2.1 What is (*)ABA?

Bobaljik (2012) provides a typological overview of the morphological patterns found in comparative and superlative forms of adjectives. The specific focus is on suppletive allomorphy, as found in the English paradigm bad, worse, worst. Bobaljik introduces a notation for suppletion based on variables A, B, C, … standing in for stems. In this notation, the stem bad can be called A, while the stem wors(e) can be referred to as the B stem. The paradigm of English bad is then as follows: A, B, B-t. Abstracting away from the endings, the pattern seen in bad, worse, worst is referred to simply as an ABB pattern. Examples of other patterns are given in Table 1. The data showing suppletion come from Bobaljik’s (2012: 106) list of adjectives meaning ‘good,’ with the non-suppletive AAA pattern added to illustrate another logical possibility.

Two out of five logically possible patterns are unattested in Bobaljik’s data, and he proposes that these patterns are ruled out by Universal Grammar.³ In other words, it is argued that in the domain of comparatives, ABA patterns are not only unattested, but unattestable.

Similar ABA-based restrictions have been observed before and since in other areas of morphology (Bobaljik 2015). Examples from Bobaljik & Sauerland (2018) include case (Caha 2009), case and number in pronouns (Smith et al. 2015; 2018), pronoun clusivity (Moskal 2017; in press), spatial deixis (Lander & Haegeman 2016), and Germanic verbs (Wiese 2004; 2008; Starke 2009). Another example comes from work on gender syncretism (Johnston 1997, and the conflicting proposal by Kramer 2015a; b; 2016), but Andersson (2017) argues that there is no evidence for ABA-type restrictions in this domain. Having seen that *ABA appears to be applicable to a wide range of morphological paradigms, we will consider some of the formal explanations that have been proposed for this restriction.

2.2 Why *ABA?

Many explanations have been proposed to account for *ABA. The goal of this section is not to provide a comprehensive overview of each, but instead to draw attention to the range of competing hypotheses in this area. It should also be noted that one’s choice of explanation will not be relevant in the rest of this paper, as the arguments presented in sections 3–6 are not affected by this choice.

Bobaljik (2012) proposes to account for *ABA in terms of feature containment. For adjectives, the featural representation of the positive is a subset of the representation of the comparative, which is in turn a subset of the representation of the superlative.

Assuming the system in (1), deriving an ABA pattern becomes impossible: if we spell out [[ADJECTIVE] COMPARATIVE] with a suppletive stem, that will automatically affect the superlative, which also contains this structure. This assumption of feature containment is also found in work on *ABA restrictions in Nanosyntax (Caha 2009; Starke 2009; for more nanosyntactic work on *ABA, see De Clercq & Vanden Wyngaerd 2017).

Bobaljik & Sauerland (2018) revise the feature containment hypothesis, and propose that *ABA is a subclass of a general pattern of restrictions on morphological systems. They propose a particular system for the notation of feature inventories, and a particular way of partitioning inventories into sets. This is applicable to paradigms of any size, but for three-cell paradigms, ABA patterns cannot be generated in their system. For the exact notation used, the reader is referred to Bobaljik & Sauerland (2018). Caha (2017) builds on aspects of Bobaljik & Sauerland’s (2018) explanation in order to explain a generalization from Blansitt (1988), who observes an ABA restriction in the case sequence dative-allative-locative. Rather than proposing a feature containment analysis [[[DATIVE] ALLATIVE] LOCATIVE], Caha uses a so-called overlapping decomposition of features. In such a system, the leftmost member of a sequence is marked with some feature X, the rightmost one with some feature Y, and the middle one with the overlapping features XY. Spellout of non-terminal nodes is permitted, and used to derive the *ABA restriction. Consider a form M appearing in both contexts X (dative) and Y (locative). We assume XY (allative) has the structure [Y [X]]. M must be specified for both features, and so we may write M ↔ [Y [X]]. But to derive an ABA pattern, we now need a different form N which only appears in the context XY (allative). Again the features XY must obviously be included, and again we would write N ↔ [Y [X]]. But now M and N are identical, so we have failed to derive an ABA pattern, where M appears in one set of contexts (X and Y), and N in another (XY).⁴

Kramer’s work on gender syncretism (see references in 2.1 above) illustrates a third way of deriving *ABA, although she does not explicitly discuss her generalization as an example of a *ABA pattern. Kramer argues that certain masculine-feminine syncretisms are impossible in three-gender languages without natural gender, while masculine-neuter and feminine-neuter syncretisms are possible. Placing the genders in the order masculine-neuter-feminine (or in the reverse order) captures these patterns (compare Johnston 1997), as the non-adjacent masculine and feminine cannot be syncretic. However, Kramer does not propose feature containment, nor an overlapping decomposition. Instead, she adopts the system in (2).

