Lenition and morphology in Finnish

This paper investigates an interaction between consonant lenition and morphology in Finnish. The language has a process of consonant lenition whereby underlying geminate consonants at syllable boundaries lenite (degeminate) when the addition of an affix makes the post-geminate rime bimoraic. A small class of possessor agreement affixes do not condition lenition, even if they create the appropriate phonological environment. A puzzling interaction emerges when possessor agreement affixes are stacked on top of certain lenition-conditioning affixes. I account for this interaction in a way that improves on Kiparsky’s (2003) analysis. In doing so, I extend Pater’s (2010) method for modeling exceptional phonology via lexically-indexed constraints.

1. Background and the puzzle.Finnish has a process of consonant lenition (also termed gradation in the Uralic literature) whereby underlying geminate stops at syllable boundaries degeminate when the post-geminate rime becomes bimoraic (Antilla 1997, Kiparsky 2003).Various types of rime are bimoraic in Finnish, but in this paper I'll focus only on VC rimes for simplicity.(See Kiparsky 2003 for an analysis of the interactions between stress, syneresis, and moraicity in the case of surface VV rimes.) Examples ( 1a) and (2a) below 1 provide a baseline, where an underlying geminate surfaces faithfully.The (b) variants illustrate the lenition process, with the conditioning environment resulting from the addition of /-C/ or /-CCV/ affixes.
1.2 AFFIX STACKING AND LENITION: THE GENERAL CASE.What happens lenition-wise when a POSS affix is stacked on top of a more internal, lenition-conditioning affix?Usually, things go as expected given the data presented above.
From a derivational perspective, however, this might look like an under-application of lenition.Why don't the affixes that usually instantiate GEN.SG and NOM/ACC.PL (/-n/ and /-t/) condition lenition on the root?We saw from example (6) that affixes situated between the root and POSS can indeed condition lenition on the root geminate, so long as they provide the appropriate environment by making a post-geminate rime bimoraic.The GEN.SG and NOM/ACC.PL case affixes in ( 7) and (8) occur in the same context (that is, they're situated between the root and POSS), and they also have the appropriate shape to condition lenition: /-C/ (cf. 1b).Kiparsky (2003: p. 153) makes an important observation about the data in ( 7) and ( 8), namely that facts [1] and [2] seem to be related.That is, perhaps the absence of an exponent for the internal affixes bleeds lenition on the root geminate.For example, if there were no phonological exponent of the GEN.SG in (7), then there would be nothing to condition lenition on /hAttu/, since /-mme/ doesn't condition lenition anyway.
In the next section I review Kiparsky's (2003) particular strategy for tying together facts [1] and [2].I offer two related arguments against his analysis, and in §3 I offer an alternative.
2. Kiparsky's (2003) Stratal OT analysis.We turn attention now to Kiparsky's solution for the puzzle from §1.3; his analysis implicates a morpho-phonological conspiracy (see Kisseberth 1970).A phonological markedness constraint (the Stem Constraint, introduced below) drives omission of the morphemes /-n/ and /-t/, which usually instantiate the GEN.SG and NOM/ ACC.PL case categories in examples like ( 7) and ( 8).The same markedness constraint drives a substantively different syncope process in other contexts; a discussion of one of these contexts will come below.The morpho-phonological component is assumed in Kiparsky (2003) to have a Stratal OT architecture (Stratal Optimality Theory; see Kiparsky 2000 for an overview).In Stratal OT, morphs are indexed to particular strata.At stratum n, morphs indexed to n are inserted, and a stratum-specific phonology -with a statum-specific constraint ranking -applies.The output of stratum n then feeds stratum n+1, where morphs indexed to n+1 are inserted, and a potentially distinct phonology applies once again.This process applies cyclically through all the strata in a given language.
