Complementizer agreement is not allomorphy: A reply to Weisser (2019)

Weisser (2019) reanalyzes the Breton rannig, Busan Korean interrogative complementizer alternations, and West-Germanic complementizer agreement as allomorphy instead of agreement, and proposes a set of diagnostics to distinguish allomorphy from agreement. While the cases for Breton and Busan Korean are convincing and the results coherent, West-Germanic complementizer agreement diverges in several respects. In this squib, I review the three case studies and show that the argument for analyzing West-Germanic complementizer agreement as allomorphy does not hold up. In addition, I discuss three new data points on complementizer agreement that cannot be analyzed with a PF account. Accepting that complementizer agreement is not allomorphy allows us to revise the diagnostics to distinguish allomorphy from agreement to a stricter set.


Introduction
Recent years have seen a number of attempts to analyze phenomena that have been regarded as agreement in different terms. For instance, many cases of purported object agreement have been argued to be clitic doubling (Preminger 2009;Nevins 2011;Kramer 2014). In a recent paper, Weisser (2019) proposes to reanalyze the Breton rannig, Busan Korean complementizer alternations, and complementizer agreement in West-Germanic as allomorphy. In addition, based on his reanalysis of these phenomena, he proposes a set of diagnostics to distinguish agreement from allomorphy. In this squib, I show that while Weisser's reanalyses of the Breton rannig and the Busan Korean complementizer alternation in terms of allomorphy are convincing, West-Germanic complementizer agreement differs from these phenomena in several ways, which forces Weisser to weaken his proposed diagnostics. I will present novel and little discussed data on complementizer agreement that is not amenable to an analysis in terms of allomorphy, or in fact any PF account of complementizer agreement. Accepting that complementizer agreement is not allomorphy allows us to formulate a stricter set of diagnostics to distinguish allomorphy from agreement. More generally, this squib contributes to the long-standing debate on the status of complementizer agreement as Agreement (Carstens 2003; van Koppen 2005) or resulting from a PF operation (Ackema & Neeleman 2004;Zwart 2006;Fuß 2014) (see van Koppen 2017 for a recent overview), explicitly arguing against the latter approach.
While Breton is typically analyzed as a VSO language, main clauses are verb second. The first position can be taken by a variety of different constituents. In these verb second main clauses, the verb is preceded by a particle which is called the "rannig". The form of the rannig is a or e, depending on the syntactic category of the element in the preverbal position. Specifically, the rannig is a when the preverbal element is nominal, and e when the preverbal element is not nominal. See (1) for an illustration. The rannig is glossed with r. (1) Anderson ( While there have been Agreement-based analyses of the rannig (Rezac 2004), Weisser (2019) points out that the behaviour of the rannig differs from the behaviour of agreement in a variety of respects. First, the sensitivity to nominal status of the preverbal element is uncommon for agreement in general. Second, φ-agreement in Breton is only visible when the agreement trigger is dropped. The rannig shows exactly the opposite effect, since in order for the rannig to occur, the trigger also needs to be overt. Third, the form of the rannig is determined exclusively based on the element that is in the preverbal position, whereas agreement is determined not by the position of the agreement goal in the clause but rather by other properties, such as case. The final point also entails that the rannig is always linearly adjacent to the preverbal element. In order to account for these facts, Weisser proposes that the form of the rannig is determined by an allomorphy rule that is sensitive to the presence of a nominal category in the preverbal position. That is, a is inserted in the context of a nominal, and e is inserted elsewhere.
Next, we turn to Busan Korean. In the Busan dialect of Korean, the interrogative complementizer shows a four-way distinction between -na, -no, -ka, and -ko, exemplified in (2).
(2) Barrie & Lee (2017: 17- Barrie & Lee (2017) argue that the complementizers in fact consist of two independent alternations. The vowel alternation between -a and -o tracks whether the question is a wh-or polarity question. The consonant alternation between -k-and -n-tracks whether the predicate of the clause is nominal (-k-) of verbal (-n-). This latter alternation is the one of our current interest. While Barrie & Lee (2017) argue for an Agree-based analysis of the complementizer alternation, Weisser (2019) points out that the behaviour of the Busan Korean interrogative complementizer diverges from agreement on several points, as was the case with the Breton rannig. First, again, the alternation is sensitive to the nominal or verbal category of the clause, which is not common for agreement. Second, if an element, such as a tense marker, linearly intervenes between a verbal clause and the complementizer, insertion of -k-is blocked; instead, -n-is used (3). Under a standard theory of Agree, such an intervention effect is unexpected, since there is no reason the tense marker should prevent agreement from being successful.
(3) Barrie & Lee (2017: 19) Ce salam-i Mincwun-i-yess-n/*k-a? that man-nom Mincwun-cop-past-comp 'Was the man Mincwun?' Weisser's analysis of these facts is that the consonant alternation in Busan Korean interrogative complementizers is the result of an allomorphy rule that ensures insertion of -k-under adjacency with a nominal category, and -n-elsewhere, much like the Breton rannig facts.
To summarize, the Breton rannig and the Busan Korean complementizer alternation have two main properties in common. First, both alternations require linear adjacency to a particular feature. Second, the features that trigger the alternations are not typical agreement features; rather, the alternation is triggered by a categorial feature.

