The typology of the distributional restrictions of a feature: occlusion and bipositionality

2.1 Element theory basics

Element Theory (Backley 2011; Harris 1990, 1994; Harris and Lindsey 1995; Kaye 2001; Kaye et al. 1985; Scheer 1996) provides an alternative view on the phonological structure of segments, developed within the general framework of Government Phonology (Harris 1994, Kaye et al. 1985, 1990; Scheer and Kula 2018). It diverges from standard feature theory in that it uses elements as phonological a.k.a. melodic primes, instead of binary features (Chomsky and Halle 1968; Clements 1985; Dresher 2008, 2009; Halle 1992; Sagey 1986).

Elements are monovalent, phonetically interpretable primitives, and acoustically defined targets that determine the internal structure of segments. Broadly split into “place” and “manner”, these primitives can be realized as simplex or complex melodic expressions. For example, in English the elements |I| and |A| may be realized as such, or may fuse into either [e], as in send, or [æ], as in action, depending on which of the two elements is the head of the expression (Backley 2011: 43). Likewise, |U| may be realized alone, as in push or combined with |A|, as in caught. The same reasoning holds for consonants: |H| (frication) and |ʔ| (occlusion) can be realized separately, like in horse, and put [puʔ] in the glottalizing variety of English. They can also combine with other elements to form, among other consonants, a coronal stop [t] = |ʔ,H,A|, a sibilant [s] = |H,A| or a rhotic [ɾ] = |A| (Backley 2011: 133). The reader is referred to the abovementioned work, which provides common Element Theory phonological break-downs for various sound classes (see also Harris 1994; Harris and Lindsey 1995).

As it stands, Element Theory allows accounting for opposing phonological processes such as lenition and fortition. The first is viewed as the loss of element(s) in a melodic expression (segment): e.g. the loss of |A| in /s/ turns the Spanish word tos ‘cough’ into [toh] in the Andalusian variety. Fortition refers to an enhancement or a preservation of the melodic content, as in Classical Arabic and Farsi word-initial empty onsets, which host a glottal stop (e.g. Classical Arabic [ʔalf] ‘a thousand’, Farsi [ʔotriʃ] ‘Austria’, Jaferian 2024: 180). The reader is referred to Harris (1990, 1994) and Harris and Lindsey (1995) for a thorough analysis of lenition trajectories within this framework. The next section expands on occlusion in relation to bipositionality.

2.2 Defining occlusion and bipositionality

Occlusion (Edge) is a formal phonological feature, it is typical of stops and affricates. It is also found in some sonorants (so called ‘hard sonorants’), though there is some language variation, nasals generally pattern with stops and it is common, though highly optional, for laterals (a pattern we also see in acquisition Rose 2024). It is almost always absent in fricatives, rhotics, glides and vowels. As mentioned in the previous section, Element Theory marks Edge by a single monovalent feature: |Ɂ|, as opposed to [-cont] in standard feature theory.

Bipositionality is an autosegmental condition by which a feature (bundle)/segment is linked to more than one position in the syllable structure; it is equivalent to heterosyllabicity. Most transparently, this is seen with geminate consonants (e.g. [bː], [dː]), but it is also a feature of partial geminates, where only some of the features of a cluster are shared across the two positions (e.g. [mb], [nd]). These transparently constitute a many-to-one relationship between segments and the skeleton (Goldsmith 1976, 1990; Leben 1973).

(1)

One-to-one association (monopositional) vs. many-to-one (bipositional)

As shown above, heterosyllabic clusters involve bipositionality: kan.na and kam.ba (1b-c), as opposed to various unattested things like *ka.tta, *katt.a, *ka.nta.^[1] All rhyme-onset sequences have the same structural configuration as geminates (Charette 1990; Harris 1990, 1997).^[2]

In Government Phonology, the C2 of these heterosyllabic clusters is taken to be the head (Charette 1990) and it is featurally never less complex than C1 (Harris 1997). We assume that the head of a C1C2 heterosyllabic cluster is C2 even in the Strict CV representation shown beneath (cf. Cyran 2010). This headedness is indicated with the arrows in (2a-b).

(2)

Geminate vs. rhyme onset sequences (coda-onset)

a.

Standard Syllabification

b.

The Strict CV translation of the above representation.

The representations in (2a) refer to standard Government Phonology, a general framework to which the Strict CV model is directly linked. The latter will be used for the purpose of our analysis. Its basic tenets are outlined below.

2.3 Strict CV and the word-initial position

Strict CV is an autosegmental approach to syllable structure which holds that the syllabic level of phonological representations consists of alternating C and V positions, with no rhymes, no codas, and no constituent nodes (Lowenstamm 1996, Scheer 2004). Government and licensing are lateral relations that allow deriving various syllabic and segmental configurations, including consonant clusters, complex onsets, long vowels and geminates. Within this model, a word like flattery is represented as in (3).

(3)

[flat(ə)rɪ]

Note the empty V positions that are interposed between the first and last two consonants. They are said to remain empty under the effect of proper government of the adjacent vowels. Furthermore, according to Scheer (1999, 2004: 37), genuine branching onsets, whose vocalic position remains constantly empty, involve a specific type of government, termed infra-segmental government. The initial cluster in (3) is a perfect illustration of this. Its members form a governing domain, headed by /l/. Such a domain allows explaining the absence in English-like languages of words beginning with falling sonority clusters. It further supports the hypothesis that an empty initial site sits at the left edge of the word. The facts in Tümbisa to be discussed in Section 4.2.2.1 will point to this specific hypothesis.