For Kiparsky, Finnish possessor agreement affixes are indeed to the word stratum (ω); all other affixes relevant here are indexed to the stem stratum (α).The output of the stem stratum feeds the word stratum.Consonant lenition is considered a stem-level phonological process, but not a word-level process.That is, at α, the markedness constraint driving lenition outranks the faithfulness constraint militating against lenition (M> >F), but at ω this ranking is reversed (F> >M).
I revisit example (6), repeated below as (9), to illustrate how this system works in practice.
The first-pass derivation I offer here fails to capture the observed surface form in two respects.First, it predicts lenition on the root geminate.If /hAttu/ and /-n/ are inserted at α, with an M> >F ranking, root lenition is predicted, since the GEN.SG affix /-n/ would make the second syllable's rime bimoraic.However, no root lenition is observed (fact [1] from §1.3).Second, an exponent of the intermediate affix (here GEN.SG /-n/) is predicted to appear, while in fact no such exponent is observed (fact [2] from §1.3).To resolve the mismatch between predicted and observed forms in examples like (7), Kiparsky puts forward that the GEN.SG category is not realized as an affix at all -rather, it is omitted morphologically (p.151).Morphological omission is driven by the Stem Constraint (*-C] α ), a markedness constraint which penalizes stems not ending in a vowel.
(13) DEP-CAT = A grammatical category must be morphologically expressed.(Kiparsky 2003: 154) Violations of this constraint are incurred when a grammatical category (in this instance, case) is omitted from the derivation: that is, not granted a morphological representation to begin with.By the ranking *-C] α , > >DEP-CAT, categories otherwise expressed as an affix with a /-C/ shape will have no expression the Stratal morpho-phonological derivation at all.Since *-C] α outranks DEP-CAT, including an underlying /-n/ in the derivation of ( 7) is non-optimizing to begin with, as it incurs a violation of the Stem Constraint.Thus the derivation of ( 7) is, for Kiparsky, represented as ( 14). ( Since there are no morphemes to instantiate the GEN.SG, NOM.PL, and ACC.PL categories (/n/ and /-t/ are, in his words, "simply not present", p. 151), the absence of lenition on the root geminate and the non-exponence of GEN.SG and NOM/ACC.PL categories are simultaneously derived.Kiparsky's intuition that non-exponence is related to the lack of root lenition is crucial, and is something we'll make use of in the proposed re-analysis in §3.However, the exact strategy by which he relates these facts is less than ideal, as I spell out below.

PROBLEMS WITH MORPHOLOGICAL OMISSION. The Stem Constraint is considered by
Kiparsky to be undominated, and its status as such means that Finnish stems will always be vowel-final.One way the grammar enforces this is through morphological omission under *-C] α > >DEP-CAT, as detailed above.But there are other process that enforce conformity to the Stem Constraint as well.
One such process is as follows.When a multi-segmental, consonant-final affix is situated left-adjacent to a POSS affix in Finnish, it surfaces without its final consonant (Kiparsky 2003: p. 149).In contrast to (14) above, the entire morpheme is not omitted.Some examples are offered below.
In the (a) examples, when no possessor agreement is present, the morphs surface with their underlying final consonants.But in the (b) examples, when left-adjacent to possessor agreement, the final consonant does not surface.Kiparsky treats this as a case of final consonant deletion: consonant-final affixes like /-hin/ and /-siin/ "lose their last consonant before possessive suffixes" (p.149).Though the faithfulness constraint he has in mind is not stated explicitly, the relevant ranking seems to be *-C] α > >MAX(C).
Recall, however, that the non-exponence of the mono-consonantal case affixes (/-n/ and /-t/) was derived through a different ranking: *-C] α > >DEP-CAT.This treatment divorces the behavior of multi-segmental, consonant-final affixes from the behavior of mono-segmental, consonant-final affixes.The former are subject to phonological syncope, while the latter are subject to morphological omission.For Kiparsky, independent processes are acting on these two types of affix, resulting in a morpho-phonological conspiracy driven by the Stem Constraint.There's nothing wrong with analyzing a set of facts as a conspiracy per se; indeed, one of the great advantages of Optimality Theory over rule-based approaches is that the former can model conspiracies in a satisfying way.Still, a more parsimonious account of the Finnish facts would be one where the behavior both types of affix (mono-consonantal and multi-segmental) was derived from a single M> >F ranking.
