2.1 Breton rannigs

The first case study comes from the Celtic language Breton. Breton clauses are typically analyzed as having a Celtic-typical VSO-structure except for matrix clauses which seem to have verb-second like properties similar to many Germanic languages. The preverbal position in Breton matrix clauses can be occupied by a number of different constituents (see i.a Anderson 1981; Schafer 1995) and is usually associated with interpretations such as focus, contrastive focus or topic (Schafer 1995; Rezac 2004).

In (1a), we have a subject in preverbal position but the same position can also be occupied by the direct object in (1b), an adverb in (1c), a possessor in (1d), a predicative adjective in (2a) or a preposition phrase in (2b). Even verb phrases and bare verbs can be moved to the first position. In this case, the base position of the verb is usually filled with a dummy verb.

The element of interest to the discussion is the little particle referred to as the ‘rannig-verb’ glossed as R in the examples in (3) and (4). The element is generally claimed to be a part of the verbal complex and always immediately precedes the main verb of the clause in verb-second contexts. In verb-initial contexts, the element does not occur.

As can be seen the element shows some variation in form, varying between a and e. This variation is demonstrably not phonologically conditioned. Since it affixes to the verb (rannig-verb translates to ‘small part of the verb’), we might expect the alternation to track phonological features of the verb but it clearly does not since we find a in (3a) and e in (3c), both with the same verb lenn. Similarly, it can be shown that the alternation does not track the phonological features of the element preceding it. Rather, the standard analysis is that the rannig tracks the syntactic category of the element in preverbal position (see e.g. Anderson 1981; Urien 1999; Rezac 2004; Jouitteau 2007 and others). If the element in preverbal position is a nominal category, then the rannig surfaces as a and if it is not, it surfaces as e.

In (3a), (3b) and (3d), we have a nominal element in preverbal position, namely a nominal subject, direct object or a possessor. In (3c), (3e) and (3f), we have elements in first position that are not nominal: An adverb, an adjective and a preposition phrase. Thus, in the former cases, the rannig surfaces as a and in the latter cases it surfaces as e.

A minor complication comes in when looking at verb fronting constructions such as the ones in (4). Here, apparently, the fronted constituent, i.e. an infinitival verb or verb phrase counts as nominal for the purposes of choosing the rannig. This might seem unusual at first but Jouitteau (2007) argues extensively that infinitival verbs and verb phrases count as nominal for other processes in Breton as well.³

So, once we accept that infinitival verbs and verb phrases count as nominal for the purposes of choosing the rannig, we can arrive at a fairly simple generalization to capture the distribution of its forms: The rannig occurs as a when it follows a nominal element and e elsewhere.

Rezac (2004) analyzes this alternation as an instance of syntactic agreement. Under his analysis, the rannig is the realization of T and probes for a categorial feature of an XP in its complement domain. The syntactic category of X is then copied onto T and the element moved to SpecTP. This is illustrated in (5).⁴

However, there are quite a number of reasons why deriving the alternation of the rannig as the result of syntactic agreement is unsatisfactory and why deriving it as an instance of allomorphy is a better choice. First, we note that syntactic agreement with respect to syntactic category of an element is very uncommon. Elements in the verbal spine are known to probe for ϕ-features of certain arguments but they do not probe for categorial features. Allomorphy on the other hand is frequently subject to categorial distinctions. A nominalizer in English for example takes different forms depending on whether it nominalizes an adjective or a verb.

Second, we can observe that actual ϕ-agreement in Breton patterns differently. ϕ-agreement in Breton always tracks the subject, regardless of which element moves to the first position. Further, visible ϕ-agreement in Breton is, as is the case in Celtic languages in general, subject to the complementarity effect. This means that agreement is only visible when the trigger of agreement is dropped. This is clearly not the case with the rannig. Crucially, the rannig is only overt if its trigger is overt.

Finally, the relation between the rannig and the element in the prefield differs from regular agreement in that it is limited to a specific position, which is obligatorily adjacent. In languages with a certain amount of word order flexibility such as many V2 languages, agreement is typically not limited to a certain position in the tree. Jouitteau (2007) observes the same noting the fact that “this type of agreement can never be made at a distance, which designs it as a very peculiar Agree relation.”

All of these arguments, some of which are similar in spirit to the ones given by Bobaljik & Harley (2017) for Hiaki, suggest an analysis according to which the choice of the rannig is an instance of allomorphy. The Vocabulary Items in (8) simply associate a functional head with two allomorphs. The a-allomorph is inserted in the context of a nominal element, and the e-allomorph is inserted elsewhere. I have abbreviated the functional head they realize as R since the exact locus of the rannig is subject to a bigger debate. Schafer (1995) locates it in C⁰ while Jouitteau (2007) assuming a more elaborate architecture of the left periphery assumes the rannig to be a realization of Fin⁰.

