO 7 00812 PAL - - - all consonants NOTO 7 00822 VEL + / - pal-alv frics ALSO 5 00912 PAL + / - all consonants NOTO 8 00922 ASP + + + vl. plosives + + + + 3 01000 000 0 0 0 0 0 0 0 0 01100 000 0 0 0 0 0 0 0 1 01200 000 0 0 0 0 0 0 0 1 01311 ASP + + + vl. plosives + + + + 3 01400 000 0 0 0 0 0 0 0 1 01511 PHA - - - ONLY 9 01612 ASP + + + vl. plosives + + + + 4 01622 BRV + + + vd. non-conts + + + + 4 01712 ASP + + + vl. plosives + + + + 4 01722 BRV + + + vd. Plosives + + + + 4 01812 PAL - - - ONLY 6 01822 ASP + + + vl. plosives + + + + 4 01911 ASP + + + vl. plosives + + + + 4 02000 000 0 0 0 0 0 0 0 1 02100 000 0 0 0 0 0 0 0 1 02212 ASP + + + vl. plosives + + + + 4 02222 EJE + + + vl. plosives + + + + 4 05000 000 0 0 0 0 0 0 0 1 05100 000 0 0 0 0 0 0 0 1 05200 000 0 0 0 0 0 0 0 1 05300 000 0 0 0 0 0 0 0 0 05400 000 0 0 0 0 0 0 0 1 05500 000 0 0 0 0 0 0 0 1 05612 PAL + + + labials + + + + 0 05622 PAL + + + coronals + + + + 8 05700 000 0 0 0 0 0 0 0 1 05811 ASP + + + vl. plosives + + + + 3 05911 ASP + + + vl. plosives + + + + 3 06011 PAL + + + cor non-conts + + + + 3 06111 VEL + + + laterals + + + + 9 06211 ASP + + + vl. plosives + + + + 3 06300 000 0 0 0 0 0 0 0 1 06400 000 0 0 0 0 0 0 0 1 06500 000 0 0 0 0 0 0 0 1 06611 ASP + + + vl. plosives + + + + 3 06700 000 0 0 0 0 0 0 0 1 06800 000 0 0 0 0 0 0 0 1 06900 000 0 0 0 0 0 0 0 1 07012 ASP + + + vl. plosives + + + + 4 07022 LAR + + + obstruents + + + + 4 07100 000 0 0 0 0 0 0 0 1 10000 000 0 0 0 0 0 0 0 0 10100 000 0 0 0 0 0 0 0 1 10200 000 0 0 0 0 0 0 0 0 10311 LAB + + + velar plosives + + + + 0 10400 000 0 0 0 0 0 0 0 0 10500 000 0 0 0 0 0 0 0 1 10600 000 0 0 0 0 0 0 0 0 10700 000 0 0 0 0 0 0 0 1 10800 000 0 0 0 0 0 0 0 0 10900 000 0 0 0 0 0 0 0 0 11000 000 0 0 0 0 0 0 0 1 11100 000 0 0 0 0 0 0 0 0 11200 000 0 0 0 0 0 0 0 0 11300 000 0 0 0 0 0 0 0 0 11411 ASP + / - vl. plosives NOTO 3 11512 LAB - - - ONLY 6 11522 ASP + + + vl. plosives + + + + 3 11614 PAL + / - bilabial plosives ALSO 9 11624 LAB + / - dorsals ALSO 9 11634 ASP + + + vl. plosives + + + + 3 11644 BRV + + + vd. plosives + + + + 4 11712 LAB - - - ONLY 6 11722 ASP + / - vl. plosives NOTO 4 11800 000 0 0 0 0 0 0 0 1 11900 000 0 0 0 0 0 0 0 0 12000 000 0 0 0 0 0 0 0 0 12100 000 0 0 0 0 0 0 0 1 12200 000 0 0 0 0 0 0 0 1 12311 ASP + + + vl. plosives + + + + 4 12411 ASP + + + vl. plosives + + + + 4 12511 PRN + + + vd. plosives + + + + 4 12612 ASP + + + vl. plosives + + + + 3 12622 EJE + + + vl. plosives + + + + 4 12700 000 0 0 0 0 0 0 0 1 12800 000 0 0 0 0 0 0 0 0 12912 LAR - - - nasals NOTO 0 12922 PRN + + + vd. plosives + + + + 0 13000 000 0 0 0 0 0 0 0 0 20000 000 0 0 0 0 0 0 0 0 20100 000 0 0 0 0 0 0 0 1 20200 000 0 0 0 0 0 0 0 0 20300 000 0 0 0 0 0 0 0 1 20400 000 0 0 0 0 0 0 0 1
* We thank the anonymous Linguistics reviewers for their extremely helpful comments on earlier versions. We are also grateful to Harry van der Hulst, Beth Hume, Grazyna Rowicka, and Erik Jan van der Torre for discussion of previous versions. Any remaining errors are of course our own responsibility. Correspondence address: Frans Hinskens, Meertens Instituut, P.O. Box 94264, 1090 GG Amsterdam, The Netherlands. E-mail: email@example.com.
