2.1 Regular [+ATR] harmony in Assamese

The language has 8 vowels which are given in (3). As can be seen, there is only a single low vowel /ɑ/ but high and mid vowels have a back and front counterpart. In addition, only /u/, /o/, and /e/ have a [–ATR] counterpart. The distribution of mid vowels in Assamese¹ follows best from assuming that there are no [+ATR] mid vowels underlyingly and that [e] and [o] exist only in derived environments.² In fact, we will see many mid vowels [e] and [o] as result of (exceptional) vowel harmony in section 2.2 and 2.3.

This asymmetric vowel inventory can be described with different feature systems and the choice for one or the other crucially influences the theoretical account of the exceptional vowel harmony. To strengthen the claim that a purely phonological account is easily possible for Assamese that does not hinge on specific background assumptions, we will present accounts based on the two possible feature systems in (4) and (5). The first analysis relies on a full specification for all feature dimensions (4) whereas the second one relies on a contrastive feature system where only those features are assumed that are contrastive (5). More concretely, the fully specified system is based on the binary features [high, low, back, round] and [ATR]³ and assumes a specification for each of these features for all vowels. The low vowel /ɑ/ is hence both [–round], [–ATR], and [+back] although there is no [+round], [+ATR], or a [–back] low counterpart. That it is taken to be a back vowel is consistent with the assumption and the phonetic facts in Mahanta (2008) that show a lower F2 frequency for [ɑ] (1547) that is closer to the F2 frequencies of [u, ʊ, o, ɔ] (707-927) than to the ones of [i, e, ɛ] (2418-2642) (Mahanta 2008: 61+62). In contrast, the feature system in (5) only assumes feature values for each vowel that are minimally necessary to contrastively specify it. In addition to not being specified for [ATR], the low vowel /ɑ/ lacks a specification for both [±back] and [±round].

We will return to a discussion of these different feature systems and their different theoretical consequences for an account of Assamese in section 4. For the first part of the analysis in section 3, both are perfectly equivalent since only height and [ATR]-features are relevant for the exceptional raising.

The language employs a pattern of regressive vowel harmony for the feature [+ATR] as can be observed in (6). If a [+ATR] vowel is present in a word, all preceding [–ATR] vowels are realized as their [+ATR] counterpart. This vowel harmony affects stem vowels (6a) as well as affix vowels (6b).⁴

It can be seen in the data in (7) that this harmony system treats [+ATR] as the dominant feature. Here, vowels specified for [–ATR] do not trigger vowel harmony for preceding vowels. This excludes, for example, a [–ATR] harmonic form like *[b^hʊtɛ] for underlying /b^hut–ɛ/. The data also show that the harmony only proceeds leftwards and is therefore regressive: The [+ATR] vowels in (7) do not trigger a change for a following vowel excluding, for example, *[b^hute].

The examples in (6) and (7) all involve morphologically complex contexts but the vowel harmony also shows its effect as a morpheme structure constraint: there are no stems with a non-low [–ATR] vowel and a following [+ATR] vowels (cf. Mahanta (2008: 66f) and Mahanta (2012: 1111). Morphologically simplex stems with a [+ATR] vowel followed by a non-low [–ATR] vowel, on the other hand, are indeed attested (e.g. [xitɔl] ‘cool’ or [ob^hinɔb] ‘new, extraordinary’, Mahanta (2008: 68+82)).

The only low vowel /ɑ/ has no counterpart in the phonemic inventory that differs only in [±ATR] and does not undergo [+ATR] harmony. As can be seen in (8), it is in fact opaque to the harmony process and blocks a following [+ATR] feature from spreading to a preceding vowel. This opaque /ɑ/ can be in a stem (8a) or a suffix (8b).⁵

2.2 Exceptional raising

There are two exceptional suffixes in Assamese that trigger an unexpected behaviour: /–ijɑ/ and /–uwɑ/.⁶ whenever the closest preceding vowel to those two suffixes is the usually opaque /ɑ/, this vowel is unexpectedly raised to a mid vowel and undergoes [+ATR] harmony. As expected, any [–ATR] vowel potentially preceding this /ɑ/ undergoes [+ATR] harmony as well (9b).

Crucially, this unexpected raising can only be observed on an opaque vowel that is directly adjacent to one of the two exceptional suffixes. As can be seen in (10), low vowels that are separated by another (low or non-low) vowel from the exceptional suffix, remain unchanged and undergo neither raising nor [+ATR] harmony.