However, in order for the Vocabulary Items to derive the correct pattern, the specifier of R must be accessible contextual information when the form for R is inserted.⁵

Nothing else needs to be said. There is no need to posit crosslinguistically uncommon ϕ-probes in the clausal spine if Vocabulary Insertion of R can consider featural properties of SpecRP when inserting R.

Also, I would like to note that the element triggering the allomorphy is derived by movement, it necessarily occupies a specifier. Elements inside the complement can never affect the choice of the rannig. The locality condition on context-sensitive spell-out in Bobaljik & Harley (2017), which excludes specifiers from the set of possible allomorphy triggers alltogether, is clearly incompatible with the pattern at hand.

2.2 Busan Korean interrogative complementizers

The second case study comes from the Busan dialect of Korean, which unlike standard Korean exhibits an interesting alternation in the interrogative complementizer. The following minimal pairs illustrate the attested four-way distinction.

Traditionally, the four forms are listed as four different complementizers: -na, -no, -ka and -ko. But Lee & Barrie (2017) convincingly argue that the forms can be subanalyzed as consisting of two independent alternations, the consonant alternation between -k- and -n- and a vowel alternation between -a and -o.

They show that both alternations are governed by completely independent triggers and should thus be analyzed independently of each other. The vowel alternation between -o and -a encodes whether the question at hand is a wh-content question or a polarity question. -o indicates a wh-question ((10b) and (10d)) and -a indicates a polarity question ((10a) and (10c)). The consonantal alternation however, independently encodes whether the predicate of the clause is a nominal or a verbal predicate. -n- indicates a verbal predicate ((10a) and (10b)) and -k- indicates the presence of a nominal predicate ((10c) and (10c)).

The use of the -k--complementizer unsurprisingly also extends to embedded nominalized clauses as can be seen in (11). In other words, nominalized clauses count as nominal for the purposes of complementizer choice.

In order to do justice to their insight that the complementizer in Busan Korean is actually a bimorphemic complex, Lee & Barrie (2017) assume a minimally split CP consisting of the two heads Int⁰ and Force⁰. The vowel alternation is located in Force⁰ and the consonant alternation is part of Int:⁶

Based on this analysis, Lee & Barrie (2017) assume that the alternation of the Int⁰-head arises as the result of AGREE of the Int-head with the syntactic category of the predicate lower in the tree. As in Rezac’s analysis of Breton, the probe does not look for a specific feature, any categorial feature will be good enough to value Int⁰’s probe. The following tree illustrates the derivation proposed by Lee & Barrie (2017):

The Int⁰-head probes for a syntactic category feature and, ignoring T for some reason, finds v, whose categorial feature is then copied onto Int⁰. Similarly, a nominal feature n of a nominal predicate would have been copied. Upon Vocabulary Insertion, the featural distinction is locally available in Int⁰ itself and can thus be referred to by the specification on the vocabulary items.

This analysis certainly works but has a number of severe downsides. First, we again must note that agreement with respect to syntactic category is relatively unusual. In fact, Lee & Barrie (2017) note that their case study and Rezac’s study on Breton are the only ones which make use of a probe looking for a syntactic category feature. Second, it is quite unclear why the probing indicated in (13) is able to skip T, and maybe even more importantly, the syntactic category of whatever is in SpecTP. Both elements have syntactic category features that should be able to satisfy the probe on Int. And according to the any definition of Minimality, both elements should definitely be closer to Int and thus bleed AGREE between Int and v. This is especially problematic for the subject position in SpecTP, which is, in the vast majority of cases, presumably filled with a nominal element. This nominal element should then trigger the nominal variant of Int even in cases where there is a verbal predicate below.

Thirdly, we find familiar cases of blocking effects where any element intervening in between the nominal predicate and the complementized actually bleeds the insertion of -k- into Int. Consider the following examples. In all cases, we have a nominal predicate, hence we would expect Int⁰ to be realized with -k-. However, since there are various elements intervening in between the copula and complementizer, insertion of -k- is blocked and we resort to what seems to be the default marker -n-. In (14), there is an overt tense morpheme intervening in between the copula and the morpheme that realizes Int⁰. In the previous examples, we had a zero realization of T, which did not block the occurrence of -k-. In (15), we have an evidential marker intervening between the copular and the complementizer and finally, in (16), we have a volitionality marker intervening.

