2.1. Distributions that do not require MSCs

Consider the textbook example of vowel nasalisation in which nasal vowels occur only in assimilation environments but are banned otherwise: [pa], [mã] vs. *[pã], *[ma] (see, e.g., McCarthy Reference McCarthy2002a: 83). In a rule-based account following Chomsky & Halle (Reference Chomsky and Halle1968), nasal vowels might be assumed to be systematically absent (i.e., banned) from URs (/pa, ma/, but not */mã/) and derived by a contextual nasalisation rule. In OT, UR bans are not needed for analysing such cases. Surface distributions are a matter of markedness and faithfulness rankings only. The OT grammar in (4a) derives the surface inventory [pa, mã] from the rich base /pa, ma, pã, mã/. The general schema for complementary distribution is in (4b):

One attraction of this approach is that it uses formally simple constraints: the general constraint legislates feature co-occurence, and the specific constraint governs bigram sequencing. Another attraction is that both constraints are perceptually grounded and receive robust typological support. Third, the ranking tells an intuitive story: ‘no nasal vowels, except adjacent to nasal consonants’. In contrast to rule-based treatments that insist on setting up a unique UR, the account in (4) does not have to decide which vowel is the underlying one, sidestepping the often difficult logic that is usually external to the analysis. In a rule-based framework, it is just as easy to set up a rule deriving oral vowels from nasal-only URs, so the decision to posit oral UR vowels usually recruits considerations such as typology.

Pertinent to the concerns of this article, the analysis in (4) also avoids the duplication problem, because it does not need to assume a constraint against nasal vowels in the UR, and it also does not need to sneak in limitations on inputs via representational assumptions (e.g., ‘only consonants can be specified for nasality contrastively in the input’). There is no duplication here; the analysis works and is elegant.

Positional neutralisation is similarly unproblematic under the rich base assumption, provided the affected contrasts do not have gaps. For example, in Nancowry, stressed syllables contrast oral and nasal versions of [i, e, ɛ, æ, u, ə, a, u, o, ɔ], but unstressed syllables allow only [i, e, ə, a, u] (Radhakrishnan Reference Radhakrishnan1970: 19). McCarthy (Reference McCarthy2002a), 88) posits the following ranking for Nancowry:

Under this ranking, inputs such as /batã/ and /bata/ will map faithfully (assuming final stress), but /bãta/ will neutralise to [batá]. There is no need to restrict the distribution of nasal and oral vowels in the input, or to guess what becomes of the hypothetical nasal vowels in unstressed positions. Since we do see vowels in those positions, and they are always oral, the direction of neutralisation is determined by the system. As we will see in the next section (and also in §5.5), this is not true when the system contains gaps.

2.2. Positional neutralisation with gaps

The problem arises in cases where a contrast involves a gap that is filled in assimilation environments. A classic case is nasal place assimilation, which often creates allophones confined to assimilation contexts: [m] and [n] have a free distribution except in place-assimilated clusters (*mt, *np, etc.), but [ŋ] occurs only before [k, g]. Languages with this pattern include Standard Italian (Bertinetto & Loporcaro Reference Bertinetto and Loporcaro2005), Lithuanian (Kenstowicz & Kisseberth Reference Kenstowicz and Kisseberth1979: 216) and Turkish (Kornfilt Reference Kornfilt2013: 486). As with the vowel nasalisation example, analysing the basic distribution does not require MSCs, as illustrated in (6).

In the absence of alternations, we may speculate that in non-assimilation environments, underlying /ŋ/ maps to, say, [n], as in (7):

But this analysis is incomplete: it is silent on the direction of assimilation. Why not change the oral consonant’s place instead, as in /pan-ka/ $\rightarrow $ [panta]? Jun (Reference Jun1995) argues that plosives carry place cues better than nasals, since the plosive is in prevocalic position. This perceptual property is encoded in a positional faithfulness constraint Ident-Onset (Beckman Reference Beckman1997, among others). Ident-Onset must outrank *[ŋ] to allow assimilation to create velar nasals – otherwise assimilation would be progressive if and only if the onset is underlyingly velar, as in candidate (8c).

The problem is that this ranking is incompatible with the assumption that /ŋ/ maps to [n] except in assimilation contexts. Once the analysis is augmented with positional faithfulness, onset [ŋ] sneaks back in from the rich base:

The obvious way to save the situation is to plug the gap with more positional markedness, for example, by ranking a constraint against onset [ŋ] above Ident-Onset. On the upside, *[ŋ]/onset might be independently necessary: [ŋ] often has a distribution different from other nasals; English is a familiar example (Chomsky & Halle Reference Chomsky and Halle1968 et seq.). On the downside, this approach does not generalise. Nasal place assimilation often creates segments such as [ɱ] or [ɲ], which have less subtlety to their distribution than [ŋ]. They are absent except in assimilation environments. This analysis would have to use multiple markedness constraints just for the gapped segments in the environment protected by the positional faithfulness constraint, and rank them in an order that mirrors the positional $\gg $ general faithfulness order:

If the argument for the rich base is that it avoids the duplication problem, then cases like this defeat it; surely it is more elegant to have a single MSC against velar nasals at the UR level. One of the selling points of Optimality Theory is that ‘the constraints provided by Universal Grammar are simple and general; interlinguistic differences arise from the permutations of constraint-ranking’ (Prince & Smolensky Reference Prince and Smolensky1993 2004: 6). But in this case, constraints on restricted allophones cannot be simple or general. Intuitively, the distribution is simple: [ŋ], etc., are banned except where nasal place assimilation requires them. In this analysis, however, the prohibitions on these segments must be stated twice: both below and above positional faithfulness.

