4.1 Phonology

A phonological analysis is made unlikely by the existence of near minimal pairs such as the ones in Table 3.

Here the members of each pair present similar phonological environments, yet behave differently with respect to the choice for truncation or containment in the comparative. Thus the first two roots end in h, whereas the latter two end in t, the final consonant in either case being preceded by a vowel. Yet the former member of each pair shows truncation, the latter containment, suggesting that the distribution of the augment in the comparative is not determined by phonological factors. Admittedly, the quality of the vowel preceding the consonant is different, yet this has generally no influence on the truncation/containment difference, as Tables 1 and 2 show.

As an anonymous reviewer points out, another phonological issue to consider is the following. In principle, it could be the case that the two classes of adjectives do not differ in the presence/absence of -k in the morphosyntactic structure, but only differ in what allomorph of the comparative they require. In particular, the suggestion is that the truncation class requires the comparative marker -š, whereas the containment class selects -ejš, as shown in (15a). In the truncation class, a CVC root is followed by the augment -k and the comparative -š, yielding a triconsonantal cluster Ckš. In the containment class, now characterised by the presence of the -ejš allomorph in the comparative, no such cluster arises. In the case of the truncation class, so the reviewer suggests, this would give rise to a phonologically triggered process of consonant cluster simplification through deletion of k, yielding the surface contrast in (15b).

Under this analysis, the truncation of the positive would be a purely phonological, and not a morphosyntactic, effect.

There are two arguments against this approach. First of all, it is incompatible with the phonological account of the distribution of the -š and -ejš allomorphs, which states that -ejš occurs after clusters, and -š elsewhere (see note 3 above). Suppose now that we abandoned that account in order to allow for (15a), we expect there to be examples of simplex adjectives that end in a consonant cluster, and select the -š allomorph, as shown in (16a). Assuming a process of triconsonantal cluster simplification, we further expect to find this process at work in CCš clusters, as shown in (16b).

However, there are no cases that fit this description, as far as we are aware. While there are many adjectival roots which end in two consonants (e.g., príkr-y ‘steep’, hust-ý ‘dense’, čistý ‘clean’, rýchl-y ‘fast’), these all select the -ejš allomorph. Roots ending in three-consonant clusters also exist (e.g., bystr-ý ‘witty/smart’, ostr-ý ‘sharp’), but none of these feature the consonant cluster simplification that is allegedly at work with the truncating k-adjectives. Finally, we note that there is at least one case in the Slovak National Corpus that features the sequence Ckš, namely jorkšírsky teriér ‘Yorkshire terrier’, a fact which casts further doubt on the existence of the consonant cluster simplification rule supposedly responsible for the truncation pattern.

The second argument against (15) is that Slovak has also an augment with the phonological shape VC, namely -ok. This augment is found with four adjectives, three of which show a truncation pattern, while the fourth shows a containment pattern, as shown in Table 4.

In this case, it becomes difficult to derive the difference between the truncation and the containment class from a phonological truncation rule, since neither class has a triconsonantal cluster created by the addition of the comparative -š after -ok. One would moreover not only have to assume the deletion of -k but also of the vowel preceding it, a kind of deletion which lacks any plausible phonological motivation. Therefore, we do not consider the phonological account sketched in (15) as a plausible alternative to a morphosyntactic account that we pursue here.

For reasons of space, we will not pursue here the analysis of -ok in any detail. The conclusion we want to draw from Table 4 is that the truncation pattern is not a quirk of a single marker, or a single phonological environment; it looks like a structural property of the Slovak grammar that shows at multiple places in the language. Moreover, the pattern is diachronically stable, going back at least a thousand years. Specifically, it has been present already in Old Church Slavonic, the language of the oldest Slavic manuscripts. As Lunt (2001: 78) notes, already back then “[t]he positive [showed] a derivational suffix in -k which is lost in the comparative.” A morphosyntactic truncation analysis has been also adopted in the reference grammar of Slovak (Dvonč et al. 1966: 211).

