1 Introduction
Allomorphy, roughly defined as the situation in which a given feature is spelled out as two or more distinct exponents (see Aronoff Reference Aronoff1976, Anderson Reference Anderson1992, Embick Reference Embick2010, Bonet & Harbour Reference Bonet, Harbour and Trommer2012, Trommer Reference Trommer2012 among many others), remains an intriguing phenomenon in many theoretical frameworks. Although morphological theories propose radically different approaches to allomorphy, there is a general consensus on the fact that the systematic vs. the unsystematic application of a phonological rule plays a crucial role in determining the limits of this phenomenon (see Kiparsky Reference Kiparsky and Singh1996). Thus, the traditional view distinguishes between phonologically conditioned allomorphy and morphologically conditioned allomorphy. Phonologically conditioned allomorphy is predictable from the phonological rules of the language while morphologically conditioned allomorphy is not.
Somali, an Afroasiatic language belonging to the Lowland Eastern branch of the Cushitic group, is characterized by a rich inflectional morphology in both its nominal and verbal systems. Nouns are divided in various inflectional classes on the basis of a few morpho-phonological parameters (Andrzejewski Reference Andrzejewski1964, Reference Andrzejewski1979; Puglielli & Siyaad Reference Puglielli and Siyaad1984; Banti Reference Banti, van der Hulst and Smith1988; Saeed Reference Saeed1993, Reference Saeed1999; Orwin Reference Orwin1995; Godon Reference Godon1998). Verbs are traditionally cast in three main conjugations (Saeed Reference Saeed1993, Orwin Reference Orwin1995) plus a fourth conjugation, also known as the ‘hybrid conjugation’ (Andrzejewski Reference Andrzejewski1969, Banti Reference Banti2012, Puglielli & Mansuur Reference Puglielli and Mansuur2012). These conjugations correspond to different verb classes, characterized by specific suffixes (e.g. causative, autobenefactive). It is assumed that these conjugations must be distinguished because the relevant suffixes trigger a morphologically conditioned allomorphy. In other words, allomorphy is taken to have a central role in the Somali verb morphology: it has a classificatory function, leading to the distinction of three main inflectional classes.
In this paper, we argue against this view. More precisely, we show that alleged cases of morphologically conditioned allomorphy can be reduced to phonologically conditioned allomorphy: surface forms result from the application of regular phonological rules to underlying structures. As a consequence, the notion of conjugation can be dispensed with. This analysis relies on a careful examination of the internal phonological structure of the relevant verbal suffixes. Endorsing an approach that decomposes the morphemes ‘all the way down’, we propose articulated phonological representations, and show how general phonological rules apply to these representations.
Our analysis shows that abstract phonological representations have a direct advantage on the way allomorphy can be analyzed. It also has a consequence on the status of ‘paradigms’ as defined by wide morphological literature (Aronoff Reference Aronoff1994; Carstairs-McCarthy Reference Carstairs-McCarthy, Spencer and Zwicky1998, Reference Carstairs-McCarthy, Booij and van Marle2005; Blevins Reference Blevins2006, Reference Blevins and Baerman2015; and references therein). Current morphological theories discuss whether paradigms are active linguistic items. We participate in this debate in showing that the notion of paradigm is not relevant in the Somali verb system: contrarily to what appears at first sight, Somali displays a unique verbal conjugation.
Our work relies on data taken from the relevant standard literature as well as our own fieldwork with native speakers of Somali who have French as their second language. The analysis is couched within the general theoretical framework of Autosegmental phonology (Goldsmith Reference Goldsmith1978) and its particular versions known as Government Phonology (Kaye, Lowenstamm & Vergnaud Reference Kaye, Lowenstamm and Vergnaud1985, Reference Kaye, Lowenstamm and Vergnaud1990) and CV-phonology (Lowenstamm Reference Lowenstamm, Durand and Laks1996).
The article is organized as follows: Section 2 introduces Somali verbal morphology as described in the literature. Sections 3 and 4 analyze conjugations 2a/b and 3a/b, respectively. We show that there is no need to consider each conjugation as a separate inflectional group. In Section 5, we conclude and present further issues raised by our analysis.
2 Somali verbal morphology
2.1 General information
Somali has a predominantly suffixal morphology. In particular, all inflectional markers (png (person, number, gender) and tam (tense, aspect, mood)) are suffixes. Any verb form can be decomposed as follows:Footnote [2]
As a representative example, consider the past and present tense paradigms of the verb keːn ‘bring’ in (2).
A form like 
(present, 2p) can be cast in template (1) as in (3).
There are two tam markers: -aj - (past) and 
(present). png markers are identical across tense categories, and appear in (4):Footnote
[4]
Finally, lex (lexical morpheme) corresponds to markers that derive a verb from a verbal, nominal or adjectival stem, and convey various grammatical values. Some representative examples are given in (5):
2.2 Three classes
Some lex markers trigger complex alternations, affecting both the stem and the png markers. For this reason, Somali verbs are traditionally divided into three main morphological classes, depending on the identity of lex and on its ability to trigger morphologically conditioned allomorphy:Footnote [5]
As an example, consider the verb mar ‘pass, tie up’. Mar can be suffixed by three distinct lex elements: causative -i and autobenefactiveFootnote
[6]
-so and -o , yielding the following three derived verbs: mari ‘make pass, rub with’ (class 2a), marso ‘finish off, consume, dress up’ (class 3a) and maro ‘be finished/empty, become used up’ (class 3b). The present tense paradigms of these verbs appear in (7). To complete the picture, we illustrate class 2b with the verb 
‘anger’, derived from the noun ʕaro ‘anger’.
The matrix in (7) gives, for each class, the citation form (imperative second person singular (or: imperative 2s)), the infinitive, the progressive presentFootnote [7] and the present tense paradigm. Past forms are parallel to present forms, the only difference being that they bear the past marker -aj instead of the present marker -aː . In particular, the allomorphic phenomena are identical in both tense paradigms. For this reason the data in (7) are sufficient for our purpose.Footnote [8]
2.3 Allomorphic alternations
Class 1 verbs differ from class 2 and 3 verbs in that they either have no lex marker or bear a lex marker that trigger predictable stem- and png-alternations. By contrast, class 2 and 3 verbs exhibit intriguing patterns of alternations that cannot at first sight be derived by the general phonological rules of the language.