In contexts without an adjacent opaque low vowel, the two exceptional suffixes behave as every other suffix with a [+ATR] vowel and trigger regular [+ATR] harmony for all preceding [–ATR] vowels as can be seen in the data in (11). This pattern gives us an example of an exceptional trigger: the two suffixes have the arbitrary lexical property of triggering a process (=raising) that is unexpected in this phonological context.⁷

2.3 Exceptional backness harmony

The second level of exceptionality in Assamese vowel harmony reveals itself with a closer look at the exceptionally raised low vowels. As can be seen in (12), those vowels always agree in backness with a preceding mid vowel. In all the contexts in (12), the opaque low vowel /ɑ/ is directly adjacent to one of the exceptional triggering suffixes /–ijɑ/ or /–uwɑ/ and is raised to a mid vowel and specified for [+ATR]. In the words in (12a), this vowel is preceded by back mid vowel and it surfaces as the back mid vowel [o]. In (12b), on the other hand, it is preceded by a front mid vowel and surfaces as the front mid vowel [e] as well. This therefore looks like a progressive vowel harmony pattern for the feature dimension [±back].⁸ We will term this process ‘exceptional backness harmony’ in the following.

This backness harmony is parasitic for height (Cole 1987; Krämer:2003): Two vowels only agree in backness if they are both mid vowels. This is illustrated with the data in (13). A preceding high front vowel does not trigger a change in the [–back] feature of a following derived mid vowel.

Crucially, this progressive backness harmony only targets phonologically derived mid vowels that result from an original low vowel, never for underlyingly mid ones. This can be seen in the data in (14) where underlyingly mid vowels with different values for [±back] surface. The examples in (14a) are contexts where a suffix is added and potentially [+ATR] harmony applies and the examples in (14b) are monomorphemic roots.⁹ The example /kɛwɔl–ijɑ/ (14a) is especially telling since it shows that even a disharmonic sequence of underlying mid vowels preceding one of the exceptional suffixes surfaces faithfully. The application of backness harmony thus crucially relies on vowel raising — only the presence of the exceptional trigger is not sufficient.

There are at least two possible characterization of these facts. One is to describe it as a phonologically Derived Environment Effect (=pDEE): Only vowels that are phonologically changed via raising can undergo fronting. On the other hand, one might argue that low vowels are simply representationally different from mid vowels and are consequently able to undergo more processes. Both these perspectives will be taken up in the two possible reanalyses in section 4.1 and 4.2 respectively.

A final restriction for the exceptional backness harmony is the fact that it is never triggered by prefixes. This is shown with the data in (15) where two mid vowels remain disharmonic for backness although the second vowel is derived from an underlying /ɑ/.

We will not discuss this restriction any further and assume that this invisibility of prefixes is due to the morphological structuring. It is a well-known phenomenon attested in, for example, many Bantu languages that prefixes are often less phonologically integrated and fail to trigger/participate in processes that suffixes and roots easily engage in (Hyman 2008). Note that prefixes still undergo regressive [+ATR] harmony; i.e. the domains for the two harmony processes are different.

2.4 Summary of the empirical facts

The depictions in (16)–(19) summarize the crucial empirical facts about vowel harmony in Assamese. A regressive [+ATR] harmony pattern (16) is blocked by the opaque low vowel /ɑ/ (17). Two exceptional suffixes /–ijɑ/ and /–uwɑ/, however, trigger raising of an adjacent low /ɑ/ to a mid vowel that also undergoes [+ATR] harmony (18). These exceptionally raised mid vowels are also subject to an additional progressive backness harmony (19). Crucially, this latter process only applies between mid vowels if the second one is derived from an underlying low /ɑ/. Whereas the exceptional triggering for raising can therefore be described as a truly morpheme-specific process that only applies in immediate adjacency to certain lexical items, the exceptional undergoing of backness harmony is subtly different since it is restricted to vowels that were underlyingly low and are raised to mid vowels. Whereas all cases of exceptional backness harmony thus imply the presence of the exceptional suffixes since this is the only context for raising, the presence of the two exceptional suffixes does not imply exceptional backness harmony in all contexts. Most notably, an exceptional suffix that precedes underlyingly mid vowels does not trigger backness harmony. The exceptional backness harmony is therefore not an instance of an exceptional triggering for a process but involves exceptional undergoers: Only raised vowels undergo exceptional backness harmony.