This proposal can be generalized: if an ABA restriction holds when three cells of a paradigm are in the order X-Y-Z, this can be accounted for by leaving Y featurally unmarked, while X and Z differ in specifications for some feature [F]. Two anonymous reviewers point out that Kramer’s proposal is, in a sense, the mirror image of Caha’s (2017) overlapping decomposition. For Caha, the middle B category has a superset of the features of the A categories, while for Kramer, the middle category has a subset of the A categories’ features.

The fourth and final account we will consider is that of Graf (2017), which differs from the above proposals in not referring to features at all. Graf’s proposal relies on transductions between graphs, and attempts to provide a unified account for ABA restrictions and the Person-Case Constraint (PCC; Perlmutter 1971; Kayne 1975; Bonet 1991; 1994; Haspelmath 2004; Anagnostopoulou 2017, among many others). Graf relies crucially on an ordering between cells, rather than relationships between features, and shows that *ABA is impossible to derive assuming certain properties hold of transductions between graphs representing relationships between cells of a morphological paradigm. Specifically, transductions must be weakly non-inverting, and such transductions are computationally simple compared to other types of transductions. Both PCC-violating and *ABA-violating patterns are thus more computationally complex in Graf’s system, meaning that the restrictions follow from simplicity considerations.

2.3 Why does (*)ABA matter?

The explanations in 2.2 are all designed to rule out ABA patterns as impossible. *ABA is then not only a descriptive statement about the languages in a given sample; as Bobaljik points out, “no language can undergo a change that yields a genuine […] [ABA pattern], since no learner could posit a grammar that would accommodate such a pattern” (Bobaljik 2012: 40). Such an approach assumes a tight fit between unattested languages and unattestable ones: if a property is unattested in a large database, it is likely that it is universally prohibited. This is usually stated in terms of avoiding overgeneration, or as capturing all and only the attested patterns.

This type of reasoning stands in contrast to theories which assume that there are many unattested patterns which are nevertheless attestable. The formal linguistic system regularly overgenerates, and the absence of unattested patterns is explained by external factors such as diachrony, acquisition, usage frequencies, processing, etc. In recent years, a number of linguists have begun adopting the idea that such extralinguistic explanations have a more important role to play than previously assumed. Examples include Evolutionary Phonology (Blevins 2004 et seqq.), Substance-Free Phonology (Hale & Reiss 2008; Samuels 2011; Bale & Reiss 2018; Reiss 2018, and others), and work following the so-called three factors approach within Minimalist syntax (see Biberauer & Roberts 2015 for one example, and Chomsky 2005 for the general theory). In morphology, Anderson (1992) is a famous example of this approach (see also Anderson 2016), and recent overviews of morphological theories which make similar assumptions can be found in Haspelmath & Sims (2010), and Blevins (2016). In the spirit of Blevins (2004), we can call this area of research Evolutionary Morphology.

(*)ABA is relevant to this debate as a proposed linguistic universal which has received much attention in recent years (see references in 2.2). Should it receive a formal, linguistic explanation, or an extralinguistic one? This paper provides arguments that the latter approach is preferable for Germanic verbs, thereby contributing to the growing body of work on extralinguistic, and especially diachronic, morphological explanation (Ackerman, Malouf & Blevins 2016; Anderson 2016; Andersson 2017; 2018a). A general account of proposed ABA restrictions is well beyond the scope of this paper. However, it is hoped that an explanation in the domain of the Germanic verb will be able to shed some light on (*)ABA, and add to the theoretical debate concerning extralinguistic explanations of linguistic patterns.

4.1 The third conjugation

The third conjugation is a small class of Swedish verbs, whose stems end in long stressed vowels. The stems are generally monosyllabic, but the addition of prefixes can create polysyllabic verbs. There are no subclasses within the third conjugation, and all verbs of the class are conjugated as in Table 6. The suffixes /-r/, /-t/, and /-dɛ/ are used to form the present, past participle, and preterite respectively. The suffix /-dɛ/ changes the pitch accent of the word from first (not indicated) to second (indicated with a superscript 2). Both /-t/ and /-dɛ/ cause the long vowel of the stem to shorten, and this shortening brings with it changes in vowel quality. Consonants after short stressed vowels lengthen by a phonological rule (Witting 1977; Linell 1978; Andersson 2018b). Andersson (2018b) provides external evidence from L1 acquisition that the vowel shortening is a morphological peculiarity of these and a handful of other suffixes, and the idea that these forms involve shortening is also found within other theories of Swedish quantity (Riad 2014). This means that vowel shortening in the past participle and preterite can be used to diagnose the presence of these suffixes, something which will become relevant in section 4.2. If the underlying form always contains a long vowel, subject to morphological shortening, this also means that third-conjugation verbs generally form an AAA pattern, as shown in Table 6.