West-Germanic complementizer agreement
Let us now turn to complementizer agreement (CA) in West-Germanic. As is well known, in several West-Germanic varieties complementizers seem to inflect for features of the subject. An example is given in (4). In the remainder of this squib, I will refer to the morpheme that attaches to the complementizer as the "CA morpheme", without committing to a particular analysis of this morpheme. In the same vein, I use the term "complementizer agreement" purely as a way to refer to the empirical phenomenon, and not as a means to claim that it should be analysed in terms of Agree.

(4)
Aalten Dutch (van Koppen 2005: 33) a. as ieleu zo losbandig leeft if you.pl so lawless live.pl 'if you live so lawlessly' b. az-e wie sober leeft if-1pl we sober live.pl 'if we live frugally' Weisser (2019) discusses two types of varieties with CA, that pattern rather differently when it comes to disrupting the linear adjacency between the complementizer and the subject. The first type is exemplified by Hellendoorn Dutch. In Hellendoorn Dutch, the CA morpheme is dropped when the linear adjacency between the complementizer and the subject is disrupted (5).
(5) van Koppen (2005: 127, 147) a. darr-e wiej de westrijd wint that-1pl we the game win.1pl 'that we win the game' van Alem: Complementizer agreement is not allomorphy Art. 44, page 4 of 10 b. dat zölfs wiej de westrijd wint that even we the game win.1pl 'that even we win the game' Weisser analyzes these data with an allomorphy rule that says that the complementizer is sensitive to the features of the neighbouring element, which successfully captures the data. Note, though, that van Koppen (2005) provides a fully syntactic analysis of the Hellendoorn Dutch data, showing that an allomorphy analysis is not necessary. I will not have much to say about Hellendoorn Dutch in the remainder of this paper, but anticipating the conclusion that other types of CA are not allomorphy, van Koppen's analysis shows that we can uniformly treat CA as not resulting from allomorphy or other PF operations.
The second variety that Weisser (2019) discusses is Bavarian. In contrast to Hellendoorn Dutch, Bavarian maintains CA when the complementizer and the subject are not adjacent: a. Fuß (2008: 85) obwoi-st du ins Kino ganga bist although-2sg you to-the cinema gone are.2sg 'although you went to the cinema' b. Fuß (2014: 56) dass-sd bei dem Brachdwedda seibsd du in den Biargoadn gehsd that-2sg in this splendid.weather even you in the biergarten go.2sg 'that even you went to the biergarten in this splendid weather' In order to account for this pattern in terms of allomorphy, Weisser proposes that CA in Bavarian is triggered by structural adjacency of the head that hosts the agreement and T. Since T is (assumed to be) head-final in German, many elements can linearly intervene between the complementizer and the subject, without disrupting the local structural relation between the head hosting the CA morpheme and T. 1 This way, the allomorphy analysis of CA can be upheld. Note that Bavarian CA crucially differs from Breton and Busan Korean: CA is not sensitive to linear adjacency, and is triggered by canonical agreement features.