There's another, related problem for Kiparsky's analysis as well.Consider the following question: why are mono-consonantal affixes subject to morphological omission, while multisegmental, consonant-final affixes are subject to phonological syncope?In principle, phonological syncope could satisfy the Stem Constraint for either type of affix, as could morphological omission.To resolve this question, Kiparsky recruits another constraint, MAX-MORPHEME.
(17) MAX-MORPHEME = A morpheme must be phonologically expressed.(Kiparsky 2003: 153) Crucially, violations of this constraint are incurred only when a morpheme already present in the morpho-phonological derivation surfaces with no exponent.That is, violations are not incurred if a morpheme is omitted entirely.
The following tableaux contain Kiparsky's ranking of these constraints6 and illustrates how they interact with some of the examples I introduced above.Note that candidates are now input-output pairs of morphemes and surface forms.MAX-MORPHEME is abbreviated at MAX-M, and each morpheme is marked with a subscript indicating the stratum it's indexed to.Tableau (18) illustrates how morphological omission is derived.While this analysis seems to work at first glance, there's a lookahead problem lurking in the shadows.Recall that the output of stem stratum feeds the word stratum.What this means is that stem-level constraint evaluation should have no information available about whether any word-level affixes will be added later in the derivation.But consider the contrast between example (1b) and the derivation in (14/18c).On Kiparsky's treatment, the GEN.SG category in (1b) is expressed as the morpheme /-n/, while in (14/18c), the morpheme corresponding to the GEN.SG category is omitted.Omission, or lack thereof, is sensitive to whether there will be a POSS morpheme added at a derivationally later stratum.
The central issue with the constraints utilized in tableaux (18-19) is that violations of *-C] α , MAX-MORPHEME, DEP-CAT, and MAX(C) are being evaluated across strata.To my understanding, this contradicts the cyclicity assumption of Stratal OT, and instead gestures toward a new analysis of these facts.My goal in the next section is to offer such an analysis.
3. Re-analysis: exceptional phonology and lexically-indexed markedness.In this section I offer a re-analysis of the Finnish lenition facts.This analysis unifies the treatment of multisegmental, consonant-final affixes with the treatment of mono-consonantal affixes, in that I derive the fact that POSS affixes always surface right-adjacent to vowels with a single M> >F ranking.(Recall that Kiparsky had two: *-C] α > >MAX(C) and *-C] α > >DEP-CAT.)My analysis obviates the need for morphological omission and avoids the lookahead problem of Kiparsky's system.Instead, I employ and expand on Pater's (2010) means of capturing exceptional phonological behavior with lexically-indexed constraints.
3.1 CONSONANT DELETION.The first task is to unify the treatments of multi-segmental, consonant-final affixes (like /-hin/ and /-siin/ from ( 15) and ( 16)) with the treatment of monoconsonantal affixes (like /-n/ and /-t/ from ( 7) and ( 8)).Since both types of affix only lose a consonant when left-adjacent to POSS affixes, I propose that we should consider the POSS affixes themselves as the culprits responsible for deletion (rather than the Stem Constraint).In candidate (b), the /n/ input segment is present in the output.This candidate incurs a violation of ALIGN, and does not incur violation of MAX(C).In candidate (a), a violation of MAX(C) is incurred, but ALIGN is satisfied.Since ALIGN outranks MAX(C), candidate (a) is optimal.This particular ranking (ALIGN> >MAX(C)) will be important later, but for the moment, it will be put aside.
3.2 EXCEPTIONAL NON-TRIGGERS.Next, we need a way of capturing the lenition facts.Specifically, we need a way of modeling the contrast between, one the one hand, POSS affixes, which exceptionally do not condition (or trigger) lenition, and on the other, the rest of the af-fixes of the language, which do trigger lenition.I make use of Pater's (2010) model, which captures exceptional phonological behavior by way of lexically indexing constraints.
