2.1 The *A–B–A Generalization

The first generalization has to do with the restrictions on suppletion in the superlative degree of an adjective. The generalization is given in (2), and in this review article, it will be referred to as the *A–B–A Generalization (or simply as the *A–B–A). Bobaljik (example (1) on page 2) calls it the Comparative–Superlative Generalization I, and credits Ultan (Reference Ultan1972) for a similar observation.

The *A–B–A label (used also by Bobaljik) is related to the fact that the generalization excludes (hence the use of an asterisk) a particular abstract pattern, where the superlative has the same root as the positive (the two As), and the comparative shows a different form (the B in the middle). All other patterns are consistent with the generalization (even if not all are attested, a fact to be discussed in Section 3.1 of the current paper).

Empirically, this means that one can find triplets such as (3a) (A–B–B) or (3b) (A–B–C), but never triplets such as (3c) (A–B–A). The roots are in bold.

Bobaljik’s empirical discussion of such patterns occupies the whole Chapter 4 of his book, and it is not possible to review all the details here. I highlight only several qualifications that Bobaljik makes. These are important due to the fact that there are potential counterexamples to the generalization. However, if certain independent assumptions are adopted, the counterexamples can be explained away.

First, it is necessary to take into consideration locality, a topic I discuss in some detail in Sections 3 and 4 of this paper. Second, Bobaljik argues for a distinction between suppletion (which involves two separate lexical items, e.g.  vs. ) and irregular phonology ( and have the same root, according to Bobaljik). The consequence is that what sometimes looks like an *A–B–A violation is actually the result of an interference caused by these independent constraints (see especially Sections 3.3.3 and 5.2 in the book).

With the generalization established, Bobaljik uses it as a first piece of evidence for the structural decomposition presented in (1) above. Specifically, unlike many alternatives, the structures in (1) are able to explain the *A–B–A Generalization. The reasoning that leads from the structures to the generalization presupposes a realizational view of morphology, and it has two branches, depending on what the ‘realization rule’ for the comparative looks like. The next two sections are devoted to the discussion of these two options.

2.2 Non-terminal lexical entries

Under one possible analysis (which I start with here although it is introduced only in Chapter 5 of the book), the suppletive comparative form is the realization of the whole meaning of the comparative (created by combining the adjective and the Cmpr head). An example of such a rule is in (4b), using the English worse. This entry says that worse is the realization of a non-terminal node containing the two terminal nodes Adj and Cmpr. The realization relation between the node and the phonology is indicated by the arrow (). Such rules are called Vocabulary Items in Distributed Morphology.

The plausibility of an approach along the lines of (4b) is enhanced by the fact that worse (unlike bett-er) actually lacks the comparative marker -er, the regular exponent of Cmpr. Its absence can be understood as a consequence of (4b), which pronounces the whole non-terminal (see McCawley Reference McCawley, Darden, Bailey and Davidson1968, Weerman & Evers-Vermeul Reference Weerman and Evers-Vermeul2002, Neeleman & Szendröi Reference Neeleman and Szendröi2007, Caha Reference Caha2009, Starke Reference Starke, Svenonius, Ramchand, Starke and Taraldsen2009).

The reason why worse out-competes the regular analytical formation *bad-er is because lexicalization at non-terminal nodes is preferred to spelling out individual heads in case both options are made available by the lexical entries of a particular language.

The crucial question in deriving the *A–B–A Generalization is what happens in the superlative given the entries in (4). The underlying structure of the superlative is as in (1c). The first thing to note is that even though Bobaljik relies on the traditional Subset Principle, the entry for worse cannot apply to the whole SprlP, even though it corresponds to its subset, and could be expected to spell it out. The overapplication of (4b) is prevented by the so-called Vocabulary Insertion Principle (example (197)), adopted from work by Radkevich (Reference Radkevich2010). The principle ensures that (4b) cannot pronounce any node larger than the one containing Cmpr and Adj.Footnote ^[2]

The next thing to note is that the ‘comparative’ constituent containing the Adj and the Cmpr is present inside the structure (1c). Worse is inserted at this node, and -st spells out the Sprl head. Crucially, since there is a principle which prevents the derivation of *bad-er in the comparative, the same principle applies in the superlative. That is because the comparative structure is contained inside the superlative, which is the core of the proposal in (1a,b,c). In sum, it is impossible to produce an A–B–A pattern with the entries in (4) (or any other set of entries).