It is a well-known feature of positional faithfulness constraints that they determine positions of static contrast as well as direction of assimilation. This is a feature, not a bug, and trying to solve the problem by getting rid of positional faithfulness is unlikely to work (see §§5.4 and 5.5.3). Nor is it likely that the answer lies in splitting Ident[place] into separate constraints for different manners of articulation, as suggested by a reviewer: NasPlaceAssim, Ident[place]_obstruent $\gg $ *[ŋ] $\gg $ Ident[place]_nasal. There are various issues that this alternative would have to resolve. The generalisations about direction of place of assimilation in NC clusters hold pretty robustly even when the consonants in the clusters are not faithful to their underlying manner features. For example, in Tswana, onsets determine the place of nasals even when the onsets are unfaithful to their underlying manner: [ɸula] ‘shoots’ $\sim $ [m-pʰula] ‘shoot me’, [rut’a] ‘teaches’ $\sim $ [n-tʰut’a] teach me’ (Gouskova et al. Reference Gouskova, Zsiga and Tlale2011). In Japanese, nasal codas assimilate to following onsets even when the codas are underlyingly non-nasal: /job-/ [jo\uline{b}-u] ‘call-pres’ $\sim $ [jo\uline{n-d}a] ‘call-past’, cf. [ʃin-u] ‘die-pres’ $\sim $ [ʃin-da] ‘die-past’ (Martin Reference Martin1975). Onsets can even determine the place of a following nasal, as in German [geb-] ‘give’, [tʀɑ:g-\textsyllabic{ŋ}] ‘carry’ (Wiese Reference Wiese1996). Positional faithfulness to onsets predicts all of these patterns.

Is there a way to reconcile the analysis of complementary distribution and this positional gap problem? The alternative I advocate is to bring back MSCs. Removing /ŋ/ from the input in an Italian-style grammar removes the need to guess as to its fate in the system, and it also simplifies the analysis of neutralisation. In the analysis in §3, adding MSCs allows for an elegant and unified analysis of Russian and Polish: they have the same input–output grammars (regressive assimilation, final devoicing), even though their inventories differ.

While I will devote effort to motivating MSCs in the case of positional neutralisation with gaps, I will not offer a definitive proposal on whether MSCs should be used in other cases, such as complementary distribution or positional neutralisation without gaps (though see Rasin & Katzir Reference Rasin and Katzir2016). MSCs are justified whenever the analyst has to guess about the fate of unseen segments. But they are not necessary in Nancowry-style positional neutralisation, since the grammar suggests the direction of assimilation. In §5.5.3, I discuss root–affix positional asymmetries, whose analysis also benefits from entertaining hypothetical inputs of affixes that contain segments seen in roots (a limited rich base). Likewise, MSCs are not needed for complementary distribution, even if they do no harm. Indeed, if we assume that MSCs are drawn from the pool of plausibly universal markedness constraints, they solve the problem that lacks a principled solution in rule-based analyses: how to decide which segments to rule out from URs. Typologically, nasal vowels are marked, justifying both the MSC and the surface markedness constraint *\~{V}. If URs are limited to oral vowels only, /pa, ma/, the ranking *NV_oral $\gg $ Ident easily derives the distribution [pa, mã, *pã, *ma]. The difference between MSCs in rule-based theories and this updated approach to MSCs is that the MSCs can be drawn from the set of markedness constraints. If a constraint is an MSC, it bans certain things from the lexicon, without specifying how to remove them. If it is a constraint in the grammar, it interacts with faithfulness.

2.3. Morpheme structure constraints, not rules

Early discussions of MSCs debated whether they should be construed as rules or constraints. Rules supply instructions for removing the offending structure. Constraints simply ban the structure, leaving multiple avenues for removing it. Halle (Reference Halle1959) and Stanley (Reference Stanley1967) disagreed on this point: in Halle’s original formulation, morpheme structure rules were formally similar to regular phonological rules that legislated redundancies (e.g., if [+strident], then [–voiced]). But Halle’s MSCs could also be feature-changing (e.g., if [+strident, +voiced], then become [–voiced]). Stanley proposed to formulate all MSCs as ‘redundancy rules’, which are essentially constraints: they specify what feature combinations and sequences may occur, but not instructions for removing offending structures. Translating this into OT terms, we would say that MSCs are markedness constraints that do not interact with faithfulness, because they hold at a level where no mappings happen.

An alternative to this view is Stratal OT (§5.5), where the role of MSCs is subsumed into the stem level. There, markedness and faithfulness constraints filter the rich base. My approach and this alternative agree on the need to filter the input before neutralisation applies. The disagreement lies in whether the analysis takes a guess about what happens to the illicit segments. My analysis is agnostic of their fate: I would argue that neither the analyst nor the learner has a way of figuring out what happens to /ŋ/ in Standard Italian, or to /ʤ/ in Russian. The evidence for learners not knowing the fate of ROTB inputs is inconsistent behaviour in borrowing, which I attribute to conventional rules (§4). Russian cannot lexicalise a loanword with /ŋ/, but it also does not have a single set of grammatical instructions for removing it. As I show in §5.5, however, the specifics of Russian preclude an internally consistent analysis in Stratal OT. Both the native phonology and the loans involve heterogeneous sets of segments; no one mapping can be justified analytically. There is evidence for a constraint, but no evidence that it lives in a coherent ranking.

3.1. Halle’s argument

The duplication problem in Russian voicing phonology was first spotted by Halle in his criticism of the structuralist phonemic level (Halle Reference Halle1959; Anderson Reference Anderson1985). As shown in (11), Russian voicing assimilation affects all obstruents, whether they contrast for voicing (e.g., /k, g/) or not (e.g., /ʧ/). In a structuralist treatment, contrastive distinctions are represented differently from non-contrastive ones. So, Halle points out, the voicing assimilation rule must be stated twice: once at the phonemic level, to cover contrastively voiced phonemes, and then again, to cover non-contrastively voiceless segments such as /ʧ/. A simple generalisation is missed:

Halle’s (Reference Halle1959: 61–63) analysis includes a morpheme structure rule that requires /ʦ, ʧ, x/ to be unspecified for contrastive voicing at the UR level, as well as a phonological rule of voicing assimilation for obstruent clusters. My analysis is similar to Halle’s: I also assume that at the UR level, certain segments are banned (viz., */ʣ, ʤ, ɣ, ʒʒ/), although this claim requires some caveats (see §§3.3.3 and 3.3.4). Two-consonant sequences such as /dz, dʐ/ are allowed in URs.

Halle’s argument for MSCs is strikingly similar to the one advanced by Prince & Smolensky ([Reference Prince and Smolensky1993] 2004) for Richness of the Base (and by extension, against MSCs). MSCs pose a duplication problem in cases where morphemes obey the same restrictions as words. As I will show, though, the popular account of Russian voicing in OT also runs into a duplication problem. The difference lies in where the duplication happens: in the constraint/rule system, or between the rule system and the MSCs.

