In defence of underlying representations: Latin rhotacism, French liaison, Romanian palatalization

4.1 The problem of split-base formations in French adjectival liaison

Steriade (1999: §4-§5) adduces French adjectival liaison as evidence against Cyclic Containment (1). The goal of her own account is to generate the masculine singular form of a French adjective in liaison, i.e. when immediately followed by a head noun beginning with a vowel: e.g. mon petit ami [mɔ̃.pə.ti.ta.mi] ‘my.m.sg little.m.sg friend.m.sg’. To this end she posits a set of optimality-theoretic Lex constraints which work by consulting the surface realizations of the adjective in nonliaison masculine singular environments and in the feminine singular; these surface realizations are assumed to be stored in memory owing to their relatively high token frequency. Examples of the paradigms under discussion are provided in (21). ^[7]

In Steriade’s account, a masculine liaison candidate exhibiting a surface property p satisfies the Lex constraint on p if p is present in either (or both) of the masculine and feminine citation forms. Through interaction with other constraints, this enables the rise of split-base formations that blend phonological properties of both citation forms. As instances of split-base formation, Steriade highlights nonnormative liaison realizations like [pʁəmjeʁ], [sot], and [sɛʁtɛ̃n]: in her analysis, for example, [pʁəmjeʁ] copies its final [ʁ] from the feminine output form [pʁəmjɛʁ], whilst the preceding [e] originates in [pʁəmje], which is the surface realization of the masculine in nonliaison contexts. This explanation is incompatible with Cyclic Containment insofar as the feminine is not a morphosyntactic constituent of the masculine liaison form. The challenge to cyclic frameworks is all the more pressing because, in French dialects where adjectival liaison remains active, ^[8] the productive pattern for adjectives like premier, sot, and certain is nonnormative split-base liaison: see Tranel (1981: 108–9).

It is eminently plausible that both the masculine and the feminine citation forms of a French adjective influence its realization in liaison. It by no means follows, however, that this influence is best described by means of constraints evaluating direct relationships between surface forms. Rather, in line with proposition (3), we may alternatively hypothesize that surface tokens in nonliaison environments exert their influence on the liaison form through the process whereby learners construct underlying representations. As required by the approach to syntagms and paradigms set out at the start of Section 1, once underliers have been established, derivations make no reference to paradigmatic information and Cyclic Containment is fully respected.

An account of French adjectival liaison centred on underlying representation discovery makes the following crucial predictions. First, the surface realization of an adjective in liaison can combine properties of the masculine and feminine citation forms only if the learner has set up a single underlying representation incorporating features of both:

In turn, this is possible only if the learner’s grammar includes phonological processes that correctly derive the masculine and feminine alternants from a shared single underlier. Split-base formations are thus predicted to arise only through the mediation of productive phonological generalizations. In contrast, blending is predicted to be impossible in cases of suppletion, in which there is no shared single underlier:

This resulting prediction is stated in (24).

Cyclic Containment, therefore, compels us to analyse the nonnormative split-base liaison forms in (21,f) as being phonologically derived from single underliers shared by the masculine and feminine citation forms. In turn, this entails that productive rules in the synchronic phonology of present-day French must control the following alternations: (i) the consonant~zero alternations in premier, sot, and certain; (ii) the vowel tenseness alternations in premier and sot; (iii) the vowel nasality alternation in certain. I now proceed to show that all of these predictions are correct.

4.2 The grammar of latent consonants

For the consonant~zero alternations in premier, sot, and certain, I adopt the classical autosegmental analysis of latent consonants (Tranel 1995) as implemented by Wetzels (1987, 2002). In this analysis, consonants that do not alternate are underlyingly fixed; consonants that alternate nonsuppletively are underlyingly latent. A fixed consonant is underlyingly anchored to a skeletal unit; a latent consonant floats in the underlying representation, and surfaces only if it is able to link up to a skeletal unit during the derivation. ^[9] Thus, the surface forms of adjectives like joli, lisse, and petit (21a,b) derive from underliers such as the following ^[10]:

Following Wetzels (1987: 296), I further assume that, when the lexicon does not store a suppletive feminine form, agreement with feminine targets is marked by means of an inflectional suffix consisting of a bare skeletal slot. As a result, no final consonant in a feminine form ever remains floating in the output of the word level. This accounts for the fact that there are no liaison alternations in feminine forms.