2.3 Contextual specification

However, Bobaljik also observes that while an analysis along the lines of (4) is viable (or even required) for some cases (like worse), it does not work well for all cases. That is because suppletive adjectives are commonly accompanied by regular comparative morphology, as in the case of bett-er. For Bobaljik, this means that there should be a second type of a lexical entry for a suppletive adjective. This second type of entry provides phonology for the Adj node alone, but applies only in contexts where that node is embedded under the Cmpr head. Such a rule is given in (6b).

The rule says that Adj is spelled out as B in a particular context (introduced by the symbol /). The relevant context is such that the locus of exponence (denoted by _) is an immediate daughter of a node (]) that has Cmpr (the trigger) as its sister. However, in the positive, the Cmpr node is absent and (6b) does not apply. The positive form of the adjective thus has a separate entry, which is construed as a context-free rule of insertion at the Adj node (see (6a)).Footnote ^[3]

Note that in the comparative (1b), both rules (6a, b) qualify as contenders for the spell-out of the Adj node. That is because the rule (6a) says that Adj can be spelled out as A regardless of the context, and hence, it is applicable in the comparative structure as well. In cases where more than one rule is applicable, competition arises. The winner is determined by the so-called Elsewhere Condition, going back at least to Kiparsky (Reference Kiparsky, Kiparsky and Anderson1973). Bobaljik (example (12) on page 12) states it as follows:

With (7) in place, we can determine the winner in the comparative. The range of the application of (6b) (only when Cmpr is present) is contained in the context for the application of (6a) (applies in any context), and hence, the rule (6b) wins. The result is that we get A in the positive, and B in the comparative.

The crucial question is what we get in the superlative. Due to (1), the superlative contains the comparative node which triggers root suppletion whenever present (by the reasoning presented above). Thus, the rule introducing B is chosen once again. Hence, when we get A in the positive, and B in the comparative, we cannot fall back on A in the superlative. This way, the *A–B–A Generalization is derived for both ways of construing the comparative entry B.

What Bobaljik shows in addition is that this is not a trivial result. Specifically, if the superlative did not contain the comparative, the generalization would not be accounted for. I skip the detailed reasoning here for reasons of space.

The result (as Bobaljik points out) is a powerful argument for the existence of Universal Grammar. In any single language (such as English), suppletion targets only a handful of adjectives. As a consequence, the overt morphological evidence for any particular learner to posit (1) as the underlying structure is minimal. Yet, the fact that no language has the A–B–A pattern shows that all learners posit such a structure, no matter the scantiness of the evidence.

2.4 Containment

Independently of the *A–B–A Generalization, the proposed containment relations among the adjectival degrees are sometimes reflected in overt morphology, so that the superlative is derived on the basis of the comparative (Section 3.2 of Bobaljik’s book). For instance, in Persian, the comparative is derived by the suffix -tær added to the adjectival root. The superlative is derived from the comparative by an additional affix -in, so that we end up with a morphologically complex superlative adj -tær-in.

At the same time, it is also the case that the reverse containment is never observed in Bobaljik’s sample: There are no comparatives that contain the superlative. Such observations provide an independent support for the proposal in (1). With the structure in place, Bobaljik turns to some additional observations and their theoretical implications.Footnote ^[4]

3.1 The Adjacency Condition and the *A–A–B Generalization

Concerning the number of nodes that may intervene between the locus and the trigger, Bobaljik suggests that one is the maximum. What that means is that in a Vocabulary Item like (8), the Adj root may only be sensitive to the Cmpr head, but not to the Sprl head, which is two nodes away, recall (1c). Bobaljik notes that this is similar to the traditional Adjacency Condition (Siegel Reference Siegel and Stein1978), which was postulated by Siegel in order to explain the difference between ungrammatical sequences of un-dis-, as in *un-dis-honest, and grammatical sequences of un-dis, as in un-dis-heartened. In the former case, un- is structurally adjacent to dis-, which means that in the assumed bracketing [un[dis[honest]]], they are separated only by a single bracket. When they are separeted by a single bracket, ungrammaticality arises due to their incompatibility. However, in the grammatical cases such as un-dis-hearten-ed, the bracketing is different, namely [un[[dis-hearten]ed]], where (minimally) two brackets intervene between un- and dis-. Siegel’s Adjacency Condition encodes the observation that once un- and dis- are separated by two or more brackets, they are no longer sensitive to each other. Similarly, Bobaljik’s proposal says that there must be a local relationship between the locus and the trigger such that no more than a single node may intervene between them.

In the book, Bobaljik initially suggests that the Adjacency Condition is needed in order to capture the generalization in (9), which he calls the Comparative–Superlative Generalization II. In this paper, I will refer to it as the *A–A–B Generalization (or *A–A–B for short).

