6.1 Epenthesis and assimilation in Lithuanian

In Baković (Reference Baković2005) I argued for a rather unusual kind of process interaction in a set of prefix alternations in Lithuanian, resulting in a very different sort of overapplication opacity. The relevant facts are as presented in (21) and (22). (All Lithuanian examples cited here are drawn from Dambriūnas et al. (Reference Dambriūnas, Klimas and Schmalstieg1966), Kenstowicz & Kisseberth (Reference Kenstowicz, Kisseberth and Kisseberth1971), Kenstowicz (Reference Kenstowicz1972), Mathiassen (Reference Mathiassen1996) and Ambrazas (Reference Ambrazas1997) ; very special thanks are due to Solveiga Armoskaite for helpful elucidation and amplification of the data set.)

One process is assimilation: the consonant of each of the distinct verbal prefixes /at/ and /ap/ assimilates to an adjacent stem-initial consonant in terms of voicing and palatalisation.Footnote ¹³

1. (21) Lithuanian voicing and palatalisation assimilation

   

The other process is epenthesis: if the initial consonant of a stem is identical to the prefix-final consonant or differs from it only in terms of voicing or palatalisation (or both), /at/ and /ap/ surface as [atʲi] and [apʲi] respectively (with automatic palatalisation before the epenthetic [i]; see Footnote note 13).

1. (22) Lithuanian vowel epenthesis

   

It is of course possible to describe the relevant Lithuanian facts with a conventional SPE derivation. We can take as our starting point the analysis discussed in textbook form by Odden (Reference Odden2005: 113–115). Odden offers the two rules in (23). (Odden's transcriptions and discussion of assimilation do not include palatalisation, and I follow suit in my discussion of Odden's analysis except where noted otherwise.)

1. (23) Rules for Lithuanian epenthesis and assimilation (Odden 2005)

   1. a. Epenthesis

      

   2. b. Assimilation

      obstruent→voiced_voiced obstruent

Note that epenthesis is stated to apply between homorganic obstruent stops – effectively, between consonants that differ at most in voicing (and, in the full analysis, palatalisation as well). By ordering epenthesis before assimilation, as Odden does, we have a bleeding-on-environment interaction between the two rules. A relevant example is derived in (24). Epenthesis applies first between the homorganic obstruent stops. Since these stops differ in voicing, this earlier application of epenthesis crucially bleeds the later assimilation rule by destroying the context necessary for the application of assimilation.

1. (24) Epenthesis must precede and bleed assimilation

   

Crucially missing in this descriptively satisfactory analysis, however, is a successful explanation of the fact that the assimilation rule is bled in precisely those contexts where it would otherwise create pairs of completely identical adjacent consonants. The paragraph just before Odden's statement of the epenthesis rule (unintentionally) reveals precisely this teleological weakness of the analysis: ‘Lithuanian does not allow sequences of identical consonants, so to prevent such a result, an epenthetic vowel is inserted between homorganic obstruent stops’ (Odden Reference Odden2005: 115).

An anonymous reviewer asks whether a feature-geometric representation would help to express the homorganicity requirement on the epenthesis rule in (23a).Footnote ¹⁴ Identity could be required on adjacent Place (or C-Place) nodes, which reside on a tier distinct from laryngeal features like voicing and secondary articulations like palatalisation. There are two significant problems with this idea. First, standard models of feature geometry house the features most likely to be involved in palatalisation under the relevant Place node, leaving little justification for this reviewer's suggestion that it is ‘generally assumed [that] palatalisation is represented under V-Place, in which case it will be independent of specifications for primary place of articulation (on C-Place)’. In Clements & Hume's (Reference Clements and Hume1995: 292) model, for example, V-Place is (indirectly) dominated by C-Place; an identity requirement on C-Place would thus entail identity of V-Place as well. Similarly, in the model proposed by Halle et al. (Reference Halle, Vaux and Wolfe2000: 389), Place dominates the Tongue Body and Tongue Blade nodes which between them include all of the features ([back], [anterior], [coronal], etc.) that are most likely to be involved in palatalisation (see also Halle Reference Halle1995: 2). Again, requiring identity on Place entails requiring identity on these nodes and all of their dependent features.

Second, and more importantly, my point here and in Baković (Reference Baković2005) is not that a homorganicity requirement is representationally inexpressible, it's that such a requirement fails to relate this requirement to anything else about the grammar of Lithuanian. The question is why certain features are ignored in the determination of identity for the purposes of epenthesis; my proposed answer is that it is because of the interaction of epenthesis with assimilation. The feature-geometric answer suggested by the reviewer is that it is because C-Place is an independent node in the representation of a segment, and that epenthesis (or any process) can require identity of any node. (And if not any node, then the question becomes: which nodes?) But even if a C-Place node that includes all and only the right features can be motivated independently – and as noted in the paragraph immediately above, this is certainly not the generally accepted view – it remains a complete accident that, in Lithuanian, identity for the purposes of epenthesis is required for the node that just so happens to exclude just those features that independently assimilate.Footnote ¹⁵

In sum, the epenthesis rule as stated in (23a) misses the generalisation that voicing is ignored not as an arbitrary function of the rule, but specifically due to its interaction with the assimilation rule in (23b). The epenthesis rule should instead be stated such that it applies between adjacent identical consonants only. The fact that an epenthetic vowel also surfaces between relevant near-identical consonants – that is, the fact that epenthesis overapplies in the (2b) sense – must somehow follow from some interaction between epenthesis and assimilation. In SPE, the only possible interaction is rule ordering, but, as shown in (25), the right result cannot be reached with either of the two possible orders between epenthesis (stated so as to apply between adjacent identical consonants only) and assimilation (still stated as in (23b)).