The initial CV dates back to Lowenstamm’s (1999) work on word-initial clusters and proclitics in French and Biblical Hebrew (see also Scheer 2004: 95, 2012: 146, Lahrouchi 2001, 2003, 2018a; Seigneur-Froli 2006). In an attempt to capture the difference between languages that display only sonority-rising clusters at the beginning of the word (e.g. French: plateau ‘tray’, bras ‘arm’, English: true, flat; Spanish: fruta ‘fruit’, trabajo ‘job’), and languages where the initial clusters are made sonority-free (e.g. Maghrebi Arabic: gləʕ ‘remove’, lga ‘find’; Biblical Hebrew: qraβim ‘midsts’, rqahim ‘spices’), the author proposes that the left edge of the word is phonologically marked with an empty CV that hosts various morphophonological operations on condition that it is licensed. This pits two types of languages against each other: (a) Type I languages where the initial site is always licensed (e.g. French, English), (b) Type II languages where the initial site is not always licensed (e.g. Maghrebi Arabic, Biblical Hebrew).^[3]

Proper government is one type of licensing that emanates from the following vowel, and leads to the cliticization process shown in the French examples below (the initial site appears in bold).

(4)

le tapis ‘the carpet’ vs. le plateau ‘the tray’

In these examples, the initial CV is licensed by the vowel /a/. The small arrow in (4b) denotes infrasegmental government from /l/ to /p/, a relation whereby the two consonants form a closed domain that the following vowel straddles in order to license the initial CV. By contrast, in Biblical Hebrew, the initial site can be either licensed (5a) or not (5b).

(5)

hakklaβim ‘the dogs’ vs. haːrqahim ‘the spices’

A uniformity convention allows the initial CV to be unlicensed throughout the language, including in words that begin with a falling sonority cluster, Lowenstamm (1999: 164) argues. Vowel lengthening (5b) and gemination (5a) are viewed as the result of the identification of the initial site.

The postulation of the initial CV further allows explaining the strength of word-initial consonants, according to Scheer and Ségéral’s Coda-Mirror theory. Indeed, word-initial consonants sit in a position that is only licensed, while government (viewed as a weakening force) targets the initial CV. By contrast, the consonants that appear in word-internal or final coda and in intervocalic position are said to be prone to weakening because they are either subject to government and licensing or they escape both forces. The representations in (6), adapted from Scheer (2004: 140) and Ségéral and Scheer (2001: 144), illustrate these situations:

(6)

With these theoretical premises in hand, we now turn to the core issue of this paper: occlusion.

2.4 Previous analyses: occlusion as bipositionality

In the alternations between the singular and plural nouns of the West Atlantic language Firdu Fula (Gamble 1958), Jensen (1994) noticed a relationship in the patterning between bipositionality (as defined above) and occlusion. As shown in (7), the alternations result from the insertion of an empty prefix in the plural. In each of these strong/weak pairs, the addition of the empty position in the plural corresponds with the hardening of the initial consonant of the stem, in all cases, it seems one has to add occlusion, Edge.^[4]

Since stopness by hypothesis is bipositionality, and you can get stops word-initially in the strong ‘grade’, Jensen ascribes the initial stopness of this grade as initial bipositionality an empty prefix + stem (7b); effectively, word-initial gemination. Outside of this grade, one obtains only the monopositional weak forms of (7a).

Jensen hypothesizes that the bipositional/geminate structure (8b) provided by the empty prefix is enough to generate the strong version of the consonant. That is, he proposed that in this language, the occlusion feature (|Ɂ|) can be entirely recast as the bipositional relationship itself. Jensen’s (1994) definition of Edge as bipositionality is shown in (8).

(8)

Occlusion as bipositionality (Jensen 1994)

This argument is central for GP2.0 (Pöchtrager 2006; Pöchtrager and Kaye 2013) and they develop their own version of this hypothesis. The claim of GP2.0 is that |Ɂ| is not a feature at all, instead it is a structure, specifically a structure which involves more than one x-slot. Jensen (1994) had a more traditional rhyme-onset structure, and GP2.0 instead place the two x-slots that make up the former |Ɂ| element stacked on top of each other. This fits the modus operandi of the framework, where an element |Ɂ| is argued not to be a feature at all, rather it is ‘more structure’, in this case a bipositional configuration of x-positions. As shown beneath, xO is the head of the sound and in combination with an x-position one obtains a fricative (9a-b). There is a further distinction of voicing that is not relevant to us, but notice that if one has another ‘layer’ with another x-position, one obtains a stop (9c-d). In this way, the framework recasts and defines what was previously a feature (|Ɂ|), as a structure involving multiple skeletal/x-slots/positions.

(9)

Occlusion as structure, Pöchtrager (2006)

3.1 Occlusion and bipositionality interplay

Ulfsbjorninn and Lahrouchi (2016) show, however, that the typology of the distributional restrictions on occlusion cannot be reduced to equating occlusion with bipositionality. This is because there are actually two kinds of languages in the restriction, and they have to be expressed bidirectionally. This directionality has different implications for the overall distribution of the feature in question, as illustrated below in (10).