The label *A–A–B is once again motivated by the abstract pattern which the generalization rules out. A hypothetical example of A–A–B (using English morphemes) is shown in (10) below: The positive and the comparative share the same root, but the superlative root stands out. According to the *A–A–B Generalization, such patterns are never found.

According to Bobaljik’s initial suggestion, such examples may be ruled out by the Adjacency Condition stated over the terminals of the structures in (1) above. Specifically, in (1c), the Sprl head is not adjacent to the root (it is a sister to a node two steps upwards from it). Since it is not adjacent to Adj, it cannot influence suppletion on that terminal, which is (in part) what the *A–A–B Generalization is about.

However, in Section 5.3, Bobaljik notes that while this reasoning successfully rules out any A–A–B pattern, it also (incorrectly) rules out A–B–C patterns, which are attested, as shown in (11).

The example in (11) is problematic because the root opt shows up only in the context of Sprl -imus (which is not adjacent to Adj in (1c)). Hence, if one is to be able to derive the Latin pattern, the specific version of adjacency that counts intervening nodes between the root (as the locus) and Sprl (as the trigger) should be abandoned.

The alternative conception that Bobaljik offers is based on insertion under non-terminal nodes. Once non-terminal insertion is taken into consideration, there are actually two ways to state an adjacency-respecting entry that spells out the Adj node and, at the same time, shows sensitivity to the Sprl node. These are as follows:

In (12a), the entry spells out a non-terminal node with the Adj, the Cmpr and the Sprl nodes inside. In (12b), the entry spells out a node with the Adj and the Cmpr in it, and it is context-dependent on a sister of that node. Both of these solutions allow Sprl to interact with the morpheme that spells out Adj. In effect, the solution gives up on adjacency evaluated on the basis of terminals hosting the trigger (Sprl) and the locus (Adj). The rules in (12) replace this conception by item-relativized adjacency, such that what matters is the structural distance between a node that undergoes spell-out and the features of the trigger.

Coming back to the *A–A–B Generalization, Bobaljik notes that with the new conclusion in hand, the initial explanation for the *A–A–B Generalization becomes unavailable. If any of the entries in (12) is combined with the entry in (13), which only spells out the root, an A–A–B pattern emerges.

Hence, once item-relativized adjacency is adopted (to allow for the A–B–C pattern), we have to conclude that such a notion of adjacency does not – on its own – derive the lack of an A–A–B pattern. What rules it out is a separate principle:

The purpose of this rule is to ensure that when there is an entry that involves the root, the Cmpr and the Sprl (a context-specific rule for the node [Adj Cmpr]), then there must be an entry that involves only [Adj and Cmpr] (this counts as a context-free rule for the same node). And recursively, since there is now a rule for [Adj Cmpr], there must also be a separate entry for Adj. This way, the set of lexical entries comprising (13) and (12a) or (12b) is incomplete, and must be enriched by a special form responsible for spelling out just Adj + Cmpr. In effect, when Sprl has a suppletive form, we end up with three roots and an A–B–C pattern.Footnote ^[5]

However, returning to locality, it seems that potentially, it is the principle in (14) alone that is sufficient here, and a notion of locality – while interesting in its own right – is not needed. To determine whether this is indeed the case, let me turn to an example raised by an anonymous JL referee. Suppose we have the entries (15a, b):

Nothing else assumed (i.e. neither adjacency nor (14)), these entries derive an A–A–B pattern: (15b) provides a special superlative form, and (15a) is an elsewhere form. In addition, it seems that (14) does not help here, because both rules target the same node, namely Adj. Therefore, we must add the number-of-nodes locality into the picture, and that prevents us from having a rule like (15b).

However, this reasoning depends on the proper interpretation of (14). Recall that in terms of (14), the entries (12a, b) must be treated equivalently, namely both are rules that involve CmprP (meaning a node containing Adj and Cmpr), even though (12a) does not target that node for pronunciation but a different one. Thus, the equivalence of the rules (12a, b) is not based on what node they actually pronounce, but on what they ‘involve’, as stated in (14). Bobaljik does not provide a formal definition of ‘involve’, but it seems reasonable to me to proceed along the lines of the Vocabulary Insertion Principle in (5) above: An operation involves a node if that node is contained in a minimal constituent that contains all the features that the rule is specified for. Under this definition of ‘involve’, the rule (15b) involves CmprP because CmprP is contained in SprlP, which is the smallest constituent containing all the features of (15b). Moreover, it is a context-sensitive rule involving CmprP, because it only applies when Sprl is present also. Thus, by (14), the set of rules in (15) is incomplete, and a context-free rule for CmprP must be added. Once added, we generate an A–B–C pattern even without invoking adjacency.