The autosegmental approach receives powerful support from the observation that it is possible to insert a pause or intonational phrase boundary before a latent consonant surfacing through liaison (Morin 1986: 198, Tranel 1990). In contrast, an intonational phrase boundary cannot ever precede a word-final linked consonant, whether the latter is underlyingly fixed (25,b) or it is underlyingly latent but becomes linked by the word level through feminine inflection (26).

This reflects the different prosodic status of floating and linked consonants in the input to the phrase level: the generalization is that intonational phrase boundaries may not fall inside strings prosodified together in the input to the phrase level (Wetzels 2002: 301).

In present-day French, latent consonants are drawn from a relatively narrow set, the most frequent being /z, n, t/ (Durand and Lyche 2008: 58). For this reason, French speakers do not interpret all consonant~zero alternations in adjectives as reflecting the presence of an underlying latent segment. The alternation between [ʃ] and [∅] in the citation forms of franc (21,d), for example, is analysed as suppletive: the corresponding lexical entry is shown in (29). The diacritic S represents a lexically stored instruction to perform syntactically driven allomorph selection during exponence: /fʁɑ̃ʃ/ is inserted only if the exponendum is feminine, and /fʁɑ̃/ behaves as the elsewhere exponent. Thus, because /fʁɑ̃/ has no latent consonant, it does not alternate in liaison.

As noted by Lamarche (1996: 149), the behaviour of franc speaks strongly against nonmodular theories like those of Tranel (1996) and Perlmutter (1998), in which a phonological constraint driving liaison (say, Onset) interacts in parallel with a morphosyntactic constraint requiring gender agreement (say, Concord): the ranking Onset ≫ Concord incorrectly predicts liaison realizations like franc entretien *[fʁɑ̃.ʃɑ̃.tʁə.tjɛ̃] ‘frank discussion.m’, instead of [fʁɑ̃.ɑ̃.tʁə.tjɛ̃].

Similarly, the alternation between [nuvo] and [nuvɛl] (21,e) is undeniably suppletive, and so cannot reflect the presence of a latent consonant in a shared underlier. In this case, the fact that the allomorph /nuvɛl/ appears both in feminine and in masculine liaison contexts is derived simply by preventing /nuvo/ from realizing [+fem] exponenda and otherwise deferring allomorph selection until the phonology (see Bermúdez-Otero 2012: 52 and references therein). This is indicated by the diacritic P in lexical entry (30), which instructs morphology to insert both allomorphs (as long as syntactic subcategorization requirements are met) and to defer allomorph selection to the phonology.

In autosegmental terms, moreover, the allomorph /nuvɛl/ must end in a fixed consonant, as shown in (30), since this /l/ is able to surface before feminine nouns that do not begin with a vowel. This inference automatically explains Morin’s (1986: 198) and Tranel’s (1990) observation that the liaison form of nouveau, unlike that of petit, rejects the insertion of an intonational phrase boundary before its final consonant: cf. (27) and (31).

Moreover, French does not have a productive phonological process devoicing word-final consonants. In consequence, the alternation between the masculine liaison form [ɡʁɑ̃t] and the feminine form [ɡʁɑ̃d] in (21,c) must be suppletive too. Yet, since the masculine alternates between [ɡʁɑ̃t] in liaison and [ɡʁɑ̃] elsewhere, its final /t/ must be underlyingly latent: i.e. /ɡʁɑ̃(t)/.

Once again, the autosegmental approach leads to exact predictions about intonational phrase boundaries: the masculine liaison form of grand (33) behaves like that of petit (27) and unlike that of nouveau (31).