As an illustration of the type of phenomenon lexical indexing is supposed to model, consider Finnish /A-i/ from Pater (2010).When /A/ and /i/ are brought together at morpheme boundaries, they can surface as either [Ai] or [oi], the choice of which depends arbitrarily on the morpheme containing /i/.Some morphs containing /i/, such as the conditional /-isi/, allow /A-i/ to surface faithfully as [Ai].When other other morphs containing /i/ are involved, such as the past tense /-i/, [oi] surfaces: The contrast between examples ( 23) and ( 24) can be captured if there are actually two markedness constraints involved in their derivations.The first is a general constraint that incurs violations for any and all [Ai] strings.The second, *[Ai] L , incurs violations only when a member of a lexically-indexed (L) class of morphs is part of the [Ai] string.( In tableau ( 27), no morpheme indexed to L is present, so violations of *[Ai] L cannot be incurred.The IDENT constraint outranks the general *[Ai] constraint, so a change in vowel height, as in candidate (b), is non-optimizing.In (28) by contrast, a morpheme lexically indexed to L is present, namely /-i/.Candidate (b) here incurs violations of both markedness constraints (but crucially *[Ai] L ); candidate (a) wins.Pater's system can model cases where an exceptional class of morphs are triggers of a process (or non-undergoers, if faithfulness constraints are lexically indexed).To model the lenition facts from §1, additional machinery will be needed; the reasons for this deserve some comment.
A process in Optimality Theory can be modeled with a markedness-over-faithfulness ranking.When a select class of morphs are exceptional triggers of a process, a markedness constraint indexed to that class can be posited above the relevant faithfulness constraint; the process will apply when the relevant structural configuration is met so long as it contains a morpheme indexed to the exceptional class.If the general (i.e., non-indexed) markedness constraint is ranked below the faithfulness constraint, the process cannot apply to strings not containing an exceptional morph.In this characteristic M L > >F> >M ranking, the indexed markedness constraint, crucially, outranks the non-indexed markedness constraint.Now recall the lenition pattern from §1: all morphs except the possessor agreement affixes can trigger lenition.Here, an exceptional class of morphs are non-triggers of a process.No ranking containing a lexically-indexed markedness constraint, a non-indexed (general) markedness constraint, and a faithfulness constraint will derive this pattern; the specific/indexed vs. general/non-indexed dichotomy is too narrow.For illustration, assume that the exceptional (lexically-indexed) class contains all and only POSS affixes.If the ranking were M L > >F> >M, lenition would only occur under conditioning by POSS affixes, the exact opposite of the attested pattern.If the ranking were M> >F> >M L , lenition would apply everywhere, again contra fact.
This problem stems from the fact that Pater's (2010) lexically-indexed markedness constraint definition features only existential quantification over morphs (as I put it, "one violation for every output [Ai] string that contains some morpheme lexically indexed to L. .."8 ).That is, the original schematic for lexically-indexed constraints is something like the following (cf.(26) above): (29) *[S] ∃,n = one violation for every output string [S] such that some (∃) morpheme contained in [S] is indexed to class n Here S stands for the locus of violation -some marked string.Subscript ∃ indicates that this markedness constraint is sensitive to whether some morpheme of a particular class falls within the string, while subscript n tracks class membership (I use n instead of L here because in what follows I'll be indexing morphemes to multiple classes.) Intuitively, what's needed to model Finnish lenition is for the process to apply just in case all morphs in the marked string are not POSS morphs.That is, the constraint definitions need to be able to reference universal quantification -whether all morphs contained in a string are of a certain variety.If they are, lenition should apply; if they aren't, it shouldn't.Accordingly, I propose that the following type of constraint definition should also be possible.