In the following section, we present a theoretical account of these patterns that is based on independently motivated purely phonological mechanisms. First, we present a floating features account for the exceptional raising (section 3) and, second, we show that the exceptional backness harmony falls out from either constraint cumulativity (section 4.1) or underspecification (section 4.2), depending on the specific feature system one assumes for Assamese.

3.1 Background assumptions

The following analyses are implemented in the framework of Harmonic Grammar where violable constraints are weighted instead of ranked (=HG; Legendre et al. 1990; Smolensky & Legendre 2006; Potts et al. 2010). This assumption is vital for the proposed gang effect in subsection 4.1 whereas the remaining analyses in 3 and 4.2 could be implemented with the assumption of constraints in a fixed hierarchy (Prince & Smolensky 1993/2004).

The essential difference from classical Optimality Theory (=OT) with ranked constraints (Prince & Smolensky 1993/2004) is that constraints are assigned a numerical weight instead of being ranked with respect to each other. The evaluation of candidates in HG is based on a harmony score (=H) which is the sum of all its weight-violation products. This mechanism is briefly illustrated with the abstract toy example in (20): a constraint Cons_A has a weight of 4, a lower-weighted constraint Cons_B has only a weight of 3, and an even lower-weighted constraint Cons_C has a weight of 2. An imaginable candidate Cand₁ violates Cons_A once and a candidate Cand₂ violates Cons_B and Cons_C. The harmony score for every candidate is calculated by taking all constraint violations times the weight of the respective constraint and summing up all these numbers. Since constraint violations are counted as negative numbers, the candidate with the highest harmony score is the optimal one: (20b) in our abstract example. One important prediction of HG that sets it apart from standard OT can already be seen in the toy example in (20): Multiple violations of lower-weighted constraints can gang up and result in a harmony score that is worse than the one of a candidate violating a higher-weighted constraint. More concretely, although Cand₂ violates only lower-weighted constraints Cons_B and Cons_C in (20), it has a worse harmony score than Cand₁ only violating a higher-weighted constraint Cons_A. The two lower-weighted constraints Cons_B and Cons_C ‘gang up’ against one higher-weighted constraints Cons_A. As is discussed in detail below, exactly this effect can predict pDEE — among other things — and will therefore be crucial in our account of exceptional undergoers in Assamese.

Our constraint set will be largely based on autosegmental representations for features inside Correspondece Theory (McCarthy & Prince 1995) where MAX(F) and DEP(F) preserve feature specifications.¹⁰ Schematic definitions are given in (21).

3.2 Regular [+ATR] harmony

Before we can turn to the analysis of the two exceptionality patterns in Assamese, an account for the regular [+ATR] vowel harmony needs to be established. For the account we present, it is in fact not crucial which constraint types derive such a well-attested system of dominant regressive vowel harmony. The only important fact is that the constraint interaction predicts that a [+ATR] feature spreads regressively and links to multiple vowels. Such a pattern can in principle easily be predicted by a constraint of the ALIGN-type (Kirchner 1993; Akinlabi 1994; Pulleyblank 1996), by a negative SPREAD constraint (Walker 1998; Padgett 2002), a positive SPREAD constraint in the Harmonic Serialism account of Kimper (2011), or by a constraint of the SHARE/AGREE-class (Bakovic 2000; Pulleyblank 2004; Mahanta 2012).

In the following, we adopt a constraint of the latter class since it is easily compatible with the assumption of autosegmental feature structures which is central to our proposal and since it easily predicts harmony which is parasitic on another feature. The specific SHARE constraint we adopt does not only explicitly refer to the linking of autosegmental features (in contrast to AGREE), it also easily allows us to specify a sub-class of participating vowels (McCarthy 2009; Mullin 2011; Zaleska 2018). More concretely, regular [+ATR] harmony is triggered by the SHARE constraint (22a) that demands all pairs of non-low vowels in adjacent syllables to share the same [ATR] feature. Given the vocalic feature specifications for Assamese in (4) and (5), this constraint refers to all mid and high vowels.