My analysis is cast in OT for several reasons. First, within rule-based analyses, the need for MSCs is usually taken for granted, and the argument would be preaching to the converted. Not so in OT approaches, which either take Richness of the Base for granted or explicitly argue for it (Davidson et al. Reference Davidson, Jusczyk, Smolensky, Kager, Pater and Zonneveld2004; Jarosz Reference Jarosz2006a, Reference Jarosz, Wicentowski and Kondrak2006b; Tessier & Jesney Reference Tessier and Jesney2014). Second, the OT analysis of Russian voicing is appealing: it has robust typological support and is grounded in the phonetics of contrast. Consequently, it is a rare point of near-universal agreement in OT. And yet this analysis has a problem with facts known since Halle (Reference Halle1959), something that has largely gone unnoticed in the literature on voicing neutralisation (one notable exception is Hall Reference Hall and Radišić2007, discussed in §5.2). One of my goals is to revive the argument, and to contribute some descriptive depth to it. Ultimately, I believe the point about MSCs holds regardless of whether the grammar of mappings is construed in rule- or constraint-based terms.

3.2. A constraint-based analysis of Russian voicing with MSCs

The analysis of voicing neutralisation in Russian follows similar lines in many OT treatments (Lombardi Reference Lombardi1999; Steriade Reference Steriade1999; Padgett Reference Padgett2002; Rubach Reference Rubach2008; Beckman et al. Reference Beckman, Jessen and Ringen2009; Padgett Reference Padgett, Borowsky, Kawahara, Sugahara and Shinya2012). It is uncontroversial that the voicing contrast is limited to presonorant position.Footnote ² This arises through the interaction of *ObsVoice (defined in (12a)) with a positional and a general Ident (defined in (12b)–(12c)).

The ranking is in (13a); voiced stops contrast in presonorant position, as in (13b), but devoice word-finally, as in (13c).

To enforce assimilation, Agree[voice] must dominate Ident[voice] and *ObsVoice, assuming that each obstruent in a cluster such as [db] violates *ObsVoice (see (15)). Agree is undominated: voicing always assimilates. Ident-Preson[voice] is also undominated: it ensures that the presonorant obstruent determines voicing of the cluster. The ranking derives the right results for sequences of multiple obstruents (/mozg bɨ/ $\rightarrow $ [mozg bɨ] ‘brain irr.’; /vosk bɨ/ [vozg bɨ] ‘wax irr.’; /mozg to/ [mosk to] ‘brain topic’) without embellishments.

Where I depart from Lombardi (Reference Lombardi1999); Padgett (Reference Padgett, Borowsky, Kawahara, Sugahara and Shinya2012), and everyone else is in accounting for forms such as [noʤ bɨ]. I argue that inputs with /ʤ/ are disallowed by the morpheme structure constraint in (16), */ʤ/, which prohibits voiced affricates in the lexicon.

As explained in §2.3, the MSC */ʤ/ does not interact with faithfulness or markedness constraints. It might seem intuitive that */ʤ/ is in conflict with Agree[voice], but Agree does not dominate */ʤ/. The affricate [ʤ] is created by assimilation on the surface, where the MSC does not apply. The only way to satisfy */ʤ/ is not to have lexical entries with the proscribed segments (we will see this idea applied in §4). I show below why it cannot be the case that /ʤ/ maps to [ʧ] or to [dʐ] in hypothetical forms such as /ʤip/. The claim is that the grammar does not know how to map such forms to outputs; all it knows is that they are not legal inputs.

In my analysis, there is just one constraint on voiced affricates in the grammar of Russian, and it holds of lexical entries only, where it is fully satisfied. This analysis would need to be augmented by some additional constraints such as */ɣ/ and */ʒʒ/ to be complete; I consider those in §3.3.

The factorial typology of this constraint set predicts the seven systems in Table 1 (cf. Lombardi Reference Lombardi1999, Petrova et al. Reference Petrova, Plapp, Ringen and Szentgyörgyi2006).Footnote ³

Table 1 Factorial typology of the standard Positional Faithfulness account.

MSCs are by nature language-specific, since lexicons are language-specific. But MSCs are not completely typologically inert, despite not being part of any factorial typology. The presence of the MSC will have consequences, restricting the distribution of the segment in question. While Russian and Polish have the same input–output mappings, the presence of the MSC */ʤ/ in Russian means [ʤ] is not freely distributed. Polish lacks an analogous MSC, and so it has a full two-way voicing contrast in stridents (Gussmann Reference Gussmann2007, among others). Limiting the input changes the surface contrast possibilities but does not affect the alternation patterns predicted for voicing.

My analysis includes just one constraint against /ʤ/, but suppose there is a redundant set of markedness constraints identical to MSCs. If so, the output constraint *[ʤ] must be dominated by Agree and Ident-Preson[voice] – otherwise /noʧ bɨ/ could not map to [noʤ bɨ] (as shown in (18)). As the tableau shows, *[ʤ] favours no winners and does no work in the analysis. There is no reason to include it in the analysis if the MSC is assumed – the statement about gaps only needs to be made once.Footnote ⁴

Crucially, however, the ranking in (18) would yield the wrong result if a rich base were assumed. Agree[voice] and Ident-preson[voice] must dominate *[ʤ]: voiced affricates are tolerated when the alternative is voicing disagreement or devoicing a presonorant obstruent. But this ranking creates a problem under standard OT assumptions: while the ranking works for the inputs considered so far, it rules in the wrong forms if the base is rich. Input /ʤ/ is predicted to map faithfully in presonorant position, which is incorrect.

The problem is a general one for Richness of the Base. When a structure does not occur on the surface in a language, the usual OT explanation is to say that markedness outranks faithfulness. But in the Russian case, the native faithfulness ranking derives the wrong outcome: the wrong thing is ruled in. If we rejigger the ranking to map /ʤ/ to something else – say, a sequence such as [dʐ], which is how it is pronounced in loanwords (see §4) – then it breaks the account of native assimilation, since the option to decompose the affricate into a CC sequence now needs to be ruled out for native voiced affricates in contexts where they are derived by regressive assimilation.