2.3.1 Deviant patterns in class 1 are triggered by general phonological rules
In class 1, stem- and png-alternations derive exclusively from general phonological rules. We briefly review these rules below.
Consider first the verb hadal ‘talk’ (class 1) in (8a).
The 2s present form of this verb /hadal + t + aː/ surfaces as hadaʃaː , *hadaltaː . This results from the application of the following rule: /l + t/
$\rightarrow$
[ʃ]. As illustrated in (8b), this rule applies in various morphological contexts and is known as a very general rule of the language (Saeed Reference Saeed1993: 26, 301).
The second source of phonological allomorphy in class 1 results from the application of a voicing rule affecting the dental plosive tː /t/
$\rightarrow$
[d]Footnote
[9]
/ V_V. For instance, the 2s present form of the verb ʕi ‘cry out’ (class 1), /ʕi-t-aː/, surfaces as [ʕidaː], as seen in (9a).
Again, this rule applies in other contexts, too; for example, when the feminine article -ta is suffixed to a noun ending in a vowel, as in (9b) (Saeed Reference Saeed1993: 302).Footnote [10]
Third, class 1 CVCVC verb stems display vowel
${\sim}$
zero alternations across their present and past tense paradigms: they surface as CVCC when followed by a vowel and as CVCVC elsewhere, as seen in (10).
For instance, the second u of gudub ‘cross’ (class 1) appears in the citation form (gudub 2s imperative) and in forms involving a consonant-initial suffix (e.g. gudubtaː 2s present), only. The same vowel
${\sim}$
zero alternation occurs in the same context in nouns, as shown in (10b) (Saeed Reference Saeed1993: 27–28, 295).
The last two relevant phonological rules are neutralization rules. The first one targets final /t/ and /d/. The contrast /t/ vs. /d/ is neutralized in final position: both /t/ and /d/ surface as [d] in word-final position. For instance, /gunut/ ‘tie a knot’ (class 1) surfaces as gunt - when it is followed by a vocalic suffix (e.g. guntaː 1s present) but as gunud in final position, as in (11a), on a par with hurud ‘go to sleep’ (class 1) in (11b); see Saeed (Reference Saeed1993: 30).
The second one targets /m/ and /n/. Both /m/ and /n/ surface as [n] in word-final position and before consonants. Hence doːn ‘dredge’ (class 1, 2s imperative), doːntaː (2s present) vs. doːmaː (1s present). The same situation obtains in nouns, as shown in (12b) (Saeed Reference Saeed1993: 301).
Since the application of the rules above is fully predictable on the basis of the phonological context, the verbs of the type mentioned in this section are correctly analyzed as belonging to the same class – class 1.
2.3.2 Deviant patterns in classes 2 and 3 cannot be derived from general phonological rules
In classes 2 and 3, both lex and png morphemes display allomorphic alternations that cannot be easily derived from general phonological rules. The table in (13) summarizes the identity of the png markers involved in class 1, 2a and 2b, and 3a and 3b. png markers are identical within a given class: class 2a and 2b on the one hand, class 3a and 3b on the other hand, have the same png markers. The shaded cells contain the forms that are unexpected.
The png markers involved in class 2a/b differ from those of class 1 in the following respect:
At first sight, the png markers involved in class 3a/b are identical to those of class 1. This apparent identity, however, hides an intriguing fact:
We now turn to the allomorphy of lex. Class 1 verbs display no unexpected allomorphy. Thus, the table in (16) represents the neutral case. The shaded cells in (16) as well in all remaining tables correspond to forms that do not exist.
Classes 2a/b and 3a/b involve various allomorphs, whose segmentation is not straightforward.
Following the traditional view (presented in (6) above), we assume that lex = i in class 2a. Then, we isolate the png and tam markers and arrive at the table in (17). The column headed ‘?’ (between lex and png) contains the material that is left unattributed.
In class 2b, lex has two allomorphs: -eː- and -aj -, yielding the following segmentation:
The appearance of -j (j)- under ‘?’ in (17) and (18) could be considered a phonotactically conditioned j-insertion (hiatus resolution). Along these lines, we should not posit an allomorphy in 1s/3ms/3p. However, this cannot be right. There is indeed a clear contrast between i-final class 1 verbs, e.g. bari ‘spend the night’ and class 2a verbs, e.g. mar-i . Class 1 1s /bari-Ø-aː/ surfaces as [barjaː]. By contrast, class 2 1s /mar-i-Ø-aː/ may surface as [marjaː], [marijaː] and [marijjaː]. Crucially only class 2 1s may be realized as [marij(j)aː]. We take this fact to indicate that the underlying structure of class 2 1s is not parallel to that of class 1 1s: it involves allomorphy.
Turning to class 3, we present in (20) and (21) the distribution of the allomorphs of lex as well as the material left unidentified under ‘?’.
The tables in (17), (18), (20) and (21) set the terms of the debate: what is the relation between the material appearing under ‘?’ and the allomorphs of lex? We will take a position on this question in the following sections.