Take for example Type A. In this language type, Edge can only be found in bipositional structures (geminates, and partial geminates) and it is lost in all monopositional positions, even word-initially (10b). However, bipositional structures are free to contain sounds with or without Edge. Conversely, in the Type B language, bipositional structures must have Edge. However, there is no restriction on Edge in monopositional structures; these are allowed to have Edge or not.

In order to capture the typology in (10), the distributional statements between Edge and bipositionality must be statable bidirectionally: Bottom up and Top down. This is shown in (11) beneath. To account for Type A (11a), the restriction is stated as a condition on the feature (M) relative to the structure (S): Bottom up. In contrast, in (11b) we see the restriction must be stated as a condition on the structure (S) relative to the feature (M): Top down.

(11)

a.

Type A – Bottom up (M → S)

b.

Type B – Top down (S → M)

The directionality is crucial because it has implications for the distribution of the feature relative to other structures. This typology necessitates that occlusion is a feature which has a licensing type relationship with syllable structure, it cannot just be that structure itself (pace GP 2.0).

The fact that MSLCs are bidirectional (Bottom up, Top down) is a formal property of MSLCs, a possibility made available by Phonological Universal Grammar. This shapes the possibilities space of the typology of particular languages.

Counterfactually, it could have been the case that languages would only license a feature in relation to a structure (all Top down), or a structure in relation to a feature (all Bottom up). Instead, we learn that phonological UG must be able to have MSLCs in either direction.

Additionally, we see that there is an implicational relationship between occlusion and bipositionality. No language known to us has a prohibition on Edge in bipositional structures, while allowing (or restricting) Edge only in its monopositional structures: e.g. [fː] *[pː], [mf] *[mb], [p], [f]. Such a system would have posed a serious problem to our MSLCs. A restriction on Edge being licensed only by singletons and not bipositionally is impossible to formally state, using the tools currently available to us.

3.2 Types of strength

We take the implicational relationship just depicted to follow from the fact that there are three distinct types of strength in phonology (Scheer 2004), briefly outlined below. In none of these types would monopositionality be stronger than bipositionality.

The first kind of phonological strength is prosodic. This can be characterized by languages where lenition is conditioned by their position relative to stress (Harris 1997), (see Scheer 2000 for further discussion of this type of strength in light of Strict CV).

A case of this involving Edge in stops is found in Tripura Bangla (Das 2001:170) or Ibibio (Harris 2023; Harris and Urua 2001). For instance, in Ibibio, stops are spirantized in intervocalic foot-medial position (the round brackets indicate feet): e.g. [(dɨ́p)] versus [(dɨ́βe)] ‘hide/hide oneself’. However, those that are intervocalic but foot-initial, or across a foot-boundary, are crucially not lenited: e.g. [ú-(tɑ́ŋ)] *[u-(ɾɑ́ŋ)] ‘NMLZ-plaiting’ and [(dɑ́pːá)-ké] *[(dɑ́pːá)-ɣé] ‘not dream’.

In such languages, monopositionality and bipositionality are simply not critical to the distribution of Edge in stops. However, Harris’s account of lenition would preclude neutralization of stops to fricatives in metrically strong positions, since fricatives have less features than stops (Harris 1990, 1994; Harris and Lindsey 1995). Therefore, the type of language with only fricatives in the weak position of feet, at the exclusion of stops, is ruled out on these grounds.

The second type of phonological strength is purely positional. It is modeled by the ‘Coda’-Mirror (Ségéral and Scheer 2001). As mentioned in Section 2.3, a position which is governed is weak, such as the intervocalic position. Conversely, an ungoverned but licensed position is strong (post-coda, and word-initial positions).

Colloquial Tuscan Italian would be a perfect example of a language where Edge in stops and affricates obeys a distributional restriction described purely by Government. Positional strength and weakness is shown in (12) and exemplified with Florentine Tuscan spirantization ‘Gorgia Toscana’ (Marotta 2008), see also: Ulfsbjorninn (2017) and Kenstowicz (2017).

In (12b) we see all stops and affricates become fricatives in positions in which they receive + Gov according to the Coda Mirror. They are strong only in positions that are -Governed, such as initial (12a) and post-coda (12c) positions. Similar facts, but with some interesting complications, are reported in Campidanese Sardinian (see Chabot 2023).

In Tuscan, there is a connection between monopositionality, bipositionality and Edge, but it is defined through Government. Bipositional structures with Edge such as geminates and the post-coda position uniformly have the head of the cluster (C2) in an ungoverned position. Any monopositional structure/C-slot that is governed will have to lose Edge. This clearly shows that the pattern is not one caused by bipositionality, rather the casual factor in lenition of Edge is purely positional and dictated by Government.

Because of how strength works in the Coda-Mirror, any bipositional structure will have to have its head be ungoverned (strong), thereby independently ruling out as a possibility the type of system where bipositionality repels Edge. Being able to exclude this possibility on principled grounds is a particular success of this phonological model of positional strength.