1. (25) Possible ordered interactions between epenthesis and assimilation

   

The reason this analysis fails is because the SPE derivation of a form proceeds one step at a time, with each rule taking as its input the representation that is the output of the immediately preceding rule in the order (or the underlying representation, in the case of the very first rule). Some work in the SPE framework, most notably by Kiparsky (Reference Kiparsky and Fujimura1973), attempted to restrict the possibilities of ‘looking back’ to some earlier point in the derivation (e.g. blocking in non-derived environments) and for ‘looking forward’ to some later point in the derivation (e.g. blocking of the immediate output of a rule). But the problem in the present case is neither one of looking back nor one of looking forward, but rather one of looking over to a counterfactual result along a separate, parallel derivational path. An attempt to represent what this analysis of Lithuanian would have to look like is given in annotated form in (26), clarifying the term that I choose to use for this type of pattern: cross-derivational feeding.

1. (26) Cross-derivational feeding in Lithuanian

   

   ℚ=assimilation counterfactually applies, resulting in disallowed adjacent identical consonants

   ℙ=epenthesis overapplies between near-identical consonants, resulting in surface representation

   ℱ=cross-derivational feeding relationship

Rule ℚ applies in a counterfactual derivational step, the output of which is, as a result, subject to rule ℙ. Rather than ℚ directly feeding ℙ, however, ℙ alone applies along a separate, actual derivational path. This economically renders both ℙ and ℚ surface-true; however, ℙ overapplies because the motivation for its application was the counterfactual application of ℚ.Footnote ¹⁶ This type of example, which is technically a case of cross-derivational feeding-on-environment (see §7.3), is problematic for SPE because it involves the crucial feeding relation ℱ across parallel derivations, one of them counterfactual. Like looking back and looking forward, the problem presented by cross-derivational feeding-on-environment is one of globality, and as discussed in §§6.2–§6.3 below, this type of global interaction has a straightforward analysis in OT.

Interestingly, the standard bleeding-on-environment analysis of Lithuanian assimilation and epenthesis in (24) formed a crucial part of Kenstowicz & Kisseberth's (Reference Kenstowicz, Kisseberth and Kisseberth1971) challenge to Kiparsky's (1968) hypothesis that was noted at the end of §2. Kiparsky's hypothesis was that ‘rules tend to shift into the order which allows their fullest utilization in the grammar', one consequence of this hypothesis being that bleeding orders are ‘marked’ and are expected to become ‘unmarked’ counterbleeding orders over time. What Kenstowicz & Kisseberth argue is that bleeding-on-environment order, as in the SPE analysis of Lithuanian in (23), is unmarked and diachronically stable, and that it is instead the opposite, counterbleeding-on-environment order that is marked (and probably unattested; see §6.4 below).

Kiparsky (Reference Kiparsky and Dingwall1971) replied to this challenge with the opacity hypothesis, which reverses the markedness relation between bleeding and counterbleeding rule orders. But, as I have argued above, the real interaction between assimilation and epenthesis in Lithuanian actually involves overapplication opacity. To the extent that my argument is correct, and to the extent that the claim that the Lithuanian pattern is unmarked is also correct, the consequences for the theory of opacity are obvious: the hypothesis that all opaque generalisations as defined in (2) are hard to learn is at best too strong. (I return to this issue in §7.3 below.)

6.2 Analysis of cross-derivational feeding in OT

In OT, the derivation of a form is computed by generating and comparing multiple complete derivations (=output candidates) of that form in parallel, typically all but one of which are counterfactual (=non-optimal). This fundamentally different property of OT allows for the kind of global interaction that is needed in order to properly account for examples of cross-derivational feeding. The different derivations involved in cross-derivational feeding are simply competing output candidates in the OT analysis, and correct adjudication of the conflict between them in the crucial set of near-identity cases falls out from the ranking that is independently required for the analysis of each of the two processes in other cases.

The core constraints necessary for the analysis of the Lithuanian pattern are as given in (27). I ignore here certain necessary but contextually irrelevant refinements of the analysis: the direction of assimilation, for example. I also ignore the automatic palatalisation of the prefix consonant before the epenthetic high front vowel, and assimilation elsewhere in the word, in the assessment of candidates.