In addition, the suppletive analysis of the voicing alternations in grand and gros (21,c) correctly entails that these alternations are not productively extended to other adjectives ^[11]:

4.3 The loi de position and the phonotactics of vowel nasalization

Having dealt with consonant~zero alternations in adjectival liaison, we are now ready to account for the productivity of the nonnormative split-base patterns in (21, f). To be explained is the fact that the liaison form [pʁəmjeʁ] combines features of both citation forms: the vowel [e] of masculine [pʁəmje] and the final consonant of feminine [pʁəmjɛʁ]. Why do speakers generate such novel alternants in liaison, as reported by Tranel (1981: 108–9), whilst they reject *[ʃot] and *[epɛz] (34)? As shown in (22) and (23), an account driven by underlying representation discovery predicts that the mixture of properties in [pʁəmjeʁ] can arise only because the learner sets up a single shared underlier for the masculine and feminine citation forms, and this, in turn, is possible only because there are productive phonological generalizations mapping this single underlier to the surface alternants known to the learner.

The productive phonological generalization at work in the paradigm of premier is the well-known loi de position: simplifying somewhat, this states that the unmarked or default realization of mid vowels is tense (mid-high) in open syllables and lax (mid-low) in closed syllables (Valdman 1972; Tranel 1988). In (35) I show that the loi de position applies at the word level. I use data from Midi French, where the loi de position sustains no exceptions (Durand and Eychenne 2014: §3.1); the relevant facts are identical in other dialects. Crucially, linked consonants syllabified in coda position at the word level trigger closed syllable laxing, and this component of the loi de position is counterbled by phrase-level resyllabification when the following word begins with a vowel. In contrast, latent consonants do not trigger closed syllable laxing because they remain floating in the output of the word level and so do not behave like codas. ^[12]

Now consider a learner who has been exposed to the citation forms [pʁəmje] and [pʁəmjɛʁ]. Faced with the [e]~[ɛ] and [∅]~[ʁ] alternations, she may take them to be morphologically controlled, and so she may subsume the adjective under the pattern of nouveau (21,e). This will result in a disjunctive underlier with /pʁəmjɛʁ/ as the allomorph unmarked for gender: i.e. {/pʁəmje/^[-fem], /pʁəmjɛʁ/}^P. Such a suppletive representation will yield single-base [pʁəmjɛʁ] in liaison: cf. (30). Alternatively, the learner may interpret the [∅]~[ʁ] alternation as reflecting an underlying latent consonant that becomes linked in the feminine by suffixation of a skeletal slot (26) and, concomitantly, she may interpret the [e]~[ɛ] alternation as derived by the loi de position from an underlying /E/ unspecified for tenseness, as in (35). In that case, there will be a single shared underlier after the pattern of petit (25,c). The shared underlier analysis is shown in (36,a). By regular morphology and phonology (36,b), the split-base form [pʁəmjeʁ] will emerge in liaison. ^[13]

Without further stipulation, this analysis also derives the judgements about intonational breaks provided by Plénat (2008: §3.1), who reports that normative single-base liaison with [pʁəmjɛʁ] requires a break to follow the [ʁ], whereas nonnormative split-base liaison with [pʁəmjeʁ] requires a break to precede the [ʁ]:

This is because single-base liaison involves a suppletive allomorph ending in a fixed consonant, like nouvel éléphant (31), whereas split-base liaison involves a floating consonant in the input to the phrase level, as in petit homme (27) and grand ami (33).

The case of certain is closely analogous. Single-base liaison in [sɛʁtɛn] arises when the learner posits a disjunctive suppletive underlier: {/sɛʁtɛ̃/^[-fem], /sɛʁtɛn/}^P. In contrast, split-base liaison in [sɛʁtɛ̃n] occurs if the learner subsumes the masculine and feminine citation forms under the single underlier /sɛʁtɛ̃(n)/ and derives the surface alternation by means of the phonotactic constraint systematically banning [+nasal] vowels in rhymes closed by nasal consonants:

4.4 Underlying representation discovery gives rise to variation

An analysis of split-base liaison centred on the discovery of underlying representations enjoys another salient advantage. We have seen that setting up underliers for French adjectives necessarily involves deneutralization (Nevins and Vaux 2008). The underlying opposition between latent and fixed consonants, as in /pəti(t)/ vs /lis/, is neutralized morphologically in the feminine by the suffixation process shown in (26). The underlying opposition between latent consonants and /∅/, as in /pəti(t)/ vs /ʒɔli/, is neutralized phonologically in masculine citation forms through the failure of floating consonants to dock. When setting up underliers, the learner seeks to undo these neutralizations; but, since her linguistic experience provides her with widely different amounts of information for different adjectives, her choice of underliers will show lexical variation, and so will the liaison patterns generated by the grammar.

