It is a significant cross-linguistic tendency that the target of allomorphy is conditioned by structurally lower material. In the case of phonologically-conditioned allomorphy, such a tendency is a near-universal with few convincing exceptions (Anderson 2008) that haven’t yet been reanalyzed (Deal & Wolf 2017; Kiparsky 2021). This paper discusses data on the definite suffix across Armenian lects. The distribution of the suffix’s surface forms is phonologically-conditioned, but their phonological trigger can vary from being structurally lower, structurally higher, or on a separate word. On the surface, this distribution looks like outwardly-sensitive allomorphy, with multiple underlying allomorphs for the suffix. I however argue that the range of data can be elegantly analyzed with a more abstract analysis that utilizes a single underlying form with latent or floating segments.

The Armenian language is an isolate within the Indo-European language family. It is made up of two standard lects, and multiple non-standard lects that are spread out across the Caucasus and the Middle East. The lects have varying degrees of mutual intelligibility or un-intelligibility. This paper discusses data from the two standard lects (Western Armenian and Eastern Armenian) and one non-standard lect (Iranian Armenian). The focus is on the definite suffix.

Consider Table 1. For words in isolation, the suffix shows a relatively simple and inwardly-conditioned distribution of allomorphs. The suffix is -n after vowels, and -ə after consonants. But between a consonant and a vowel-initial clitic, Western Armenian uses the -n surface form. This form is conditioned by the following clitic. However, if the following V-initial segment is an independent lexical word, then the suffix resorts back to the -ə form. Eastern and Iranian display further forms of variation. In Eastern, the -n form is conditioned both by subsequent clitics and subsequent lexical words, unlike in Western. In Iranian, a separate surface form -ən is used between consonants and vowels.

The wide array of surface forms at first suggests that the definite suffix is an outwardly-conditioned suffix, and that its underlying form is chosen post-lexically (Henderson 2012). I argue against this impression. Such a late spell-out analysis would create paradoxes in other areas of the Armenian grammar. Instead, I argue that this entire range of surface forms is all derived from a single underlying form with floating segments: /-(ə)(n)/. This suffix is spelled out in the lexical level. But the latent segments compete to get anchored or docked in the post-lexical level. The lects vary in the ranking of phonotactic and prosodic constraints. Such rankings control the distribution of surface realizations.

This paper is organized as follows. §2 discusses data on the possessive suffixes in Western Armenian, as spoken in Lebanon. These suffixes occupy the same inflectional slot as the definite suffix, and display a similar distribution of surface forms. I show that these surface forms are inwardly-conditioned and not true allomorphy. The distribution of forms displays cyclic opacity once we add clitics and reach the post-lexical level. Based on this backdrop, §3 discusses the definite suffix in Western Armenian, and how the definite suffix behaves transparently with the post-lexical level and displays outward-sensitivity. I analyze the Western data with floating segments in §4. The analysis is then extended in §5 to capture dialectal variation with Eastern Armenian and Iranian Armenian. §6 discusses the diachronic origin of the definite suffix, and the role of sociolinguistic variation. I conclude in §7.1

Cross-linguistically, phonologically-conditioned allomorphy tends to display three types of conditions: input-sensitivity, internal directionality, and cyclic opacity (Paster 2006; 2009). First, the choice of allomorph depends on the phonological structure of the input base, not of the output structure. Second, allomorphy is sensitive to the phonological structure of morphemes that are spelled-out below (inside) the allomorphs, not above (outside) them. Third, cyclic spell-out can cause the loss of the trigger of allomorphy. All three of of these tendencies are reflected in the Western Armenian possessive suffixes.2 I go through the data to set up the baseline architecture of the morphology-phonology interface of Armenian.