To conclude, in order to derive an A–B–C pattern, we must relativize the Adjacency Condition with respect to nodes that undergo lexicalization. Subsequently, the *A–A–B Generalization must be derived with the help of a principle like (14). Whether we need to add a locality condition in addition to (14) is unclear, as this depends on the technical definition of ‘involve’ in (14).Footnote ^[6]

3.2 The need for Adjacency and Adjacency as Constituency

The current section presents additional observations from the book which show that adjacency is needed also beyond the *A–A–B Generalization; and thus, there is still a need for it in suppletion even if *A–A–B is ultimately derived by some other means.

The evidence comes from two types of examples. The first type corresponds to cases where adjectives are derived by a suffix. Here Bobaljik observes that the derivational suffix blocks potential root suppletion triggered by a comparative morpheme. To illustrate that, he considers the adjective good-ly. Potentially, in the comparative form, one could expect *bett-li-er, which is not found, and good-li-er is the only option. The explanation for this can be easily provided by (the lack of) adjacency.

The second case where locality is apparently involved concerns adverbial formations. The starting point is that comparative adverbs (as in e.g. run faster ) are often transparently derived from the corresponding adjectives and that in such cases, the adverbial marker is usually found outside of the comparative; see (16). In (16), rutto-mpa- corresponds to an oblique stem of the comparative adjective.

The distribution of -h in the adverbs suggests that comparative adverbs are not derived from positive adverbs, but rather from adjectives in the comparative degree:

This structure is relevant for the behavior of adjectival roots that have a suppletive form for their adverbial meaning, such as English well. For such a root, one could potentially expect to find a form such as *well-er (intended to mean: ‘in-a-more-good-way’). Instead, one finds the form bett-er in this meaning. This shows that inside a comparative adverb, there is an asymmetry between the two potential triggers (Cmpr and Adv), which can be understood as related to the underlying structure (17). Since *well-er is blocked here, it seems that locality is the right notion to explain the asymmetry.

An important question is how this type of locality should be stated. Intuitively, what the two examples (*bett-li-er and *well-er) show is that intervening lexical items (-ly and -er, respectively) block the relationship between the locus and the trigger. Bobaljik’s suggestion is that these examples fall under the Adjacency Condition (relativized with respect to the node spelled out by the locus). In neither of these examples are the relevant nodes adjacent. Hence, adjacency correctly rules them out.

Alternatively, one could also propose that suppletion is restricted by constituency: In neither of the examples do the locus and the trigger form a constituent to the exclusion of other material. We could call this the Constituency Condition on allomorphy. Bobaljik does not discuss this explicitly, but the book contains a footnote with a particular formalism for context sensitivity that would make the Constituency Condition follow. Specifically, in footnote 15 of Chapter 5, it is suggested that context-sensitive rules may in fact be a special type of a non-terminal spell-out rule. Bobaljik’s suggestion (for the Latin suppletive comparative mel-ior ‘bett-er’) is as follows:

The rule (somewhat reminiscent of Trommer’s Reference Trommer, Retoré and Stabler1999 proposal) treats context-sensitive Vocabulary Insertion as a type of a rule, which rewrites a particular non-terminal node in the manner specified in the second part, with Cmpr still left to be spelled out. From a theoretical point of view, this is an elegant proposal: If this way of stating context sensitivity is on the right track, it follows from the very format of the allomorphy rule that the process of context-sensitive spell-out is restricted by constituency. This is interesting because the Adjacency Condition does not follow from anything. It is an independent statement whose only purpose is to restrict the generative power of the ‘slash’ rule format. And conversely, the condition is needed only because the formal power of the slash format goes beyond what is attested, as the examples discussed indicate. Bobaljik’s suggestion given in (18) above changes the format, and the restriction follows.

The question (as always) is whether the move from adjacency to constituency can be maintained given additional data. This is not a simple empirical question. The answer depends in part on the extent to which one is willing to re-think issues that are affected by the decision to treat allomorphy as reflecting constituent structure, a task that needs further research.Footnote ^[7] For the time being, I conclude with Bobaljik that there is a role for a locality condition in suppletion, and that this condition requires that the trigger and the node spelled out by the locus are either ‘adjacent’ or ‘form a constituent’.