1. (27)

   1. Constraints for analysis of the Lithuanian pattern (Baković Reference Baković2005)

      1. a. Agree[voi], violated by adjacent obstruents that differ in voicing.

         Agree[pal], violated by adjacent consonants that differ in palatalisation.

      2. b. Ident[voi], violated by changes in voicing from input to output.

         Ident[pal], violated by changes in palatalisation from input to output.

      3. c. NoGem, violated by adjacent identical consonants (geminates).

      4. d. Dep[V], violated by vowel epenthesis.

Assimilation between adjacent consonants differing in voicing or palatalisation requires that each Agree[x] constraint dominate its corresponding Ident[x] constraint, as shown in collapsed form in (28) for [adʲ-bʲekʲtʲi] ‘to run up’. Note that since vowel epenthesis is in principle an alternative to assimilation that would also satisfy both Agree[x] constraints, Dep(V) must also be ranked above both Ident[x] constraints.

1. (28) Assimilation between adjacent disagreeing consonants

   

In order to enforce epenthesis between adjacent identical consonants, NoGem must dominate Dep(V). The necessity of this ranking is illustrated in the tableau in (29), with the conflict between the fully faithful candidate [ap-putʲi] in (29a) losing to the candidate with epenthesis, [apʲi-putʲi] ‘to grow rotten’ in (29b).

1. (29) Epenthesis between adjacent identical consonants

   

In this case, a change in voicing as in [ab-putʲi] in (29c) would in principle also be an alternative way to satisfy NoGem. Because it was shown in (28) above that Dep(V) must be ranked above both Ident[x] constraints, this candidate must be suboptimal due to the Agree[x] constraints dominating Dep(V).

Given the ranking in (29), epenthesis in the case of adjacent near-identical consonants is automatically expected. This is shown in (30). The crucial suboptimal candidate in which assimilation applies but epenthesis does not is shown in (30b); this candidate fatally violates NoGem, the markedness constraint otherwise responsible for epenthesis between completely identical consonants. Remaining faithful as in (30a) is of course not an option, because that violates one or both of the Agree[x] constraints otherwise responsible for assimilation. The only remaining option is epenthesis, the optimal candidate in (30c), [apʲi-bʲerʲtʲi] ‘to strew all over’.

1. (30) Epenthesis, not assimilation

   

This is a completely straightforward OT analysis that properly describes the right generalisation about the distribution of the epenthetic vowel in Lithuanian. Other examples of cross-derivational feeding-on-environment that I am currently aware of also involve interactions between consonantal assimilation processes and epenthesis to avoid adjacent identical consonants, as in Lithuanian; perhaps the most familiar example is the case of inflectional suffix allomorphy in English (see Baković Reference Baković2005). As I show in §7 further below, there also exist cases of cross-derivational feeding-on-focus.

6.3 The crucial role of globality

Cross-derivational feeding requires an unprecedented type of global mechanism, one in which the consequences of a crucially counterfactual derivational path – that is, a mapping from an underlying representation to a distinct representation that does not actually surface in the language – must be considered in order to properly define the conditions necessary for a separate derivational path from the underlying representation to the actual surface representation. These separate derivational paths, one actual and one counterfactual, are equally necessary in order for the correct generalisation to be expressed.

In the OT analysis detailed in §6.2, the endpoints of these two derivational paths simply constitute two of the competing candidate outputs derived from the underlying representation, one optimal and the other suboptimal. The choice between these two candidates is correctly and automatically made by the rankings independently responsible for each of the interacting processes (assimilation and epenthesis).

The success of this OT analysis is not at all surprising, given that global interactions of several kinds succumb naturally to unified explanation in OT in ways that have not been possible in other frameworks of process interaction.Footnote ¹⁷ In this section I discuss how cross-derivational feeding fits into a larger typology of global interactions involving blocking (‘do something except when’), a relatively simple case of which was already noted in §1. A more complex class of blocking patterns, conspiracies (following Kisseberth Reference Kisseberth1970), is discussed by Pater (Reference Pater, Kager, Hulst and Zonneveld1999) and McCarthy (Reference McCarthy2002: 95ff). Cross-derivational feeding is essentially a special case of a conspiracy in the typology of blocking patterns, as described in (31).Footnote ¹⁸

1. (31) Typology of blocking patterns

   

Consider this typology in terms of counterfactual derivations, as described and schematically represented in (32). Note that the schematic representations allow the closer relationship between conspiracies and cross-derivational feeding to stand out particularly clearly.

1. (32) Typology of blocking patterns in terms of counterfactual derivations

   

The blocked counterfactual path in each case is the failed mapping /X/→*|Y|, and the actual path is /X/→[X] in the case of simple blocking and /X/→[Z] in the case of conspiracies, including cross-derivational feeding. The difference between cross-derivational feeding and other types of conspiracies concerns the relationship between the counterfactual and actual outputs: there is a feeding relationship between *|Y| and [Z] in the case of cross-derivational feeding, as indicated by the horizontal ℱ arrow, and no such relationship in the case of other types of conspiracies.