The usual avenues for saving a theory are to question the facts (e.g., by questioning the phonetic reality of assimilation) or to modify the theory by complicating it (e.g., by adding elaborate constraints, or entire derivational levels as in Stratal OT). The first strategy does not seem promising. All sources that discuss the assimilation pattern in sufficient detail agree on the facts (Halle Reference Halle1959; Comrie et al. Reference Comrie, Stone and Polinsky1996; Garde Reference Garde1998, among others): Russian has non-contrastively voiced allophones in assimilation contexts (see also Figures 1–4). This contrasts with other claims made about Russian voicing, such as sonorant transparency, which has engendered far more controversy (see Kulikov Reference Kulikov2013 for a review). The second strategy of complicating the theory comes in many flavours, as I discuss in the section on alternatives in §5. The problem with these, I argue, is that they either introduce the duplication problem or do not succeed on OT terms. Before discussing alternatives, I consider the full complexity of the Russian facts: the other inventory gaps and loanword adaptation.

Figure 1 Left: voiced [ʤ] derived by voicing assimilation in /vraʧ bɨl starɨj/ ‘the doctor was old’. Right: the same speaker’s [ʧ] before a sonorant in /vraʧ nina/ ‘doctor Nina’.

Figure 2 Voiced [ʣ] derived by regressive voicing assimilation in /liʦ ʐe/ ‘persons c.foc ’.

Figure 3 Voiced [ɣ] derived by regressive voicing assimilation in /nikakix bɨ/ ‘none irr. ’.

Figure 4 Voiced [ʒʒ] derived by regressive voicing assimilation in /pomoʃʃ budet/ ‘help will be’.

3.3. Russian voicing assimilation in detail

The consonant inventory of Russian is in Table 2. (Orthographic representations are given in angle brackets where Russian/Slavicist readers might find them useful.) There is a [ $\pm $ back] (palatalisation) contrast for most consonants (Rubach Reference Rubach2000; Padgett Reference Padgett2003). There is also a voicing contrast for most obstruents.Footnote ⁵ The strident and velar series have gaps in voicing and backness (in the shaded regions of the table). Both affricates are voiceless, and they lack same-place backness counterparts: [tsˠ, tʃʲ] but not *[tsʲ, ʈʂˠ]. The consonant [ʃʃʲ] lacks a robust voiced counterpart; [ʒʒʲ] is at best marginal (see §3.3.5). Finally, there are no retroflex/velarised affricates, *[ʈʂ ɖʐ] (this is unlike Polish, whose strident affricate and fricative contrasts have no gaps in place or voicing; see Gussmann Reference Gussmann2007; Padgett & Żygis Reference Padgett and Żygis2007; Żygis & Padgett Reference Żygis and Padgett2010).

Table 2 Russian contrastive consonants (Padgett Reference Padgett2003; Padgett & Żygis Reference Padgett and Żygis2007).

Table 2 includes transcription details that are often omitted, and I suppress them in the remainder of the article (thus I follow the convention of marking palatalisation but not velarisation, and only where there are robust contrasts; see Padgett Reference Padgett2003; Iskarous & Kavitskaya Reference Iskarous and Kavitskaya2018; Gouskova & Stanton Reference Gouskova and Stanton2021).Footnote ⁶ I transcribe vowels phonemically, except for [i] and its back allophone, [ɨ], which occurs after velarised consonants. (I ignore the analogous retraction of [e], [kˠófʲe] ‘coffee’ vs. [kˠafˠé̠] ‘cafe’.)

The paradigm in (20) shows assimilation patterns for a fuller range of consonants. The unpaired consonants appear in the last four sets of forms, in (20e)–(20h).

3.4. Local summary

Four Russian obstruents can be argued to lack a systematic voicing contrast: [ʧ, ʦ, x, ʃʃ]. All four have voiced allophones in regressive assimilation. But in other contexts, the consonants vary. On the surface, [dz] does occur outside the assimilation environment, although it might be analysed as a CC sequence there. As for [ɣ] and [ʒʒ], they definitely occur in ambient speech, but their variable presence and restricted stylistic/morpho-phonological distribution complicate their analysis. By contrast, [ʤ] alone is found only in regressive assimilation contexts; it is this allophone, therefore, that presents the clearest argument for an MSC. Its special status is also supported by loanword phonology, as I show next.

4.1. Borrowing data

The preceding discussion argued for MSCs on analytic grounds: an insightful analysis of Russian voicing neutralisation must rule out gaps at the UR level. MSCs allow for a simple, cross-linguistically valid analysis of voicing neutralisation in the grammar of input–output mappings. Unlike some OT alternatives in §5, the MSC account uses formally simple constraints that can be motivated substantively.

In this section, I consider the role of MSCs in loanword adaptation, with a somewhat narrow focus on Russian. The broader question is what Richness of the Base and MSCs predict for the handling of segments that a language lacks entirely. The usual OT explanation is that inventories are determined by markedness and faithfulness rankings; if a segment is missing, it violates an undominated markedness constraint. But which faithfulness constraint is violated in the mapping from a rich input to the output? This question rarely receives a clear answer.

In the case of Russian, we will see that foreign [ʤ] is adapted as though the phonology cannot even represent it as a single sound, and the handling of loan [ʤ] leaves few avenues for saving an analysis of voicing neutralisation that does not rely on MSCs.

A striking feature of Russian loanword adaptation is that some segments are borrowed as consonant clusters. For example, Russian lacks a velar nasal [ŋ], even in place assimilation contexts, where many languages require it. Thus, in borrowings from English and German, [ŋ] maps to [ng] or [nk], as in the examples in (31a). But Chinese and Korean loans follow a different convention: source [ŋ] maps to [n], as in (31b).

Borrowing [ŋ] as [nk] could be orthographically motivated: German and English lack a single letter to write [ŋ], so Russian speakers render the orthographic cluster as a sequence of two sounds. If they were basing their pronunciations on perceptual input, they might be expected to render [ŋ] as [nʲ]. I suggest an explanation for this differential adoption pattern in §4.3.