Having seen what the third conjugation looks like, it is time to turn to the verb ‘have.’

4.2 ‘have’ in Swedish

In Table 7 we see the forms of the Swedish verb ‘have,’ presented in the surface form and the orthography, together with the underlying forms I will defend later in this section.

As can be seen from Table 7, I propose that ‘have’ should be analyzed as an ABA verb, where the stem /hɑː/ is used for the present and the preterite, while the stem /haf/ is used for the past participle. Note that apart from the short vowel /a/ and the /f/ in the past participle, ‘have’ looks just like other third-conjugation verbs. The preterite triggers vowel shortening and the appearance of the second pitch accent, and all endings are exactly as expected for the third conjugation. However, the ABA analysis is not the only one available. In Table 8 I give seven other possible sets of underlying forms for the data in Table 7, none of which involve an ABA pattern.

Some of these analyses are more plausible than others. For example, analysis 7, where the historical /v/ of this word is synchronically present, is relatively easy to rule out. Deletion of /v/ in the present and preterite is unexpected, especially since there is no such deletion in the participle. There would have to be a lexically-specific rule of /v/ deletion applying to ‘have’ only. But if we can apply lexically-specific rules to the present and the preterite, independently of the past participle, ABA patterns can be derived, and we would be left with no explanation for why they are so rare. Several more arguments against analysis 7 can be found: why is the imperative of ‘have’ [hɑː] rather than the predicted *[hɑːv], when no other consonant-final verb in Swedish shows this pattern? Why is the infinitive of ‘have’ [hɑː] rather than the predicted *[²hɑːva], when no other consonant-final verb in Swedish shows this pattern?

Also implausible are analyses 1 and 2, where the preterite has an underlying short vowel /a/ rather than a long vowel /ɑː/. In other words, it is assumed that the short vowel in this form is entirely unrelated to the fact that it precedes a suffix which triggers vowel shortening for all other verbs of the conjugation. There is no more reason for thinking that ‘have’ is unusual with respect to vowel shortening than there is for thinking that any other third-conjugation verb is unusual. Adopting this analysis for ‘have’ alone is therefore unmotivated.

In analyses 5 and 6, there is an extra formative between the stem and the past participle suffix. Such a formative could be /f/, as seen in the surface form with [f], or /v/ undergoing voicing assimilation (Riad 2014: 102–106). These approaches would have to explain what features are being spelled out by this formative, and argue that its insertion in only this one verb is well-motivated. Even if such arguments could be found, there is a potential locality problem in having the suffix /t/ trigger vowel shortening of the stem across this intervening morpheme (see Embick 2013 and references therein). Phonological vowel shortening in this environment is not predicted by any theory of Swedish phonology (Andersson 2018b).

One way of fixing analyses 5 and 6 is seen in analyses 3 and 4. They treat the /ft/ or /vt/ sequence as a single formative, which could be an allomorph of the /t/ seen in other third-conjugation verbs, or a new suffix altogether. If it is a new suffix, why does it trigger the exact same morpheme-specific process of vowel shortening as the unrelated suffix /t/? If it is an allomorph of the regular third-conjugation participial suffix, we would again have to argue that this allomorphy, limited to this one verb, is well-motivated. The vowel shortening would also have to apply to the V in a VCC sequence. This is only found in two other verb forms in the language, köp-t ‘bought, past participle’ and behöv-t ‘needed, past participle,’ and for both verbs only some speakers allow shortening, and shortening may be register-specific. If the shortening in the participle of ‘have’ represents the same phenomenon, why is it obligatory for all speakers?

Because of the problems with the alternative analyses in Table 8, I propose that the analysis in Table 7, where ‘have’ has an ABA pattern, is preferable. If this analysis is correct, it would mean that ABA patterns cannot be prohibited from appearing in Germanic verbs. This means, in turn, that the reason for the rarity of ABA patterns in this domain cannot be a universal restriction, as assumed by the explanations in section 2.2. Below I show that Swedish is not unique, and that ABA patterns can be found elsewhere in Germanic.