4.2 Semantics

Let us now turn to the possibility that the distribution of -k (including the distinction between the containment and the truncation pattern) is determined not by phonology, but by the abstract meaning of the adjective. Following this avenue, we find that there is indeed a semantic regularity that unifies all adjectives with -k. In particular, as we shall discuss in more detail below, all adjectives with -k are scalar. What we mean by this is that adjectives with -k are gradable, tend to have antonyms and their interpretation is context-dependent. For example, the adjective ľah-k-ý ‘light,’ seen in Table 3, has a synthetic comparative, it has an antonym (ťaž-k-ý ‘heavy’), and its meaning is context-dependent: a light/heavy handbag has a different comparison class (and consequently a different meaning) than a light/heavy truck. These facts confirm our initial decision to treat -k as a marker that spells out POS and other projections of gradable adjectives.

At the same time, the relationship between scalarity and morphology is imperfect. To show that, we list a representative sample of adjectives with the augment -k in the first column of Table 5. The first thing we want to focus on are their antonyms (horizontally in the same row).

What we see is that most of the time, the antonyms lack the augment despite the fact that they too are gradable, context-dependent and have an antonym. Hence we must conclude that the correlation between -k and scalarity is only a one way implication: -k entails scalarity, but scalarity does not necessarily lead to -k.¹⁰

Let us now turn to the division between truncation and containment, represented in the table by the horizontal line in the middle. The specific question we address is whether scale structure can be made responsible for the bifurcation between the two patterns. It turns out that it cannot. To show that, we have selected the adjectives in the first column from different sub-classes of scalar adjectives. Consider, for instance, the adjectives ‘heavy’ (line 1) and ‘short’ (line 2), both belonging in the truncation class. They represent the so-called open-scale adjectives in the terminology of Kennedy & McNally (2005). Both have antonyms and both are context-dependent (a short giraffe is a different kind of short than a short grasshopper). However, they differ in that ‘heavy’ is associated with the positive extent on the scale, while ‘short’ is a negative adjective (cf. How heavy/#short is he?). This shows that the augment -k is not associated with adjectives of a particular direction (positive or negative), and that both positive and negative adjectives may exhibit the truncation pattern.

A similar state of affairs is found in the containment class (lines 6–10). The adjective kreh-k-ý ‘fragile’ is a negative adjective with an open scale, while brit-k-ý ‘sharp’ is a positive open-scale adjective. The important conclusion is that both positive and negative open-scale adjectives may exhibit or fail to exhibit k in the positive degree, and may belong to either the truncation or the containment class in the comparative.

The adjective vlh-k-ý ‘wet’ on line 4 is a so-called minimum standard adjective (in the terminology by Kennedy & McNally 2005): it suffices that an object (like a chair) has a minimum amount of wetness for it to count as (a) wet (chair). Another distinguishing property is that the comparative entails the positive: if A is wetter than B, it follows that A is wet. This is not the case for the open scale adjectives on line 1 and 2: for example, A is heavier than B does not entail that A is heavy. Despite the difference in scalar structure between ‘wet’ and ‘heavy,’ they both belong in the truncation class.

Moreover, we find minimum standard adjectives (associated to partially closed scales) in the containment class as well. To see that, consider the adjective sliz-k-ý ‘slim-y,’ which behaves semantically like ‘wet’: if A is slimier than B, it entails that A is slimy. The main difference with wet seems to be the viscosity of the material rather than the type of scale. Despite the fact that ‘slimy’ has abstractly the same type of scale as ‘wet,’ they pattern differently in the comparative.

Finally, lines 5 and 10 show two instances of taste adjectives. These adjectives do not have antonyms: being sweet does not, for instance, entail not being bitter (consider caramelised sugar). However, they are context dependent: a sweet beer is a different type of sweet than a sweet cake; a cake with the same degree of sweetness as beer would not count as sweet. The main point is, however, that one of the taste adjectives (‘sweet’) belongs in the truncation class, the other (‘bitter’) belongs in the containment class. So once again, it seems that the type of scale or even the type of an adjective do not, as far as we can see, allow us to distinguish what kind of morphological pattern the relevant adjective will show. The bifurcation between containment and truncation thus seems to be then best understood as an arbitrary property of the relevant root. This fact is reflected in our analysis sketched above, which relies on a lexically arbitrary property of the root, namely its size in the lexicon.