4.1 The Root Suppletion Generalization and Bobaljik’s account of it

The type-of-node locality says that only head nodes may intervene between the locus and the trigger, whereas phrasal nodes block suppletion, as schematically represented in (19).

The empirical motivation for the proposal is a generalization which says that suppletion is limited to morphological words. Equivalently, one may say that suppletion shows a rather strong tendency not to be triggered by phrase-level degree morphology. The generalization is stated as follows:

Bobaljik follows DM in understanding words (i.e. morphological comparatives) as complex heads. Hence, the proposal (19) (intended to capture the generalization in (20)) says that suppletion is limited to complex heads only. Equivalently, phrasal nodes (XP) may not intervene between the locus and the trigger, which is the exact content of (19) above.

The generalization in (20) rules out hypothetical patterns, where an adjective such as good has a comparative *more bett. The predictive strength of the generalization can be best seen in languages where morphological and periphrastic comparatives may be formed side by side (Bobaljik mentions Georgian, Greek, Romance and Abkhaz). Here follow some relevant examples:

In all such examples, we can see that the morphological (affixal) formations show suppletion, while the periphrastic (non-affixal) forms show no suppletion. Clearly, such regularity should be captured and Bobaljik’s proposal in (19) above does the job.

4.2 Scrutinizing the number-of-nodes locality

There are two important questions I want to address in relation to Bobajik’s proposal. The first question is whether (20) is the only way to state the regularity (and if not, what the alternatives are). This question bears on the status of words in grammar. For instance, Williams (Reference Williams, Ramchand and Reiss2007) defends the Lexicalist Hypothesis by collecting a range of facts which show that the internal composition of words cannot be influenced by the word’s broader syntactic environment, and uses this evidence against the famous dictum of Distributed Morphology – ‘syntax all the way down’. It is therefore slightly paradoxical if Bobaljik’s work, rooted in the spirit of Distributed Morphology, should end up adding to Williams’ list one more property, namely that syntax cannot trigger suppletion inside a word.

The second question is that even if (20) is empirically adequate, it is not clear whether making recourse to complex heads (as in (19)) is the correct way to encode the description theoretically. This is because a strand of research has argued for relaxing the relationship between words and complex heads. For instance, Julien (Reference Julien2002) argues that units which qualify as words by standard tests do not always correspond to complex heads. Koopman (Reference Koopman2005), Muriungi (Reference Muriungi2008) and Kayne (Reference Kayne, Bok-Bennema, Kampers-Manhe and Hollebrandese2010) suggest that traditional words correspond to phrasal objects quite regularly.

In this section, it is my goal to argue that the evidence for suppletion being restricted by wordhood is not very strong, and that the relevant facts can be captured without relying on the notion of word or complex head. Instead, adjacency may be used to explain the facts without entering into the tricky issues as to what is a word, and whether that word corresponds to a head.

Let me begin scrutinizing Bobaljik’s proposal by pointing out an apparent implicit assumption, namely that the head/phrase distinction is the only difference between the analytic and periphrastic forms. However, it seems that this is not always so. In order to show that there are additional differences, I start by putting brackets in the examples for the words for ‘good’ presented in (21) above. The result is in (22).

The way I put the brackets in the Greek and Georgian positive degree is based on the understanding that there is a separate node in the morpho-syntactic representation that hosts agreement features (encoded by -os and - $i$ , respectively). Whether the agreement node is present in the syntax or created later in the mapping to morphology is, strictly speaking, irrelevant for the discussion, as long as the creation of that node precedes Vocabulary Insertion. Starting with the morphological comparatives, the Cmpr morphemes that trigger suppletion (-ter in Greek, and -es in Georgian) attach to the root directly, i.e. inside the agreement markers -os and - $i$ . As a consequence, the trigger and the root are adjacent/form a constituent, and we expect no locality effect here; suppletion takes place. Note that in Georgian, the prefix u- ‘is identical with the versional marker of a third person in finite verbal forms and refers to the object of the comparison’ (Gippert Reference Gippert2000: 37), so it is analyzed as external to the whole inflected form.

In the periphrastic formations, it is a priori not clear what the bracketing should be. However, based on the fact that the morphemes -os and - $i$ do not have the distribution of phrasal affixes in other contexts, there are reasons to parse them as a unit with the root. As a consequence of this morpho-syntactic parsing, the Cmpr markers (pjo and upro) are not adjacent to the root (do not form a constituent with the root to the exclusion of the agreement morphemes). Hence, even without invoking any new locality condition, it is expected that suppletion is blocked in such examples.