4.3 Further examples of ABA verbs

Swedish is not the only Germanic variety showing ABA patterns in verbs. I give two further cases below, from Low German and Gammalsvenskbymålet, which are not as complicated as the Swedish example in terms of morphophonological alternations. For these languages one usually cites the infinitive rather than the present form in paradigms, a tradition which I follow here. However, this convention is of little import linguistically. Low German generally has the same stem for the present and the infinitive, with the exception that, as in High German, there are sometimes vowel changes in the 2^nd and 3^rd person singular (Thies 2007: §2.1.3.3, §2.1.12; for the vowel changes, see footnote 5). The same pattern of the infinitive and present sharing a stem holds for Gammalsvenskbymålet (Karlgren 1953: 26).

We will begin with Low German; the data come from the dictionary of the Institut für niederdeutsche Sprache (The Institute for the Low German Language; INS-Lex 2.0), where the verb nehmen ‘take’ shows an ABA pattern. The data below are also confirmed by Thies (2007: §2.1.12.2.3). In Table 9, the AAB verb ‘come,’ which is of the same conjugation and therefore shows exactly the same endings, is shown for comparison.

Another case of an ABA pattern is found in Gammalsvenskbymålet, a critically endangered Germanic variety descended from 18^th-century Estonian Swedish and spoken today by a handful of elderly speakers in Zmiivka, Ukraine (see Forsman 2015 for a description of the history of the people and their language). The variety of the last generation of speakers has been documented in a series of articles by Mankov, whose studies of the morphology have been published in several languages (Mankov 2012; 2013 in Russian; Mankov 2015 in English; Mankov 2017 in Swedish). The verb ‘sleep’ is documented by Mankov with an ABA pattern, as shown in Table 10. Again the identically-conjugated ‘come’ is shown for comparison.

The three examples of ABA verbs that we have seen are summarized in Table 11.⁷

It appears that ABA patterns are not prohibited in Germanic verbs. There are relatively few examples, and it appears that the cases which do exist do not have any particular properties in common. In the next section, we will see a historical explanation for the rarity of ABA, and I will also argue that there is a principled diachronic reason for the appearance of ABA patterns in the verbs in Table 11.

6.1 Supporting evidence

We will begin by noting that there are other patterns in Germanic verb systems which receive a natural explanation in terms of Proto-Germanic, and which are not accounted for by a *ABA constraint. These patterns are important as they provide independent evidence for the approach in section 5. The new data concern the types of vowel alternations found in the modern Germanic languages, and their relative type frequencies. A verb like sing will be referred to as having a vowel pattern i-u-a, or /ɪ/-/ʌ/-/æ/, based on the vowels found in sing, sung, sang. Now consider Table 13, showing asymmetries in the type frequencies of different cognate vowel patterns in modern Germanic.

Why are there consistently more verbs with a pattern i-u-a than with a pattern e/ä-u-a? This fact, which bears no obvious relationship to ABA patterns, can be understood with reference to Proto-Germanic. Both of these patterns are the continuations of Proto-Germanic class 3 strong verbs, and this class was split into two by a phonological rule raising *e to *i before tautosyllabic nasals (Ringe 2017: 269). The i-u-a pattern is the outcome of the *i subclass (earlier *e + nasal), while e/ä-u-a is the outcome of the *e subclass (earlier *e + non-nasal). In Proto-Germanic, the *i subclass contained more verbs than the *e subclass (Ringe 2017: 269–270). It seems that the frequency asymmetry between these two subclasses has been faithfully retained into the modern languages over a period of approximately 2,500 years. Faithful retention from Proto-Germanic can be used to explain the rarity of ABA patterns, as well as the frequency asymmetries. Even if one wants to defend a constraint-based analysis of ABA, the frequency facts must be explained somehow. If one uses retention from Proto-Germanic, the need for a constraint *ABA in this domain is significantly reduced: we would be accounting for the same data twice if we use both a synchronic constraint and a diachronic explanation. There is no need to duplicate the history of Germanic languages synchronically in this way, especially since the proposed synchronic constraint is not universal (see sections 4 and 6.2). In other words, since a diachronic explanation appears to be necessary to account for the facts in Table 13, I propose that we also use it for the rarity of ABA patterns. In this way, seemingly unrelated facts about Germanic verbs receive a unified explanation.

6.2 Where ABA patterns are found

In this section I will argue that diachrony can tell us where ABA verbs are more likely to be found, in the rare cases where they have been innovated. Based on the Proto-Germanic forms given in Table 12, some classes will be more likely than others to develop into ABA patterns. Consider classes 4 and 6, repeated for convenience in Table 14 with only the default preterite.

It would be surprising if class 6 developed into an ABA pattern, since this would require the *a of the present changing to *ō, with the past participle *a being completely unaffected. By contrast, in class 4 we see that the present and (default) preterite differ only in length, and that both cells differ from the past participle in terms of height, backness, and rounding. This suggests that if we want to look for further ABA patterns in Germanic verbs, we would be wise to begin our search in class 4.