In OT, these different blocking patterns result from particular rankings of constraints that lead to the choice of X over Y in the case of simple blocking, and to the choice of Z over (both X and) Y in the case of a conspiracy, including cross-derivational feeding. This is shown in the ranking schemas in (33), where is a markedness constraint, is a faithfulness constraint and ℂ is either of these. Subscripts denote what is penalised by each constraint; ‘Y/…’ indicates the contextual restriction on the markedness or faithfulness of (a change to) Y that must exist in order for the blocking of /X/→*|Y| to be evident.

1. (33) Basic constraint ranking schemas for blocking patterns

   

An example of the simple blocking schema in (33a) was already illustrated by the blocking of extrametricality in Latin monosyllables in (1). In that particular case, the markedness constraint _X otherwise responsible for the blocked process of extrametricality is Non-finality, and the markedness constraint _Y/… doing the blocking is Lx≈Pr, requiring that words be prosodified and stressed.

The cross-derivational feeding schema in (33c) is of course just a special case of the conspiracy schema in (33b); the unspecified ℂ_Y/… is simply further specified as being a markedness constraint (_Y/…) in the case of cross-derivational feeding. The relevant tableau from the Lithuanian analysis is (30), where there are two markedness constraints _X otherwise responsible for the blocked assimilation processes (Agree[voi], Agree[pal]). The markedness constraint _Y/… causing assimilation to be blocked is NoGem, the faithfulness constraint _Z penalising the optimal candidate with epenthesis is Dep(V), and the faithfulness constraints _Y penalising the blocked assimilation candidate are the Ident[x] constraints.

In other examples of conspiracies, the unspecified constraint ℂ_Y/… is a faithfulness constraint (_Y/…), as exemplified by Pater's (Reference Pater, Kager, Hulst and Zonneveld1999) analysis of the pattern of nasal+voiceless obstruent avoidance in Si-Luyana. Nasal+voiceless stop sequences are avoided via coalescence of the nasal and stop, satisfying a markedness constraint against nasal+voiceless obstruent sequences (*N) and violating the faithfulness constraint Linearity; this justifies the ranking *NLinearity (=_XY). Deletion could have also been employed to satisfy *N, justifying the ranking MaxLinearity (=_ZY), as shown in (34) for [nabi] ‘prince’.

1. (34) Conspiracy I:

   coalescence of nasal+voiceless stop sequences in Si-Luyana

   

*N also penalises nasal+voiceless fricative sequences, but coalescence in these cases would further violate the faithfulness constraint Ident[cont]. Here deletion occurs instead, violating Max and justifying the ranking *N, Ident[cont]Max (=_X, _Y/…Z). This is shown in (35) for [supa] ‘soap’.

1. (35) Conspiracy II:

   deletion in nasal+voiceless fricative sequences in Si-Luyana

   

An important distinction between the analysis of cross-derivational feeding in (30) and that of other types of conspiracies in (35) is whether optimal derivations of the form /Y/→[Z] are otherwise predicted to exist. In the case of cross-derivational feeding, such derivations are necessarily optimal. In Lithuanian, for example, NoGem enforces epenthesis even when assimilation is not relevant (e.g. in /ap-putʲi/→[apʲi-putʲi] ‘to grow rotten’), because it dominates Dep(V). In the case of other conspiracies, by constrast, /Y/→[Z] derivations are not necessarily expected to be optimal. In Si-Luyana, for example, deletion of the nasal is not necessarily expected unless coalescence of the nasal with a voiceless fricative is at stake, since it is only the possibility of coalescence that invokes the blocking constraint Ident[cont]. This is not to say that deletion of a nasal cannot be otherwise motivated, of course; what is relevant here are the necessary expectations from the ranking schemas in (33): cross-derivational feeding requires other independently motivated /Y/→[Z] mappings; other types of conspiracies do not.

Examples discussed in the literature under the banner of ‘conspiracies’ tend to be of the type that fit McCarthy's (Reference McCarthy2002) slogan ‘homogeneity of target, heterogeneity of process’ – examples in which the avoidance of a single marked configuration (or set of related configurations) is responsible for more than one underlying–surface disparity. This is also true, to an extent, of cross-derivational feeding: in Lithuanian, the avoidance of disagreeing clusters is responsible for assimilation and, in certain cases, for epenthesis as well. In the case of epenthesis, however, this responsibility is crucially shared with the avoidance of another marked configuration: adjacent identical consonants, which are specifically in danger of resulting from assimilation. The conspiracy in this case thus involves the simultaneous avoidance of both marked configurations with one of two underlying–surface disparities: assimilation when only disagreement is at stake, and epenthesis when adjacent identical consonants are (also) at stake.

In his discussion of conspiracies and their relation to opacity, Kiparsky (Reference Kiparsky and Fujimura1973: 80–81) concludes:

The results of this section suggest a more nuanced conclusion: conspiracies lead to increased surface truth, and this can happen at the expense of surface apparentness – though perhaps only of the kind that does not involve unnecessary rule application or gratuitous faithfulness violation. This appears to be true for cross-derivational feeding and also for non-gratuitous feeding, as discussed with respect to (18) in §5.