Abkhaz shows the adjacency/constituency effect even more clearly. In the periphrastic form, the root ( $\mathit{bz}\unicode[STIX]{x2202}\mathit{ja}$ ) is linearly separated from the suppletion trigger (jejHa ‘more’) by $i$ -, which is a (reduced) relative clause marker. The surface position of the morphemes in the string clearly reveals that $i$ - indeed intervenes between the trigger and the locus. (Note that the final - $w$ /- $u$ in the Abkhaz examples are not comparative markers, they are non-finite predicative elements (be).) If this is the correct parsing, we do not need any new explanation (beyond adjacency/constituency locality) for the fact that suppletion is blocked in such cases.

To sum up, I add a schematic representation of how to explain the core facts using adjacency/constituency as the only relevant locality condition:

The prospect of reducing the type-of-node locality to number-of-nodes locality is theoretically interesting and its implications will be explored in Section 6. Hence, the question is whether an account based on adjacency/constituency can be extended to all the relevant cases where Bobaljik uses the head restriction. Specifically, one wonders about languages where no agreement is found on adjectives. I turn to such an example in the following section. Ultimately, the discussion will lead us to eliminating the type-of-node locality in a general way, no matter whether the language does or does not have agreement.

4.3 RSG: A counterexample from Bulgarian

Consider first one example from Bulgarian, problematic for the RSG, which Bobaljik discusses in Sections 4.3.1 and 4.3.2 of his book. To provide the relevant background, let me mention that in this language, comparatives are formed by the marker po, which precedes the (agreeing) adjective. Thus, from the adjective dobr-a ‘good-fem’ we have the comparative po-dobr-a ‘more-good-fem’.

Despite the fact that po is written as a prefix, it has its own stress and it can modify full phrases, see (24). Bobaljik therefore analyzes po as a phrasal marker.

According to the proposal in the previous section, the phrasal nature of po would lead one to expect that the structure of comparatives is as shown in (23b), i.e. [cmpr [[adj] agr]]. If that is so, we expect no suppletion. According to Bobaljik’s head-locality account, po is not expected to trigger suppletion either. There is no difference in predictions on this point.

This expectation is largely borne out, and Bulgarian indeed has no suppletion with ordinary adjectives. The lack of suppletion with agreeing adjectives is non-trivial; Bulgarian differs in this respect from other Slavic languages with morphological comparatives (where comparative markers are affixes that attach inside agreement).

Importantly, we find one particular case in Bulgarian, given in (25), where po apparently triggers suppletion.

This example is problematic for the RSG, bacause an analytic/phrasal degree marker located outside of the complex head triggers suppletion. In order to make it compatible with the RSG, Bobaljik suggests an account where po-veče ‘more’ has a silent synthetic suffix in addition: po veče-ø. This silent suffix (for which Bobaljik provides some justification by looking at related data from Macedonian) is the true comparative marker, and triggers suppletion. Po is only an obligatory reinforcer according to this explanation.

The alternative account (which relies on adjacency/constituency locality) fares better with this example. Specifically, quantificational expressions such as mnogo ‘many’ or the suppletive po-veče ‘more’ are exceptional in showing no agreement with the noun. As a consequence, there is no intervener between the Adj root and the Cmpr marker in the surface string. If that is so, Adj and Cmpr are adjacent/form a constituent in [po-veče], and suppletion is possible. This is so even though they are separated by a phrasal boundary.

To conclude, minimally some of the facts that serve as a motivation for restricting suppletion to complex heads in Bobaljik’s work can be attributed to the number-of-nodes locality alone. In fact, to the extent that the isolated example from Bulgarian is representative of the underlying principles, it seems that the number-of-nodes type of locality makes better sense of the data so far. However, there are some complications that need to be looked at in more detail.

4.4 The interaction of the RSG with the *A–B–A Generalization

The Bulgarian data seen up to now nicely support the account of RSG based on adjacency/constituency. But presenting the full data set leads to a puzzle for that account:

What we see here is that the superlative is based on the same root as the positive degree, and hence, that Bulgarian (apparently) violates the *A–B–A Generalization.

For Bobaljik, there is no puzzle. As the reader may recall from Section 2.1 above, Bobaljik makes a number of caveats for the range of the *A–B–A Generalization. One of them is that the Generalization holds of triplets that differ only in their degree and all else is kept constant. Specifically, all members of each triplet must be either morphological or periphrastic. If the Bulgarian comparative is really covertly morphological (po veče-ø), as Bobaljik argues, and the superlative is periphrastic (naj mnogo), the two examples differ not only in their degree (i.e. the presence/absence of Sprl) but also with respect to locality (suppletion is blocked in periphrastic expressions). Since Bobaljik independently shows that the distribution of naj is indeed like the distribution of a phrasal modifier, the surface counterexample to the *A–B–A Generalization is explained by recourse to the type-of-node locality in (19) above.