In order to derive these cases, Lee & Barrie (2017) resort to the stipulation that these heads (overt tense, evidentiality, and volitionality) in fact have a syntactic category feature that blocks AGREE between Int⁰ and the copula or the nominal predicate down below. Moreover, all of these heads seem to be syntactically verbal, thereby copying a v-feature onto Int⁰. In other words, overt T has a v-feature but covert tense does not. It is clear that this analysis is merely a stipulation that tries to account for a morphological blocking effect by means of syntactic intervention. Again, as in the Breton case, the analysis works, but it makes a number of undesirable stipulations and leaves a number of questions unanswered.

I would like to submit that a morphological analysis where the consonant alternation of the interrogative complementizer in Busan Korean is modelled in terms of context-sensitive spell-out is much more straightforward and captures the attested pattern with far less stipulations. I follow Lee & Barrie (2017) in assuming that the complementizer is actually a bimorphemic complex and the syntactic structure they propose. Unlike in their analysis, I do not assume any agreement probes on Int⁰ but simply assume that upon Vocabulary Insertion, the realization of Int⁰ is sensitive as to whether it is adjacent to a nominal category. As mentioned in the previous section, allomorphy with respect to syntactic category is very widespread. I thus assume the following entries for the realization of Int⁰:

The -k--allomorph is inserted when right-adjacent to a nominal category. I have abbreviated this with the label n trying to capture the natural class between (a) nominal predicates and (b) nominalized verbal predicates. It is possible that future research will reveal a more fine-grained distinction here but I assume that nonetheless there must be a common feature that unifies these two contexts.⁷

The -n--allomorph, which is mainly used with verbal predicates counts as a default exponent, thereby accounting for cases where it occurs with nominal predicates and overt material in between n and Int⁰ as in (14), (15) and (16). These cases would be linearized as (18) and this linearization would not allow for the insertion of (17a).

Abstractly speaking, the case studies from Breton and Busan Korean share a number of properties. In both cases it is the categorical feature (nominal vs non-nominal) which governs the form alternation of a functional head in the verbal spine. And in both cases, the use of the nominal allomorph required linear adjacency to the nominal element in question. The major difference was whether the nominal element triggering the alternation on the functional head is in fact part of the same word or not. But apart from that, the alternations seem to be quite parallel. That suggests that there is no crucial underlying difference between allomorphy withing words and across word boundaries.

2.3 Complementizer agreement

The final case study that I want to highlight in this paper is about the well-known cases of complementizer agreement in certain dialects of Dutch as well as Bavarian German. Complementizer agreement has gained a lot of attention during the early years of this century as at seems to provide evidence for the existence of a ϕ-feature set on C and various modellings of feature inheritance from C to T (see amongst many others Haegeman 1992; Carstens 2003; van Koppen 2005; Richards 2007; Chomsky 2008; Haegeman & van Koppen 2012). At the same time however, a number of people remarked that complementizer agreement seems to behave fundamentally different with respect to a number of properties and that it looks much more like a surface phenomenon that is to be dealt with on PF rather than in syntax proper (see e.g. Ackema & Neeleman 2004; Fuß 2008; Miyagawa 2009; Fuß 2014).

I will follow the arguments presented in favor of the second analysis, ultimately arguing that complementizer agreement is not an instance of agreement at all. Rather it is an instance of allomorphy as it shows all the hallmark properties of context-sensitive spell-out: (Linear) Adjacency requirements, insensitivity to hierarchical structure and a very restricted set of alternating forms.

I will not provide entirely new arguments for the PF-approach but I will show how recent discussions about the restrictions on context-sensitive spell-out can nicely derive the attested variation between Dutch and Bavarian German. In Section 3, we will see how the inclusion of complementizer agreement phenomena informs the discussion about the general properties of context-sensitive spell-out.

In the first subsection, I will look at complementizer agreement in some Dutch dialects and provide the general evidence for a treatment of complementizer agreement as a PF phenomenon. In the second subsection, I will then highlight some interesting differences found in Bavarian German. The discussion presented in this section draws heavily on the treatments in van Koppen (2005; 2007; 2012); Haegeman & van Koppen (2012) and Fuß (2014).