3.3 Feature sharing strength and edge/bipositionality

A third type of strength is ‘sharing makes us stronger’ (Honeybone 2002, 2005a, 2005b). This is a case where bipositionality renders a feature inherently strong.

The most famous example of this is geminate inalterability (Hayes 1986, Schein and Steriade 1986). This is partly modeled by the positional strength type (coda-mirror). However, one can distinguish between true positional strength and bipositional strength (‘sharing makes us stronger’) because the latter crucially requires feature sharing to induce strength.

In positional strength, the initial position is parametrically strong (Lowenstamm 1999; Scheer 2004), and, even if it is not a phonological geminate. So is a post-coda position which retains Edge, as in Italian scar p a ‘shoe’.

However, in some languages, lenition/strength patterns care specifically about the doubly linked status of a feature/segment. In a true bipositional-strength case, the initial position and the post-coda position without feature sharing will be classed as weak. An example of this is the v-b pair in Neapolitan (pace recent loanwords from Italian): [v]arca ‘boat’ (cf. tre [bː]arche ‘three boats’), [v]ar[v]a ‘beard’ (Russo and Ulfsbjorninn 2020). Here we see the weak variant of the labial consonant in initial and the post-coda position simply because they do not geminate or share features. Positionally of course, these are strong positions, however, the lenition pattern in traditional Neapolitan is dictated by another type of strength, one defined by feature sharing.

Clearly in these ‘sharing is strength’ systems, Edge would always be weak in monopositional structures and always strong in bipositional structures. Again, from this formulation of phonological strength and weakness, one cannot make a system where Edge survives only in monopositional, but not bipositional structures.

4.1 Hard and soft sonorants

Occlusion is parametrically found in some sonorants, these are anecdotally referred to as ‘hard sonorants’. In fact, it is typical for nasals to be hard, and variably laterals can also be shown to be hard in some languages (Backley 2011: 116; Harris 1994: 125; Harris and Lindsey 1995: 73).

The hardness of sonorants depends on how their occlusion, which phonetically decreases from nasal to lateral to rhotic to glide, is phonologized. There is an implicational hierarchy of sonorant hardness.

The difference between occlusion in non-sonorants and sonorants can be expressed with headedness, a contrast based on the prominence of an element in the melodic expression (Charette 1994; Charette and Göksel 1998; Kaye 2001; Ploch 1999); for more recent work on this see Breit (2013, 2017, 2023), Faust (2017), Backley (2011, 2017), Ulfsbjorninn (2021). Headed elements are typically notated by being underlined: |ʔ| ‘headed Edge’ referred to as Edge* versus |ʔ| ‘unheaded Edge’ referred to simply as Edge. Stops can be argued to have a headed Edge feature (Edge*), whereas hard-sonorants can be ascribed the non-headed feature (Edge).

Berber, presented in UL as a Type A language, provides an interesting example of a Bottom up MSLC, not only for stops but also for a subset of hard sonorants. As we will see in §4.2.2.2, Tarifit, a variety spoken in Northern Morocco, bans the stops as well as the lateral sonorant from a monopositional structure, while Tamazight does so only for stops (§4.2.1), in sharp contrast with Tashlhiyt in which occlusion is systematically retained.

4.2 A four-way typology: directionality and headedness

Since MSLCs are bidirectional, and we have two features to consider: Edge* (head) and Edge (non-head). It follows that our system actually predicts four patterns of featural distributional restrictions. Languages can either have Bottom up or Top down MSLCs which refer to either Edge* only (stops, affricates) or Edge generally (stops, affricates and hard sonorants).

(14)

In addition to the Ontena Gadsup/Tamazight versus Kingi Soninké/Somali,^[5] this leaves us to find examples of a Bottom up MSCL that applies to all occlusion (Edge), including that of hard sonorants (14a), and Top down MSLC that applies only to headed occlusion (Edge*) (stops, affricates) but without any implication for hard-sonorants (14d).

4.2.1 Bottom up (M → S) for Edge*

Some Berber languages demonstrate the positional restriction on occlusion, colloquially the ones that are known as ‘spirantizing varieties’. An example of a spirantizing variety is Tamazight, spoken in the Central Atlas Mountains of Morocco. This can be compared against a non-spirantizing variety such as Tashlhiyt, which closely preserves the occlusion distribution of Proto-Berber. As we see below, singleton stops have become fricatives in Tamazight in all positions, even word-initially (Kossmann 1995; Saïb 1976).

(15)

Tashlhiyt	Tamazight
akabar	açaβar	‘caravan’
aglːid	aʒəlːið	‘king’
tirgin	θirʒin	‘embers’
akuz	açuz	‘weevil’

Non-sonorant occlusion can only be found in bipositional structures such as geminates, or in sonorant-obstruent clusters where occlusion can be shared across adjacent consonant positions. Given that occlusion is not permitted in singletons, and morphological gemination is common in the language, this generates many synchronic alternations, where occlusion can only be maintained in the geminate sequence (see (16)), or where occlusion can be shared with a preceding hard-sonorant such as a nasal or a lateral, not a rhotic (see (17)).

(16)

Tamazight
Aorist	Imperfective
nçər	nəkːər	‘stand up’
mʒər	məgːər	‘harvest’
fθəl	fətːəl	‘roll’
rβəl	rəbːəl	‘ramble’

Beni Iznassen Berber (El Kirat 1987), another spirantizing variety to the North-East of Morocco, provides evidence for occlusion surviving in post-hard sonorant position; these correspond to the sharing of Edge.