In contrast to simple SHARE constraints demanding sharing of a certain feature, this constraint has a context specification. Such a contextual markedness constraint predicts well-attested parasitic harmony patterns (cf., for example, Jurgec (2011; 2013): Only elements that are already similar with respect to some feature(s) strive to also agree in an additional feature value.¹¹ Another relevant markedness constraint is *[–ATR] (22b). Its weight ensures that SHARE(ATR)_–low only triggers a feature spread if this spread reduces markedness by keeping [–ATR] features from being realized. It predicts that [+ATR] is the dominant feature in the Assamese vowel harmony. We follow Bakovic (2000) in assuming that dominant-recessive vowel harmony systems can be construed of as assimilation to the unmarked feature value. Independent evidence on the markedness of [ATR] values in Assamese is inconclusive, going against the crosslinguistic tendencies observed in (Casali 2014) for the languages of Africa. For high vowels, [+ATR] seems to be the unmarked value (cf. the inventory gap for /I/ and the additional restrictions on /ʊ/ (Mahanta 2008: 65)). As for the mid vowels, [–ATR] is less marked, since [+ATR] vowels only occur in derived environments. [–ATR] mid vowels and [+ATR] high vowels freely show up in affixes.

[+ATR] harmony in Assamese is always regressive. This finally follows from the fact that the directionality constraint *SPREAD-R_ATR that prohibits rightwards spreading of ATR features has a very high weight.¹² In the following, we refer to these constraints penalizing spreading in a certain direction as faithfulness constraints. Its counterpart constraint *SPREAD-L_ATR (22d), on the other hand, has a very low weight in Assamese. Since spreading of an [ATR] feature results in deletion of an underlying [ATR] feature, MAX(ATR) (22e) is also relevant and will be violated by all spreading processes. All constraints are given with their assumed weight in Assamese. It has be kept in mind that these numbers are of course only one exemplifying weight assignment and the crucial assumption are the relative weighting arguments of constraints to each other. Those are given below each tableaux in the following.

An example evaluation of how these constraints and their weight predict regular regressive ATR-harmony is given in (23) for a stem with a [–ATR] vowel to which a suffix with a [+ATR] vowel is added. The faithful candidate (23a) violates both SHARE(ATR)_–low and *[–ATR] since the two vowels are associated to different [±ATR] features and one [–ATR] feature is present. The spreading candidate (23b) violates both MAX(ATR) and *SPREAD-L_ATR since it deletes one underlying [–ATR] feature and spreads [+ATR] leftwards to a new vowel. The combined weight of SHARE(ATR)_–low and *[–ATR], however, outweighs the combined weight of these two faithfulness constraints and (23b) emerges as the optimal candidate. Candidate (23c) where [–ATR] spreads progressively is sub-optimal because it violates not only MAX(ATR) and *[–ATR] but also the very high-weighted *SPREAD-R_ATR. The relevant weighting arguments established here are the ones in (24).

In all following tableaux, autosegmental structures are simplified and all relevant features are simply associated to an IPA symbol representing the resulting sound.¹³ To keep the autosegmental structures in the following as readable as possible, we excluded correspondence-theoretic indices but marked all inserted (non-underlying) association lines as dotted lines and all epenthetic material with a grey background. To make the morphological structure clearer, all affix material is given in boldface.

Tableau (23) illustrates the [+ATR] harmony for a mid stem vowel, but it is clear that the same effect is predicted for high vowels. In an underlying form like /gʊl-i/ (cf. (6)), for example, a non-low [–ATR] vowel again precedes a non-low [+ATR] vowel and SHARE(ATR)_–low is violated if both are not linked to the same [ATR] feature.

Note that in (23), both the directionality constraint *SPREAD-R_ATR and *[–ATR] disfavour candidate (23c). Tableau (25) proves the relevance of both constraints and their respective weights independent of each other. In (25), a [+ATR] vowel precedes a [–ATR] vowel, a violation of SHARE(ATR)_–low hence arises (25a). Regressive spreading of [–ATR] (25b) avoids this violation but does not help reducing the markedness of [–ATR] vowels — *[–ATR] is still violated. The crucial weighting argument (26a) thus shows a gang effect: SHARE(ATR)_–low alone is not sufficient to trigger ATR-spreading in Assamese. Only the combined force of SHARE(ATR)_–low and *[–ATR] can justify the faithfulness violations that arise from spreading the feature (cf. the weighting argument (24a)). The second spreading option (25c) avoids both violations of SHARE(ATR)_–low and *[–ATR] but induces a new violation of the very high weighted *SPREAD-R_ATR since spreading is regressive. In a context where only progressive [+ATR] harmony or regressive [–ATR] spreading are possible, a disharmonic structure violating SHARE(ATR)_–low (25a) thus surfaces as optimal.