3 The allomorphy of the causative marker
We first focus on class 2a (causative -i) and class 2b (causative -eː) and propose a unique representation of the causative suffix in both classes, thus unifying classes 2a and 2b under a single inflectional class. We further show that the alternations observed on the surface result from the application of regular phonological principles to this underlying representation. Our line of reasoning necessitates first a detour into the status of -s in class 3a, in Section 3.1. Sections 3.2 and 3.3 are dedicated to the analysis of -i and -eː, respectively.Footnote [11]
3.1 The identity of -s in class 3a: Allomorphy of the causative suffix in derivational morphology
A comparison of the tables in (20) and (21) above shows that class 3a and class 3b basically differ in the presence vs. the absence of -s- in lex.Footnote [12] In order to gain insight into the status of this segment, let us consider again the three verbs derived from mar ‘pass, tie up’ in (22a): mari ‘make pass’ (causative, (22b)), maro ‘be finished/empy, become used up’ (autobenefactive, (22c)), and marso ‘finish off, consume, dress up’ in (22d):
Class 2a verbs (e.g. mari ) are derived from class 1 verbs by suffixation of -i , and they are causatives (23b). Class 3b verbs (e.g. maro ) are derived from class 1 verbs by suffixation of -o , and they are autobenefactives (23c). Class 3a verbs (e.g. marso ) are derived from class 1 verbs by suffixation of -so and they have both an autobenefactive and a causative meaning (23d). The relationship between maro and mar is identical to that between marso and mari : maro is the autobenefactive counterpart of mar , marso is the autobenefactive counterpart of mari . We conclude that -s in class 3a is an allomorph of the causative: autobenefactive verbs in -so are derived from causative verbs by ‘replacing’ -i by -s and adding autobenefactive -o :
The derivation of class 3a verbs on the basis of class 2a verbs is very productive; a representative set of examples is given in (24).
The derivational paths involving class 1, 2a, 3a/b verbs are summarized in (25):
An additional argument in favor of the analysis of -s as an allomorph of the causative comes from the behavior of CVCVC verbs, e.g. gudub ‘cross’. As mentioned above, in (10), these verbs display a vowel
${\sim}$
zero alternation conditioned by the context: imperative 2s gudub
${\sim}$
present 1s
gudbaː . Following Barillot (Reference Barillot2002), we assume the stem of such verbs to be /C
$_{1}$
VC
$_{2}$
C
$_{3}$
/. The stem vowel propagates onto the position between the second and the third stem-consonant when the phonotactic constraints of the language require this position to be identified. This is the case in two configurations: (i) if the verb stem is in final position (i.e. 2s imperative), e.g. gudub
${\sim}$
*gudb , CC# clusters are prohibited in Somali; and (ii) if the verb stem is followed by a C-initial suffix (i.e. 2s/3fs/1p/2p), e.g. gudubtaː
${\sim}$
*gudbtaː , CCC clusters are prohibited in Somali.Footnote
[13]
If the stem is followed by a V-initial suffix, no vowel needs to surface between the second and the third stem-consonants, e.g.
pres 1s gudbaː .
Given this distribution, we expect a class 2a causative derived from a CVCVC verb to have the following shape: CVCCi . The causative derived from gudub ‘cross’ is indeed gudbi ‘send across’. Consider now the autobenefactive derived from a CVCCi causative. We expect -so to attach to the stem /gudb/: *gudbso . The ban on CCC clusters prevents this form from surfacing. Two equally possible strategies are conceivable:
There are 57 class 3a verbs derived from a CVCVC stem (see Agostini, Puglielli & Siyaad Reference Agostini, Puglielli and Siyaad1985 and Zorc Reference Zorc1993). Fifty-two of them are of the type CV
$_{i}$
CV
$_{i}$
Cso, seen in (27a) below. The remaining forms are distributed as follows: four are CVCCiso verbs, illustrated in (27b), and there is one verb with both CV
$_{i}$
CV
$_{i}$
Cso and CVCCiso, with the same meaning, seen in (27c). We take the sequence -is- in (27b–c) to be an allomorph of the causative. This allomorph is selected to prevent a CCC cluster from surfacing.
A similar situation obtains in the derivation of stative verbs. Consider as an example marsan and maran in (28a): maran ‘be tied up’ is semantically related to mar ‘tie up’, whereas marsan ‘be rubbed with s.t.’ is built on mari ‘rub s.o. with s.t.’. Since -an is the stative suffix, the -s- of marsan is an allomorph of the causative suffix. The elements -s- in marsan and -s- in marso are one and the same allomorph of the causative suffix. In the derivation of statives from CVCC stems, the strategies illustrated in (27) above obtain, giving CV
$_{i}$
CV
$_{i}$
Csan in (28b), CVCCisan in (28c), and in one case, again, both strategies result in two forms with the same meaning, as shown in (28d).
To conclude, the causative suffix has three allomorphs, -i, -s- and -is-,Footnote [14] whose distribution is as follows:
We now have a complete description of the causative allomorphy and are in a position to propose a representation for this marker.
3.2 The representation of the causative marker
In this section, we argue that the underlying representation of the causative marker is /It/. First, we introduce Element Theory (Kaye et al. Reference Kaye, Lowenstamm and Vergnaud1985, Reference Kaye, Lowenstamm and Vergnaud1990), which we exploit to decompose the vocalic segments, and the CV-framework (Lowenstamm Reference Lowenstamm, Durand and Laks1996), which we adopt to represent the skeletal tier of phonological representations. Then we propose our representation of the causative marker. Finally, we focus on three inflected forms, 1s, 2s and 1p, which illustrate how the underlying material is computed into surface forms by the phonology of the language.
3.2.1 Element Theory and the representation of the causative marker
Element Theory (Kaye et al. Reference Kaye, Lowenstamm and Vergnaud1985, Reference Kaye, Lowenstamm and Vergnaud1990) is a theory of segmental representations. In this framework, segments are the surface realizations of underlying Elements, rather than features. Elements can be combined by an operation called ‘Fusion’. Fusion is an asymmetrical operation, involving a head (underlined in the representations) and an operator (Kaye et al. Reference Kaye, Lowenstamm and Vergnaud1985: 309). Following Backley (Reference Backley2011), we assume the Elements associated with vowel structure to be |A|, |I| and |U|. The internal structure of consonants will not be relevant in this article; we thus simply posit the segment, as a shortcut to a more complex underlying Element structure. Abstracting away from harmony phenomena, which are irrelevant for our purposes, we derive the Somali underlying five-vowel system as in (30); see Backley (Reference Backley2011: 41).