(17)

Tashlhiyt	Beni Iznassen
lmdint	θamdimt	‘town’
taqbilt	θaqbilt	‘tribe’
ultma	ultma	‘my sister’

Nasals and laterals behave as hard sonorants in Beni Iznassen Berber, like in many other neighboring varieties (Iazzi 2018: 367), ultimately leading in some cases to gemination as a result of the sharing of Edge and place of articulation. This is shown in the examples in (18) taken from Chibli and Bensoukas (2020: 66).

(18)

Ait Sgougou Tamazight

a.

/nt/, /nd/ → [nn]
/t+adun+t/	[θaðunn]	‘fat’
/imndi/	[imnni]	‘seeds’

b.

/lt/, /ld/ → [ll]
/t+amllal+t/	[θaməllall]	‘white.FM’
/ildi/	[illi]	‘catapult’

A noteworthy point about the Beni Iznassen variety is that Edge is allowed to branch across adjacent consonants that contain this feature. This is permitted in our model because the Edge* is in the head-position of the bipositional structure.

(19)

/tamdimt/ [θamdimt] ‘town’

In (20), we see occlusion being lost in post-consonantal position. This cannot be due to positional strength, since the Coda Mirror makes this a strong position, as explained in §3.2 and exemplified with Tuscan spirantization. However, rhotics do not contain Edge, and, as such, they cannot share Edge with the neighboring stop, hence /g/ realized as [ʒ]; this occurs also in the initial position, where singleton /t/ spirantizes into [θ].

(20)

[θirʒin] ‘embers’

4.2.2 Bottom up (M → S) for Edge and Edge*

The gap in the typology (14a), repeated beneath in (21), describes a system, which is effectively the same as that of the Tamazight Berber, but which applies not only to Edge* (stops and affricates) but to all Edge (including hard sonorants). This would be a pattern where nasals and laterals would undergo Edge-lenition, alongside the stops. Moreover, where their lenition pattern cannot be described as positional strength (Government, Coda mirror), but has to be due to branchingness strength (dictated by an MSLC). The next section presents data from Tümbisa that fill the typological gap just described; Tarifit Berber does this only partially, as we will see in §4.2.2.2.

(21)

Bottom up (M → S)

4.2.2.1 Tümbisa

Tümbisa (Panamint Shoshone) is a recently extinct member of the Numic branch of Uto-Aztecan. It was spoken by hunter gatherers in the Mountains valleys and deserts East of the Sierra Nevada in what is today Southeastern California and Southwestern Nevada, including what is known today as Death Valley. The processes described in the language that we reference here were all productive when the language was recorded (Dayley 1989).

The language has the following phonological inventory (Dayley 1989: 401). It has phonemic stops including a glottal stop, an affricate (ts), hard sonorants (nasals), glides and fricatives including the glottal fricative /h/. The language only had one or two instances of /l/, always initial, borrowed from English (Dayley 1989:402).

(22)

Inventory of Tümbisa
p	t	ts	k	kw	ʔ
m	n		ŋ	ŋw
	s				h
	j			w

The language has contrastive gemination, additionally /s/ and /ŋ/ are inherent geminates. In addition to these heterosyllabic consonant clusters, the language also has underlying consonant clusters of the shape: NC and HC: e.g. /nampe/ ‘foot’, /ohpin/ ‘mesquite bean’.

(23)

Inventory of clusters in Tümbisa
pː	tː	tsː	kː	kwː
mː	nː		ŋː	ŋwː
	sː
mp	nt	nts	ŋk	ŋkw
hp			hk	hkw
hm	hn
	hj			hw

As shown above, /h/ is not permitted to form a cluster before the inherent geminates /sː/ and /ŋː/, which is consistent with a heterosyllabic analysis of these h.C clusters: *hŋː. There is a gap with a missing: /ht, hts/, we take this to be accidental; any other non-inherent geminate consonant can form a consonant cluster by being preceded by /h/.

Turning to the distribution of occlusion in Tümbisa, we note that this language has lenition of occlusives in singleton positions, except word-initially. Word-medially, singletons surface as fricatives/approximants (lenis fricatives, rhotics, glides). Crucially, in Tümbisa, the lenition of occlusion affects both headed and headless Edge, in stops, affricates and nasals (Kirchner 1998). The weak variant is shown voiced here but it becomes voiceless next to a voiceless vowel.^[6],[7]

(24)

		Strong	Weak (no Edge)
Edge*	p	U, H, ʔ	β/ɸ
	t	R, H, ʔ	ɾ/ɾ̥, ð/θ (after front vowels)
	ts	R H, ʔ	z/s or ʃ/ʒ (palatalization)
	k	H, ʔ	ɣ/x
	kw	H, ʔ + U	ɣw/hw
Edge
	m	U, L, ʔ	w̃/w̥̃
	n	R, L, ʔ	j̃/ j̥̃ (after front vowels)
	ŋw	H, L, ʔ + U	w̃/w̥̃ (after back vowels)

Crucially, the strength and weakness of stops in the language is not defined purely positionally, because stops after the fricative coda are lenited, despite being positionally strong. Instead, the MSLC does not permit Edge in these non-branching positions.