3.1 Triggers

Firstly, I would like to submit that, in many cases, agreement and allomorphy can be distinguished on the basis of the trigger of the alternation in question. Agreement is typically restricted to elements with a specific property. This property is usually modelled as to whether a potential goal X has a certain feature or not. With ϕ-agreement for example, this property is often thought to be case marking (following the logic of Bobaljik 2008 and Preminger 2014). ϕ-agreement targets elements which bear a certain case marker such as nominative case. More abstract notions of agreement, where sequence of tense effects or negative concord are modelled in terms of Agree behave similarly. Negative concord items target elements which have a polarity feature and probes on embedded tense heads naturally target heads with tense features.

Allomorphy, I would like to submit, behaves differently. Allomorphy does not target specific features but rather targets specific positions which may be, as we have seen in the case studies of the previous section, linearly or structurally defined. In other words, allomorphy scans a given position for a certain feature whereas agreement targets elements with a specific type of feature for the feature value.

For this reason, we find that allomorphy often shows linear intervention effects with elements that are of completely different types. Recall the intervention effects in Busan Korean where any kind of element – a tense marker, or a evidential or a volitionality marker – could intervene between the predicate and the complementizer. With agreement however, we only ever find relativized intervention effects where an element only intervenes if it is sufficiently similar to the target in terms of the feature that was probed for.

It is clear though that not in all cases we can distinguish between the two types of targets. If a feature such as nominative case is uniquely associated with a position in the tree (e.g. SpecTP), then this diagnostic fails to apply since we will never be able to tell whether it is the position that is targeted or the feature.

3.2 Adjacency

Adjacency is another diagnostic that is frequently used to distinguish between syntactic agreement and allomorphy (or PF-effects, more generally). If the inclusion of adverbs or adverb phrases, focus markers, any kind of parenthesis affects the alternation on a given head, this is often taken to indicate that we are not dealing with syntactic agreement as such but rather that PF plays its role too (possibly in the guise of allomorphy).

That said, linear adjacency requirements alone are not indicative of PF-effects in general. In a given context, the syntax might simply not provide a possible position for an adverb or an adverbial phrase. In the case of Breton for example, the rannig was always immediately adjacent to the element that triggered its form as /a/ or /e/. However, since we do not expect adverbs in between heads and their specifiers, this does not tell us much about whether we are dealing with allomorphy or agreement.

Another question regarding adjacency effects is how strict these effects have to be to serve as a convincing argument for or against an analysis in terms of adjacency. For cases of complementizer agreement, van Koppen (2005) and Haegeman & van Koppen (2012) emphasize that not all dialects of Dutch show the same kind of adjacency requirements. Some dialects such as Tegelen Dutch disallow any kind of material in between the inflected compelementizer and the subject, others allow for example focus markers but disallow adverbs.⁹

Ackema & Neeleman (2003; 2004) who propose that the domain for context-sensitive spell-out is the prosodic domain, argue that some smaller elements are easier to integrate into a prosodic domain that others. If we pursue this line of argumentation for the cases at hand, we can submit that focus particles can be integrated into a prosodic domain including the complementizer and the subject but full adverbial phrases cannot. The issue might be further complicated by the proposal made by Wagner (2005) that prosodic domains are possibly also recursive entities that are built up bottom-up.

So, in clear cases where the intervention of a non-homogeneous set of elements intervene between the trigger and the target of an alternation and lead to a default realization, it is plausible to believe that we are dealing with instances of allomorphy rather than with actual agreement.

3.3 Inventory of alternating forms

Another descriptive difference between agreement and allomorphy seems to be that the latter exhibits a significantly lower inventory of alternating forms. Allomorphs usually show two alternating forms, in some cases three. Agreement alternations often yield complete paradigms referring to different combinations of ϕ-features, which are furthermore relativized to a subset of contexts. Many fusional languages for example employ different sets of agreement forms for different tense or aspect specifications.

The same does not seem to be true for allomorphy. In the Breton case, we saw two alternating rannigs, indicating whether the element in the prefield is of nominal type or not. The same was true for the complementizer alternation in Korean. With the cases of complementizer agreement, possibly a borderline case, we saw that many dialects only show an alternation in marked person number combinations such as second person singular or plural.¹⁰ In the Bavarian case, we find that second singular and second plural have their distinct forms.

It is not per se clear why instances of allomorphy show less distinctions that agreement paradigms. Certainly, nothing in the modelling of these two alternations reflects this. In a standard DM-system, for example, each inflectional affix is provided by a single Vocabulary Item stored independently in the lexicon. In a similar way, each allomorph is stored as a single Vocabulary Item. The only difference is presumably that allomorphs are distinguished by contextual features whereas inflectional affixes are distinguished by the actual feature content they realize. One may capitalize on this distinction by positing that contextual features are harder to store or learn in the first place.