In sum, OT's facility with cross-derivational feeding follows from the global properties of the theory, and SPE's difficulty with this type of pattern follows from its lack of such properties. This fact has potential consequences for refinements of OT such as those by Wilson (Reference Wilson2000) and by McCarthy (Reference McCarthy2007), both of which significantly affect OT's predictions of globality. Aspects of either or both proposals may turn out to be necessary in some form, and they have the added benefit of being able to describe some of the types of opacity discussed in this article that are otherwise problematic for OT. In assessing these proposals, however, the positive aspects of globality noted in this section must be kept firmly in mind.

6.4 More on gratuitous faithfulness violations

Consider again the analysis of cross-derivational feeding in Lithuanian in (30), repeated as (36) but with another candidate added in (36d), one in which there is both assimilation and epenthesis.

1. (36) Cross-derivational feeding: Lithuanian prefixes, based on (30)

   

It would of course be problematic for OT if (36d) were the actual output in Lithuanian, because this candidate is harmonically bounded by the optimal candidate in (36c): the additional faithfulness violation incurred by (36d) in the Ident[x] column is gratuitous. Indeed, because the standard SPE analysis of Lithuanian involves bleeding – recall Odden's (Reference Odden2005) rules in (23) and the derivation in (24) – a simple reordering of the rules under such an analysis results in (36d) and would be a case of counterbleeding.

1. (37) Lithuanian bleeding (based on (24)) and anti-Lithuanian counterbleeding

   

But note that the codicil on epenthesis that is necessary under the bleeding order of rules in (37a), specifying that the flanking consonants may ‘differ at most in voicing (and palatalisation)’, would no longer be motivated under the opposite ordering of the rules in (37b). This is because the prior application of assimilation would in all relevant cases guarantee that the flanking consonants agree in voicing (and palatalisation). So, the result in (37b) is due to a self-destructive feeding interaction: assimilation provides the necessary environment for the subsequent application of epenthesis, which destroys the context that made the application of assimilation possible in the first place. (This is reflected in the constraint-violation profiles of the candidates in (36); compare counterbleeding in (7) and self-destructive feeding in (14).)

This kind of interaction between assimilation and epenthesis appears to be completely unattested, as originally suggested by Kenstowicz & Kisseberth (Reference Kenstowicz, Kisseberth and Kisseberth1971). The one apparent counterexample to this claim that is known to me comes from New Julfa Armenian (Vaux Reference Vaux1998, Odden Reference Odden2005). Since epenthesis in this case applies regardless of featural (dis)agreement, this is a standard counterbleeding interaction.

1. (38) Epenthesis and assimilation in New Julfa Armenian

   

Vaux writes (1998:216, emphasis added): ‘since intervening -ə- but not full vowels are transparent to [assimilation], we can say that the rule applies before epenthesis, which produces all schwas in the New Julfa dialect’. While this is technically true, the facts simply do not justify such an analysis. Even if all predictable information must be absent from underlying representations – see Orgun (Reference Orgun2000: 141) on this point, specifically with regard to the analysis of schwa epenthesis in Armenian – the predictability of the distribution of schwa only allows assimilation to be ordered before epenthesis. Vaux presents no specific argument for this ordering analysis, empirical or otherwise; a more plausible alternative is implicit in Vaux's own proposed statement of the rule specifying that schwa is transparent to assimilation.

Support for the schwa-transparency analysis comes from the urban Utrecht dialect of Dutch, which has regressive voicing assimilation across non-epenthetic schwa (van Oostendorp Reference Oostendorp, Taeldeman and van Oostendorp2002, personal communication). As shown in (39), the masculine definite determiner /də/ is devoiced before stem-initial voiceless obstruents and the infinitival marker /tə/ is voiced before stem-initial voiced obstruents (39a); the underlying distinction between these two inflectional morphemes surfaces before stem-initial sonorants (39b).

1. (39) Assimilation across schwa in urban Utrecht Dutch

   

Assimilation of laryngeal features across reduced vowels should not be surprising, given the likely articulatory configuration involved: the gestures of consonants flanking a reduced vowel no doubt overlap more than they do when flanking a full vowel, if they overlap in the latter case at all. If the sense of consonantal adjacency relevant to assimilation is (or at least can be) defined in terms of (degree of) gestural overlap, then the New Julfa Armenian and urban Utrecht Dutch patterns above are expected.Footnote ¹⁹

The ordering solution suggested by Vaux (Reference Vaux1998) entails the possibility of a counterbleeding interaction between epenthesis and assimilation: a pattern in which epenthetic vowels are distributionally distinguishable from otherwise identical underlying vowels (e.g. with evidence from morphophonological alternations), and in which assimilation applies only across the epenthetic vowels. This hypothetical pattern, unattested to the best of my knowledge, would be impossible to replicate in OT except by brute force.Footnote ²⁰ To the extent that the type of pattern considered in this subsection is indeed unattested, OT has a clear advantage over SPE, in which the analysis of this unattested pattern is as straightforward as any other.