But orthographic borrowing cannot explain what happens to [ʤ].Footnote ¹⁵ Russian borrows [ʤ] from a variety of languages, including ones where it has no single consistent spelling (English) or where orthography is unlikely to have been the main mode of contact (Turkic, Arabic via Persian). Almost without exception, [ʤ] is borrowed as [dʐ]:

Word-finally, the sequence becomes [tʂ], not [ʧ]. This is systematic and does not depend on the source language:

By contrast, voiceless [ʧ] and [ʦ] are borrowed as affricates:

There are comparatively few borrowings with [dz], and the Russian string’s status is unclear (recall §3.3.3). One source is Polish, whose [ʣ, ] are borrowed as [dz] and [dzʲ], respectively. Another source is Japanese, whose affricated /d/ before [i] is borrowed as [dzʲ], with palatalisation disagreement. It is unclear whether this pattern is guided by perceptual similarity or convention (see Kang Reference Kang, Oostendorp, Ewen, Hume and Rice2011 for more).

While [ʤ] is usually borrowed as [dʐ], there are exceptions (see (36a)). Some of these are very recent (‘gender’, ‘digitiser’). Some older borrowings from English have [ʐ]ː [ʐokej] ‘jockey’, [ʐuri] ‘jury’, [piʐama] ‘pyjamas’, [sufraʐɨzm] ‘suffragism’.Footnote ¹⁶ Vasmer (Reference Vasmer1958) speculates that [ʐokej] was borrowed from English via French, which would explain the [ʐ], but it is not clear that this holds for the other examples. This kind of inconsistency is not unusual; recall [ŋ] in (31). Consider also [h], which is [g] in older borrowings but [x] in contemporary ones (see (36b)). The glide [w] is borrowed as [v] or [u] in contemporary Russian, sometimes in the same word (see (36c)). In places where English orthography is under-informative, for example, as to the voicing of [s, z], Russian sometimes borrows [s] as [z] (see (36d)). This would be surprising if borrowing happened via perception, since Russian has both [s] and [z], and speakers should be able to distinguish voicing contrasts. I think a better explanation is that borrowing is agrammatical, and I analyse it as such in the next section.

Another argument against a perceptual account is that English, the source of the vast majority of recent borrowings into Russian, does not have true voicing in [ʤ]. The English contrast is one of aspiration. If Russian speakers were using their perception, as opposed to orthographic and metalinguistic conventions, we would expect them to (occasionally) borrow [ʤ] as [ʧ], and they do not.

The main significance of the borrowing facts is that [ʤ] is not borrowed as an affricate, even in cases where it could be devoiced to a native sound (in words like image). Most of the time, it is decomposed into a stop and a fricative, and sometimes it is borrowed as other sounds. The most troubling aspect of this pattern for an OT account is that the ranking suggested by the native voicing alternations cannot be reconciled with loanword adaptation, as I explain in §4.3. I argue instead that [ʤ] is mapped to /dʐ/ at the point of lexicalisation, by a conventional mapping rule. This rule exists to enforce the MSC, but it is not part of the grammar of voicing neutralisation.

4.2. Evidence for analysing [dʐ] as a CC sequence

The argument that loanwords are a problem for Richness of the Base needs some evidence that [dʐ] is indeed a CC cluster. This is not a foregone conclusion: the only Western study that addresses [ʤ] loans into Russian, Benson (Reference Benson1959), characterises [dʐ] as an affricate. This is, in my opinion, incorrect.

First, consider the place of articulation change in borrowing [ʤ]. If viewed as a phonological mapping, the treatment of place of articulation is inconsistent and puzzling. Russian systematically maps [ʃ] to a retroflex [ʂ] when borrowing from English, German, French and other European languages. When borrowing from Japanese, the sound variably transcribed as [ʃ] and [ɕ] maps to Russian [sʲ]: [xirosʲima] ‘Hiroshima’, [xonsʲu] ‘Honshu’. When Russian borrows English [ʤ], it renders the affricate as a sequence of dental and retroflex articulations. By contrast, Russian borrows [ʧ] without major alteration (it is phonologically palatalised, but the minor place distinction is probably too subtle to detect in the acoustics; see Jongman et al. Reference Jongman, Wayland and Wong2000; Żygis Reference Żygis2003). It is certainly not [tʂ] or [ʈʂ] in loanwords. The simplest explanation for the fate of [ʤ] is that it is conventionally mapped to stand-alone sounds available in Russian, [d] and [ʐ]. The heterorganicity of this sequence is therefore one of the best arguments for its analysis as a CC cluster.

Next are some language-internal arguments for analysing [dʐ] as two segments, following Trubetzkoy (Reference Trubetzkoy1939). First is the diagnostic of phonotactics. Phonotactic arguments work well in languages like Fijian, in which words cannot start with CC sequences, but they may start with [mb] and [nr]; the analysis of the phonotactics is simpler if these sequences are prenasalised consonants rather than CC clusters. In Russian, by contrast, phonotactic patterns are rather permissive: words can start with many different sequences; even [ʧ] and [ʦ] cannot be distinguished from stop-fricative sequences on this basis (Gouskova & Stanton Reference Gouskova and Stanton2021). Thus, phonotactics is of little help.

But phonotactic permissiveness no doubt facilitates the borrowing of sequences that do not occur in native morphemes. This is relevant to another of Trubetzkoy’s (Reference Trubetzkoy1939) diagnostics: affricates are supposed to be freely distributed within morphemes, while CC clusters might be more likely to occur at morpheme boundaries only. In native Russian words, [dʐ] occurs across morpheme boundaries (e.g., [pod-ʐog] ‘arson’), but not morpheme-internally. The native portion of the Russian lexicon suggests that [dʐ] must be two consonants. Benson (Reference Benson1959), by contrast, speculates that occurrence at morpheme boundaries facilitates borrowing of [dʐ] as an affricate. If this were a legitimate pathway towards borrowing non-native sounds, there would be no debate in English phonology about the status of [ts], given words like ou\uline{t-s}ide and ca\uline{t-s}. As it is, most analyses of English treat [ts] as a CC sequence partly because it almost never occurs morpheme-internally (see Gouskova & Stanton Reference Gouskova and Stanton2021). Thus, morphological distribution diagnoses [dʐ] as a CC sequence.