As for the skeletal level of phonological representations, we adopt the CV-framework (Lowenstamm Reference Lowenstamm, Durand and Laks1996, Scheer Reference Scheer2004), according to which the skeletal tier consists of a strict alternation of non-branching nuclei (C) and non-branching onsets (V). In this framework, the distribution of empty V positions is constrained by the Empty Category Principle and Proper Government (pg), as defined in Kaye et al. (Reference Kaye, Lowenstamm and Vergnaud1990: 219) and subsequent work, e.g. Charette (Reference Charette1990: 236):
The notion of licensing has been discussed in various contexts (Charette Reference Charette1989, Reference Charette1991; for details within the CV-framework, see Scheer & Ségéral Reference Scheer and Ségéral2001a: 138). Licensing determines whether a segment may be realized or not: typically, an unlicensed segment has a very restricted phonetic expression, or none at all. Licensing was introduced to encode the observation that there is a dependency between the onset and the nucleus: the segmental expression of an onset crucially depends on the ability of its nucleus to license. More specifically, the ability of a nucleus to license is constrained as follows:
Finally, following Bendjaballah & Haiden (Reference Bendjaballah and Haiden2008) and Lowenstamm (Reference Lowenstamm2008), we assume that skeletal units can be the exponents of grammatical features. The minimal skeletal unit being CV, this means that a CV unit can be a morpheme.
Equipped with these tools, we argue that the Somali causative marker has the following representation:
This structure contains (i) the minimal skeletal unit, CV, and (ii) two objects at the segmental level: |I| and t. As can be seen from (34), we assume that |I| and t are floating, i.e. they are not lexically linked to the skeleton. The question is thus whether their association to the skeleton is constrained by some principles, and if it is, by which ones. We assume that the association of |I| is constrained by the Empty Category Principle and Proper Government. Let us now consider t. The table in (35) gives a survey of the distribution of [t] in Somali.
The table is based on an exhaustive examination of three of the most comprehensive Somali dictionaries: Agostini et al. (Reference Agostini, Puglielli and Siyaad1985; Somali–Italian, 36,276 items), Zorc (Reference Zorc1993; Somali–English, 25,987 items) and Keenadiid (1976; monolingual Somali, 14,497 items). It turns out that [t] is almost always followed by a vowel, as seen in (35a, b). [t] appears before a consonant or in final position in a small number of words, only (40 out of 37,000 words).Footnote
[15]
Moreover, all of them are loanwords (
$_{L}$
in (35b)), e.g. matħaf ‘museum’
${<}$
Arabic mutħaf, batroːl ‘petrol’
${<}$
English.
We thus assume that [t] must be followed by a vowel in Somali. We take this distributional observation to indicate that, in order to be linked to a C position, t must be followed by a phonetically interpreted V position.Footnote [16] In our framework, this generalization is formulated in (36):Footnote [17]
We now review the different configurations involving caus and show how the surface forms are derived.
3.2.2 Causative allomorphy (I): Derivational morphology
The causative marker appears either in isolation or before the autobenefactive marker -o. In isolation, caus always surfaces as -i:
The three configurations exemplified in (37) are illustrated in (38).
In all cases, the final V position of the verb stem is not properly governed, and it is identified by caus |I|. The segment t cannot be associated with the skeleton: it is not followed by a vowel, and is not licensed.
If caus is followed by the autobenefactive marker, it surfaces either as -s-, as in (39a), or as -is-, as in (39b):
We start with the derivation of caus + autoben from a CV(V)C base in (39a). As can be seen in (40), -o (resulting from A U) identifies the final V position.Footnote [18]
This has two consequences: (i) this position licenses t to be associated with the preceding C position, and (ii) it governs the penultimate V position (V
$_{3}$
) with the effect that |I| does not need to be associated with that position. We claim that t is palatalized to [s] by underlying |I|, and caus surfaces as [s].
Consider now the derivation of caus + autoben from a CVCVC base in (39b) above, as illustrated in (41).
As in (40), -o on the final V position (V
$_{4}$
) licenses t to be associated with C
$_{4}$
. It also governs V
$_{3}$
. In this configuration, two options are available: either V
$_{3}$
, being governed, remains empty, and V
$_{2}$
must be identified, yielding qurunso , in (41a), or |I| is associated with V
$_{3}$
, and V
$_{2}$
remains empty, which yields ergiso , in (41b). The former option consists in the propagation of the stem vowel between the last two stem-consonants and is attested in 52 cases (e.g. qurunso ‘cause to rot’, (26a) and (27a) above). The latter option is attested in only four verbs (e.g. ergiso ‘take s.t. in trust’). Finally, one verb displays both strategies: koroɖso , korɖiso ‘increase for o.s.’.Footnote
[19]
Before turning to the allomorphy of caus in inflectional morphology, we believe that it is important to reaffirm our hypothesis on how underlying /t/ is palatalized in (41b). An anonymous JL referee points out that, since [it] sequences are attested in Somali, |I| cannot both palatalize and be associated with the skeleton. This would imply that only floating |I| can palatalize and (41b) is not well-formed. We do not agree with this conclusion. First, note that the strategy presented in (41b) involves four verbs only. Secondly, an exhaustive count of the lexical entries given in two reference dictionaries (Agostini et al. Reference Agostini, Puglielli and Siyaad1985, Zorc Reference Zorc1993) reveals that there are 108 occurrences of [it] in Somali. These 108 occurrences can be divided into three groups:
-
(i) In 57 cases, -i- and -t - are heteromorphemic: 29 sequences involve the concatenation of -i and the derivational suffix -taːn , e.g. fur ‘open’ vs. furitaːn ‘(act of) opening’; 15 sequences result from several derivational processes, e.g. baħ ‘exit, leave, go out’, biħi ‘take out, remove’ and biħitin ‘leave, make a trip’; and 13 sequences correspond to compounds, e.g. rabbituːg ‘prayer to God’
${<}$
rabbi ‘Master, the Lord’ + tuːg ‘pray’. -
(ii) In 32 cases, the forms are clearly loanwords, e.g. kitaːb ‘book’
${<}$
Arabic kitaːb , (ʕ )isbitaːl ‘hospital’
${<}$
English hospital . -
(iii) Finally, there are 19 items in which the sequence -it - does not seem to be analyzable in terms of morphological structure, e.g. abitej ‘kite’, fiːto ‘pimple, wound’ and hitiq ‘walk slowly’.