This phenomenon has been recognized by many scholars: e.g. Bonami and Boyé (2003: §2.2), Féry (2004: 26), Plénat (2008: §2.1). Crucially, our analysis predicts that lexical irregularity will extend to adjectives capable of displaying split-base liaison like premier and certain, even though the split-base pattern is currently dominant. Strikingly, this prediction proves accurate: Féry’s (2004: 34–35) informants, for example, displayed strongly divergent preferences for dernier and léger, favouring normative single-base liaison for the former and nonnormative split-base liaison for the latter:

Féry suspects that her findings reflect a bias towards vowel harmony. This may be so: the data in (39) may thus reflect competition between the loi de position and vowel harmony in northern dialects (Durand and Lyche 2004; Nguyen and Fagyal 2008: 23). However, there is another consideration that carries greater weight: dernier has higher token frequency in prenominal environments than léger, and this enhances the survival chances of the more conservative liaison realization (Wetzels p.c.).

In sum, an account of French adjectival liaison centred on the discovery of underlying representations explains why split-base liaison is possible only when the masculine and feminine citation forms are related by synchronically productive phonological processes: see (36) and (38). The analysis also predicts the fact that the incidence of split-base liaison is lexically specific (39). Through its connection with the autosegmental approach to latent consonants (Wetzels 2002), it further yields a strikingly accurate and entirely nonstipulative characterization of the behaviour of single-base and split-base liaison forms before intonational phrase boundaries (37). But, most importantly for our current purposes, this analysis impeccably complies with Cyclic Containment (1).

4.5 Problems for surface-to-surface computation: the suppletive blending pathology

We may now compare our account with one based on Lex constraints. To explain the variation shown in (21,f), Steriade (1999: 261) posits two antagonistic Lex constraints: LexP-∀ requires the liaison form to be fully and completely identical to a single citation form, whereas Lex(V́,gender) just requires the liaison form to copy the accented vowel of the citation form of the same gender (the masculine). The ranking LexP-∀ ≫ Lex(V́,gender) generates consistently normative single-base liaison; the opposite ranking Lex(V́,gender) ≫ LexP-∀ predicts liaison to be uniformly nonnormative and split-base. In the case of premier, for example, Lex(V́,gender) requires the masculine liaison form to copy the accented vowel of the masculine citation form [pʁəmje], whilst leaving it free to avoid hiatus by also copying the final consonant of feminine [pʁəmjɛʁ]. ^[14]

It becomes immediately apparent, however, that the ranking Lex(V́,gender) ≫ LexP-∀ predicts that French dialects with [pʁəmjeʁ] will choose *[nuvol] as the liaison form for nouveau. Yet *[nuvol] is completely unattested and entirely unacceptable. To avoid the problem, Steriade posits another Lex constraint, LexP-∀_loose, which requires the liaison form to be fully and completely identical to a single citation form except in respect of nonsalient properties. Steriade (1999: 264) asserts that the [ɛ]~[e] alternation in [pʁəmjɛʁ]~[pʁəmjeʁ] is nonsalient because mid-vowel tenseness is largely phonologically predictable by the loi de position, whereas the [ɛ]~[o] alternation in [nuvɛl]~*[nuvol] would be too salient because it would not be bridged by a productive phonological generalization. Thus, with LexP-∀_loose ranked above Lex(V́,gender), *[nuvol] loses to [nuvɛl].