(30) *[S] ∀,n = one violation for every output string [S] such that all (∀) morphs contained in [S] are indexed to class n I'll make use of both existential and universal varieties in the analysis below (though the exis-tential variety of the constraint could be substituted for a general/non-indexed constraint without complications).Before going there, we need to spell out exactly what the n and S variables in the constraint definition schematic above should correspond to for the task at hand.n is a variable over classes of morphs.I assume the following classes of Finnish morphs: class 1 contains all morphs except the possessor agreement morphs, while class 2 contains all and only possessor agreement morphs.9(31) Class 1 = {x | x is not in Class 2} Class 2 = {-ni, -si, -mme, -nne, nsA, -Vn} S is a variable over types of phonological strings.Following Antilla (1997), the type of phonological string that drives lenition is assumed here to be [σ µµ σ µµ ] -i.e., a string consisting of adjacent, bimoraic syllables.10This is a "quantitative dissimilation constraint" (Kiparsky 2003: 131) (Jarosz 2014).The pattern in (37) could be captured a different way if phonological optimization is interleaved with morpheme insertion (add an affix, lenite, add another affix, lenite, etc.).This sort of architecture is utilized in, e.g., Wolf (2008) andBaković (2002).Such an approach would require reference to both morphology and phonology, so is a bit less parsimonious than Directional HS.

(
20) ALIGN(L, POSS, R, V) = one violation for every left edge of a POSS affix that is not aligned to the right edge of a vowel If the ALIGN constraint outranks a faithfulness constraint militating against consonant deletion, the final consonants of affixes left-adjacent to POSS affixes will be subject to syncope.The relevant faithfulness constraint is defined in (21), and tableau (22) illustrates the ranking required to derive (16b).

(
21) MAX(C) = one violation for every input consonant not present in the output

(
25) *[Ai] = one violation for every output [Ai] string (26) *[Ai] L = one violation for every output [Ai] string that contains 7 some morpheme lexically indexed to L, the exceptional class of morphs Assume past tense /-i/ belongs to the lexically indexed class.With a faithfulness constraint in the mix (in this case IDENT([LOW])), the attested pattern can be derived on the ranking given in tableaux (27-28) below.
).The privative itself contains another geminate, which does not lenite because the following rime is closed by a possessor agreement affix.
hA.tum.me]* * c. /hAttu α -Ø α -mme ω /∼[hAt.tum.me]* Here candidate (a) expresses the category GEN.SG with the morpheme /-n/.This /-n/ is not subject to phonological syncope, so no violation of MAX(C) is incurred.The output of the stem level for this candidate is [ha.tun], which incurs a violation of *-C] α , the Stem Constraint.Candidate (b) likewise expresses the category GEN.SG with the morpheme /-n/, but now this morpheme is subject to syncope, incurring a critical violation of MAX-MORPHEME (and incurring a violation of MAX(C) as well).Candidate (c) opts for morphological omission, and does not express GEN.SG at all.This incurs a violation of DEP-CAT, but satisfies the higherranked constraints *-C] α and MAX-MORPHEME.Now consider (19), whose winning candidate features the other process implicated in the conspiracy -phonological syncope.-i α -siin α -si ω /∼[vA.pAi.siin.si] * b. /vApAA α -i α -siin α -si ω /∼[vA.pAi.sii.si]* c. /vApAA α -i α -Ø α -si ω /∼[vA.pAi.si]* Candidate (a) expresses the category LOC with the morpheme /-siin/.The final consonant of this candidate is not subject to phonological syncope, so no violation of MAX(C) is incurred.The output of the stem level for this candidate is [vA.pAi.siin], which incurs a critical violation of *-C] α , the Stem Constraint.Candidate (b) also expresses the category LOC with the morpheme /-siin/, but with the final consonant being subject to syncope.This incurs a violation of MAX(C), but crucially not MAX-MORPHEME, as the morpheme is still phonologically expressed.Candidate (c) opts for morphological omission and does not express LOC at all; this incurs a violation of DEP-CAT.