7.1 Gliding and deletion in Cibaeño Dominican Spanish

The final type of process interaction involving obscured generalisations that I examine in this article are what I call concealed free rides. Concealed free rides are of interest here for a couple of reasons. First, there are two reasonable ways to analyse a concealed free ride: (i) in terms of self-destructive feeding, which is possible in SPE but not in OT, or (ii) in terms of cross-derivational feeding, which is possible in OT but not in SPE. Second, the types of obscured generalisations that arise from concealed free rides turn out not to be classifiable as opaque according to the standard definition of opacity in (2).

Consider the following example from the Cibaeño variety of Spanish, spoken in the north-central region of El Cibao in the Dominican Republic (Golibart Reference Golibart1976, Guitart Reference Guitart, Contreras and Klausenburger1981).Footnote ²¹ In Cibaeño, there is a process of liquid gliding whereby the liquids /r/ and /l/ become the palatal glide [j] word-internally before consonants (40a) and word-finally in a stressed syllable (40b). (Guitart Reference Guitart, Contreras and Klausenburger1981 argues that the context of liquid gliding is not, despite appearances, reducible to syllable-final position. I simply adopt Guitart's position in that follows, as the issue does not affect the point being made here.)

1. (40) Liquid gliding in Cibaeño

   

Word-final liquids in unstressed monosyllabic words also glide, but only before consonant-initial words: compare el mes [ejˈmes] ‘the month’ with el año [eˈlaɲo] ‘the year’. Guitart (Reference Guitart, Contreras and Klausenburger1981: 227) reduces this to the word-internal preconsonantal vs. prevocalic cases in (40a). Guitart argues that the relevant unstressed monosyllables are all determiners and prepositions with a plausible analysis as proclitics, and so the final liquids of these morphemes is word-internal in the relevant cases (/el+mes/→[ejˈmes], /el+aɲo/→[eˈlaɲo]) and thus conforms to the word-internal preconsonantal vs. prevocalic distinction.

Some underlying liquids in the relevant positions delete instead of surfacing as glides. Golibart (Reference Golibart1976) and Guitart (Reference Guitart, Contreras and Klausenburger1981) demonstrate that deletion in these cases is motivated by either one or both of two independent conditions, one phonotactic and the other prosodic. Consider first the phonotactically motivated case. Preconsonantal or word-final liquids after /i/ delete instead of surfacing as glides, as would otherwise be expected given the liquid-gliding rule. But surface sequences of a high vowel and homorganic glide ([ij ji uw wu]) are independently excluded in Spanish, a phonotactic fact that can be reasonably attributed to an independent glide-deletion rule. Golibart (Reference Golibart1976) and Guitart (Reference Guitart, Contreras and Klausenburger1981) thus analyse liquid deletion as a two-step process: liquid gliding feeds the glide-deletion rule. This is shown in (41).

1. (41) Liquid gliding feeds phonotactically motivated glide deletion

   

The independent motivation for this glide-deletion rule is, as noted above, the systematic absence in Spanish of high vowel and homorganic glide sequences like [ij]. There is no further evidence for the rule in Cibaeño, but Guitart (Reference Guitart, Contreras and Klausenburger1981: 224), following Golibart (Reference Golibart1976: 51–54), suggests that 2nd person plural familiar verb forms, found exclusively in Peninsular varieties of Spanish, provide some external motivation for the rule. The suffix (or suffixal complex) that instantiates this morphosyntactic feature set is [j]-initial after 1st (42a) and 2nd (42b) conjugation verbs, which are marked by the theme vowels /a/ and /e/ respectively. 3rd conjugation verbs (42c) are marked by the theme vowel /i/, and the underlying sequence /…i+j…/ arguably undergoes the phonotactically motivated glide-deletion rule in (41) to surface as [i].

1. (42) 2nd person plural verb forms (tv=verb conjugation theme vowel)

   

Now consider prosodically motivated deletion. When a word with non-final stress ends underlyingly with a liquid – the residue of the prosodic contexts classified by the liquid-gliding context in (40b) – there is again deletion rather than a surface glide. Stems ending in closed unstressed syllables are generally considered exceptional in analyses of Spanish stress (see e.g. Harris Reference Harris1995), but glide-final stems are unexceptionally stressed on the final syllable. As Guitart (Reference Guitart, Contreras and Klausenburger1981: 225) puts it, ‘in Spanish in general there are absolutely no polysyllabic words ending in unstressed vowel plus glide’.Footnote ²² Guitart attributes the absence of these words to a prosodically motivated deletion rule, summarising its statement as follows: ‘delete [y] ([j]) at the end of a polysyllabic word whenever it is preceded by an unstressed vowel’. Guitart's formal statement of the rule is in essential respects equivalent to the rule stated in (43). Like the phonotactically motivated rule, this prosodically motivated glide-deletion rule is crucially fed by liquid gliding.