Distinguishing allomorphy from agreement
Based on his three case studies, Weisser (2019) proposes the following diagnostics to distinguish allomorphy from agreement.

(7)
Diagnostics for identifying allomorphy and agreement a. Allomorphy is triggered by the properties of an element in a defined (linear or structural) position, while agreement is triggered by the properties of an element with a defined feature. b. Alternations that are sensitive to linear adjacency are allomorphy. Alternations that are not sensitive to linear adjacency are either allomorphy or agreement. c. A large inventory of alternating forms is indicative of agreement. d. Alternations triggered by non-canonical agreement features are allomorphy.
Alternations triggered by canonical agreement features are either allomorphy or agreement. e. An alternation that is bled by post-syntactic operations is likely to be postsyntactic itself (e.g. allomorphy). f. An alternation that does not obey the regularities of agreement in a given language is likely not to be agreement.
Note that diagnostics (7b) and (7d) are not very strong. (7b) states that if an alternation is sensitive to linear adjacency, it is not agreement, but it cannot say anything about cases in which an alternation is not sensitive to linear adjacency. Similarly, (7d) can be used to exclude an agreement analysis if an alternation is triggered by non-canonical agreement features; if an alternation is triggered by canonical agreement features, it can either be agreement of allomorphy, so in that case the diagnostic is of no use. These two diagnostics are relativized in order to include CA in the allomorphy analysis. Regarding (7b), the idea that structural adjacency between C and (head-final) T is underlying to CA in Bavarian requires that the trigger and target for allomorphy can be linearly non-adjacent. The main piece of evidence for the relevance of T for Bavarian CA comes from Right Node Raising structures. Weisser (2019) gives the example in (8) However, van Koppen (2017: f.n. 39) notes that some speakers of Bavarian have exactly the opposite judgements, i.e. they find (8a) good and (8b) ungrammatical. Under an account that connects the presence of CA to the presence of T, this alternative judgement is inexplicable. In any case, the varying judgements on the Right Node Raising data exclude it as the source of a strong argument in favor of the idea that the trigger for CA is a linearly non-adjacent T.
Regarding (7d), in contrast to Breton and Busan Korean, where the alternation is triggered by categorial features, CA (in both Bavarian and Hellendoorn Dutch) involves features that are run-of-the-mill agreement features, namely person and number. In order to be able to analyze CA as allomorphy, Weisser needs to allow allomorphy also to be sensitive to canonical agreement features. In fact, CA morphemes are identical to verbal agreement morphemes. The allomorphy analysis of CA has to treat this overlap as a coincidence.
Concluding, based on his case studies of the Breton rannig, the Busan Korean complementizer alternation, and West-Germanic CA, Weisser proposes six diagnostics to distinguish agreement from allomorphy. Because West-Germanic CA behaves differently from the other two case studies in certain respects, the diagnostics need to be relativized in order to analyze CA as allomorphy. I have shown that the evidence for these relativizations is not justified, or raises new questions.