So far, we have shown how the regressive dominant vowel harmony pattern follows in HG from a simple interaction of two markedness and standard faithfulness constraints. Since *SPREAD-R_ATR has such a high weight and is never violated in any optimal structure of Assamese, we will not consider candidates spreading [±ATR] rightwards in the following tableaux anymore.

Tableau (27) shows that the [+ATR] the analysis matches the data in producing iterative harmony in the output in Assamese and changing all non-low [–ATR] vowels preceding a [+ATR] vowel to [+ATR]. Here, a stem with two [–ATR] vowels precedes a suffix with a [+ATR] vowel. In candidate (27a), three non-low vowels are present, all associated to a [ATR]-feature on their own: Two violations of SHARE(ATR)_–low arise. Since two of the vowels are [–ATR], two violations of *[–ATR] arise as well. Spreading the [+ATR] feature to only one of the vowels as in candidate (27b) already improves the harmony score since one violation of *[–ATR] is avoided. Similarly, spreading of [–ATR] avoids one SHARE(ATR)_–low and one *[–ATR] violations (27c). However, these candidates still violate SHARE(ATR)_–low once since there is still a pair of adjacent non-low vowels with different [±ATR] specifications. Candidate (27d) where [+ATR] harmony proceeds iteratively through the word avoids this fatal violation and has the best harmony score. The relevant weighting argument (28) is identical to the one we established in (24): Every pair of SHARE(ATR)_–low and *[–ATR] violations is worth another pair of MAX(ATR) and *SPREAD-L_ATR violations.

It has to be noted that an additional candidate with association of [–ATR] to all three vowels would also avoid all SHARE(ATR)_–low violations but would induce an additional violation of *[–ATR] and would replace one *SPREAD-L_ATR violation with *SPREAD-R_ATR which has a considerable higher weight (cf. tableau (23)). Spreading of [–ATR] hence always results in a worse harmony score than spreading of [+ATR].

The opacity of the low vowel /ɑ/ with regard to the [+ATR] harmony is a straightforward consequence of the definition of the SHARE(ATR)_–low constraint that only demands parasitic sharing of an [ATR] feature if two vowels have matching [–low] feature specifications (cf. (22a)). The tableau (29) shows a relevant (abbreviated) context where a vowel /ɑ/ is followed by a [+ATR] vowel. Crucially, this configuration does not violate SHARE(ATR)_–low since one of the vowels is [+low] and therefore excluded from the scope of the constraint. The faithful candidate (29a) consequently only violates *[–ATR]. This violation is avoided by spreading the [+ATR] feature in (29b). The violations of MAX(ATR) and *SPREAD-L_ATR induced by this repair, however, induce a worse harmony score than the faithful candidate.

The important weighting argument established here is that only a violation of *[–ATR] is not a good enough reason to induce violations of MAX(ATR) and *SPREAD-L_ATR, only a combination of *[–ATR] and SHARE(ATR)_–low (24). This is yet again an instance of constraint ganging, i.e. a cumulative constraint interaction in HG.

Low vowels thus never undergo [+ATR] harmony in Assamese due to the context specification for SHARE(ATR)_–low. Crucially, this also implies that they will always block [+ATR] harmony for any vowel further right in a word as in, for example /bɛpɑr–i/ (cf. (8)). For one, a [+ATR] feature can not associate to another preceding vowel across an intervening low vowel since association lines may never cross (Goldsmith 1976; Sagey 1986). Second, turning a vowel preceding a low vowel into a [+ATR] vowel as in *[bepɑri] would not help to avoid a SHARE(ATR)_–low violation since the two now [+ATR] vowels are not in adjacent syllables.

Our account also predicts straightforwardly that monomorphemic stems in Assamese mirror the distribution of [±ATR] vowels we find in morphologically complex forms (cf. section 2.1): A stem with a disharmonic sequence of a non-low [–ATR] vowel followed by a [+ATR] vowel is predicted to be neutralized to a harmonic [+ATR] sequence (cf. tableau (23)) whereas a non-low [–ATR] vowel followed by a [+ATR] vowel followed by a non-low [–ATR] vowel is realized faithfully (cf. tableau (25)).