At first blush, the adjacency/constituency alternative I explore here has a problem with the superlative in (26) above. In the preceding section, I have mentioned that the quantifiers such as mnogo, as in (26), are rare in Bulgarian because they have no agreement, and this is what allows suppletion in the comparative to take place to begin with. As a consequence, since there is no agreement, no locality effect is expected in the superlative either, contrary to fact. However, it is still possible to account for this piece of data without invoking the type-of-node locality. In order to show how this works, two observations must be put in place.

Let me start by looking at Tikhvin Karelian data, see (27a), which are similar to Bulgarian in that they present a potential *A–B–A violation. (I will consider the relevance of (27b) shortly.)

The *A–B–A violation in Karelian is not surprising any more, given that the comparative and the superlative differ in their morphological/analytic status. What is to be noted, however, is the contrast with Votic in (27b). Specifically, the difference is whether the superlative is based on the form of the positive (Tikhvin Karelian) or the form of the comparative (Votic). Given the universality of the structures in (1) above, this cannot translate into the presence/absence of Cmpr in the Tikhvin Karelian superlative. Instead the difference must be stated in terms of how the Cmpr head is expressed in the superlative. In Votic, it is expressed by a separate suffixal morpheme and triggers root suppletion. For Tikhvin Karelian, Bobaljik proposes that it is expressed together with Sprl as a portmanteau. If that is so, the Cmpr terminal is expressed inside the phrasal modifier and outside of the complex head where the root resides. For Bobaljik, this automatically means that it triggers no suppletion as a consequence of the type-of-node locality (19).

The lesson we learn from this is that the Cmpr terminal in Tikhvin Karelian is expressed at different places in the comparative (suffixal on the root) and in the superlative (a part of a phrasal pre-modifier). In general terms, Bobaljik’s system allows one and the same terminal node to be found at different places in the comparative and the superlative. I will use this new finding as one part of my solution for cases where suppletion is blocked even in the absence of agreement, as in (26).

5.1 Evidence for morphologically complex comparatives in Czech

The example comes from Czech. For most adjectives, the comparative is derived by the addition of -ějš- to the root. The examples in (28) illustrate this. The final morpheme in each example is agreement, obligatory for Czech adjectives.

Czech descriptive grammars (reflected partly in Bobaljik’s own analysis) conclude that we are looking at a single marker, because with the class of adjectives given above, the adjective always has to include the full string -ějš. However, under scrutiny, it seems that -ějš should be segmented into -ěj (Cmpr1) and -š (Cmpr2, the obligatory reinforcer). The most straightforward indication of this is that the reinforcer is obligatory only in adjectival uses, but it is missing in comparative adverbs, as seen in (29).

It seems implausible that its absence has anything to do with phonology, since the adverbial marker (- $i$ ) is virtually identical to the agreement marker (-í), which triggers no phonological readjustment.

The second indication that we should decompose -ějš is that the Cmpr1 part (-ěj) may also be missing in the comparative of a small class of adjectives, some of them shown in (30), while the reinforcer is preserved.

All suppletive adjectives belong in this class, but also some which are not suppletive (‘old’). Even if non-suppletive, they have to be somehow listed as belonging to this special class, because membership in this class is unpredictable.

Once again, there should be nothing phonological going on here; the examples in (28) and (30) are chosen so that the roots in the Cmpr are similar (e.g. the attested hloup-ěj-š-í ‘more stupid’ is phonologically similar to the potential form *lep-ěj-š-í ‘better’, and červen-ěj-š-í ‘more red’ is similar to *men-ěj-š-í ‘smaller’; in sum, the ungrammatical forms do not seem to be ruled out by an obvious phonotactic restriction).

Taking the bi-morphemic analysis for granted, let us turn to the reasons why none of the markers can be analyzed as ‘phrasal’ specifier of the other. For -ěj, the reason is that the morpheme triggers palatalization, as can be seen for instance on the last row of (28), where $r$ turns to ř. In Czech, palatalization only happens inside units that fulfill the language internal criteria for wordhood (single main stress, inseparability, etc.), and hence it seems unlikely that -ěj would be some sort of a clitic in the Spec of Cmpr.