This solution is open to two objections. First, it is unclear how LexP-∀_loose gains access to the necessary information about salient differences: crucially, reference to distance in acoustic space will not be enough, since the relevant notion of salience derives from systemic patterns of contrast. The underlier discovery approach, in contrast, deduces the same result from first principles, without repackaging phonological predictability as lack of salience: the learner can conflate two surface forms under a single underlying representation only when their alternations are predictable by phonological rule (22).

Secondly, the crucial ranking LexP-∀_loose ≫ Lex(V́,gender) holds only by stipulation; nothing prevents a minimally different language, French′, from having the opposite ranking. If so, the grammar of French′ will productively mix features of suppletive masculine and feminine adjectival forms to create the impossible liaison realizations in (40).

This is a very serious problem: the theory of Lex constraints turns out to predict the existence of grammars that generate new allomorphs automatically and deterministically by blending submorphic phonological properties from different cells in suppletive paradigms, contrary to (24). Yet suppletive blending is at best a sporadic phenomenon; it is not known ever to apply regularly. More generally, the blending of submorphic properties is at best a highly peripheral mechanism of morphophonological innovation. The true engines of morphophonological change are morphological and phonological reanalysis: morphological reanalysis involves the recutting (Diertani 2011) and redistribution of morphs; phonological reanalysis involves the restructuring of phonological input representations, leading to well-understood phenomena such as domain narrowing, rule scattering, morphologization, and lexicalization (Bermúdez-Otero and Trousdale 2012: §2; Bermúdez-Otero 2015: §2.3; Ramsammy 2015). As shown in (23), frameworks incorporating underlying representations avoid the suppletive blending pathology: the submorphic properties of suppletive exponents do not mix because, by definition, suppletive exponents are not subsumed under a single underlier.

In sum, both the masculine and feminine citation forms of a French adjective play a crucial role in determining the adjective’s behaviour in prenominal liaison. This role is best understood as taking effect during the discovery of underlying representations. This explains why the generation of split-base liaison forms always involves productive phonological processes and why split-base liaison shows lexical irregularity. In combination with an autosegmental approach to latent consonants, moreover, this analysis predicts the behaviour of liaison forms before intonational phrase boundaries with striking elegance and accuracy. In contrast, theoretical frameworks in which the computation of the liaison form refers directly to the surface realizations of the masculine and feminine incur the suppletive blending pathology. These results indicate that the facts of French adjectival liaison are perfectly consistent with Cyclic Containment.

5.1 The problem of inflection dependence in Romanian palatalization

Steriade (2008) launched a further challenge to Cyclic Containment (1) with evidence from palatalization in Romanian; here I focus on her arguments from the behaviour of the velar stops /k, ɡ/. According to Chitoran (2001: 184ff), Romanian velar plosives are subject to two separate palatalization processes: one, which Chitoran calls ‘phonological’, applies across the board before front vocoids and yields stops with a secondary articulation, i.e. [kʲ, ɡʲ]; the other, which she calls ‘morphological’, applies before /i, e/ in certain morphological environments and generates palatoalveolar affricates, i.e. [ʧ, ʤ]. ^[15] Steriade’s argument concerns the latter process. Morphological palatalization is perfectly systematic before the inflectional suffixes /-i/ and /-e/: the former, usually surfacing as desyllabified [i̯], marks second person singular in verbs, and plural number in a large subset of nouns and adjectives, including all masculine forms; the latter marks third person singular in certain verbs, and plural number in some feminine and neuter nouns and adjectives. ^[16]

Steriade observes that a stem-final /k/ or /ɡ/ becomes [ʧ] or [ʤ] before derivational suffixes beginning with /i/ only if the consonant is subject to regular morphological palatalization before the marker /-i/ or /-e/ in the stem’s inflectional paradigm. Thus, the derivational suffix /-ist/ causes palatalization only in base nouns with plurals in /-i/ or /-e/; base nouns with plurals of other types resist palatalization: cf. (42,a) and (42,b). In denominal verb formation, in turn, nouns with plurals in /-i/ or /-e/ preferentially combine their palatalized stem allomorphs with the default verbalizer /-í/; other nouns, in contrast, select less frequent verbalizers, like /-á/ or /-uí/, whose initial vowels are non-palatalizing: cf. (42,c) and (42,d).