In (i), a morphological boundary intervenes between -i and -t. Such is not the case in the causative suffix. We assume that palatalization is blocked by the presence of a morphological boundary in group (i). As for group (ii), it is well-known that, under certain conditions, loanwords resist full integration into the native language phonology (see, among many others, Kang Reference Kang, van Oostendorp, Colin, Hume and Rice2011: 2261, and references therein). Finally, we are left with 19 monomorphemic occurrences of [it], for which we have no explanation for the absence of palatalization. Such a situation is not uncommon. Consider Italian, for example. In Italian, palatalization is triggered by the plural suffix -i and affects velars, as in e.g. [amiːko] ‘friend.m.sg’ vs. [amiːtʃi] ‘friend.m.pl’. However, in certain comparable structures, palatalization does not apply, e.g. [pikko] ‘peak.m.sg’ and [pikki] ‘peak.m.pl’. In Italian as in Somali, a regular phonological process (palatalization) may not apply under certain (lexical) conditions. In Somali, one possibility would be to assume that this condition is the presence of inherent skeletal material. The causative suffix introduces its own templatic space, the CV unit in (34) above (Bendjaballah Reference Bendjaballah1998, Reference Bendjaballah1999: 191ff.). In (40), as in (41), |I| associates with /t/, yielding /s/. In both configurations, C
$_{4}$
is licensed by -o in V
$_{4}$
. As a consequence /s/ is linked to C
$_{4}$
. In (40) and (41a), |I| remains floating. In (41b), it is associated with V
$_{3}$
and surfaces as [i]. The difference between (40) and (41a) on the one hand and (41b) on the other hand is that in (41b) [i] must surface because its absence would result in an ungrammatical CCC cluster. Both configurations are consistent with previous analyses of palatalization in Government Phonology and CVCV (see e.g. among many others, Charette Reference Charette1989, Scheer & Ségéral Reference Scheer and Ségéral2001b, Cristófaro Silva Reference Cristófaro Silva and Ploch2003). The same referee seems to suggest that only floating |I| may palatalize. If we followed this suggestion, we would have to claim that -i - in ergiso is an epenthetic vowel (i.e. it would be similar to [i] in loanwords: ʃifti
${<}$
shift (English), waqti
${<}$
waqt ‘time’ (Arabic)). However, it is obviously not the case that in Somali only floating |I| may palatalize. Consider for instance the pairs noːg ‘be tired’
${\sim}$
noːtʃi ‘tire’ and daːq ‘graze’
${\sim}$
daːtʃi ‘make graze’. In all comparable contexts, both palatalization of
$\{\text{g},\text{q}\}$
to [tʃ] and realization of causative -i are observed (see Bendjaballah Reference Bendjaballah1998, Reference Bendjaballah1999 for analysis). We therefore maintain our analysis of -i - in ergiso as the causative -i .
3.2.3 Causative allomorphy (II): Inflectional morphology
Recall from Section 2.3.2 above that allomorphy involves 1s/3ms/3p -j (j )- and 1p -n - in the column labeled ‘?’ in the tables in (17) and (18). We select the following representative forms 1s, 2s and 1p and show that -j (j )- and -n - result from the application of regular phonology in the linearization process.Footnote [20]
We start with 1p karinnaː ‘we cook s.t.’ in (42).
The stem kar is followed by caus /It/, followed by png n and tam |A|. We assume each of the two categories png and tam to be equipped with a single CV unit (see the typology of spell-out proposed by Bendjaballah & Haiden Reference Bendjaballah and Haiden2008). The V position of caus, V
$_{3}$
, is empty. As a consequence (i) t cannot associate with its docking site C
$_{3}$
, and (ii) since V
$_{2}$
is not properly governed, |I| associates with V
$_{2}$
. We are left with an empty C
$_{3}$
V
$_{3}$
sequence, which is identified by the propagation of 1p
n to C
$_{3}$
.
Consider now 1s karij (j )aː ‘I cook s.t.’ in (43).
As in 1p, the V position of caus, V
$_{3}$
, is empty. As a consequence (i) t cannot associate with its docking site C
$_{3}$
, and (ii) since V
$_{2}$
is not properly governed, |I| associates with V
$_{2}$
. We are left with an empty C
$_{3}$
V
$_{3}$
C
$_{4}$
sequence. The only strategy making it possible to avoid such a sequence is the propagation of |I| to C
$_{3}$
and C
$_{4}$
, hence [jj].Footnote
[21]
Finally consider 2s karisaː ‘you cook s.t.’ in (44).
V
$_{3}$
being empty again, the same consequences obtain: (i) since the causative t is not licensed, it cannot associate with the skeleton, and (ii) since V
$_{2}$
is not properly governed, |I| associates with V
$_{2}$
. The crucial difference with (43) is that png is not zero, but t . This t associates with C
$_{4}$
(which is licensed by |A| in V
$_{4}$
), and propagates onto C
$_{3}$
. In other words, we have an underlying /tt/. /tt/ is palatalized by |I|, and surfaces as [s]. Note that [s] is a latent geminate in the sense that it is phonologically geminated, but phonetically realized as a single consonant (Ségéral & Scheer 2001).Footnote
[22]
The forms 1s, 2s and 1p are parallel in the sense that they involve an underlying geminate. The difference lies in the realization of this underlying geminate: it is always phonetically realized as a geminate in 1p ([karinnaː], *[karinaː]), it is optionally realized as a geminate in 1s ([karijjaː], [karijaː]) and it is never realized as a geminate in 2s ([karisaː], *[karissaː]). This result is coherent with the general behavior of n , j and s in Somali: [n] is always phonologically short, whereas [j] and [s] can be either short or long.