3.3 Exceptional raising and floating features

The exceptional triggers in Assamese are taken to be representationally different from regular non-triggers in Assamese. More concretely, the two exceptional affixes /–ijɑ/ and /–uwɑ/ triggering exceptional raising contain a floating [–low] feature that strives to associate to a preceding segment. Association of the floating feature is ensured by a set of constraints adapted from Wolf (2007). *FLOAT (31b) penalizes any unassociated feature in the ouput whereas MAXFL (31a) penalizes deletion of an underlyingly floating element. Since an underlying [+low] specification is potentially overwritten by this floating feature, the MAX constraint (31c) preserving an underlying [±low] feature is also relevant.¹⁴

If this underlyingly floating feature associates to a preceding vowel, it makes simultaneous realization of an underlying [–low] feature on this vowel impossible: A low vowel is raised and can now undergo regular [+ATR] harmony. This floating feature will only have an effect for a preceding low vowel: Preceding mid and high vowels are underlyingly already specified for [–low] and realization of the floating feature is possible without any surface effect. As is discussed in more detail below, there are in fact multiple autosegmental structures that will result in the same absence of a surface effect for a non-low vowel that precedes the floating [–low] feature.

An example derivation for an exceptional trigger is given in (32). The exceptional suffix /–ijɑ/ contains an additional unassociated feature [–low] that precedes all other phonological structure of this morpheme. Candidate (32a) is excluded because the floating feature is not associated to any vowel under violation of *FLOAT. Candidate (32b), on the other hand, is sub-optimal because the floating feature has been deleted which incurs a violation of the high-weighted constraint MAXFL. Candidate (32c) consequently wins even though the [+low] feature of the underlying /ɑ/ is deleted under violation of MAX(low). This tableau illustrates the high weight of MAXFL and *FLOAT that is crucially higher than the weight of MAX(low): overwriting of underlying features will therefore always be optimal since not realizing the floating feature results in a far worse harmony score.

If the floating feature follows a vowel that is already [–low], four surface-identical options arise: The floating feature could overwrite the preceding [–low] feature (34a), it could associate to the vowel resulting in a redundant double specification (34b), it could remain floating (34c), or it could delete (34d), briefly illustrated in (34).

The choice between these strategies mainly depends on the weighting of the markedness constraint against multiple specifications for the same feature. We assume for now that it has a lower weight than MAXFL, *FLOAT, and MAX(low) which results in redundant double association. However, nothing hinges on this assumption and a weighting resulting in any of the other surface-identical options would also correctly predict the Assamese pattern.

It follows that only an /ɑ/ adjacent to a triggering suffix can be raised since the suffix and its floating feature follow all base material and any reordering is excluded by high-weighted LINEARITY. We hence assume that morphemes are linearized in the input and the constraint LINEARITY penalizing metathesis demands preservation of this order (McCarthy & Prince 1995; Horwood 2002; McCarthy 2003b). LINEARITY constraints exist for all tiers of the phonological representation. The relevant one for Assamese is one preserving the order of height features (35). Tableau (36) shows briefly how its high weight ensures local realization of the floating feature. In candidate (35a), the floating [–low] feature is in its original position where it follows all height features of the stem and can only associate to the stem-final vowel.¹⁵ In candidate (36b), however, it shows up in a linear position before the stem-final height feature and induces a fatal LINEARITY violation.

Since LINEARITY is a violable constraint, languages are predicted where floating features have non-local or infixing effects. Morphological labialization in Chaha is one example where such non-locality is apparently borne out (McCarthy 1983; Akinlabi 1996; Rose 2007; Banksira 2013). The third person object in Chaha is marked by labialization of the rightmost labial or dorsal consonant. It potentially skips several non-labializable coronal consonants. In contrast to the Assamese grammar with a very high-weighted LINEARITY, the markedness constraints against labialized coronal consonants will hence have a higher weight than LINEARITY in Chaha. However, it is apparent that in the typology of featural affixation (Akinlabi 1996; Zoll 1996), edge-realization is by far the most frequent pattern and Chaha is one of few examples for non-local realization of features. A more restrictive theory would assume that phonologically-driven reordering of elements concatenated in the morphology is inherently impossible. Infixation then results from affixation to a certain anchor or pivot position of a base that might potentially be inside the base (Yu 2002; 2007). It has been argued in Zimmermann & Trommer (2013) and Zimmermann (2017) that this restricted theory is empirically correct for affixation of prosodic nodes.