1. (43) Liquid gliding feeds prosodically motivated glide deletion

   

Note that the only independent motivation for this prosodically motivated glide-deletion rule is the systematic absence of glide-final polysyllabic words with non-final stress in Spanish; there is no further evidence for it, even of the external type given for phonotactically motivated glide deletion in (42).

The Cibaeño pattern just described is an example of a concealed free ride. More generally, free rides are so named because they can be analysed in SPE terms as follows (see Zwicky Reference Zwicky1970). Suppose an analysis of some phonological pattern involves input–output mappings of the form /X/→[Z], and that these mappings can be broken down into two serially ordered derivational steps, /X/→|Y|→[Z]. If the |Y|→[Z] step is independently motivated in some way – most convincingly, but not necessarily, by other observed mappings of the form /Y/→[Z] – then the process responsible for the /X/→|Y| step can ‘take a free ride’ on the independently motivated process responsible for the |Y|→[Z] step.Footnote ²³

In the Cibaeño example, liquid gliding takes a free ride on glide deletion, resulting in derivations of the form /r l/→| j |→. (I refer to the two glide-deletion rules together simply as ‘glide deletion’ in much of what follows.) The reason I call it a concealed free ride is because glide deletion completely conceals the result of liquid gliding. This concealment makes the SPE analysis of a concealed free ride, as illustrated by the derivations in (41) and (43), remarkably similar to the analysis of self-destructive feeding in Turkish discussed in §4: the | j |→ step taken by glide deletion destroys the result of the serially prior /r l/→| j | step taken by liquid gliding, but of course the result of liquid gliding was necessary to motivate the destructive application of glide deletion in the first place. One might think, then, that the analysis of a concealed free ride is problematic for OT for the same reasons that self-destructive feeding is.

But the interaction between the rules involved in a concealed free ride can also be understood in cross-derivational feeding terms, much like the interaction between assimilation and epenthesis in Lithuanian discussed in §6. The essence of a cross-derivational feeding analysis of the Cibaeño example can be summarised as follows. Preconsonantal and word-final liquids become palatal glides unless either (i) the result would be a phonotactically impermissible sequence (*[ij]) or (ii) the result would be a prosodically impermissible form (*σC[j]#). In these two classes of cases, preconsonantal and word-final liquids delete instead; this is because if they were to counterfactually become glides, they would be subject to glide deletion anyway. A schematic representation of this cross-derivational feeding interaction is given in (44).

1. (44) Cross-derivational feeding in Cibaeño

   

Deletion is a thus a back-up strategy to gliding under this analysis: liquids delete whenever gliding would otherwise be blocked. This makes the OT analysis of a concealed free ride straightforward.

7.2 Analysis of concealed free rides in OT

Both gliding and deletion are possible responses to a markedness constraint against preconsonantal or word-final liquids, here simply abbreviated No[r,l]{C,#}. Those responses are regulated by particular faithfulness constraints ranked below No[r,l]{C,#}; in order to ensure that gliding as opposed to deletion is the default response, Max(C) must be ranked above any and all Ident[x] constraints, where x is a feature that distinguishes the liquids /r l/ from the glide /j/. To keep things simple, I assume that liquids and glides are primarily distinguished by the feature [±consonantal] and I refer to the relevant set of Ident[x] constraints as Ident[cons]. The preference for gliding over deletion is realised in examples where the phonotactic and prosodic conditions are not at stake, as shown in (45).

1. (45) Gliding by default

   

The phonotactic constraint responsible for the lack of sequences of high vowels and homorganic glides is abbreviated here NoV_iG_i. When NoV_iG_i is at stake, the preference for gliding is subverted and there is deletion instead, so long as NoV_iG_i is ranked together with No[r,l]{C,#} above Max(C).

1. (46) Deletion under phonotactic duress

   

Likewise, when the prosodic constraint responsible for the lack of polysyllabic glide-final words with non-final stress is at stake, the preference for gliding is again subverted and there is deletion instead. This prosodic constraint is abbreviated here as No-σC[j]# (probably a constraint on quantity-sensitivity, activated by independent considerations that ensure the heaviness of word-final syllables ending in glides).

1. (47) Deletion under prosodic duress

   

The rankings in (46) and (47) are clear instances of the basic constraint-ranking schema for cross-derivational feeding (33c), repeated below as (48). The markedness constraint _X is No[r,l]{C,#}, otherwise responsible for liquid gliding; the markedness constraint _Y/… causing gliding to be blocked is NoV_iG_i in (46) and No-σC[j]# in (47), the faithfulness constraint _Z penalising the optimal candidate with deletion is Max(C) and the faithfulness constraint _Y penalising the blocked gliding candidate is Ident[cons].

1. (48) Cross-derivational feeding ranking schema

   

This cross-derivational feeding analysis of the fate of Cibaeño liquids in glide-deletion contexts makes sense of the fact that deletion is a response to independent constraints in the language that conflict with the preferred liquid-gliding strategy. Of course, the prediction of this analysis is that deletion should be independently observable given an appropriate context; i.e. a glide arising after a homorganic high vowel (recall (42)), or one arising finally in a form with non-final stress. This is not necessarily so of the self-destructive feeding analysis in (41) and (43), where the liquid-gliding rule could in principle be preceded by rules that make changes other than deletion that also lead to conformity with the relevant phonotactic and prosodic conditions; e.g. a rule shifting stress to the final syllable in glide-final words.