Western Armenian is the lect of Armenian spoken in various Middle Eastern countries, including Lebanon, Syria, and Turkey. Data is based on my native judgments for Lebanese Western Armenian. Western Armenian has two possessive suffixes: 1SG [-(ə)s] and 2SG [-(ə)tʰ] (Table 2). These suffixes surface as the lone consonants [-s, -tʰ] after vowel-final (V-final) bases: ɡɑdu-s ‘my cat’. They surface with a schwa after consonant-final (C-final) bases: kʰumɑɾ-əs ‘my amount’. It doesn’t matter if the final C is part of simplex vs. complex coda: ɑʃɑɡeɾd-əs ‘my student’.3

The distribution of the surface forms of the possessives is not true allomorphy. The underlying form of the 1SG possessive is /-s/, and that of the 2SG is /-tʰ/. The schwa is epenthesized to prevent the suffix from syllabifying as part of a complex coda with the root-final consonant. However, these surface forms do show the three cross-linguistic properties of input-sensitivity, inward-directionality, and cyclic opacity, as we will show.

First, the presence of the schwa after C-final bases is not purely phonological but is morphophonological (Ղարագյուլյան 1974: 172; Vaux 1998: 29; Baronian 2017: 7). Complex codas generally show falling sonority. Within roots, the consonants -/s/ and /-tʰ/ can form complex codas: hɑɾs ‘bride’. It is thus the presence of the possessive morpheme boundary which triggers schwa epenthesis: d͡zɑɾ-əs instead of *d͡zɑɾs ‘my tree’. The data in Table 3 illustrates (near)-minimal pairs between root-final [Cs] and [Ctʰ] sequences, versus epenthesis in possessive [C-əs] and [C-ətʰ].

Note that complex codas are permitted in suffixes, such as in the present 1PL suffix: kʰeɾ-e-ŋkʰ ‘we scratch’, glossed as [√-th.1pl].

We capture these facts with the following constraint. Complex codas cannot be formed across a morpheme boundary.4 That is, within a complex coda, the two segments must be part of the same morpheme. This constraint is undominated and it outranks the anti-epenthesis Dep-ə constraint. We include a low-ranked constraint NoCoda because it will be useful later. This constraint is defined such that it assigns one violation for a syllable with a single coda, and two violations for a syllable with a complex coda.

The above *C-C]_σ constraint can be formulated in multiple ways, such as in terms of morphological colors (Trommer 2011; 2015), alignment (Bonet & Lloret 2005), or phrasal syllables (Cardinaletti & Repetti 2009). We use the above simple form for illustration. We show a derivation for the V-final and C-final bases. No schwa is needed for V-final bases (2a), while a schwa is needed in C-final bases to resolve the constraint *C-C]_σ (2b). We set aside the issue of deciding the location of schwa epenthesis.5

In the above data, the schwa was needed because the consonant would otherwise form an illicit complex coda over a morpheme boundary. We see cyclically-induced opacity in epenthesis once we add V-initial clitics (Table 4). When added to unsuffixed forms, clitics trigger glide epenthesis after V-final bases: ɡɑdu[j]=e ‘is cat’. After C-final bases, they trigger resyllabification: kʰumɑ.ɾ=e ‘is amount’.6 I use = to mark clitic boundaries. This data requires the ranking Onset >> Dep-j.

Stress is on the rightmost non-schwa vowel in the prosodic word (Vaux 1998; Dolatian 2021a). The above morphemes are thus considered clitics because they are unstressed.7 I do not show stress markings after this section.

When these clitics are added after the possessive suffixes (Table 5), the schwa is still maintained for C-final bases even though the suffix is now syllabified as an onset: kʰu.mɑ.ɾə.se, *kʰu.mɑɾ.se ‘is my amount’.

Before clitics, the presence of the schwa in C-final bases is phonologically unneeded because the possessive is syllabified with the clitic. Schwa epenthesis thus looks like opaque overapplication. A monostratal non-cyclic derivation cannot derive the cliticized forms of C-final bases. Instead, the above rankings would prevent any epenthesis.8

Furthermore, if we instead look into the prosodic structure of these cliticized words, the paradox remains (Figure 1). Structurally, clitics form either recursive PWords or are part of a clitic group that is post-lexically formed (Nespor & Vogel 1986; Selkirk 1996; Booij 1996). The clitic takes the preceding consonant as its onset, regardless if that consonant is part of the root or possessive. Thus MWords and PWords are not isomorphic. If we redefined *C-C]_σ such that it only applies within the minimal prosodic word, we still have epenthesis unmotivated.