We will leave this issue for future research and assume for now that LINEARITY ensures that the [–low] feature cannot leave its base position. It also cannot dock to a non-final vowel of its base from this underlying position. This follows from an inviolable ban on crossing association lines (Goldsmith 1976; Sagey 1986); a well-motivated principle about well-formed autosegmental associations (Goldsmith 1976). The depiction in (37) illustrates this for an exceptionally triggering suffix that follows a bisyllabic stem with an initial low vowel. Raising of this initial vowel triggered by the floating feature of the exceptional suffix implies that the [–low] associates across an intervening [–low] specification of a vowel (37b).¹⁶

The locality restriction that only a vowel which is adjacent to the exceptional suffix is ever raised in Assamese hence straightforwardly falls out from the high weight of LINEARITY and standard assumptions about autosegmental structures in our account. This point will become important again in section 5 where we emphasize that this locality restriction follows from an additional assumption specific to locality in the alternative account based on lexically indexed constraints.

We propose a representational solution to the exceptional raising in Assamese: The two suffixes are exceptional triggers for raising a low vowel since their underlying representation is different from non-triggering suffixes and contains additional phonological material that needs to be realized. Such an account is reminiscent of the many arguments that morphological feature mutation is due to floating or unassociated phonological material (Lieber 1992; Zoll 1996; Akinlabi 1996; Wolf 2007). This makes this account purely phonological in the sense that the phonological grammar or constraint ranking makes no specific reference to these exceptional morphemes: The constraint inventory refers solely to phonological information.

A reviewer points out that under this assumption it is entirely accidental that the suffixes triggering exceptional raising are of a rather similar shape. Both end in a sequence of a glide and /ɑ/ and — more importantly — both start with a high vowel. One might therefore be tempted to analyse this pattern as an exceptional raising harmony triggered by those high vowels. In contrast, we are convinced that the floating feature account is warranted in general and hence reasonable for Assamese if one considers the broader typological picture. There does not seem to be a general correlation between the segmental content of an affix and its potential to trigger an exceptional feature change (cf., for example, Lieber (1987) or Gleim et al. (in progress). In the closely related language Bengali, for instance, the perfect participle suffix -/e/ exceptionally raises mid vowels (e.g. in [∫on-a] ‘to hear’) to high vowels (cf. [∫un-e] ‘heard’ (David 2015)), even though it does not contain a high vowel itself. Furthermore, this exceptional raising can be triggered by consonantal suffixes that do not include any vowel at all. A prime example is the future tense suffix -/b/ that derives the third person future tense form [∫un-b-e] from the third person present tense form [∫on-e].

The more general typological picture also confirms this picture. A quick survey of exceptional vowel raising/lowering in thirteen language (excluding Assamese) reveals that in the majority of nine languages the triggering affix does not necessarily contain a vowel specified for the feature value that the process triggers on a neighbouring vowel. The triggering suffix can also include only vowels specified for the opposite feature value. This is the case in three of these languages which can be classified as morpheme specific dissimilations. In only one language, it is generally true that the triggering affix includes a vowel with the feature value it triggers on a neighbouring vowel.¹⁷ Even though the sample size is very small, it confirms the claim that morphologically conditioned exceptional height changes can be idiosyncratic with respect to the vowel specifications of the triggering affix. Since a mechanism like featural affixes is needed for those languages in any case, it is only natural to extend this proposal to cases like Assamese.

4.1 Fully specified feature system: A gang effect

The first possible feature system describing the Assamese vowels we introduced in 2.1 is repeated in (38): It assumes a value for each of the 5 features for each vowel. It is a substantive feature system that assumes a full specification for each articulatorily defined feature (Kiparsky 1965; Chomsky & Halle 1968).

Under this feature specification, if the [+low] value of /ɑ/ is overwritten with [–low], a mid non-round back vowel [ɤ] would result which is illicit in Assamese (cf. (3)). Raising of a mid vowel thus implies a change of either the [±back] specification resulting in [e] or the [±round] specification resulting in [o], illustrated in (39). We argue in the following that the latter realization as [o] is the default strategy but that the alternative realization as [e] follows if additional backness harmony with a preceding mid vowel can be achieved.