In Tümbisa, if an occlusive is post-consonantal/heterosyllabic, it will be strong only if it shares Edge with the preceding consonant, such as in NC clusters or geminates. Otherwise, the stops, the affricate /ts/, and even the hard sonorants (nasals), all lose their occlusion in monopositional structures.^[8] Intervocalically and post-H-coda, all Edge is lost.^[9]

(25)

Weak form intervocalically and HC clusters

a.

Edge* (stops and affricates)
/kupisi/	[kúβiʃi̥]	‘brains’
/hupapin/	[húβaɸĩ̥]	‘soup’10
/sutɨ/	[súɾ̥ɨ̥]	‘that’
/motson/	[mṍzo]	‘bear/whiskers’
/tsitoːhin/	[tsíðoːhi̥]	‘push’
/ɨm pitsiʔi/	[ɨ̃m bíʒiʔi̥]	‘your breast’
/pakatɨn/	[páɣaɾɨ̃̀] or [páɣaɾ̥ɨ̥]	‘body of water’
/paa ok w enː̥a/	[paː óɣ w ẽnː̥ ḁ]	‘water flowing’
/hupiatɨki/	[húβiàɾɨɣì] or [húβiàɾɨxi̥]	‘sing’

1. 10

   It is apparently also sometimes possible not to lenite before final voiceless vowels, though this is generally a leniting context.

b.

Edge (nasals)
/tɨmuhun/	[tɨ̃́w̃ũhu̥]	‘rope’
/senu/	[sẽ́j̃ũ]	‘therefore’
/kwinaː/	[kwĩ́j̃ãː̃]	‘eagle’
/ɨŋ w atɨn/	[ɨ̃́w̃ãɾɨ̀]	‘rain’11
/pɨe tukwanːi naːkinːa/	[pɨ́e ðúɣwanːi j̃áːɣìnː̥ ḁ]	‘it’s already getting dark’

1. 11

   /ŋ/ never lenites because it is an inherent geminate, however, /ŋ^w/ can lenite, though it sometimes appears to resist lenition, another variant of ‘rain’ for instance is: [ɨ̃́ŋwãɾɨ̀].

The lenition in HC clusters leads to two fricatives in a row. When this happens, these lenited forms surface with coalesced melody e.g. /h+p/ = [ɸ]. Although on the surface they appear to be one sound, there is ample evidence for it underlyingly (phonologically) being a consonant cluster. Firstly, the nasal harmony of the language is blocked by HC. This outcome happens equally across words and in productive affixation, where we can be sure there is a synchronic h+C sequence underlying [ɸ] (see 27 below). Elsewhere, the final /h/ surfaces transparently before vowel-initial affixes: /ɨatːɨah + a/ [ɨ́atːɨ̀ah-ḁ] ‘ranch-obj’. Secondly, the evidence for [ɸ] being the surface interpretation of an underlying /h+p/ cluster comes from its non-voicing in an otherwise ostensibly ‘intervocalic’ context. If [ɸ] was not a cluster, it would behave unlike all other obstruents. Incidentally, its non-voicing in this context cannot even be explained away using extrinsic rule ordering since, in Tümbisa, intervocalic voicing is a post-lexical process found with both derived and underlying obstruents in the language. Meanwhile, non-obstruent consonants (and exceptionally /n/) are unaffected after /h/.

(26)

HC clusters within the same word
/ohpimpɨ/	[oɸĩmpɨ̥]	‘mesquite tree’
/tahmani/	[tã́hw̃ãn̥ĩ̥]	‘springtime’
/kuhmatsːi/	[kṹhw̃atsːi̥]	‘husband’
/wihnu/	[wĩ́hj̥̃ũ] or [wĩ́hj̥̃u̥]	‘then’
/wihnumpitsːi/	[wĩ́hj̃ũmbìtʃːi̥]	‘buzzard’

(27)

HC across words12
/tɨkːapːih + pan/	[tɨkːápːi̥ɸá]	‘food + on top of’
/tɨkːapːih + tukwːan/	[tɨkːápːi̥ɾ̥ùkwː̥ḁ]	‘food + under’
/tɨkːapːih + man/	[tɨkːápːi̥hw̃ã́]	‘food + on’
/ tɨkːapːih + naː/	[tɨkːápːi̥hj̃ã̀ː]	‘food + be (groceries)’
/ɨatːɨah + ka/	[ɨ́atːɨ̀a xá]	‘ranch + on’
(cf. ɨatːɨah + a	[ɨ́atːɨ̀ah-ḁ]	‘ranch-obj’)
/juŋwah + kwan/	[jṹŋwahwà]	‘scoop + COMPL’

1. 12

   Some suffixes are unaffected the stem, we put this affix-specific behaviour down to a cyclic effect (in the vein of Scheer 2012; Newell 2021) effect, which we will not have space to explore here but would be very unlikely to prove problematic for the analysis as a whole (since these would be essentially string initial = word-initial).

The above data clearly show the loss of occlusion by singleton consonants in intervocalic, as well as in post-consonantal, position. However, this happens only when there is no bipositional Edge-sharing.