Summarising this section, we have seen that Slovak has complex adjectives in the positive, which are built up according to the template in (17):

Such adjectives either drop the augment -k in the comparative, or they keep it. We have shown that this bifurcation is in large part arbitrary, and cannot be predicted on the basis of phonology or semantics. While we did note that the adjectives with the augment are scalar, their scalar properties do not determine the particular pattern they belong in. It should be further noted that the appearance of an augment with a particular root is also in part arbitrary, simply because not all scalar adjectives have an augment, as becomes clear from observing the antonyms in Table 5.

Our goal in the next section will be to model this behaviour in nanosyntax. In this approach, lexical items are associated to syntactic constituents of various sizes, differing as to what features are contained in their lexical entry. In the next section, we present such an analysis and argue that it leads to interesting predictions for root suppletion, to be discussed in section 7.

6.1 Prerequisites

The analysis we shall present is couched in the framework of nanosyntax, a late-insertion model that has a postsyntactic lexicon, and a syntax where each feature is a syntactic head (Caha 2009; Starke 2009). The model is represented schematically in Figure 1.

Figure 1

Nanosyntactic model of grammar.

The syntax creates syntactically well-formed objects, which are then matched against the lexicon for pronunciation. The lexicon plays the role of a translator between the syntactic representation on the one hand and phonology and concepts on the other. For content words, a lexical item links a syntactically well-formed tree with a well-formed phonological object at PF and a concept at CF (Starke 2014). Functional words only map syntax onto pronunciation at PF, but do not contribute any additional conceptual meaning beyond the grammatical meaning contributed by the set of functional heads the functional morpheme spells out. A model of spellout based on these assumptions leads to a restricted lexicon in the sense that only well-formed syntactic trees (rather than any haphazard set of features) can be linked to a phonology and/or concept inside a lexical entry.

A crucial part of this model is the idea that linking syntactically well-formed objects to phonology is not restricted to terminals. Since phrases containing multiple terminals are also well-formed syntactic objects, they may be easily pronounced by a single piece of phonology (a single morpheme in the traditional sense), provided these terminals form a constituent (i.e., they correspond to a syntactic object). An example is provided in (20) by the entry for an M-root. It will become clear in the course of the discussion that in our model, spellout in fact has to target phrasal nodes and cannot target terminals.

Whether a particular syntactic tree is matched by a lexical item is determined by the Superset Principle (Starke 2009):

As a consequence of the Superset Principle, the M-root entry given in (20) is also applicable as the spellout of a DIRP in (22), since such a DIRP is contained inside the lexically stored tree.

Recall now that roots (which in our system means ‘morphemes associated to conceptual information’) come in various sizes. Therefore, it will sometimes be the case that multiple roots of different sizes match a particular structure. We shall be assuming that in such cases, there is Free Choice, and we may insert the lexical item that we ‘want to talk about’:

This statement is compatible with the idea that normal competition occurs between roots, but their choice is restricted by the concept they are attached to.

To see how this works, consider the lexical items in (24), where (24a) represents an M-root and (24b) an S-root:

If syntax produces a DIRP like the one in (22), then both of these roots are candidates per the Superset Principle (21), because DIRP is contained in both entries. What we do not want to happen is that some version of the Elsewhere Condition forces the insertion of the S-root (e.g., ‘slimy’) on the grounds that it is a ‘perfect match,’ since that would in effect prevent M-roots (e.g., ‘heavy’) from being usable as the spellout of DIRP at all. The problem with this would be that a set of concepts (associated to M-roots) would not be available at all in the given context, an unwanted result. In order not to run into a problem like this, we may simply avoid postulating the Elsewhere Condition for roots. Alternatively, we may say that the Elsewhere Condition does apply to roots, but that the set of competing roots is restricted only to those that are associated to the relevant concept in a way that ‘slimy’ and ‘heavy’ never compete. In either case, (23) will emerge as a consequence of such a setup (i.e., it is not an axiom).