(1.) See also van de Weijer and Hinskens (forthcoming) for further discussion.
(2.) "Aspiration" and "preaspiration" are supposedly never contrastive within single language systems, and neither are "breathy voice" and "breathy release". This paper is not concerned with these claims (to which labialization and velarization could be added), but rather investigates whether these MOD types occur on natural classes of segments or not.
(3.) Maddieson codes glottalized sounds as laryngealized (see Maddieson 1984:99-101).
(4.) Apart from vowels, this segmental modification type occurs only on clicks. See the last but one paragraph of section 2.1 in the main text on remarks on clicks, as well as the final paragraph of section 4.1.
(5.) A further investigation of Nootka (native name: Nuuchanulth), based on the original sources as well as other, more recent work, is presented in van de Weijer and Hinskens (forthcoming).
(6.) The sources for Yurak mentioned by Maddieson do not agree on the status of some of the palatalized consonants: they may either be phonologically or phonetically palatalized (see Decsy 1966: 72). This difference does not affect our basic point.
(7.) One suspects that the /f, v/ pair may be historically related to the (oddly missing) aspirated labial plosive.
(8.) An anonymous reviewer points out that Hausa specialists usually regard Hausa as having a single "glottalized" series, realized as ejective or glottalized implosive according to place of articulation. The underlying feature would thus be [constricted glottis] in this analysis. The laryngealized /s/ varies with /ts/ and is usually included among the stops. Laryngealized /j/, deriving from historical * [d.sup.j], is sometimes also included among the stops as it is still realized as a stop in some dialects.
(9.) These languages belong to the Indic family (Indo-European group), Turkic family (Ural-Altaic group) and Pama-Nyungan family (Australian group), respectively.
(10.) Among the very few cases where they do, are pharyngealization in Nootka (language 730 in Maddieson's survey, Wakashan, Northern Amerindian), which occurs on the glottal stop only, and labialization in Kabardian (911, Caucasian, "other families").
(11.) In the case of Nez Perce (706, Northern Penutian, Northern Amerindian), the information on the stop or affricate nature of the apico-alveolars, and hence on the nature of a subset of the ejective consonants, is ambiguous. In the case of Ket (906, Paleo-Siberian, "other families"), it is impossible to determine whether the sibilants are fricatives or affricates, which makes it impossible to decide on the distribution of palatalization.
(12.) To give a few examples: the UPSID sample contains seven languages with breathy voiced consonants. With one exception (where it occurs on voiced plosives, voiced nasals and a voiced approximant), this type of MOD only occurs on voiced plosives. In all cases the consonants with MOD have "plain" counterparts; these counterparts are invariably voiced as well, so the voicing in the consonants with MOD does not appear to result from assimilation. There are eighteen languages in the sample with prenasalized consonants. In thirteen cases, the consonants with MOD have "plain" counterparts; in the four remaining cases only some of the consonants with MOD have "plain" counterparts. In eleven eases the "plain" counterparts are voiceless, in two cases they are voiced, and in five cases there are both voiced and voiceless counterparts. One of these seventeen languages has both voiced and voiceless nasals, three languages have no nasals at all, and all remaining languages have only voiced nasals. Here, at least in a considerable number of cases, pure phonetics again does not seem to allow predictions as to the distribution of MOD.
(13.) Recall that MOD on affricates does not play a role in the question whether hypothesis 1 was upheld or not, in order to keep the evaluation of hypothesis 1 separate from hypothesis 2 (see section 2.1 after ).
(14.) Namely Chipewyan (703, Athabaskan, Northern Amerindian), Amuesha (824, Arawakan, Southern Amerindian) and Nama (913, Khoisan, "other families"). The first two of these languages also have (other) MOD types on nonaffricate segments.