The evidence comes from a wide range of morphophonological processes in Somali: the pattern of V
${\sim}$
Ø alternations in CVCVC verbs, reduplication, compounding, behaviour of n + C clusters, loanwords, etc. (see Barillot Reference Barillot2002: 197–351 for a precise review and analysis). In this article we exemplify the reasoning on the basis of the behaviour of CVCVC verbs, e.g. gudub ‘cross’ in (45a).
Recall from Section 3.1 that these verbs are /CVCC/ at the phonological level. Some CVCVC verbs, however, resist V
${\sim}$
Ø alternation, e.g. feker ‘think’ in (45b). Building on the fact that CVC
$_{\text{i}}$
C
$_{\text{i}}$
VC verbs, e.g. beddel ‘change’ in (45c) never exhibit V
${\sim}$
Ø alternation, Barillot (Reference Barillot2002) argues that intervocalic k in feker is phonologically long: /fekker/. The medial geminate forces the lexical vowel to propagate. If it did not propagate, a banned CCC cluster would surface.Footnote
[23]
Examining the behavior of the complete set of CVC
$_{0}$
VC verbs in Agostini et al. (Reference Agostini, Puglielli and Siyaad1985) and Zorc (Reference Zorc1993), Barillot (Reference Barillot2002) shows that the phonological status of Somali consonants is as summarized in the table in (46). Some examples are given in (47).
It is thus clear that intervocalic [s] and [j] can be either phonologically simple or geminate: 2s karisaː and 1s karijaː exemplify /ss/ and /jj/, respectively.
To conclude, we argued that the allomorphic material in the column labeled ‘?’ in (17) above results from the addition of a CV unit by caus. It is fully predictable on the basis of representation (34). As such, allomorphy in class 2a does not qualify as a classificatory property.
3.3 The representation of -eː
3.3.1 Class 2b verbs
We now turn to class 2b verbs and show that class 2b and class 2a are one and the same group. According to Saeed (Reference Saeed1993: 61ff.; Reference Saeed1999: 74ff.) and Orwin (Reference Orwin1995: 60ff.), class 2b verbs are formed by adding the suffix -eː to nouns and adjectives:
Observe the data in (49).
Class 2b verbs are the base of the derivation of class 3a verbs: -eː is replaced by -aj - and the suffix -so is added. This situation is similar to that described for the derivation of class 3a verbs on the basis of class 2a verbs in (23) and (24) above.
Class 3a verbs have an additional autobenefactive value. We take this observation to indicate that class 3a verbs are built on class 2b verbs by suffixation of autobenefactive -o . In other words, -ajs - is the allomorph of class 2b -eː before autoben. The table in (50) shows the parallelism between class 2a and class 2b. Class 2b has -aj - wherever class 2a has -i -, with a phonologically conditioned alternation: lex of class 2b surfaces as 
3.3.2 The decomposition of -eː
Let us start with the phonological makeup of class 2b -eː . According to Element Theory, e results from the fusion of Elements A and I: e = | A I|. The representation of -eː /aj is thus as follows:
We propose to analyze class 2b -eː as causative /It/ + floating |A|. In other words, class 2b verbs differ from class 2a verbs only with respect to the presence of |A|.
This representation immediately raises two questions: (i) Why do class 2b verbs have an additional |A|? and (ii) Where does the extra CV position in (51) come from?
Recall that -eː selects for nouns and adjectives and transforms the item it attaches to into a verb. We therefore propose to analyze the additional |A| in class 2b as the exponent of verbality (v in (53)–(58)).Footnote [24] This exponent brings its own CV unit, and the final representation of -eː /aj is as follows:
The item eː /aj differs from caus in the presence of the exponent of verbality only. Our analysis correctly describes the fact that
-
∙ the only way to derive a causative verb from an adjective is to add -eː as -i cannot be used in this case;
-
∙ it is impossible to derive an autobenefactive from an adjective, e.g. adag ‘strong, hard’
${>}$
*adko (class 3b) is ungrammatical.
This is precisely because |A| is necessary in the derivation.
3.3.3 Causative allomorphy (I): Derivational morphology
The representation of a class 2b verb involves the root, the verbalizer |A| and the caus suffix.
Consider first the case where class 2b -eː is not followed by another derivational suffix in (54).
The causative t cannot be associated because its potential docking site, C
$_{5}$
, is not licensed. |A| and |I| fuse and associate with V
$_{3}$
and V
$_{4}$
. The final empty CV, being unidentified, is not taken into account in the linearization process.
If caus is followed by the autobenefactive suffix, class 2b surfaces as -ajs - (55).
The suffix -o on the final V position (V
$_{5}$
) licenses t to be associated with C
$_{5}$
. As in (40) and (41) above, |I| palatalizes t to [s]. At the present time we have no explanation for the fact that class 2b suffix surfaces as -aj - when followed by C(V). We encode this observation by associating |A| to V
$_{3}$
and |I| to V
$_{4}$
instead of fusing |A| and |I| as in (54) above.Footnote
[25]
Since V
$_{4}$
is properly governed by V
$_{5}$
, it does not need to be identified. As a result, an entire CV unit (C
$_{4}$
V
$_{4}$
) remains empty. We assume such a structure is not well-formed: C
$_{4}$
V
$_{4}$
spells out verbality (it is the exponent of the v head), it must be identified and |I| associates with C
$_{4}$
.Footnote
[26]
3.3.4 Causative allomorphy (II): Inflectional morphology
We now turn to the representations of the inflected forms and take as a representative example the verb ʕareː ‘anger, make angry’. As above for class 2a, we focus on 1s, 2s and 1p.
We start with 1p. The basic ingredients are the root ʕar , the verbalizer |A|, the causative marker /It/, png -n- and tam |A|. The underlying sequence is shown in (56a), whereas, in (56b), we illustrate the linearization and the computation of the surface form.
The process is parallel to that illustrated for 2a causatives in (42) above. The V position of caus, V
$_{4}$
, is empty and, as a consequence, (i) caus t is not licensed and cannot associate with C
$_{4}$
, and (ii) V
$_{2}$
and V
$_{3}$
are not properly governed. |A| associates with V
$_{2}$
, whereas |I| associates with V
$_{3}$
. n identifies C
$_{4}$
V
$_{4}$
by propagation on C
$_{4}$
.