Instead, these cliticized forms need a cyclic derivation (Kiparsky 1982; 2000; 2015; Hargus 1993; Booij 1997; Rubach 1985; Bermúdez-Otero 2011; Embick 2010). The base and suffix are syllabified together in one cycle (4a). Schwa epenthesis applies. This cycle can be considered a word-level or lexical cycle. The output of this cycle is then fed to a post-lexical cycle where cliticization applies (4b).9 The schwa is not deleted because the relevant anti-deletion constraint exists, and there is no high-ranking constraint that would trigger deletion.

The schwa is thus present in the output because it was epenthesized early in the derivation, when there was no other way to syllabify the suffix. Surface cliticization makes this condition opaque. This opacity reinforces the inward-directionality and input-based nature of determining the surface forms of the possessives. Furthermore, the role of strata is further manifested throughout the rest of Armenian phonology (Dolatian 2020; 2021a; b).

So far, we’ve seen that Armenian provides independent evidence that spell-out applies cyclically, with the need for at least two levels of derivation: lexical and post-lexical. For the possessive suffixes, their surface forms are determined by the their input forms in the lexical level where schwa epenthesis can apply. The output of the lexical level is counter-bled by post-lexical phonology, generating opacity. The next section introduces the unexpected behavior of the definite suffix, which seems to show outward-sensitivity.

This section presents data on the Armenian definite. Like the possessives, the definite suffix is part of the word-level morphology. But unlike the possessives, the definite article appears to be a more concrete type of phonologically-conditioned allomorphy that cannot be modeled via a single underlying form (§3.1). Furthermore, the choice of exponents appears outwardly-conditioned because it interacts transparently with post-lexical cliticization (§3.2). This pattern appears to be a case of outwardly-sensitive allomorphy. I show how a cyclic derivation with a simple set of multiple underlying forms cannot capture the apparent outward-sensitivity. Utilizing late spell-out however is unmotivated from the morphology (§3.3).

The definite suffix seems to display outwardly-sensitive allomorphy. However, this section argues that such allomorphy is epiphenomenal. It emerges from the abstract underlying structure of the definite suffix as a sequence of floating segments (also called ghosts or latent segments). With this abstraction, we can cover the data without blurring the line between lexical and post-lexical phonology.

The basic data set is repeated below for the definite suffix. We provide the paradigm for the 1SG possessive for contrast.

To analyze the apparent outward-sensitivity of the definite, I analyze its underlying form as a sequence of floating segments: /-(ə)(n)/ where parentheses mark floating segments. There is a wide cross-linguistic literature on the existence of floating segments and floating features (Archangeli 1991; Tranel 1995; 1996; Zoll 1996; 2001; Wolf 2007; Akinlabi 2011; Côté 2011; Scheer 2011b; Zimmermann 2019). They have been used to model a diverse range of theoretically-controversial phenomena (Scheer 2016; Barillot et al. 2018; Faust et al. 2018; Perry & Vaux 2018; Ulfsbjorninn 2019; Newell 2021; Lahrouchi & Ulfsbjorninn 2022). They have likewise been utilized to model apparent outward-sensitive allomorphy (Zimmermann 2019; Lindsey 2019; Ulfsbjorninn 2020).

In its underlying form, a segment is considered floating if it does not consist of a melody that is associated to the skeletal tier. By lacking a complete association, a floating segment can vary in the position, shape, and context of its surface appearance. In terms of its internal structure, a floating segment can have a melodic feature but lack a skeletal slot (a melodic ghost), or it can have a skeletal slot but lack a melodic feature (a skeletal ghost) (Lindsey 2019). In terms of default behavior, a ghost can be specified to generally surface in the language, but be deleted in order to resolve a phonotactic constraint, i.e, a martyr ghost (Lindsey 2019) or a disappearing ghost (Zimmermann 2019). Conversely, a ghost can be specified to generally delete, but be allowed to surface to resolve a phonotactic constraint: a hero ghost or appearing ghost.

For Armenian, I analyze the surface morphs [-ə], [-n] as deriving from a single abstract underlying form /-(ə)(n)/.12 Intuitively, the definite has an underlying consonant like the possessives. This consonant however is weaker because it is floating. The surface schwa is furthermore not epenthetic, but phonologized as a floating segment in the suffix. Both the schwa and the nasal are melodic ghosts: they have melodic features but no underlying skeletal slot. I later show that the ghosts are hero ghosts or appearing ghosts.