Further issues with complementizer agreement as allomorphy
In this section, I present three data points on CA that are problematic for an allomorphy account. 2 The aim is to show that we should not regard CA as allomorphy. All data in this section come from my own elicitations with native speakers.
The first points come from Frisian. Without disrupting the linear adjacency between the complementizer and the subject, Frisian CA looks similar to Bavarian CA. However, the effect of intervention is quite different: in Frisian, disrupting the adjacency relation between the complementizer and the subject leads to ungrammaticality with and without the CA morpheme on C (9) (see de Haan 2010 for comparable data).  that also you vegetarian eat.2sg ('that you, too, eat vegetarian') Note that it is not ungrammatical per se to disrupt the relation between the complementizer and the subject in Frisian, cf. (10).
(10) dat sels ik/Jan komme soe that even I/Jan come will 'that even I/John will come' Furthermore, the ungrammaticality of (9b, 9c) cannot be due to a problem with modifying a 2sg pronoun specifically. Frisian allows for V2 embeddings, which never have CA. In a V2 embedded clause with a 2sg subject, it is fine to modify the subject with a focus particle: (11) dat ek do ytst al fegetarysk that also you eat.2sg already vegetarian 'that you, too, sometimes eat vegetarian' Thus, the ungrammaticality of (9b, 9c) seems to be crucially connected to the presence of CA in the non-modified counterpart (9a). As such, it is inexplicable under any PF account of CA. Under standard assumptions, whether a sentence is grammatical or not is determined in the syntactic part of the derivation. All that PF can do when the derivation is sent there is make alternations to the phonological shape of the utterance, which should not affect grammaticality. Since in (9b, 9c) the ungrammaticality is related to CA, and PF alternations do not affect grammaticality, we must conclude that CA is not created in the PF component of the derivation. This provides strong evidence against a PF account of CA.
The second point from Frisian is based on first conjunct agreement. While CA is obligatory with simple subjects, it is optional with coordinated subjects where the first conjunct is a 2sg pronoun. Having or not having first conjunct CA has effects on interpretation. See (12) for examples. 3,4 (12) a. Ik tink dat do en Jan de wedstriden winne sille. I think that you and Jan the games win will.pl 'I think that you and Jan will win the games.' (collective reading preferred: you and Jan are a team) b. Ik tink dat-st-o en Jan de wedstriden winne sille. I think that-2sg-you and Jan the games win will.pl 'I think that you and Jan will win the games.' (distributive reading preferred: you and Jan are each playing your own games) The first conjunct CA data are problematic in two ways. First, I see no way to posit a CA rule based on linear adjacency that captures the optionality when the subject is part of a coordinated subject and the obligatoriness when it is not. Second, more pressingly, first conjunct CA can only have an interpretative effect if the structural configuration that establishes it is already present in syntax. A PF rule, such as allomorphy, should not be able to affect semantics in this way.
The third, slightly more intricate point comes from novel data from Southern Limburgian. 5 Without disrupting the adjacency between the complementizer and the subject, Southern Limburgian is very similar to both Bavarian and Frisian. The intervention effect in Southern Limburgian is yet different, however: when an element intervenes between the complementizer and the subject, the CA morpheme appears directly to the left of the subject, cf. (13). The size of the intervening material does not matter; in (13c), a topicalized direct object and a focus particle attached to the subject intervene between the complementizer and the subject.
(13) a. dat-s-tich de westrijd geis winne dat-2sg-you the game go.2sg win 'that you are going to win the game' b. dat auch-s-tich wel is vegetarisch uts that also-2sg-you sometimes vegetarian eat.2sg 'that you, too, sometimes eat vegetarian' c. dat zo'n boek allein-s-tich in het openbaar lus that such.a book only-2sg-you in the public read.2sg 'that only you would read such a book in public' Thus, like Bavarian, the Southern Limburgian data illustrate a non-local alternation. However, in contrast to Bavarian, for which Weisser (2019) argued that CA depends on a relation between the agreement head and T, the Southern Limburgian data shows that CA must depend on a relation between (at least) the complementizer and the subject, as in this case it looks like the subject is the alternant, and not the complementizer. Since the complementizer and the subject are both structurally (13c) and linearly (13b, 13c) non-local, this cannot be modelled in an allomorphy account that relies on (structural or linear) adjacency. 6 Summarizing this section, I have presented new data points on CA that are not compatible with PF approaches to CA, such as allomorphy. I conclude that CA is not created in the PF part of the derivation. As we will see in the next section, this actually has some welcome consequences for Weisser (2019)'s diagnostics to distinguish agreement from allomorphy.