In Romanian, therefore, the behaviour of a stem-final velar stop in derivation is intimately correlated with its realization in inflection. Steriade refers to this phenomenon as ‘inflection dependence’.

Crucially, Steriade regards palatalization before the plural markers /-i/ and /-e/ as a derived phonological property. Her key argument is that palatalization in this environment is perfectly systematic and productive, applying both to loanwords (43,a) and to nonce items (43,b).

Nonetheless, /k/ and /ɡ/ may resist palatalization when followed by a front vowel across a derivational boundary (42,b) or within the same morph (44). ^[17]

Accordingly, Steriade (2008: 319, 326–7) proposes that palatalization is driven by a comparative markedness constraint (McCarthy 2003) which bans sequences of a velar plosive plus a front vocoid if they are derived by morphological concatenation, but not if they are tautomorphemic. The constraint is morphologically indexed (Pater 2009) so as to penalize only sequences created by the relevant inflectional suffixes.

If palatalization were indeed a purely derived phonological property, then the pattern of inflection dependence illustrated in (42) would refute Cyclic Containment: when computing the surface representation of the derivative [stɨnʤíst], the phonology would need to consult the output realization of the stem in the plural form [stɨ́nʤi̯], even though [stɨ́nʤi̯] is not a constituent of [stɨnʤíst]. In Steriade’s terms, the derivation would be ‘pseudocyclic’.

However, the analysis of palatalization in Romanian plurals as a purely phonological process incurs serious theoretical costs. The failure of /k, ɡ/ to become [ʧ, ʤ] before tautomorphemic /i, e/, as in (44), must be accounted as an instance of non-derived environment blocking (NDEB); but Inkelas (2011) makes a strong case for regarding morphological NDEB as a spurious category (see also Chong 2017). Inkelas finds no instances of a phonological process that applies whenever its environment is met by virtue of morphological concatenation, and not otherwise. Rather, all putative examples of morphological NDEB appear to involve restrictions to specific constructions: once these have been enumerated, there is no work left for a derived environment condition to do. Indeed, Steriade’s analysis of palatalization in plural nouns supplements comparative markedness (which captures the putative derived environment effect) with constraint indexation. The latter, however, has been criticized as compromising the empirical content, heuristic power, and learnability of linguistic theory (Bermúdez-Otero 2012); in the Four-Hypothesis Programme, it is explicitly banned by Indirect Reference (see Section 2). Replacing indexed constraints with less powerful devices can improve empirical coverage, as Zimmermann (2013) showed in her reanalysis of a parade example of morph-specific phonological behaviour: vowel deletion in Yine (cf. Pater 2009). There are thus good reasons for pursuing an analysis of Romanian plural palatalization in which morphology plays a larger and more direct role, in line with Chitoran’s (2001: 184ff) label for this phenomenon (see also Maiden 2013 for germane discussion of palatalization in Romanian verbs).

5.2 Morphological control of palatalization

Clues to such an analysis may be found in the behaviour of nouns with plurals in [-ur-i̯], like [fók]~[fók-ur-i̯] in (42,b). The key observation is that, if a noun takes /-ur-i/ in the plural, then it may also take /-ur-/ before derivational suffixes (Steriade 2008: 343):

Since plurality does not enter into the computation of the meaning of the items in (45,b), the morph [-ur-] cannot be an exponent of number; rather, it must be analysed as an extension of the stem, i.e. as a thematic element (th). This analysis is corroborated by diachronic evidence (Martin Maiden p.c.). The stem extension -ur- arose through the morphological recutting of the stems of Latin noun forms like corpor-a ‘body-n.nom/acc.pl’ (cf. sgcorpus-∅). Latin or-a became -ur-e in old Romanian. Subsequently, Romanian nouns with plurals in -ur-e were caught up in a more general process of morphological change whereby the plural marker -i enlarged its distribution at the expense of -e: e.g. roat-ăf.sg ~ roat-ef.obl.sg/f.pl>roat-ă ~ roţ-i ‘wheel’. In nouns with plurals in -ur-e, however, innovative -i replaced just -e, rather than the whole string -ur-e: e.g. vânt-ur-e>vânt-ur-i, *vânţ-i ‘wind.pl’. This confirms that speakers did not parse -ur- as part of the plural marker, but as part of the stem. Thus, Romanian nouns with plurals in [-ur-i̯] turn out to have the morphosyntactic structure shown in (46).