In the next section, we present the relevant data on the sharing of occlusion, then we provide our representational analysis of the loss of occlusion in monopositional structures. The general pattern is one of occlusion being found exclusively in bipositional structures, be that the headed occlusion of stops and affricates, or that of hard sonorants, in this case nasals.

(28)

Strong form in geminates and NC clusters13
/tɨasːɨpːɨh/	[tɨ́asːɨ̀pːɨ̥]	‘frozen’
/posːotːɨ/	[pósːotːɨ̥]	‘alkali’
/sakːa/	[sákːḁ]	‘that.OBJ’
/kimːanːa/	[kĩ́mːãnː̥ ḁ]	‘to come’
/watsːɨwitɨn/	[wátsːɨwìðɨ]	‘four’
/ukwːah/	[úk w ːḁ]	‘when/if’
/tɨmpe/	[tɨ̃́mpe̥]	‘mouth’
/naŋki/	[nã́ŋki̥]	‘ear’

1. 13

   A nasal + stop cluster will usually lead to voicing (unless the adjacent vowel is voiceless), however, these are not prenasalized (monopositional) stops. Nasal spreading shows that NCs are not prenasalized segments. It is for this reason that mb does not trigger nasalization of the following vowel: /ohpimb/ [óɸĩmbɨ/*ɨ̃] ‘mesquite tree’ (cf. /kimːakinːa/ [kĩ́mːãɣĩ̀nː̥ḁ] ‘to come here’).

In non-bipositional structures we see the loss of all Edge, including that of hard sonorants. The representation below illustrates the mono- and bipositional configurations, using /m/ as an example: the |ʔ| feature is deleted when it has access to only one C position (29a), and preserved when it is shared over two C positions (29b), that is in geminates and CC clusters.

(29)

The fact that this Bottom up MSCL applies to all occlusion, not just headed occlusion, is the big difference between this type and the Tamazight Berber type.

One wrinkle in the data has to do with the word-initial position. In Type A (Ontena Gadsup and Tamazight) we saw that word-initial singletons lost Edge*, as we would expect from a Bottom up MSCL. Tümbisa is a Bottom up language too, though applying to all Edge not just Edge*. However, unlike other Bottom up languages, word-initially, we see the retention of stops.

This point of variation has an independent explanation, the fact that in Tümbisa one gets an initial-CV and word-initial geminates. Therefore, the presence of stops in ‘monopositional’ initial position is in fact due to initial consonants in the language being bipositional.

In our framework, initial strength is independently motivated in languages by the presence or absence of an initial empty CV (Lowenstamm 1999; Scheer 2004, 2011, 2012). The reader is referred to Section 2.3 for an outline of the initial CV hypothesis.

In Berber, there seemingly is no initial CV. This allows for unrestricted consonant clusters at the left edge, and makes the initial position weak not only in terms of Government but also with regard to the number of positions that the initial consonant has access to: the absence of the initial site prevents the consonant from branching into two C positions (for an alternative view, see Lahrouchi 2001, 2003, 2018a).

In Tümbisha on the other hand, we have no clusters at the left edge of words, consistent with the presence of an initial empty CV. This alone would be enough to suspect that the bipositional strength of initial segments comes from their branching (like geminates) word-initially (cf. (30a) and (31a)). This initial CV is not present if there is any material at the left of the stem (cf. (30b) and (31b)), which is why these forms become weak if they are not string initial (e.g. see below in 30b) (Dayley 1989:514).

(30)

a.

/tonamːinːa/

[tṍnãmːìnː̥ḁ]

‘to stab’

b.

/u tonamːinːa/

[u ɾṍnãmːìnː̥ḁ]

‘to stab it’

(31)

a.

Edge is bipositional string initially

b.

Edge is monopositional when non-string initial

In Tümbisa there is additional evidence that this is the case. The language has inherent geminates, [sː] and [ŋː]. For independent reasons, the language does not allow word-initial velar nasals, so both /ŋ/ or /ŋ^w/ are never found initially, however, there is word-initial /sː/.

Unlike other sounds, /sː/ is an underlying geminate, what it demonstrates is that the language allows linking a geminate word-initially. It stands to reason, therefore, that the other sounds spread into the initial CV rendering them long (and strong) exclusively when these singletons are supplied with the initial CV.

4.2.2.2 Tarifit Berber: Edge in stops and hard sonorants

Tarifit Berber, spoken in Northern Morocco, seemingly stands on the same typological side as Tümbisa in as much as Edge is banned from monopositional structures, specifically in stops and liquids. Strangely, nasals are immune from weakening, which we would not predict (a problem we leave for future research).

In a number of items that are shared by Tarifit, Tamazight and Tashlhiyt, there is a regular change whereby any singleton lateral in Tashlhiyt rhotacizes in Tarifit, while what appears as a Tashlhiyt coda /ɾ/ undergoes deletion in Tarifit and causes compensatory lengthening of the preceding vowel (Bouarourou 2014; Chtatou 1980; Dell and Tangi 1993, Lafkioui 2006, 2007). This is shown in the data in (32a), drawn from Serhoual (2002) and Taïfi (1991), along with examples of the process of spirantization, commonly found in Tamazight and Tarifit (32b). The reader is also referred to Section 4.2.1 for further data and discussion about spirantization.