The representation of 1s ʕareːj(j)aː is shown in (57): the form after concatenation of the relevant morphemes appears in (57a), and (57b) shows how the segmental make up of the form is associated with the skelettal tier, yielding to phonetic interpretation.
V
$_{4}$
is empty. As a consequence, caus
t cannot associate with C
$_{4}$
. V
$_{3}$
and V
$_{2}$
are not phonologically governed. |A| associates with V
$_{2}$
, whereas |I| associates with V
$_{3}$
. We are left with an empty C
$_{4}$
V
$_{4}$
C
$_{5}$
sequence. The only possible repairing strategy is the propagation of |I| to C
$_{4}$
and C
$_{5}$
. |I| palatalizes preceding |A| into surface [e].
Finally, 2s is made of the same ingredients, with the exception of the person/number marker, which is -t -, as can be seen in (58).
In this configuration, again, V
$_{4}$
is empty and does not license C
$_{4}$
. As a consequence, caus
t cannot associate with its template. By contrast, since V
$_{5}$
is identified by tam A, 2s
t does associate with C
$_{5}$
and spreads onto C
$_{4}$
. Finally, |I| palatalizes underlying t , which surfaces as [s].
3.4 Conclusion
In this section, we have shown the complexity of the alternations involved in the two causative verb classes. We argued that the underlying representation of the causative marker is /It/ – recall (34) above. Regular phonological processes apply to the concatenation of the base and this suffix, which is linearized into surface forms through the principles of Government Phonology. We propose to analyze -eː of class 2b as involving the same causative marker /It/, plus an additional |A|, which, as we argue in Section 3.3.2, is a verbalizer. As a consequence, there is no need for a distinction between three inflectional verb classes 1, 2a and 2b.
4 Classes 3a and 3b
We now turn to the analysis of classes 3a and 3b, the two ‘autobenefactive’ classes. Our aim is to propose a representation that accounts for the various surface realizations of the autobenefactive suffix and show that the allomorphy observed in these classes can be reduced to regular phonological processes, hence there is no need for the assumption of a specific ‘autobenefactive’ class.
4.1 The autobenefactive marker
The analysis presented in this section is based on Barillot (Reference Barillot2002) and Barillot & Ségéral (Reference Barillot and Ségéral2005). Autobenefactive verbs are divided into two classes depending on the form of the suffix attached to the basic class 1 verb, -so (class 3a) or -o (class 3b). As shown in Section 3.1 above, class 3a verbs are derived from causative verbs and -so must be analyzed as caus -s - + autoben -o . Class 3b verbs are derived from class 1 verbs by suffixation of autoben -o . Given this analysis, the allomorphs of autoben appear in bold face in (59).
The suffix -o appears only in the imperative 2s, the citation form of verbs in Somali dictionaries and grammars. The various allomorphs of autoben, presented above in (20) and (21), are recast in (60): we abstract away from caus -s - in class 3a and filter out png morphemes.
This leads us to include both -d - (1s/3ms/3p, class 3a) and -n - (infinitive/progressive/1p, both classes) into lex. autoben is composed of a vowel -a -, which alternates with Ø in class 3b (1s/3ms/3p), and of a coronal consonant (t , d or n ). This exhausts the possibilities since no other derivational suffix may appear to the right of autoben.
The table in (60) reveals that autoben displays a very similar allomorphy in both classes: the only difference lies in 1s/3ms/3p forms, which are shaded in gray. There, two differences are found: (i) the vowel -a - appears only in class 3a, and (ii) class 3a has -d - while class 3b has -t -.
Let us start with the a
${\sim}$
Ø alternation, e.g. 1s class 3a marsadaː ‘I finish off, consume, dress up’
${\sim}$
class 3b martaː ‘I am finished/empty, become used up’. This alternation has to be compared with that in (61), involved in CVCVC class 1 verbs, e.g. gudub ‘cross’ (recall example (10a) above):
This alternation stems from the fact that CCC and CC# clusters are prohibited in Somali. To avoid such clusters, the preceding stem vowel u propagates in the relevant contexts. The a
${\sim}$
Ø alternation in autoben has the same cause: -a - surfaces to avoid either a *CCC or a *CC# cluster. In class 3b 1s, mar-t-aː ‘I am finished/empty, become used up’ does not contain any putative CCC cluster and -a- does not appear (*mar-at-aː ). In class 3a 1s, by contrast, mar-s-ad-aː ‘I finish off, consume, dress up’ contains a putative CCC cluster, *marsd-aː , and -a - shows up. In the framework of Government Phonology, this type of V
${\sim}$
Ø alternation regularly derives from Proper Government relations: a does not surface in martaː because V
$_{2}$
is properly governed by V
$_{3}$
, as seen in (62a).
By contrast, in /mar-s-d-aː/ in (62b), either V
$_{2}$
or V
$_{3}$
has to be identified in order for the representation to be well-formed. Of these two conceivable options, it is the second one which is applied.Footnote
[27]
We conclude that the a
${\sim}$
Ø alternation opposing class 3a/b verbs is strictly conditioned by the phonological context: no unpredictable allomorphy is involved.Footnote
[28]
We can now turn to the representation of autoben. The alternating vowel in class 3a/b is a . We propose that this | A| belongs to autoben. Moreover, since it alternates with Ø, it has to be considered as a floating vowel. Now let us turn to the identity of the underlying consonant: it must be either d or t . We propose it is t , and the t
${\sim}$
d alternation derives from a general voicing rule stated in (9) above, /t/
$\rightarrow$
[d] /
V_V:
To sum up: the alleged allomorphy in class 3a/b 1s/3ms/3p forms (marsadaː ‘I finish off, consume, dress up’ vs. martaː ’I am finished/empty, become used up’) is accounted for by general phonological rules. The underlying representation of autoben in (63) is the same in both classes. It includes two parts: t , which is regularly voiced in intervocalic position, and |A|, which regularly surfaces when its absence would lead to a prohibited consonant cluster.