The third Trubetzkoyan diagnostic is phonetic duration: CC sequences should be longer than single Cs (affricates). Brooks (Reference Brooks1964) shows that duration correlates with the trzy/czy distinction in Polish; the CC parts of [tʂɨ] ‘three’ are longer than the parts of the affricate in the question particle [ʈʂɨ] (though in Polish, the difference mostly affects the fricated portion, not closure). This diagnostic is problematic when applied cross-linguistically (Arvaniti Reference Arvaniti2007; Stanton Reference Stanton2017; Gouskova & Stanton Reference Gouskova and Stanton2021), but for what it is worth, it goes in the same direction in Russian as in Polish: the affricates are shorter than stop-fricative clusters. Figure 5 shows loanword [dʐ] and native heteromorphemic [d-ʐ], from recordings of the same speaker (the actor Oleg Tabakov).Footnote ¹⁷ The [d] in [dʐ] is fairly long.

Figure 5 Left: borrowed morpheme-initial [dʐ]; right: native heteromorphemic [d-ʐ].

Figure 6 shows this same speaker’s intervocalic [ʧ], which has the very short closure that appears to be typical in Russian (recall also Figure 1). One cannot draw conclusions about C vs. CC status from acoustics alone, especially when the sequences differ in voicing and constriction location, but this is still suggestive.Footnote ¹⁸

Figure 6 Native tautomorphemic [ʧ].

Fourth is a morpho-phonological diagnostic: in Russian, the allomorphy of the diminutive. The allomorph [-ok] tends to not attach to CC-final stems, while the allomorph [-ik] is found on disproportionately many CC-final stems (Gouskova et al. Reference Gouskova, Kasyanenko and Newlin-Łukowicz2015). The [-ik] allomorph occurs on one [dʐ]-final noun, [kotédʐ] ‘cottage’: RNC has 14 instances of [kotédʐ-ɨk] ‘cottage dim’, and none of [kotedʐ-ók]. This is consistent with [dʐ] being a CC sequence, although more systematic study is needed.

The last Trubetzkoyan diagnostic is inventory structure. In Russian, voice contrasts are mostly symmetrical: [b, p], [d, t], etc. Of course, many inventories have gaps – but the Russian system is less typologically odd if only [ʧ, ʦ] are affricates. Under the alternative analysis, the affricate [dʐ] has a relatively free distribution (albeit mostly in loanwords), while its voiceless counterpart, [tʂ], occurs only as a devoiced allophone (also in loanwords, such as [imitʂ] ‘image’), and at morpheme boundaries (in native words such as [ot-ʂelʲnik] ‘hermit’). That is an odd distribution. By contrast to this hypothetical, the analysis of [dʐ] as a cluster treats Russian as a typologically typical gapped system (per Żygis et al. Reference Żygis, Fuchs and Koenig2012).

To conclude, the evidence points to analysing [dʐ] as a CC cluster. When it devoices, the result is also a cluster, [tʂ].

4.3. Is this fission?

Superficially, the loanword pattern seems like a conspiracy: Russian avoids voiced [ʤ], devoicing it in native contexts and fissioning it in non-native morphemes. But this intuitive characterisation does not translate into a neat analysis under Richness of the Base. There are two problems:

Put in formal terms, Integrity (McCarthy & Prince’s Reference McCarthy and Prince1995a anti-fission constraint) has no good ranking position in the standard analysis. If Agree[voice] and *[ʤ] dominate Integrity, we expect fission in assimilation contexts (problem (37a)).Footnote ¹⁹ If underlying /ʤ/ devoices except in assimilation contexts, then we expect devoicing in loanwords as well, which is wrong (problem (37b)). Parallel OT encounters another problem, namely, fission in word-final position in loanwords such as [imitʂ] ‘image’ constitutes overkill. Regardless of the ranking of Integrity, devoicing should be enough if /imiʤ/ is the UR: *[imiʧ] changes just one feature, while [imitʂ] changes voicing, number of segments, place and [back]. Whatever is happening in loanwords is not a straightforward extension of the native pattern. This is not rare cross-linguistically, of course (Broselow Reference Broselow2004; Kang Reference Kang, Oostendorp, Ewen, Hume and Rice2011; Simonović Reference Simonović2015).

4.4. Conventional mappings

I argue that loanword [ʤ] must be decomposed before a morpheme enters the lexicon. I adopt Simonović’s (Reference Simonović2015, ch. 6) conventional mappings, which are agrammatical rules, established in the community by convention to map foreign structures to native ones. Community conventions can differ between dialects of the same language even if there are no relevant grammatical differences between the dialects. Simonović discusses Belgian and Netherlandic Dutch, which borrow English [æ] as [ɑ] and [ɛ], respectively (by using different conventional mappings). Russian motivates the following conventional mappings:

There is no direct connection between these mappings and MSCs. It so happens that all MSCs are satisfied, but this is not a necessary feature: a language may borrow a foreign segment and restructure its inventory, in which case MSCs might eventually change. Simonović has many arguments for this view of loanword adaptation, which I will not rehearse here. Conventional mappings do explain several intractable puzzles.

First, this view of loanword lexicalisation straightforwardly explains why loan [iməʤ] does not devoice to *[imiʧ]. In my account, the UR is /imidʐ/. Its mapping to *[imiʧ] is ruled out by Uniformity (the anti-fusion constraint) and Ident[back].

Second, conventional mappings can be mutually inconsistent, as in the differential adoption of [ŋ]. Since [ng] does not result from fission, we do not need to ponder why the segments appear in that order, or why [dorsal] is preserved in English/German borrowings but not Chinese ones. The agrammatical account allows for conventional mappings to arise because of different borrowing channels (with and without exposure to orthography, perhaps). It can also be influenced by extralinguistic factors suh as prestige. There is compelling evidence for such influences: in Lev-Ari & Peperkamp’s (Reference Lev-Ari and Peperkamp2014) experiment, French listeners adopt the fake loanword [ʤenna] more faithfully when presented as the name of a prestigious item (Italian ice cream) than when it is a non-prestigious Italian beer.