(32)

Tashlhiyt

Tamazight

Tarifit

a.

ul	ul	uɾ	‘heart’
ifili	ifili	fiɾu	‘string, thread’
adˤaɾˤ	aðˤaɾˤ	ðaː	‘leg, foot’
uɾ	uɾ	uː	‘not’
ʁɾ	ʁɾ	ʁaː	‘to read’

b.

akal	açal	çar	‘ground, soil’
tabɾatː	θaβɾatː	θaβɾatː	‘letter’
aglːid	aʒəlːið	aʒəʤið	‘king’
udi	uði	uði	‘rancid butter’

c.

aman	aman	aman	‘water’
ini	ini	ini	‘to say’
mani	mani	mani	‘where’

Of the Berber varieties spoken in Morocco, Tarifit is the only dialect that expanded the set of consonants undergoing Edge-lenition to the class of sonorants, with the notable exception of nasals (32c). Even if it may appear unclear why the dialect preserves edge in monopositional nasals, in contrast with Tümbisa, there is one important fact about the mechanism regulating the distribution of Edge: This feature is banned from monopositions, regardless of its function within the melodic expression, that is, when it is the head as in stops or dependent as in /l/.^[14]

In an attempt to handle these facts, Lahrouchi (2018b: 10) has proposed that the lenition process consists of one single subtraction operation triggering the Edge element. Removing this element from a stop leads to a spirant; the same operation derives a rhotic from an underlying lateral. This is illustrated in the example represented below.

(33)

Tarifit: /akal/ → [açar] ‘ground, soil’

Two observations are in order. Firstly, the Edge-deletion in Tarifit’s /l/ and Tumbisa’s /t/ (see 31) leads to the same output: [ɾ]. Secondly, leniting /l/ to [ɾ] in Tarifit yields a case of phonological opacity where two instances of [ɾ] may occur in the onset position: e.g. aɾi ‘write’ versus aɾi (from underlying /ali/) ‘climb’. This is in contrast with the coda position in which only [ɾ] derived from /l/ occurs: Compare [uɾ] ‘heart’ to [uː] ‘not’ in (32a).

Our refined typology is thus enriched by these two languages, Tümbisa and Tarifit, that represent the Bottom up MSLC on Edge, not only in stops, but also in hard sonorants, with the case of Tümbisa being consistent all the way down: any occlusion (Edge) is banned from all monopositional structures. The next section completes our typology.

4.2.2.3 Top down (S → M) for Edge*

The final missing section of the typology is one where the MSLC on Edge and Bipositionality is Top down, but where this applies only to the headed occlusion of stops/affricates (Edge*) and not that of hard sonorants.

In a language like this, partial gemination (e.g. nd C1C2) would be allowed but there would have to be either a stop or affricate in C2. Moreover, full geminates would have to be stops or affricates, the language should exclude all non-Edge* geminates, even nasal geminates would be banned here.

(34)

Top down (S → M)

A language like this does exist, though we have only been able to find one instance of this language type. The reason for this may lie in a complicated interplay between phonetics and the specific pathways of phonologization of geminates (Blevins 2008).

The language that completes our typology is Anejom̃. This is the only indigenous language of Aneityum, the southernmost inhabited island of Vanuatu, where it is spoken by approximately 800–900 people (Lynch 2000).

Crucially for us, Lynch (2000:23) shows, that in Anejom̃, only stops can be geminates. Even though there are some geminate affricates, Lynch explicitly mentions that only the stop part is geminated (2000:24).

(35)

Full geminates only contain stops and affricates (Lynch 2000:24)
nop w ːa	‘ashes’
apːei	‘scrape <taro+>’
atːeθ	‘chipped’
akːe	‘poke about in a reef’
ʔatːʃaŋ	‘sneeze’
nɪtːʃan	‘sharp/pointed thing’

For our purposes it means that in a pure geminate structure, only Edge* can sit in its head-position.^[15]

(36)

[nop w ːa] ‘ashes’

This is a robust attestation of a pattern we predicted to find, the Top down MSLC acting specifically on Edge*.

There are further aspects of the grammar that show this MSLC forming or in progress in the development of the language. Consider for instance, the historical origin of these full geminates. Interestingly, these seem to have come historically from /h + C/ sequences. Lynch (2000) reports Inglis [1882]’s transcriptions with this regular sound correspondence.

(37)

Inglis [1882]	Lynch (2000)
ahtaij	aːtːatʃ	‘they (trial)’
nohpa	nop w ːa	‘ashes’
ahpei	apːei	‘scrape (taro)’
ahtag	atːaŋ	‘come together’
ehped	epːeθ	‘castrate’

At this point in the history of the language, the consonant clusters went from glottal fricative-stop, that is non-edge and Edge* to structures that have a headed Edge shared across both parts of the cluster, that is: the branchingness of |ʔ|. This is illustrated in (38) beneath.

(38)

Rejection of bipositional structures not containing branching Edge*

a.

**nohpa16

1. 16

   * refers to an ungrammatical form. ** refers to a historically reconstructed form.

b.

[nop w ːa] ‘ashes’

This development is in line with what we would expect from our MSLC. This language therefore only has Edge* geminates, constituting the missing slot in our typology, the Top down restriction on headed Edge* only.