This structure is well-attested in other present-day Romance varieties: see e.g. Bermúdez-Otero (2013: 4–6) for a parallel in Spanish.

The morphological parses in (46) entail that, in nouns of this class, the thematic element has two allomorphs: /-∅-/ and /-ur-/. Interestingly, these two allomorphs are subject to phonologically driven selection in the presence of certain derivational suffixes (Steriade 2008: 343–4). Inside the diminutive suffix /-él/, for example, /-ur-/ is chosen if its presence is required to avoid a pretonic sequence consisting of a single syllable (47,a); otherwise, the null allomorph is selected (47,b). ^[18]

From this viewpoint, the appearance of [-ur-] inside [vɨ̀nturél] does not involve pseudocyclicity; it is a simple instance of phonologically driven theme selection. The underlying representation of the noun stem contains all the information required by the phonology, and the derivation never refers to the plural form [vɨ́nturi̯]. The latter does of course play a crucial role in acquisition by causing the learner to set up the underlier /vɨnt - {∅,ur}/. After that, however, it plays no part in the derivation.

Crucially, this approach can easily be extended to the pattern of morphological palatalization illustrated in (42). We need only assume that nouns with plurals in /-i/ and /-e/ have exactly the same morphosyntactic structure as those with plurals in /-ur-i/ (46), but select different exponents for their theme position: again, one exponent is null, the other overt. In the spirit of Generalized Nonlinear Affixation (Bermúdez-Otero 2012: 53), we are to envisage the overt exponent of th as consisting of a floating piece of melody like (48).

Combining the noun’s root with the null exponent of th results in a stem allomorph ending in a plain consonant (49,a). In contrast, concatenating the root with the overt exponent of th generates a stem allomorph subject to morphological palatalization, as the features of the thematic element seek an anchor in the root’s final consonant (49,b). ^[19]

Contrasts between derivatives like [stɨnʤ-íst] ‘leftist’ (2,a) and [stɨŋɡ-áʧ] ‘left-handed, clumsy’ (Steriade 2008: 322) can now be seen to arise through phonologically driven theme selection: both forms are based on the underlying stem /stɨŋɡ - {∅,^{[Cor,+strid,-ant]}}/ ‘left (hand)’. The derivation is analogous to that in (47).

Some of Steriade’s key observations may now be interpreted as follows. In Romanian, the exponents of plural number subcategorize morphologically for stems with overt themes ^[20]:

In consequence, all pluralizable nouns must select a thematic element possessing an overt exponent: members of the declension classes with plurals in [-i̯] and [-e] choose (51,a); most others choose (51,b) (see note 16).

Otherwise, the default theme is /-∅-/. This explains why singularia tantum, which satisfy (50) vacuously, never exhibit morphological palatalization (Steriade 2008: 345–55), as they need not abandon the default null theme for (51,a). Similarly, the perfect systematicity of morphological palatalization in plural inflection, even in loanwords and nonce items (43), reflects the operation of productive processes in the morphosyntax, not in the phonology. For example, a masculine singular noun like [pféniɡ] can only belong to the declension that takes the suffix /-u/ in the singular and /-i/ in the plural (see notes 16 and 17), but all members of this declension systematically select theme (51,a). It, therefore, follows that the plural of [pféniɡ] must be [pféniʤi̯].

5.3 Phonotactically controlled phonological properties do not exhibit inflection dependence

The fact that an approach based on theme selection looks prima facie feasible for Romanian palatalization reveals a vulnerable spot in Steriade’s (2008) argument against Cyclic Containment. As we saw above, the argument draws its force from the premise that palatalization before the plural markers /-i/ and /-e/ is a derived phonological property; but the restriction of palatalization to a narrow set of morphological environments casts doubt on this claim.