Figure 2 shows the realization rules or Vocabulary Insertion rules for the definite suffix. It is spelled out as two latent segments /-(ə)(n)/ which are not underlyingly associated to a skeletal slot (C or V). In contrast, the 1SG and 2SG possessive suffixes are spelled out as fixed consonants with an underlyingly associated C slot.13

With this floating analysis, I argue that docking is postponed to the post-lexical level. That is, in the lexical level without clitics, floating segments don’t need to associate to a timing slot. They can survive or pass from the lexical level to the post-lexical level without deleting. This argument is not cross-linguistically problematic. There are many attested cases where a floating segment displays deletion or docking based on its post-lexical environment (Côté 2011; Ulfsbjorninn 2020; 2021). Coincidentally, floating features that dock late have been called dormant features (Paschen 2018).

For illustration, I show a simple derivation table below in a rule-based format. In the lexical level, the suffix is spelled-out as a sequence of floating segments /-(ə)(n)/ (Table 10). These ghosts are not incorporated into any syllable structure; they are invisible for syllabification. Thus for a form like ɡɑdu-(ə)(n), there is no hiatus or onset violation for the schwa because it’s not part of a syllable.

Autosegmental docking is postponed to the post-lexical level. For words in isolation, the schwa docks after C-final bases: kʰumɑɾ-ə, while the nasal docks in V-final bases: ɡɑdu-n. Before a clitic, only the nasal is docked, while the schwa is deleted: kʰumɑɾ-n=e. In contrast for the possessives (Table 11), floating segments are not involved at all. The only process we have is epenthesis at the lexical level.14

I now decompose the above derivation with a simple stratal OT analysis. I utilize a lexical level and post-lexical level. We need to introduce the following core constraints on docking, adapted from Lindsey (2019). The constraint RealizeMorpheme requires that at least one of the floating segments survives in the output (Kurisu 2001).15 The constraint Dep(skel) is violated when we insert a skeletal or timing slot for a segment, such as for a floating segment or for an epenthetic segment. The constraint Max(mel) is violated when we delete the melody of a (floating) segment. The constraint Parse is violated by the presence of floating segments. The postponement of docking is done by a careful ranking of the above constraints across the two strata.

First consider a V-base word in isolation. The definite suffix is added in the lexical word-level: //ɡɑdu-(ə)(n)// (15a). The floating segments are not docked yet. They survive in the output as floating segments. This is ensured by the ranking of our constraints. RealizeMorpheme blocks the deletion of both segments. The constraints Max(mel) and Dep(skel) outrank Parse. This ensures that we never delete nor dock the floating segments.

In the post-lexical level (15b), the nasal is docked while the schwa is deleted: [ɡɑdu-n]. the constraints are reranked such that Parse outranks Max(mel) and Dep(skel). This ensures that the output lacks any floating segments. High-ranked RealizeMorpheme ensures that at least one floater is docked. As for the choice of what segment to dock, we have the choice of outputting (docking) either the schwa [-ə], the nasal [-n], or both [-ən]. For the competition between the schwa and the nasal, the nasal wins because of Onset. To prevent both forms from surfacing, Onset must outrank Dep(skel). The constraint Max(mel) is low-ranked in the post-lexical stratum, and is inactive.

For C-final bases, the lexical level adds the floating segments which are not docked yet: //kʰumɑɾ-(ə)(n)// (16a). In the post-lexical level (16b), we again have the choice of docking either the schwa [-ə], the nasal [-n], or both [-ən]. The nasal (candidate d) is blocked to minimize NoCoda violations.16 We thus need to choose between [-ə] vs. [-ən]. The latter loses (candidate b) because Dep(skel) outranks NoCoda and Max(mel). There is likewise no point in docking the nasal and then epenthesizing a schwa (candidate f). Epenthesizing a schwa violates Dep-ə and also Dep(skel); and this candidate is harmonically bounded by candidate (c) with the docked schwa.17

For the post-lexical stratum, I have found no evidence of Max(mel) outranking any other constraint. I treat this constraint as low-ranked and inactive. I do not show it in later tableaux. Because Dep(skel) outranks Max(mel), the floating segments in the definite suffix are hero ghosts (Lindsey 2019) or appearing ghosts (Zimmermann 2019). The ghosts are deleted by default. At least one surfaces to satisfy RealizeMorpheme, and the choice of segment is based on minimizing the violations of Onset and NoCoda.