(14)
Diagnostics for identifying allomorphy and agreement a. Allomorphy is triggered by the properties of an element in a defined (linear or structural) position, while agreement is triggered by the properties of an element with a defined feature. b. Alternations that are sensitive to linear adjacency are allomorphy. Alternations that are not sensitive to linear adjacency are either allomorphy or agreement. c.
A large inventory of alternating forms is indicative of agreement. d. Alternations triggered by non-canonical agreement features are allomorphy.
Alternations triggered by canonical agreement features are either allomorphy or agreement. e. An alternation that is bled by post-syntactic operations is likely to be postsyntactic itself (e.g. allomorphy). f. An alternation that does not obey the regularities of agreement in a given language is likely not to be agreement.
The first revision I propose has to with adjacency, and indirectly with the trigger for allomorphy (diagnostics (14a) and (14b)). The analysis of CA as allomorphy forces Weisser to assume that allomorphy can also apply under structural adjacency. Eliminating CA as allomorphy therefore allows us to discard this assumption and maintain a very strict locality condition on allomorphy: allomorphy is always triggered by the features of an element that is linearly adjacent to the alternating morpheme. In other words, we can formulate a stronger diagnostic than the initial proposal, which was unable to identify cases of nonlocal alternations as agreement or allomorphy; according to the revised diagnostic, nonlocal alternations cannot be allomorphy. If we accept the linear adjacency condition on allomorphy, we can also restrict the trigger of allomorphy: allomorphy only scans a linear local domain for a trigger, not a structural local domain. This has the additional benefit that structure does not have to be transferred to PF. The second revision that I propose has to do with diagnostic (14d), which states that if an alternation is sensitive to non-canonical agreement features, we are dealing with allomorphy. CA refrained us from saying anything about cases in which canonical agreement features trigger an alternation, since it is triggered by canonical agreement features, but still regarded by Weisser as allomorpy. Given the conclusion that CA is not allomorphy, we can make a stronger claim, namely that alternations triggered by canonical agreement features are not allomorphy. While in order to test this claim it is necessary to look at morpheme alternations in a large number of languages, which is obviously beyond the scope of this squib, I would still like to point out the conceptual appeal of this diagnostic, since it would prevent us from positing different mechanisms (agreement and allomorphy) in different domains (syntax and PF) that lead to highly similar outcomes. For this reason, I include it nevertheless.
The revised diagnostics are summarized in (15). The overall consequence of the revision is that allomorphy and agreement are regarded as completely non-intersective processes, operating in distinct parts of the grammar under distinct constraints. As such, the revisions reinforce Weisser's point that there are multiple types of word-external dependencies in the grammar, which should be distinguished empirically and analytically.

(15)
Revised diagnostics for identifying allomorphy and agreement a. Allomorphy is triggered by the properties of an element in a defined linear position, while agreement is triggered by the properties of an element with a defined feature. b. Allomorphy is sensitive to linear adjacency. Agreement is not sensitive to linear adjacency. c. A large inventory of alternating forms is indicative of agreement. d. Allomorphy is triggered by non-canonical agreement features. Agreement is triggered by canonical agreement features. e. An alternation that is bled by post-syntactic operations is likely to be postsyntactic itself (e.g. allomorphy). f. An alternation that does not obey the regularities of agreement in a given language is likely not to be agreement.

Conclusion
In this squib, I discussed the three cases of purported agreement that Weisser (2019) argues should be analyzed as allomorphy. While for the Breton rannig and the Busan Korean complementizer alternations, the argument is well-grounded and the results coherent, we have seen that West-Germanic CA patterns differently in some respects. In addition, I have given three more data points that are problematic for an allomorphy account of CA (and in fact any PF account of CA), based on novel and underdiscussed data. I concluded that CA is not a PF phenomenon. As it turns out, this is a welcome result, since eliminating CA as allomorphy allows us to maintain a stricter set of diagnostics for distinguishing allomorphy from agreement than Weisser's initial proposal.