(15.) The latter category consists mainly of the cases where MOD occurs on a single consonant only (section 2.1),
(16.) Inferential statistics, i.e. the type of statistics that permits one to generalize the findings for a sample of objects (such as languages) to the entire "population" of objects (in casu all natural languages) presupposes the sample to be drawn randomly. This is not the case for the UPSID sample of languages described in Maddieson (1984), which is a quota sample. "The quota rule is that only one language may be included from each small family grouping [...]. Each such small family grouping should be represented by the inclusion of one language. Availability and quality of descriptions are factors in determining which language to include from within a group, but such factors as the number of speakers and the phonological peculiarity of the language are not considered" (Maddieson 1984: 5-6). So to the extent that the (documented) languages of the world form the population of objects, the sample is not random; however, if phoneme inventories constitute the population, the sample can be said to have been drawn randomly.
(17.) The two averages for the morn frequently occurring types of MOD are 52.7 and 20.0, respectively.
(18.) Relevant as far as the three categories +, [+ or -] and--goes, that is, relevant to hypothesis 1.
(19.) But compare Chomsky and Halle's (1968) feature [-anterior], since abandoned for this purpose.
(20.) As well as on the coronal nasal /n/, so that this case represents a "mild" counterexample to our hypothesis 1.
(21.) The quotation marks around the coronals indicate that the source was not specific on the place of articulation: this might either be dental or alveolar (see Maddieson 1984: 31f., 164f.),
(22.) It obviously would make no difference to the results of our analyses of MOD in affricates if Lombardi's (1990) proposal were adopted, in which affricates have two monovalent features [stop] (instead of [-cont] and [cont] (instead of [+cont]).
(23.) In the case of Wichita, there is also an isolated velar plosive with secondary ejectivity.
(24.) The UPSID sample contains only 26 instances of MOD on fricatives as against 160 on stops (cf. Table 7). This, however, does not necessarily mean that subhypothesis 2b has a smaller chance of being supported than subhypothesis 2a. After all, smaller numbers of occurrence of MOD on fricatives (than on stops) does not imply smaller chances for the affricates to show the same MOD patterning as the fricatives (than for affricates to show the same patterning as stops). The simultaneous occurrence of MOD in fricatives and affricates has been analyzed independently of that in stops and affricates. Moreover, the subhypotheses have been tested on the basis of the proportions of the number of cases in which instances of MOD in affricates do show the same patterning as that in fricatives (or stops) and the number of cases in which instances of MOD in affricates do not show the same patterning as that in fricatives (or stops). The statistical test applied calibrates for large differences in absolute numbers.
(25.) For the sake of briefness, the figures are not presented here.
(26.) Of the 19 Australian languages in the UPSID sample only two have affricates.
(27.) Assuming that the contrasts between the dentals and the palatals are really a matter of differences in the primary place of articulation; we did not have the means to verify Maddieson's (1984) analysis of this language.
(28.) Cf. Kiparsky's (1988: 380) observation that the coalescence of a sequence of feature bundles on different tiers into a single feature bundle "is a common source of complex segments, particularly secondary articulations such as glottalization [...], aspiration [...], and, for vowels, nasalization [...], breathy voice, etc."
(29.) A closely related process is the one which makes aspirated voiceless stops shift to fricatives. Iverson and Salmons (1995: 389-390) discuss phonetic research into this phenomenon.
(30.) Or implosive, according to Salmons 1993 (after Job 1995: 239, 247-248).
(31.) As in the case of Tolowa, summarized in example (10). Of the 51 languages with ejectivization in the UPSID sample, only !Xu (918, Khoisan, "other families") has an ejectivized voiced labial stop.
Basboll, Hans (1985). Review article of Ian Maddieson, 1984, Patterns of Sounds. Phonology Yearbook 2, 343-353.
Chomsky, Noam; and Halle, Morris (1968). The Sound Pattern of English. Harper and Row: New York.
Clements, George N. (1991). Place of articulation in consonants and vowels: a unified theory. Working Papers of the Cornell Phonetics Laboratory 5, 77-123.
--(1999). Affricates as non-contoured stops. In Item Order in Language and Speech, O. Fujimura, B. J. Joseph, and B. Palek (eds.), 271-300. Prague: Charles University Press.
--; and Hume, Elizabeth V. (1995). The internal organization of speech sounds. In The Handbook of Phonological Theory, J. Goldsmith (ed.), 245-306. Oxford: Blackwell.
Decsy, Gyula (1966). Yurak Chrestomathy. The Hague: Mouton.
De Jong, Kenneth; and Gyasi Obeng, Samuel (2000). Labio-palatalization in Twi: contrastive, quantal, and organizational factors producing an uncommon sound. Language 76, 682-703.
Fallon, Paul (1998). Debuccalization. Paper presented at the Meeting of the Chicago Linguistic Society.
Gamkrelidze, Thomas; and Ivanov, V. (1973). Sprachtypologie und die Rekonstruktion des gemeingermanischen Verschlusse. Phonetica 27, 150-156.