Moving on to additional assumptions, we adopt here the cyclic spellout algorithm described in Starke (2018) (cf. Baunaz & Lander 2018; Caha et al. 2019). The gist of the algorithm is that structure building (Merge F) is intertwined with spellout, such that spellout applies after each application of Merge F. The way spellout applies is by targeting the FP (i.e. the topmost node) that has just been created by Merge F. The algorithm is phrased in such a way that the spellout of FP must actually succeed for Merge F to continue further. The simplest way for spellout to succeed is by finding a matching item for the whole FP. However, if that fails, the structure is rejected at the interface and returned to syntax. In order to save the structure, syntax must perform certain rescue operations, movement in particular. The movements happen in a pre-defined order, first trying to remove the Spec of FP out of FP, and then checking again if this helps with finding a spellout for the FP. The last step is moving the complement. These successive steps in the spellout algorithm are formulated in (25).

It is important to stress that ‘Spell out FP’ in (25a) does not equal ‘Pronounce FP.’ Spellout can be intuitively understood as ‘finding a match’ in the lexicon. When a matching item is found, this is enough for Merge F to proceed, with actual pronunciation postponed.

Keeping this in mind, consider how cyclic derivations proceed. Suppose that syntax constructs FP and spells it out, i.e. it finds a match in the lexicon. The derivation then continues by another step of Merge F, e.g., by merging F2 with FP, producing the constituent [_F2P F2 FP]. At this point, spellout applies again. Suppose that a matching item for the full F2P is found. This will lead to the previous match (at FP) being forgotten, and only the highest match survives. This is known as Cyclic Override:

Note that (26) should not be considered a principle on its own, but rather a consequence of the architecture where spellout cyclically targets higher and higher nodes. Note, however, that roots associated to different concepts cannot override each other, recall (23) (see Caha et al. 2019 for discussion). For example, an XL-root like ‘old’ cannot override an M-root like ‘heavy.’

The final tool in our spellout toolbox is backtracking, which is a last resort operation that is activated when the derivation gets stuck following the algorithm described above. We will say more about backtracking at the relevant place in the discussion.

We shall illustrate the workings of the above mechanisms and principles as we proceed. While doing so, we demonstrate that the correct pairing of roots and augments, as well as their correct ordering, can be derived using the tools described above.

6.2 The positive degree

We start by considering the derivation of the positive degree of an XL-root (e.g., ‘old’). These lack augments both in the positive and in the comparative (recall (4)). An XL-root has a lexical entry like (27a). In the first step, the syntax merges DIRP as in (27b).

Recall now that the way the syntax interfaces with the lexicon is via the Spellout Algorithm, which we repeat in (28).

According to (28), the first thing we should do is try the spellout of DIRP without any movement (28a). Since DIRP is contained in the lexical tree of the XL-root, the XL-root can spell out DIRP. We indicate successful spellout with the >> sign:

In the next step of the derivation, POINT will be merged on top of DIRP, producing POINTP as in (30a). The spellout algorithm is cyclic, and the lexicon will therefore be consulted again in an attempt to spell out the newly formed POINTP. A match will be found in the same XL-root, so that spellout at POINTP is successful, and it will override the earlier spellout at DIRP, as shown in (30b).

The same procedure will be repeated one more time, producing the whole UPP (corresponding to the positive degree), which is still spelled out as the XL-root.

Derivations with augments in the positive work similarly in the initial stages, but they differ in that at some intermediate point in the derivation, the application of Merge F will result in a tree that is no longer contained in the lexical tree of the relevant root, i.e., the syntactic tree will become too big to be realised by the root. As a result, spellout will fail, and rescue strategies will be applied.

Let us see in more detail how this works with an S-root like sliz- ‘slime’ in sliz-k-ý ‘slimy’. An S-root has a lexical entry of the size DIRP, as shown in (24b) above, repeated below for convenience.

On the first step of the derivation, when DIR and DIM are merged, an S-root can spell out this constituent. However, if POINT is merged to DIRP, creating a POINTP shown in (33a), direct spellout fails. This is because the lexical tree of an S-root like sliz- no longer contains the syntactic tree.