Conventional mappings illuminate another mystery: the inconsistent handling and lack of fission in English interdentals (see (39)). Mapping [θ] to /t/ and [ð] to /z/ is mutually inconsistent. Zooming out, if [ŋ] and [ʤ] undergo fission, then we expect fission for interdentals – perhaps to /tx/ or /dv/, which preserve [coronal] and [continuant]. Instead, [z] adds [strident], and [t] removes [continuant]. My explanation is that these are conventionalised mappings, which are under no requirement to be consistent with anything. Supporting this, Greek [θ] is borrowed as [f] in religious vocabulary (e.g., [anafema] ‘anathema’). And some English loans are inexplicable exceptions, such as the mapping of [ð] in [golsuor\uline{s}i] ‘Galsworthy’.

Conventional mappings exist because the language lexicalises morpheme representations in forms that use native segments. Loanwords show that the grammar is not equipped to handle a rich base; there is just no evidence here for a grammar where the regularities are enforced by a consistent ranking of subordinated faithfulness constraints (Broselow Reference Broselow2004).

To be clear, I would not claim that all loanword adaptation works via conventional mappings, either in Russian or in other languages. Our current understanding of loanword adaptation suggests that the nature of contact influences the mechanisms of adaptation. Some loanword adaptation patterns have roots in perception, but it is controversial whether they interact with the grammar (Silverman Reference Silverman1992; Kang Reference Kang2003; Peperkamp et al. Reference Peperkamp, Vendelin and Nakamura2008, among others). There are also numerous examples of adaptation that do not lend themselves to a grammatical explanation (see Kang Reference Kang, Oostendorp, Ewen, Hume and Rice2011; Simonović Reference Simonović2015). Loanword adaptation is likely not one thing but many things.

5.1. More on the duplication problem

The duplication problem, as framed both by Halle (Reference Halle1959) and by Prince & Smolensky ([Reference Prince and Smolensky1993] 2004), is a problem of theoretical economy. A theory misses a simple generalisation about a pattern, requiring two separate devices. Halle’s critique rests on the intuition that there should be one treatment for all segments, contrastive or not. He criticises structuralism for needing two voicing-assimilation rules, but his own account also handles non-contrastively voiceless consonants in several places. First, a morpheme structure rule requires /ʦ, ʧ, x/ to lack a voicing specification. Then, a phonological rule gives these consonants redundant features. His voicing assimilation rule is general, but the phonological system is not simple. Of course, as we have just seen, the Russian system is more complex than Halle’s presentation suggests, so a simple analysis is unlikely.

The problem for Optimality-Theoretic analyses is more dire, I think. All of the constraints in the positional faithfulness analysis have been recruited in the analyses of languages other than Russian. They are well-motivated substantively and typologically. Voiced obstruents are aerodynamically difficult, and many languages avoid voicing in stops (Westbury & Keating Reference Westbury and Keating1986, among others). The prohibition on voiced affricates is similarly well-grounded, and the Russian-style gapped inventory, where affricates are voiceless, is typologically common (Żygis et al. Reference Żygis, Fuchs and Koenig2012). So this is an analysis worth saving. It is interesting, therefore, that most OT analyses either cannot handle these facts or encounter a duplication problem. Duplication is a problem for Disalign (§5.2), positional markedness (§5.4) and comparative markedness (§5.3). By contrast, Stratal OT (§5.5) fails to supply an internally consistent account of Russian and makes some odd typological predictions, depending on the specific internal assumptions.

5.2. An alternative: Hall’s Disalign

The shape of the nasal assimilation problem in §2.2 suggests a general solution: ban the gapped segment in the environment where positional faithfulness protects contrasts. This solution is obviously duplicative: assimilation is handled once for all segments, and then again just for the gapped segments. A non-hypothetical example of such an analysis is Hall (Reference Hall and Radišić2007), who identifies the problem presented by voiced affricates in Russian and Czech (the latter only involves /ʦ, ʧ/). Hall’s solution is to augment the analysis with Disalign (40):

Unlike *[ʤ], which is simple paradigmatic feature co-occurence constraint, Disalign penalises (i) affricates as sole sponsors of [+voice], or (ii) sequences assimilated for [+voice] and [del rel]: [ʣʤ, pʤ] are bad, [bʤ, ʤb, dʤ, ʤʧ] are good. Hall is himself skeptical, noting that Disalign cannot distinguish the legitimate Czech word [le:ʤba] ‘cure’ from banned *[le:bʤa]. His fix is a directional Disalign-R, which penalises *[bʤ] but not [ʤb].

This style of approach can be criticised for any number of reasons. First, it does not escape the duplication problem for the reasons already explained. Second, the constraint is complex and stipulative; it is unclear why [ʤb] should be preferred to [bʤ] at all, or why either is better than, say, intervocalic [ʤ]. Third, an OT account should be judged on its predicted typology. The basic typology of the constraint set {Agree[voice], Ident, Ident-Preson, *ObsVoice} generates seven attested patterns (recall Table 1). By contrast, adding {Disalign-R, Disalign} predicts 19 systems, including some rather intricate patterns that are, I think, unattested. Two such systems are illustrated in (41) and (42). The first language has regressive assimilation in obstruent clusters in general (e.g., [atpa], [adba]: Agree dominates Ident), and [ʤ] occurs as a singleton (Ident dominates Disalign). But in clusters containing /ʤ/, there is wholesale devoicing. This is the opposite of Russian: [ʤ] is allowed except in assimilation contexts, and the presonorant contrast is conditional on a nearby stop. The second language has word-final devoicing and limited regressive assimilation. It is not triggered by presonorant obstruents except for /ʤ/.Footnote ²⁰

Another odd prediction of Disalign is spread of [del rel] in voice-unassimilated clusters as a way to avoid voicing assimilation. The mapping /paʤta/ $\rightarrow $ [paʤʧa] satisfies Disalign, since [ $+$ del rel] is linked to the entire cluster, while [+voice] is linked only to the first segment.

As was shown in §2.2, such solutions do not generalise. A specific Disalign constraint on affricate voicing in certain clusters might suffice for Czech, whose gaps *[ʤ, ʣ] form a natural class, but additional Disalign constraints would be needed for Russian, where the gaps are not a natural class. Russian [x, ʃʃ] lack robust voiced counterparts just like the affricates do—but the entire set of gaps cannot be isolated with one phonological feature, and it does not make much sense phonetically, as the gaps are non-contiguous in the articulatory tract (recall Table 2).