Halle, Morris (1992). Phonological features. In International Encyclopedia of Linguistics, vol. 3, W. Bright (ed.), 207-212. Oxford: Oxford University Press.
Hock, Hans Henrich (1991). Principles of Historical Linguistics, 2nd ed. Berlin: Mouton de Gruyter.
Hyman, Larry (1975). Phonology: Theory and Analysis. New York: Holt, Rinehart and Winston.
Iverson, Gregory; and Salmons, Joseph (1995). Aspiration and laryngeal representation in Germanic. Phonology 12, 369-396.
--(1996). Mixtec prenasalization as hypervoicing. International Journal of American Linguistics 62, 165-175.
Jakobson, R.; Fant, C. G. M.; and Halle, M. (1952). Preliminaries to Speech Analysis--The Distinctive Features and their Correlates, 2nd ed. Cambridge, MA: MIT Press.
Job, Michael (1995). Did Proto-Indo-European have glottalized stops? Diachronica XII(2), 237-250.
Kiparsky, Paul (1988). Phonological change. In Linguistics: The Cambridge Survey, vol. I, Frederick J. Newmeyer (ed.), 363-415. Cambridge: Cambridge University Press.
Ladefoged, Peter; and Maddieson, Ian (1995). The Sounds of the World's Languages. Cambridge, MA: Blackwell.
Lass, Roger (1984). Phonology. Cambridge: Cambridge University Press.
Laver, John (1994). Principles of Phonetics. Cambridge: Cambridge University Press.
Lombardi, Linda (1990). The nonlinear organization of the affricate. Natural Language and Linguistic Theory 8, 375-425.
Maddieson, Ian (1984). Patterns of Sounds. Cambridge: Cambridge University Press.
Ohala, John (1983). The origin of sound patterns in vocal tract constraints. In The Production of Speech, P. MacNeilage (ed.), 189-216. New York: Springer Verlag.
Padgett, Jaye (1991). Stricture in feature geometry. Unpublished PhD. dissertation, University of Massachusetts, Amherst.
Prince, Alan; and Smolensky, Paul (1993). Optimality theory: Constraint interaction in generative grammar. Unpublished manuscript, Rutgers University, NJ, and University of Colorado at Boulder.
Pulleyblank, Douglas (1998). Optimality theory and features. In Optimality Theory, An Overview, Diana Archangeli and Terrence Langendoen (eds.), 59-101. Oxford: Blackwell.
Sagey, Elizabeth (1986). The representation of features and relations in non-linear phonology. Unpublished PhD. dissertation, MIT, Cambridge, MA.
Salmons, J. C. (1993). The Glottalic Theory: Survey and Synthesis. Journal of Indo-European Studies Monograph 10. Institute for the Study of Man, McLean.
Sapir, Edward (1912). The Takelma language of southwestern Oregon. In The Handbook of American Indian Languages, Franz Boas (ed.), 1-296. Washington D.C.: Bureau of American Ethnology.
--; and Swadesh, M. (1955). Native accounts of Nootka ethnography. International Journal of American Linguistics 21 (4 part II).
Steriade, Donca (1993). Closure, release and nasal contours. In Nasals, Nasalization, and the Velum, Marie Huffman and Rena Krakow (eds.), 401-470. New York: Academic Press.
Stevens, K. N.; and Keyser, S. J. (1989). Primary features and their enhancement in consonants. Language 65, 81-106.
Trubetzkoy, Nikolaj (1969). Principles of Phonology. English translation by C. A. M. Baltaxe. Berkeley: University of California Press.
van de Weijer, Jeroen (1996). Segmental Structure and Complex Segments. Tubingen: Niemeyer.
--; and Hinskens, Frans (forthcoming). Patterns of segmental modification in consonant inventories: contrastive vs. redundant systems. In Proceedings of LP 2002, B. Palek, O. Fujimora, and S. Haraguchi (eds.). Karolinum Press, Prague and Meikai University Press.
Watkins, Calvert (1992). Indo-European languages. In International Encyclopedia of Linguistics, William Bright (ed.), 206-212. New York/Oxford: Oxford University Press.
|Printer friendly Cite/link Email Feedback|
|Author:||Hinskens, Frans; Van De Weijer, Jeroen|
|Publication:||Linguistics: an interdisciplinary journal of the language sciences|
|Date:||Nov 1, 2003|
|Previous Article:||Modern Spanish "spurious" se: a genuine case of pseudo analysis.|
|Next Article:||Prosodically constrained postverbal PPs in Mandarin Chinese *.|