Note in addition that the POINT head cannot be spelled out as a a terminal (without movement), because spellout (as per the spellout algorithm) only targets phrases. Since spellout fails, the structure is rejected at the interface and returned to syntax, which has to rescue the structure in accordance with the spellout algorithm. The first rescue strategy that should be tried is the movement of the Spec of the complement of POINT (see (28b)), but since the complement has no Spec, this option is undefined. That leads to complement movement, as mandated by (28c), yielding (33b).

At this point, the lexicon will again be consulted, and it will find a match for POINTP in the augment k, whose lexical entry is given in (34). Recall that the specification of -k has already been introduced informally in (14a).

Observe that the syntactic tree POINTP of (33b) is contained in (34) as a subtree, so that spellout can occur:

As a side effect, the underlying structure has been rearranged in a way that the correct linear order of the root and the suffix is derived.

Note also that spellout movement has simply removed the DIRP out of the POINTP, leaving no trace behind. In this property, spellout movement differs from standard feature-driven movement, which does leave a trace. The reason why spellout movement leaves no trace is the fact that unlike feature-driven movement, it does not lead to two interpretive positions, see Starke (2018) and Caha (2019: Chapter 4) for discussion. This also allows us to keep the spellout procedure matching syntactic trees with lexical ones maximally simple, since the trace does not enter into the computation of matching. We adopt this ‘traceless’ convention here, noting that it is not crucial for the system to work; the alternative would be for spellout to ignore such traces.

The derivation now continues by merging UP to (35), producing (36):

No lexical entry matches the UPP in (36), and movement of the Spec of the complement of UP applies. This moves the DIRP out of UPP, yielding (37). This derives, as the right hand member of the tree, an UPP that is identical to the lexical entry of -k in (34). The marker -k can therefore spell out the UPP in (37), overriding the earlier spellout of POINTP by the same -k.

Let us now turn to the derivation of M-roots, which is the most technically complex. To see why, consider first their lexical entry given in (24a). This entry will initially allow for a derivation where we merge features and spell out without movement, ultimately reaching the stage where POINTP is spelled out by the root, which exhausts its lexicalisation potential; see (38a). Once POINTP is spelled out by the root, the next feature (UP) is merged, yielding (38b).

Since the M-root cannot spell out (38b), rescue movements take place. The complement of UP has no Spec, so the movement of the complement of UP will apply, yielding (39):

But now the right hand part of the tree (39), i.e., the UPP, cannot be realised by the augment, since the UPP of (39) is not a subtree of the augment’s entry (34). In fact, there is no lexical entry in the Slovak lexicon that can apply to this structure, which leads to spellout getting stuck: no rescue operation produces a lexicalisable output, and Merge F cannot continue.

When a derivation gets stuck in this way, backtracking is set in motion. What backtracking does is that it undoes one Merge F operation and goes back to the previous stage in the derivation. Once there, it tries a different derivational option than the one which had previously applied – and ultimately led nowhere. This is formulated in (40):

Reverting to the previous cycle brings us to POINTP, as in (41).

At the first pass through this cycle (38a), we have directly spelled out POINTP by the M-root (the first option of the spellout algorithm (28)). However, this has led to a dead end, so (40) instructs us to try the next option. This would be movement of the Spec of the complement of POINT. However, since this is undefined, we end up with the third option of (28), namely complement movement. This leads to the structure in (42). Here we POINTP can be spelled out by -k, since it is contained in the lexical item (34).

At this point, the derivation of M-roots runs exactly in parallel to the derivation of S-roots, see (35). Just like with S-roots, we merge UP to (42). The structure then looks as in (36), and fails to spell out without movement. We therefore move the DIRP out of UPP, by Spec movement, yielding a structure identical to (37) above, except that DIRP is spelled out by the M-root. The positive degree of an S-root and an M-root therefore end up identical (both have an augment), even though their root sizes are different.

To sum up, this section showed how the derivation of the positive degree proceeds such that only XL-roots lack AUG. M-roots and S-roots need one.

6.3 The comparative

Let us start the discussion of comparatives by making explicit the lexical entry of the comparative morpheme we propose here, see (43).