Nonetheless, the distinction might serve as a simple preliminary diagnostic to distinguish the two operations but, as with all the other diagnostics as well, it is certainly not without counterexamples. As van Koppen (2005) et seq. has shown, there is a great amount of variation between different Dutch dialects some of which show more alternating complementizer forms than others.

3.4 Restrictions on types of features

Another point where allomorphy and agreement diverge is the kind of the features that govern the alternations in question. For agreement, we have a relatively good idea what kind of features play a role in determining the agreement affixes. The straightforward case are of course ϕ-features, which typically govern agreement on verbs or nouns (e.g. in the case of possessor agreement). A more abstract notion of agreement may provide for other kinds of agreement to model non-local depedencies in a clause. An Agree-based theory of negative concord for example might assume that polarity features can be exchanged between certain heads in the clause. The same holds, of course, for Agree-based theories of sequence of tense, etc. But crucially, in all of these cases, it is clear why a head in question probes for certain features.

But with allomorphy, the pattern is certainly less clear. The heads that seemingly encode the alternation are usually not (or not necessarily) the ones that are typically associated with the type of features at hand. While we have reasons to assume that an embedded tense head probes for tense features of a higher clause in order to determine its relative value, it is, for example, not clear why the verb in Breton should probe for the syntactic category of its specifier or why the complementizer in Busan Korean should probe for the syntactic category of the predicate of the clause.

The same, I would like to argue, also holds for the case of complementizer agreement in Dutch. Even though there are a number of attested examples of complementizer agreement out there, complementizers in Germanic languages are not typically associated with the features of the subjects of their clauses.¹¹

Thus, I would like to submit that the type of features that govern a given alternation might serve as a diagnostic to determine whether we are dealing with an instance of allomorphy or with an instance of agreement. As already alluded to in course of the case study about Breton, syntactic category features are well-known to govern allomorphy but do not participate in agreement.

3.5 Order of operations

Further, I want to argue that the order of operations that affect the morphological alternation in question can be used as a diagnostic. In the case of the Bavarian complementizer agreement, we saw that ellipsis of the trigger of the alternation rules makes the alternation disappear. This is not the case with standard subject-verb agreement which typically survives non-realization of its argument due to ellipsis. If operations such as ellipsis which are typically thought to apply postsyntactically affect the alternation, then this suggests that the source of the alternation is very likely postsyntactic too. It is clearly beyond the scope of this paper to propose a list of processes which could serve as a diagnostic for the postsyntactic application of rules since each of the processes in question deserves a paper of its own. The case of ellipsis destroying the context for the complementizer alternation in Bavarian is certainly a robust case since ellipsis is generally thought to be a postsyntactic process. Other processes that one might consider are various head-formation processes (head-movement, lowering or Local Dislocation) which are sometimes analyzes as applying postsyntactically. But clearly more work needs to be done to figure out what counts as a postsyntactic process.

3.6 Generalizations about agreement

The final diagnostic that I want to point out in this section concerns the fact that even though agreement seems to have very different properties from language to language, it is often the case that, within a language, agreement behaves very regular, and shows clearly defined properties in terms of direction, targets, etc. Take the case of Breton for example. We have seen that generally agreement with verbs and prepositions only ever surfaces when the trigger of agreeement is morphologically unpronounced. In the Celtic literature, this is sometimes referred to as the complementarity effect. But we also saw that the alternation of the rannig actually presupposes an overt trigger. Thus, if the alternation of the rannig really were due to agreement, it would violate a very strict rule about agreement of the language. It strikes me as empirically preferable if we manage to derive the alternation as resulting from something other than agreement, thereby maintaining the complementarity effect as a strong principle of agreement in Breton.

A similar point is made by Bobaljik & Harley (2017), who argue that verbal root suppletion governed by plural feature on one of the verb’s arguments should not be analyzed as agreement. They cite Durie (1986) who conducted a study on verbal root suppletion and found that the root suppletion is crosslinguistically sensitive to direct objects as well as subjects of intransitives, i.e. root suppletion shows an ergative-absolutive distribution. Crucially this finding is completely independent of the otherwise attested agreement pattern in the language. The related Uto-Aztecan language Huichol for example shows actual ϕ-agreement with subjects and objects, both of which has a nominative-accusative alignment, but still shows ergative-absolutive alignment for suppletion. If both processes were really the result of similar processes, it would be unclear how this difference could be accounted for.

Thus, if agreement is known to be perfectly regular with respect to a certain property but a given alternation violates this regularity, this might indicate that the alternation is actually not due to agreement but rather due to some other process such as allomorphy.