It is perhaps relevant that glide-final loanwords with non-final stress in the source language are produced with final stress in Spanish; e.g. [komˈboj]<English ˈconvoy. This could be taken as evidence for the stress-shift rule just mentioned in the actual phonological grammar of Spanish, preceding liquid gliding in Cibaeño and coincidentally having the same motivation as the glide-deletion rule in (43). Whether this is the right approach to loanword phonology is a question far beyond the scope of this article; the more relevant problem for this analysis lies behind the motivation for stress shift and glide deletion: the lack of an effective description of – much less an explanation for – the teleological unity of these rules.

7.3 Concealed free rides and opacity

Given the similarities between concealed free rides and self-destructive feeding on the one hand and cross-derivational feeding on the other, a concealed free ride might be expected to involve overapplication opacity. It is the first process in a self-destructive feeding interaction that overapplies, and this corresponds to liquid gliding in the SPE analysis of the concealed free ride in Cibaeño. By contrast, it is the actually applying process that overapplies in the case of cross-derivational feeding, and this corresponds to glide deletion in Cibaeño. Does this mean that both processes express opaque generalisations due to their interaction with each other? The matter turns out to be more complicated than one might think.

On the one hand, it is clearly the case that each of the two processes expresses a generalisation that is obscured by its interaction with the other. Liquid gliding turns liquids into glides, but this transformation is obscured by the fact that some of those glides-from-liquids are subsequently deleted. Similarly, glide deletion deletes glides, but this is obscured by the fact that some glides come from underlying liquids, which glide deletion also deletes. There is a clear intuitive sense in which glide deletion can be thought to overapply, then: glide deletion appears to delete more than just glides. Conversely, liquid gliding is obscured, but it is pretty clear that it doesn't overapply: it doesn't appear to apply to more than just liquids, nor does it result in glides in contexts other than those stated by the rule.

In fact, neither process can be classified as opaque in either of the two senses of the term according to Kiparsky's definition of opacity in (2). First of all, liquid gliding does not overapply in the (2b) sense: there are no surface instances of glides derived by liquid gliding that occur in environments other than preconsonantally or word-finally. Liquid gliding also does not underapply in the (2a) sense: there are also no surface instances of liquids that occur preconsonantally or word-finally. Likewise, glide deletion clearly does not underapply according to the definition in (2a), because there are no surface instances of glides in either of the two deletion environments (and recall that the surface truth of this fact is the primary motivation for both types of glide deletion in the first place). It proves to be somewhat trickier to square the definition of overapplication in (2b) with a deletion rule: what constitutes a surface instance of B when B is , the absence of a segment? Even assuming that there is a way to refer to ‘deleted segments’, however, glide deletion does not overapply according to the definition in (2b), because there are no instances of deleted segments derived by glide deletion in environments other than those in the two deletion environments.Footnote ²⁴

The interaction between the processes involved in a concealed free ride is thus, somewhat counterintuitively, a transparent one according to the standard definition of opacity. The source of this puzzle appears to be the formal distinction that we have encountered several times throughout this article: the concealed free ride in Cibaeño involves feeding-on-focus, while the self-destructive feeding and cross-derivational feeding examples discussed in §4 and §6 both involve feeding-on-environment. The diagnostic for overapplication in (2b) is defined in such a way that only a disruption of the environment of a process ℙ can mean that ℙ has overapplied, which thus excludes on-focus interactions.

It might be asked what consequences this result has for the theory of opacity, and the answer is unclear. The definition of opacity in (2) simply identifies surface diagnostics for opaque generalisations, and so the definition itself could in principle just be modified to accommodate concealed free rides if there were reason to believe that concealed free rides must involve opacity. But recall that the theory of opacity is based on the substantive assumption that opaque generalisations are harder to learn than transparent ones; given that it is unknown whether the obscured generalisations involved in concealed free rides are easier or harder to learn than other generalisations, it is unclear whether they should be considered opaque.

Now recall that Kiparsky's original evidence for the psychological reality of the opacity of a particular (type of) rule order was more or less the extent to which such an order could be shown to be diachronically unstable. If, for example, an innovating dialect appears to require the inverse of an inherited order between two rules, this could be an indication that the inherited order is opaque while its innovated inverse is transparent. However, note that the inverse of a concealed free ride would be a case of counterfeeding, and thus a genuine case of (underapplication) opacity. In a putative innovative subdialect of Cibaeño, glide deletion would not delete those instances of glides derived by the now later-ordered liquid-gliding rule, rendering glide deletion non-surface-true. If a concealed free ride indeed involves opacity, then, it cannot be diagnosed by some diachronic tendency for the rules involved to be reordered.