Let us now turn to the comparatives of XL-roots and M-roots, which have identical derivations. They both recursively Merge F and spell out without movement up to POINTP, see (44a), after which CMPRP is merged, see (44b).

Since the lexicon contains no roots of this size, rescue movement must be applied, and this will move the complement of CMPR to the left, giving rise to (45). The right hand branch of this tree matches with the lexical entry for the comparative suffix, so that spellout is successful, and the correct linear order is derived.

Subsequently, the feature UP is merged to (45), producing (46).

This structure fails to spell out, leading to rescue movements. Spec movement is tried first, producing (47), which spells out successfully.

This is how the problematic truncation pattern is derived: M-roots are not large enough to spell out the positive degree (since they lack UP in their lexical entry), and so they need an augment in the positive. But since the comparative is built by adding CMPR on top of POINTP, and since M-roots can spell out POINTP, no augment is needed in the comparative.

With S-roots, the derivation proceeds exactly as in the positive degree, up to the level of POINTP. This means that rescue movement will have to be applied at the merger of POINT, causing DIRP to be raised across POINT, leading to the structure in (48) (repeated from (35)).

Next, CMPRP will be merged, yielding (49).

(49) does not spell out, so the Spec of its complement (i.e. DIRP) is raised across it, creating (50).

But spellout for CMPRP will fail because no lexical item contains this syntactic tree in the lexically stored tree. We therefore need to undo this movement, revert to (49), and apply the next rescue option of the spellout algorithm, which is movement of the complement of CMPR. This will result in the structure in (51), where CMPR can now spell out successfully as the comparative suffix after the evacuation of its complement.

The final step of the derivation is to add the UP element to the structure (51). This structure fails to be spelled out, and Spec movement is therefore tried. The output of Spec movement is in (52), where UPP correctly spells out as -(ej)š. This leads to the correct result, where the comparative form of S-roots contains the augment. We moreover derive the correct linear order of the suffixes, with the augment appearing between the root and the comparative marker.

This concludes the discussion of the formal derivation of the positive and comparative degrees with the three different root types. We have shown how the derivations unfold on the basis of the different root sizes and the spellout algorithm, deriving the pattern of data discussed in the introduction: with XL-roots there is never any augment, with M-roots there is an augment in the positive but not in the comparative, whereas with S-roots, there is an augment both in the positive and the comparative.

An essential aspect of our analysis are the non-containment/gapped structures we are proposing for the relation between the positive and the comparative. We believe that these are an improvement over any type of approach that would rely on containment structures. It is hard to see how a containment view on the positive and the comparative would be capable of delivering both the truncation pattern and the containment pattern, a result that our approach does achieve.

6.4 The diminutive

Finally, let us briefly address the issue of the derivation of the diminutive forms. Our goal is to show how they are derived for M-/S-roots, and what prediction this makes for the diminutives of XL-roots. Recall first from Tables 1 and 2 that all adjectives with augments pattern alike in that the diminutive marker -un always attaches to the root, and it is always followed by the augment -k. We can capture this pattern by proposing that the diminutive marker spells out an (optional) features DIMIN, which is just below POINT, i.e., on top of DIRP in the case of positive adjectives:

In this structure, we expect that all root sizes (including XL-roots) will maximally spell out DIRP. They cannot spell out the DIMINP, because none of the roots contains the DIMIN feature. Therefore, with all kinds of roots, the DIRP will have to move to the Spec of the DIMIN projection, as in (54). We are assuming that the lexical entry for the diminutive is as in (55).

When POINT is merged on top of (54), we get (56):

There is again no way to spell out without movement. Spec movement (of DIRP) also fails, because the resulting constituent containing POINT and DIMINP cannot be spelled out (there is no lexical entry that matches). Therefore, complement movement produces the structure in (57), where POINTP is spelled out by the augment, and the correct linear order is derived.

This account of diminutives predicts that all adjectives with a DIMIN morpheme must have the augment, including XL-roots. This prediction is borne out, and the diminutives of XL roots must indeed have the augment -k. This is shown in Table 6 for two XL-roots. A diminutive without the augment is ungrammatical.