The above is for words in isolation. In cliticization, we start to see the apparent outwardly-sensitive allomorphy thanks to late docking. For a V-final base, the lexical level will add the floating suffix without any docking: //ɡɑdu-(ə)(n)//. In the post-lexical level, the clitic is added. The floating nasal is docked to provide an onset to the clitic: [ɡɑdu-n=e]. The schwa is deleted because it would create Onset violations. We show just the post-lexical derivation.

For a C-final word, again the lexical level outputs an identical input with the floating segments intact: //kʰumɑɾ-(ə)(n)//. In the post-lexical level, the nasal is docked to provide an onset to the clitic: [kʰumɑɾ-n=e]. The schwa is not docked (candidate b,c) because Dep(skel) outranks NoCoda.

As we see, the above floating analysis reduces the outwardly-sensitive allomorphy of the definite suffix into an epiphenomenal result. The outward-sensitivity allomorphy is epiphenomenal because it is an indirect consequence of lexical abstraction and syllabification. There has been a growing body of which work has argued that apparent outwardly-sensitive allomorphy (Hannahs & Tallerman 2006; Kikuchi 2006; Bonet et al. 2007; Deal & Wolf 2017; Brinkerhoff 2019) can be reduced to the interaction of floating segments (Scheer 2016; Lindsey 2019; Ulfsbjorninn 2020; Kiparsky 2021). The fact that the Armenian data can seamlessly integrate with this body of work is a further conceptual argument for this analysis.

In sum, above analysis can capture the bizarre behavior of the definite suffix in Western Armenian. It captures the apparent outwardly-sensitive allomorphy without violating our lexical-postlexical division, and without needing to postpone the spell-out of word-level morphology. The cost is the need for more abstract representations that utilize floating segments, and the need to make docking be a post-lexical process. These assumptions are not cross-linguistically problematic. The next section explores dialectal variation in post-lexical docking.

The data and analysis so far concern the allomorphy of the definite in isolation and in cliticization. We analyzed the outward-sensitivity by utilizing floating segments. In this section, we show how definite allomorphy shows dialectal variation in phrasal resyllabification between Western and Eastern Armenian (§5.1), and in the ability to dock multiple segments in Iranian (§5.2). The variation shows that the domain of docking is truly post-lexical.

The previous sections provided a large chunk of the morphophonological patterns of the definite suffix. We analyzed its surface forms via floating segments. This section briefly goes through some diachronic and colloquial data, in order to shed further light on how these floating segments behave across Armenian.

This paper discussed data on the definite suffix of Armenian, across three lects: Western Armenian, Eastern Armenian, and Iranian Armenian. Across these lects, we found that the definite suffix displayed a wide array of possible surface forms. The distribution of these forms was phonologically-conditioned, based on the preceding and following segment. The triggers for the shape of the definite could were distributed across the preceding word, the following clitic, or the following word.

On the surface, the above distribution looks like outwardly-sensitive allomorphy. However, we analyzed this entire range of data by using an abstract underlying that used floating segments. With this abstraction, we can derive the dialectal variation of the suffix. Furthermore, by using floating segments, we can maintain the cyclic nature of Armenian phonology, with strong divisions between lexical and post-lexical strata. By using cyclic spell-out, we are in fact forced into using ghost segments. The alternative would have been to abandon cyclicity and use non-cyclic spell-out. Furthermore, because docking is post-lexical, this requires that the post-lexical stratum has access to morphological information like RealizeMorpheme; otherwise, floating segments wouldn’t need to dock.

It is an open question if all cross-linguistic cases of outwardly-sensitive allomorphy can be reduced to floating segments. But at least for the Armenian definite suffix, such a reduction is viable and systematic.