Thus, in addition to requiring a specific Disalign constraint for affricates (picked out by their [delayed release] feature), this analysis would have to recruit two additional, even more specific constraints: Disalign[+voice, dor, +cont] to govern the distribution of [ɣ], and Disalign[+voice, $-$ ant, +distributed, +cont] for [ʒʒ]. This is a general problem for a surface-constraint account: no simple constraint will do.

5.3. Comparative markedness

Another alternative in the category of adding constraints was suggested to me by Andrew Lamont: McCarthy’s (Reference McCarthy2002b) Comparative Markedness. In this theory, every markedness constraint is split into two versions: old and new. In Russian, *[ʤ]-old would penalise faithful voiced affricates (e.g., /noʤ-am/ $\rightarrow $ [noʤam]). The new version, *[ʤ]-new, penalises a voiced affricate not present in the fully faithful candidate – such as one derived by voicing assimilation, /noʧ-bɨ/ $\rightarrow $ [noʤbɨ]. The Russian pattern could be analysed in terms of the ranking *[ʤ]-old, Agree[voice] $\gg $ *[ʤ]-new: old/underlying voiced affricates are not allowed, but new/derived ones are.

The question is what happens to these underlying affricates, and how the learner would ever figure this out. The ranking established so far suggests that Ident-[voice] is the crucially dominated faithfulness constraint: underlying /ʤ/ devoices. But, as McCarthy himself points out, rankings of the shape Mark-old $\gg $ Faith $\gg $ Mark-new are not learnable through basic recursive constraint demotion in phonotactic learning (see McCarthy Reference McCarthy2002b: §6.3). The only examples where old markedness transitively outranks new markedness in McCarthy’s catalogue involve counterfeeding opacity – evidence for which must come from morphophonemic alternations, not phonotactics. The nature of the Russian problem is simpler: there are several inventory gaps, but the pattern is transparent, treating all underlyingly voiceless consonants alike.

McCarthy himself anticipates the criticism that Comparative Markedness introduces duplication by splitting every markedness constraint in two. McCarthy counters that, unlike MSCs, old and new markedness constraints can compel and block alternations. But, as we saw in the discussion of loanword adaptation, in Russian, *[ʤ]-old – if it were to exist – does not compel the right kind of alternation, since borrowed affricates undergo fission rather than devoicing. Handling this would require multiple faithfulness constraints in addition to splitting markedness constraints into two, way beyond duplication.

5.4. Positional Markedness

Positional neutralisation can often be analysed either in positional faithfulness or markedness terms. I took it for granted that the positional faithfulness account is right, so here, I show that even the positional markedness alternative does not escape the duplication problem.Footnote ²¹ The obvious alternative to *ObsVoice $\gg $ Ident[voice] is to ban word-final voicing instead:

Voiced affricates could then be banned by a general constraint *[ʤ], which is outranked by Agree, as in (44):

As with nasal place assimilation in (8), we face problems in analysing the direction of assimilation. To get that right, Ident-Preson could be ranked anywhere, as it just breaks the tie between two assimilated cluster candidates (45c,d). But to get the correct results for clusters that include an affricate, Ident-Preson must dominate *[ʤ], and this leads to a ranking contradiction. If the presumed fate of /ʤop/ in (44) is to devoice, then the ranking cannot be as in (45). We would need a constraint against [ʤ] in presonorant position, and with it returns the duplication problem.

A reviewer challenges the assumption that devoicing rather than fission is the right outcome in this analysis (recall that any conclusions about these hypothetical inputs with voiced affricates are guesses, since we never see evidence for them from alternations).Footnote ²² If we suppose that underlying /ʤ/ maps to [dʐ] in the grammar in (45), it would have to be because fission is less costly than devoicing – that is, Integrity is the bottom-ranked constraint. But if that were true, then /ʧ/ would split into [dʐ] in assimilation contexts, too (recall (27) and (37)).

To summarise, the analytic role of positional faithfulness is not only to protect contrasts in certain positions but to determine the direction of assimilation. This aspect of the constraint family makes it necessary for complete analyses of assimilation patterns. Whenever these assimilation patterns involve gapped inventories – which is not rare – positional faithfulness needs to be augmented with a constraint, or several constraints, to block the gappy segments from occurring in positions protected by positional faithfulness. I argue that this situation is unsatisfactory. A simpler analysis is to ban the non-contrastive segments from URs across the board, and to understand their absence in various environments as a consequence of their derived status. If theoretical economy is the goal, then MSCs achieve it better than surface-oriented constraints.

5.5. Stratal OT

The last alternatives I consider are cast in Stratal OT (Kiparsky Reference Kiparsky2000; Rubach Reference Rubach2000; Bermúdez-Otero Reference Bermúdez-Otero, Hannahs and Bosch2018), a constraint-based descendant of Lexical Phonology and Morphology (Kiparsky Reference Kiparsky1982; Mohanan Reference Mohanan1982; Kaisse & Shaw Reference Kaisse and Shaw1985). A Stratal OT grammar usually assumes a serial interaction between morphology and phonology. A stem is phonologised in the stem level/stratum. It is then concatenated with affixes, and the result passes through another, possibly different phonological grammar. Clitics and phrasal phonology apply in the postlexical stratum. Stratal OT is argued to be a theory of inventory restrictions: the stem level enforces them.

I discuss three stratal analyses. In the first one, affricates fission into stop–fricative sequences, before either assimilation or devoicing apply. The next two options are suggested by Mackenzie (Reference Mackenzie2024): in the first, affricates become voiced fricatives, and in the second, they devoice. I will argue that fission and spirantisation fail to capture the facts I laid out earlier. The predictions of the third analysis are examined in the last subsection, where I evaluate Stratal OT as a general theory of morpheme shape. Depending on the specific assumptions, Stratal OT either makes the wrong predictions for Russian, or has nothing to say about well-documented root–affix asymmetries, which positional faithfulness to roots explains well. Worse still, various published Stratal OT accounts either explicitly or covertly assume MSCs, suggesting the stem level does not suffice as a theory of morpheme shape.