7.1 *ABA

Let us first introduce the basic facts. As Bobaljik observes, root suppletion in the triplet positive—comparative—superlative is restricted. In particular, if the comparative is suppletive with respect to the positive degree (as in good—bett-er), then the superlative is suppletive as well (i.e., there is no triplet where good—bett-er would be followed by *good-est). Bobaljik states this as in (58).

This generalisation also holds in Slovak, as the following table illustrates. First, in order to get some background on superlatives in Slovak, the topmost line shows a regular adjective where the superlative adds the prefix naj- to the comparative. This is a typical pattern, found with all comparative—superlative pairs in the language.

Against this background, the middle part of the table illustrates the pattern found with suppletive adjectives. We see here first of all that the comparative has a suppletive root compared to the positive. The very same suppletive root is found in the superlative, which, as in the regular case, is still derived from the comparative by the prefix naj-. Such a pattern of root supletion is called the ABB pattern by Bobaljik. The final line in the table makes it clear that there is no adjective in the language that would show the ABA pattern, with one root (A) found in the positive and in the superlative, and a different root (B) in the comparative.

These facts are usually interpreted in terms of structures shown in (60) (as suggested in Bobaljik 2012: 32).

(60)

a.

b.

c.

Given these structures, Bobaljik excludes the ABA as follows. Suppose that we have an adjective with one root in the positive (A) and a different one in the comparative (B). For such an adjective, the comparative root B will be made sensitive to the presence of CMPR. The positive root A (with no particular specification) does not get to surface in the comparative, because it is less specific than the dedicated comparative root B. In this setup, it is impossible to insert the non-specific root A in the superlative, since the structure of the superlative contains the very same element that triggers the presence of the suppletive root B in the comparative (namely CMPR). This makes it impossible to fall back on the default (or elsewhere) root A. The root B will be a better match in the superlative than the nonspecific root A for exactly the same reasons for which it is a better match in the comparative.¹⁴

However, note that the very same reasoning can be replicated for the non-containment structures we have been proposing here, as long as we stick to the hypothesis that the superlative contains the comparative. A simplified version of such a proposal is shown in (61). We are switching back to our original symmetric structures here, noting that they can be interpreted in multiple ways, one of which is that POS and CMPR are two different values of a single head, and another possible interpretation is the ‘gapped’ type of structure we have been investigating in the previous section.

(61)

a.

b.

c.

To see how Bobaljik’s original reasoning applies to these structures, consider it once again. Suppose first that we have a pair of suppletive roots, one in the positive (A) and a different one in the comparative (B). One analysis says that the positive root (A) is the elsewhere case, and that the comparative root (B) is sensitive to CMPR. This derives the ABB pattern, and rules out the ABA pattern, for the same reason as before: the structure of the superlative contains the element CMPR, which triggers the appearance of B. This is so even if CMPR decomposes into CMPR.UP, as we have suggested in the previous section. As long as the superlative contains CMPR and the positive lacks it, the *ABA follows.

It is interesting to note, in addition, that the non-containment structures open some new analytical possibilities. Specifically, if the trees are as in (61), we can consider the root in the comparative (B) to be an elsewhere form, while the root in the positive (A) is triggered by the POS head. Under such a scenario, the POS-specific root A will only appear in the positive, giving (again) rise to an ABB pattern, and not ABA. Under the structures in (61), this is because A is triggered by POS and neither the comparative or the superlative contain POS. Under the ‘gapped’ interpretation considered in (18) above, the positive and the comparative share a common feature (UP) at the top. In order to get an ABB in such a system, A (as the special form) must be triggered due to adjacency to UP. Such adjacency is available in the positive, but fails to obtain in the comparative, as well as the superlative. This makes A unavailable both in the comparative and the superlative. In other words, under both interpretations of the symmetric structures, *ABA follows. In sum, this means that both the scenario in (60) (suggested in Bobaljik’s work) and the one in (61) (which we are exploring here) are equal in their ability to derive the *ABA pattern.