There are reasons for thinking that also the Cmpr2 marker -š is word-internal. Consonant mutations are again one of those reasons, but there are others. As Scheer (Reference Scheer2001: 14–16) notes, when the Cmpr1 marker is missing, and the Cmpr2 marker -š attaches directly to the root, we see a templatic type of effect: The root is phonologically reduced to a (C)CVC shape, as seen in (31a–c).

The reduced root is in bold, and we can see that vowel shortening in (31a), phonological truncation in (31b), and morphological truncation in (31c) take place in order to get the right shape. The forms bear a single stress and agreement attaches outside of the comparative marker. It is also relevant to note that also in this class, the comparative adverb is formed on the basis of the comparative adjective minus the ‘reinforcer’.Footnote ^[8]

When -ěj is present, no templatic effect is observed in (31d–e). In (31d), the vowel remains long (unlike in (31b)); in (31e), the same piece of morphology that was removed in (31c) remains in its place, undergoing regular palatalization from -ok to -oč. These effects show that -š- requires the root to be in a specific prosodic shape when adjacent/forming a constituent with it, and that this effect is canceled by locality when Cmpr1 -ěj intervenes. The fact that we get such effects, and that -ěj is able to block them by intervention, shows that both -š and -ěj are members of the same unit that the tradition labels the word.

Adopting the conclusion that there are two distinct terminals in the comparative, all the facts fall out of the analysis as follows. Adjectival roots that combine with -ěj-š spell out the Adj node only. The first column in (32) below shows how such roots are used in comparative adjectives and adverbs, illustrating with the roots from (31d–e). Recall that adverbs lack Cmpr2, which is reflected in the analysis by placing Adv over Cmpr1. For the adjectives whose positive degree is morphologically complex (div-ok ‘wild’), I assume that Adj has to be further decomposed.

In the second column, we find the analysis of adjectives that lack -ěj, recall (30). These adjectives have a lexical entry that may spell out a non-terminal node consisting of Adj and When these Vocabulary Items are inserted in the structure, they spell out all the nodes contained inside Cmpr1P, which is indicated in (32) by the rectangular containing all the relevant nodes. In the adjective, this leaves only the higher Cmpr2 without phonological content, which leads to the insertion of the reinforcer -š. Since adverbs lack Cmpr2, the reinforcer disappears in the comparative adverbs exactly as in the regular cases, and we only see the root. It is not clear how to account for these facts without assuming that the comparative can be decomposed into a core part (-ěj) and an obligatory reinforcer (-š), an analysis which Bobaljik independently suggests for Bulgarian.Footnote ^[9]

5.2 Back to Bulgarian

Let me now say how splitting the Cmpr node into Cmpr1 and Cmpr2 helps with the Bulgarian examples in (26) above. Suppose first that the form veče ‘mo-’ is a portmanteau for Adj and Cmpr1. Po is then analyzed in line with Bobaljik’s analysis as an obligatory reinforcer. This is depicted in (33):

The positive root mnogo ‘many’ can then be analyzed as the spell-out for Adj. As a consequence, when no Cmpr is present in the structure, mnogo is used. When Cmpr is present, veče is inserted.

Suppose further that in the superlative, naj spells out alongside Sprl also the two comparative heads, namely Cmpr2 and Cmpr1. In order for that to be possible in Bobaljik’s system, the nodes must form a constituent. In order for that to be the case, there has to be a series of two steps of head movement (or a different operation with the same effect has to apply), which leads to the creation of a constituent that corresponds to naj. The structure and its spell-out are shown in (34).Footnote ^[10]

As a consequence of the fact that naj spells out Cmpr1, what is left for the root to spell out is just Adj. Veče cannot be used, because there is no constituent that contains Adj and Cmpr1 to the exclusion of other material. We know that Adj is spelled out as mnogo, so we derive the form naj-mnogo (not *naj-veče). In effect, we have explained the ungrammaticality of *naj-veče without making any reference to the head restriction in (19).

Note that while achieving this result, we have not added any complexity compared to Bobaljik’s overall proposal. Recall that he too must admit the existence of two separate comparative elements in Bulgarian, and hence the two analyses make use of the same number of underlying formatives. He too admits that the Cmpr head may be expressed at different places as we go from the comparative to the superlative. These two analytical tools taken together have sufficient power to derive the Bulgarian pattern in (26) without – in addition – involving a type-of-node locality implemented along the lines of (19).

On the basis of these considerations, I conclude that many – perhaps all – relevant cases which Bobaljik explains on the basis of the type-of-node locality can be explained without it, using tools that are available independently. If this is so, type-of-node locality can be eliminated and the theory simplified without losing any empirical coverage.