At first sight, this proposal seems to be immediately contravened by the 2s/3fs/2p forms of both classes. In these forms, (i) -t - surfaces in intervocalic position, and (ii) in class 3b, the floating vowel -a - surfaces even if it is preceded by a single consonant: its appearance cannot be ascribed to the necessity of preventing a prohibited CCC cluster:
However, as we turn to show, the second -a - in marataː does appear in the relevant context: C_CC.
4.2 Virtual geminates
The underlying structure of marataː ‘you tie up for yourself’ is as follows:
The presence of -a - is accounted for: it is due to the /tt/ cluster to its right. The floating vowel -a - surfaces to break up a CCC cluster (r-t-t). Moreover, it explains why intervocalic [t] is not voiced: it is a geminate. In other words, /t/
$\rightarrow$
[d] / V_V, and /tt/
$\rightarrow$
[t] / V_V.Footnote
[29]
The representations of class 3b, 2s marataː and class 3a, 2s marsataː are given in (66a) and (66b), respectively.
An interesting by-product of this analysis comes from the examination of a particular group of class 3b verbs. Recall from (46) above that intervocalic k , ʃ and w are underlying geminates, like t . We predict that if CVko , CVʃo , CVwo autobenefactive verbs exist, they will not behave like maro in (67a), but rather like marso in (67b) because their last consonant is phonologically long. This is exactly what we observe in (67c):
4.3 1P and imperative 2S
We are left with two cases of allomorphy: (i) 1p in which autoben surfaces as -an -, and (ii) 2s imp in which autoben surfaces as -o . Let us first consider -an - in 1p (marannaː ‘we are finished/empty, become used up’), infinitive (maran
‘be finished/empty, become used up’) and progressive forms (maranajaː ‘I am becoming used up’). The underlying structure of 1p can be broken down into four parts, and represented as in (68a): V
$_{3}$
being empty, it does not properly govern V
$_{2}$
nor does it license C
$_{3}$
.
This has two consequences: (i) V
$_{2}$
must be identified – autoben |A| associates with this position, and (ii) t cannot be associated with C
$_{3}$
. 1p -n - spreads and identifies C
$_{3}$
. (68b) illustrates the same 1p form for class 3a marsannaː . The only difference lies in the presence of caus to the left of autoben, which surfaces as [s].
Consider now the 2s imperative. Why does autoben surface as -o in this form, and only in this form? In order to answer this question, we have to understand the distribution of final -o s in Somali. For that purpose, consider the plural marker -o in (69a–c).
When the determiner (either [ha] masc, or [da] f) is suffixed to a plural noun ending in -o , this vowel alternates with -a -, as seen in (69a–c). The same alternation takes place in singular nouns ending in -o , seen in (69d).
We take this alternation to indicate that final short -a is banned in Somali:
This is confirmed by an exhaustive examination of Agostini et al. (Reference Agostini, Puglielli and Siyaad1985): nouns with a final short vowel overwhelmingly end in e #, i # or o #, as anticipated by (70).Footnote [30]
Returning to class 3b, we observe that the 2s imperative is the only form where autoben appears in word-final position. In all other cases, autoben is followed by either png or tam. We thus claim that -o surfaces in 2s imperative to satisfy (70). In Element Theory, [o] and [a] share Element |A|, and differ in that [o] contains |U|:
In class 3 verbs, exactly as in nouns, an additional |U| appears in order to satisfy (70): this is again a general phonological rule of Somali, and not an idiosyncratic property of class 3a/b.
The representation of the imperative 2s maro (class 3b) appears in (73).
We predict that -t does not surface: V
$_{3}$
being empty, it cannot license the preceding C position, which would be the only C-slot available to -t .
4.4 Conclusion
In this section, we have pursued our analytical path into the complex patterns of classes 3a and 3b. We observed that, abstracting away from -s - in class 3a, which is caus, class 3a and class 3b behave on a par, except for one property: in the 1s/3ms/3p, class 3a displays lex = ad , whereas class 3b has lex = t . We proposed an underlying representation for autoben and showed that the diverging patterns in class 3a and 3b can be derived from this representation by regular phonological principles. Finally, we showed that the inflected forms of both classes can be represented on a par with the inflected forms of class 1 and class 2.
5 Conclusion and further issues
We have argued that the allomorphic patterns targeting the complex derivational and inflectional morphological processes of Somali verbs are epiphenomenal. More precisely, we reduced the allomorphy exhibited by lex and png markers to a case of phonologically conditioned allomorphy. Our approach has a straightforward impact on the organization of Somali verb classes: there is no need to divide Somali verbs into 3 distinct conjugations. We claim that surface distinctions are made exclusively on the basis of the exponent selected by each verbal root.
We have established unified representations of the causative and the autobenefactive markers (see (34) and (63) above). These representations are repeated below. Our central claim is that all surface realizations of caus and autoben derive from these representations.
On the one hand, the representations in (74) share some phonological material, namely the CV unit and the consonant /t/. On the other hand, they differ with respect to the vocalic Element: caus contains |I|, whereas autoben contains |A|. We propose to interpret the phonological material shared by both caus and autoben as the exponent of a derivational marker. This marker is responsible for building a new word from a well-formed noun, verb or adjective, either of the same category or of a different one. As for the Elements, they are the actual causative and autobenefactive exponents.
The possible paths of verbal formation are schematized in (75).
Besides the combination of caus and autoben examined in this article, our analysis accounts for two additional combinations of suffixes: caus + caus (‘strong causative’ in -siː , Saeed Reference Saeed1993: 59–60) and caus + caus + autoben (-siːso ).
Finally, in our approach, we dispense with the notion of paradigm as a linguistic active object. Paradigms result as the output of phonological computation and do not need to be specified in advance in the lexical entry of each verb. Further research, drawing on additional Somali and cross-linguistic data will have to corroborate this hypothesis on the status of paradigms.