We also wish to note that our proposal does not in fact contradict the core of what Bobaljik proposes. Specifically, he clearly states that what is at stake for *ABA to follow is the containment relation between the comparative and the superlative:

This statement is obviously as much in line with our (symmetric/gapped) proposal of how to be positive (depicted in (61)) as with the Russian-doll containment structures in (60).

In conclusion, we note that there seem to be a number of ways how *ABA patterns may arise in the grammar, as a growing number of works acknowledg (see, e.g., Caha 2017; Andersson 2018; Bobaljik & Sauerland 2018, for how the so-called ‘overlapping’ structures yield *ABA; for an approach to *ABA based on symmetric relation between the first two members like the one suggested here, see Christopoulos & Zompì 2019).

7.2 Suppletion as evidence for backtracking

In this section, we would like to elaborate in more detail on how suppletion is implemented in Nanosyntax, and we would like to use this as an additional evidence for the postulation of backtracking.

To have a specific example to work with, consider, first the suppletive relation between dobr-ý ‘good’ and lep-š-í ‘better.’ The starting point of our discussion is the fact that the positive degree dobr-ý ‘good’ has no augment, and therefore, the root dobr must spell out a structure of the size UPP. This is shown in (63a). The root lep has no augment either, and it therefore spells out the whole constituent that augment-less roots in the comparative spell out, i.e. POINTP. This is shown in (63b).

(63)

a.

b.

The interesting fact that can be observed in these structures is that we do not need context specification to express the distribution of these roots, because they are already differentiated by the size of the structure they spell out. Specifically, we may easily express their difference by storing the root lep ‘bett’ as a lexical item of the size POINTP (an M-root), as in (64a). The root dobr ‘good’ is of a different size, namely UPP (an XL-root). However, dobr differs from standard XL-roots in that it is the positive degree version of lep. An ordinary XL-root like ‘old’ does not stand in a suppletive relationship to another root. The way this special relationship of dobr to lep is expressed in nanosyntax is by using the so-called pointer (Starke 2014; Caha et al. 2019). In our particular case, the entry of the root dobr ‘good’ says that it is the spellout of a structure that includes the head UP as one of its daughters, and where the other daughter corresponds to a structure that has been spelled out by the lexical item lep ‘bett’ at the previous cycle. This is what the lexical entry (64b) says.

(64)

a.

b.

A crucial consequence of this approach is that suppletive lexical items always spell out constituents of different size. It is impossible to have one and the same constituent expressed by two roots with one suppletive for the other.

With the basic understanding of suppletion in place, we want to present a piece of empirical support for the backtracking analysis of the M-roots. Recall that these lack the augment in the comparative, but have it in the positive degree. Under the derivation that we developed earlier, M-roots (which are specified for POINTP) spell out the whole POINTP in the comparative. The relevant structure is given in (65) for convenience, repeated from (47).

This contrasts with the positive. Here (due to Backtracking) M-roots spell out only DIRP. This is shown in the structure (66), repeated from (37).

In the context of the theory of suppletion sketched above, this leads to a prediction. Specifically, the difference in the size of structure spelled out by the root may be reflected in a different shape of the root. Specifically, we allow a situation where an M-sized root spells out POINTP in the comparative, and stands in a suppletive relation to a positive degree root (of the size DIRP), which must combine with an augment.

Slovak provides some evidence that bears out the prediction of the Backtracking derivation. The relevant adjectives is the adjective veľ-k-ý ‘big’, which has an augment in the positive, loses it in the comparative and simultaneously changes to a suppletive root, as shown in (67) (data from Dvonč et al. 1966: 212).

These adjectives thus support our claim that with truncating adjectives, the root spells out constituents of different root sizes in the positive and the comparative, which is a consequence of the backtracking derivation. The lexical items for veľ- ‘big, POS’ and väč- ‘big, CMPR’ are given below for completeness.

(68)

a.

b.

A further prediction that we make is that such suppletion will not arise with S-roots, which have an augment both in the positive and the comparative. These adjectives spell out DIRP in both cases, which makes it impossible to characterise such roots as differing in size. This prediction is also borne out, and we find no suppletion in the containment class of augmented adjectives.