The better to appreciate this point, consider an alternative scenario. Let Romanian′ be a language exactly like actual Romanian, with the exception that the palatalization of /k, ɡ/ to [ʧ, ʤ] before front vocoids applies across the board without exception within the inflectional paradigms of nouns with simple stems (i.e. stems that do not contain derivational suffixes). Thus, Romanian′ exhibits all the alternations in (41), but it has [ʧiloɡrám] instead of [kʲiloɡrám], [ʧéstije] instead of [kʲéstije], and [óʧi̯] instead of [ókʲi̯]; cf. (44). At the same time, Romanian′ retains the correlation between plural inflection and derivation observed in (42): i.e. palatalization still goes ahead in [stɨnʤ-íst] (cf. plural [stɨ́nʤ-i̯]), and is still blocked in [fok-íst] (cf. plural [fók-uri̯]).

If such a language existed, it would provide a devastating argument against Cyclic Containment. In modular cyclic frameworks like Stratal Phonology (Bermúdez-Otero 2018), a process of palatalization that applies across the board in all the inflectional wordforms of all simple stems must take the grammatical word as its cyclic domain, and so be word-level; but a word-level phonological process is predicted to apply across word-internal derivational boundaries too. In Romanian′, the theme selection account of (42) would be manifestly ad hoc, in that it would arbitrarily fail to subsume the palatalization of [stɨ́nʤi̯], [óʧi̯], and [ʧiloɡrám] under a single phonological generalization. But, significantly, Romanian′ does not exist.

Thus, the containment-compliant analysis of Romanian morphological palatalization as phonologically driven theme selection brings to light a deep link between the morphological control of palatalization in simple stems, manifested in the contrast between (41) and (44), and its apparent inflection-dependent behaviour in complex stems (42). The general prediction is stated in (52).

Strikingly, this prediction is supported by evidence from Romanian itself: as expected, Romanian phonological processes found to be under pure phonotactic control in the paradigms of simple stems fail to exhibit inflection dependence in derived stems. Notably, Romanian has a rich set of alternations affecting mid vowels (Chitoran 2001: ch. 7). Of interest here is a process whereby [o] diphthongizes to [o̯a] under stress when followed by a non-high vowel in the next syllable (Steriade 2008: 341) ^[21]:

Steriade observes that the [o]~[o̯a] alternation is not inflection-dependent. Denominal verbs consistently display [o] outside the diphthongization environment. Crucially, this is true not only when the inflectional paradigm of the base noun contains a monophthongal stem allomorph (54,a), but also when the base noun shows the diphthong [o̯a] in all its inflected forms (54,b).

Stratal Phonology explains this state of affairs. The distribution of [o̯a] is subject to two systematic restrictions in the native Romanian vocabulary: [o̯a] cannot occur in a stressless syllable, and it cannot occupy the final syllable in the prosodic word (Chitoran 2001: 213, Pană Dindelegan 2013: 12). ^[22] These restrictions hold across the board in the wordforms of simple native stems, as illustrated by the following verbal paradigm (Chitoran 2001: 205):

The ban on unstressed [o̯a] thus proves to be a true word-level phonotactic constraint on the native lexicon, quite unlike morphological palatalization. But native derived stems—or rather their inflected forms—go through the word-level phonology just like simple stems. It, therefore, follows that derivatives like [ɨnɡlotí] will show monophthongal [o] in unstressed syllables, even if the base noun displays [o̯a] throughout its paradigm. In contrast, a theory that describes inflection dependence pseudocyclically by means of Lex constraints overgenerates: it can describe a language, Romanian″, where [ɨnɡlotí] loses to *[ɨnɡlo̯atí] or *[ɨnɡlo̯atá] owing to a top-ranked constraint forbidding the occurrence of [o] in a derivative unless [o] surfaces in some inflectional form of the base. Like Romanian′, however, Romanian″ does not exist, and for the same reason.

I conclude that the facts of Romanian not only comply with, but in fact support, Cyclic Containment.