No Germanic language has undergone the right set of sound changes to neutralize *e and *ē regularly. In East Germanic, *e and *ē are kept distinct, represented as orthographic ai and ē respectively in transcriptions of Gothic. The former can be seen in bairand ‘they bring forth’ (Wulfila Project, The Gothic Bible, Luke 8:15), and the latter in bērun ‘they brought forth’ (Wulfila Project, The Gothic Bible, Luke 18:15). In the remaining branches, North and West Germanic, *ē differs from *e in undergoing lowering to *ā (Antonsen 1965). But despite the absence of merging of the length contrast, there are cases where ABA patterns have been created in class 4 verbs. Both the Low German and Gammalsvenskbymålet examples of ABA verbs are in fact from class 4.¹¹ The data from Gammalsvenskbymålet are of special interest because they represent a case where a subclass of class 4 verbs regularly developed ABA patterns. In the variety of Gammalsvenskbymålet used by speakers born in the late 19^th century, ‘sleep’ showed the ABC pattern in Table 15, with the infinitive and the preterite distinguished by the tenseness of the vowel (Karlgren 1953: 29).¹²

In the modern variety, however, Mankov (2012; 2013; 2015; 2017) finds no /ɔː/ vowel at all, and the data in Table 16 show that this vowel has become tense /oː/. Table 16 shows all forms from Mankov (2015) which Karlgren (1953) recorded with /ɔː/, making it clear that this is a completely exceptionless generalization.

The data in Table 17 show that original /oː/ has remained unchanged. In other words, the change ɔː > oː neutralized the historical tenseness distinction in favor of /oː/.

Given this neutralization, an ABA pattern should be found in all verbs which previously distinguished the infinitive and preterite only by the tenseness in /ɔː/ versus /oː/. Unfortunately, ‘sleep’ is the only verb patterning this way in Karlgren (1953), and in the modern variety, only five verbs of class 4 are known (Mankov 2012; 2013; 2015; 2017). As Gammalsvenskbymålet is moribund and underdocumented, we are unlikely to ever find out whether there are other ABA verbs in the language. Nevertheless, this case shows us that such verbs can arise through regular sound change.

6.3 Beyond Germanic verbs?

The analysis developed so far has only been concerned with Germanic. The general approach of extralinguistic explanation can, however, be extended to family-independent universals (see Ackerman, Malouf & Blevins 2016 for a learning-based explanation of Greenberg’s Universal 34, for example). An obvious extension of the work pursued here would be to consider whether ABA patterns in other domains can receive extralinguistic explanations, whether they are based on diachrony, usage frequency, language acquisition, or something else entirely. But staying with ABA patterns in verbs, I wish to point out another prediction of the explanation in this paper, which has not yet been tested. If there is no universal constraint against ABA verbs, it is predicted that in languages unrelated to Germanic, ABA verbs might be very common, or they might be neither common nor rare. Testing this prediction lies far beyond the scope of the present paper, which has not even attempted a systematic survey of Germanic itself. However, it is relatively easy to find out whether ABA verbs differ in frequency across families, and others who wish to confirm or disconfirm my analysis are encouraged to explore this question further. If ABA verbs are found to be systematically disfavored crosslinguistically, it is clear that the Proto-Germanic explanation would not be sufficient to account for the data. Other explanations to be considered in such a scenario might include asymmetries in usage frequency (Bybee & Hopper 2001; Haspelmath et al. 2014), semantic factors (see Cristofaro 2010 for extralinguistic explanations in semantic typology), or indeed a universal markedness constraint, relativized somehow so as to allow ABA in Germanic.

Two anonymous reviewers point out that it may be difficult to find other languages and families whose verbal systems exactly parallel what we find in Germanic. Here we can distinguish two main approaches for exploring ABA beyond Germanic. One could relax the formal and meaning-based requirements for what constitutes a Germanic-like system, allowing for easier crosslinguistic comparison. This in turn makes the predictions outlined above easier to test. Alternatively, one could maintain that once we carefully consider other verbal systems in detail, there is not really a parallel to Germanic verbs elsewhere. This is not an unreasonable position, but it does seem to call universal *ABA-based theories in this domain into question. If only Germanic exhibits the relevant verbal system, a theory based on UG seems like far too strong a hypothesis, since the only evidence we could ever hope to adduce would be from Germanic. In practice, this option becomes very similar to my approach, which is explicitly family-specific, and which also predicts that systems outside of Germanic may look very different.