If a missing feature [+round] or [–back] must be provided, there are two basic strategies: Epenthesis under violation of DEP or spreading under violation of *SPREAD-R/L_F. Additionally, both these operations violate MAX(F) since an underlying feature specification is overwritten. The full set of these three relevant faithfulness constraints for the features [±round] and [±back] are given in (40) together with their weight in our account of Assamese.¹⁸ Both these strategies imply of course an additional violation of MAX constraints (40f+g). It is shown in the following that Assamese in fact employs different strategies for providing different features: New specifications of [±round] can only be provided via epenthesis whereas new specifications of [±back] can only be provided by rightwards spreading. This asymmetry is already apparent in the constraint weights: Whereas DEP(round) has a lower weight than *SPREAD-R_round, *Spread-R_back has a lower weight than DEP(back).

A first illustration of the default state for [o] as the default after vowel raising is illustrated in tableau (41) where the context (32) of a monosyllabic stem with the low vowel /ɑ/ preceding an exceptional triggering suffix is shown again. Since no vowel precedes the raised vowel, rightwards spreading of either [–back] or [+round] is trivially not an option in such a case and the decision for which mid vowel is realized is made by the weightings of MAX(back) and DEP(back) as well as MAX(round) and DEP(round) respectively. Realizing the newly created mid vowel as [o] in candidate (41b) implies violations of MAX(round) and DEP(round) and realization as [e] in (41c) the mirror violations of MAX(back) and DEP(back). Since the former constraints have a lower weight sum than the latter two, (41b) wins the competition. The crucial weighting argument that insertion of an epenthetic [±round] is less costly than insertion of an epenthetic [±back] is given in (42). Note that we omit the feature specifications of final low vowels in the following to ease readability since those vowels are not expected to trigger or undergo any processes in our analysis.

Given that epenthesis for [–back] and spreading of [+round] are penalized by constraints with such a high weight, we will disregard these options in the following tableaux. We will also exclude the candidate (42a) that creates the illicit vowel [ɤ] — it should be clear from the high weight of *ɤ that this is always a sub-optimal option and only [o] or [e] can result from raising.

An insightful context that illustrates whether [o] or [e] surface as default vowel for a raised mid vowel is (43) where a high front vowel precedes a low vowel. It can be seen that epenthesis of [+round] resulting in [o] (43a) is indeed the default strategy to form new mid vowels that is preferred over spreading of [–back] (43b). This default state of [o] is an argument that the combined relative weightings of MAX(back) and *SPREAD-R_back are higher than the ones of MAX(round) and DEP(round), cf. (44).

We argue that spreading of [–back] to provide a licit new mid vowel is only possible in a single context in Assamese, namely when additional backness harmony with a preceding mid vowel can be achieved for a low vowel that is raised to a mid vowel.

The one context where spreading of [–back] applies to ensure backness harmony is in fact the pDEE we identified in 2.2: Only a low vowel that is raised to a mid vowel and follows a mid vowel with another [±back] specification undergoes backness harmony. This additional backness harmony as pDEE follows in our HG account from constraint cumulativity when a very low-weighted constraint demanding additional backness harmony gangs up with the faithfulness constraints MAX(round) and DEP(round) that are violated by the default strategy to provide a new mid vowel. A structure that is disharmonic with respect to backness in mid vowels is thus expected in most contexts, namely in all contexts where a low vowel is not raised to a mid one.

The relevant constraint favoring parasitic backness harmony is a SHARE constraint that requires mid vowels to agree in backness SHARE(back)_–lo,–hi (45). As we can see in the tableaux (46), its weight is relatively low and it consequently never triggers backness harmony for underlying mid vowels. More concretely, the combined respective weights of MAX(back) + *SPREAD-R_back and MAX(round) + DEP(round) are higher than the weight of SHARE(back)_–lo,–hi. All these four faithfulness constraints are violated if backness spreading applies (46b) since in addition to the change of [–back] into [+back], the vowel also needs to change its [+round] into [–round] to form a licit vowel [e]. As was established above, this additional change can only happen via epenthesis and implies a violation of MAX(round) and DEP(round). The candidate (46a) that violates SHARE(back)_–lo,–hi is therefore the optimal one.

In a context where an underlyingly low vowel is exceptionally raised to a mid vowel, however, the situation is crucially different. As was shown in (41), the expected default mid vowel for the exceptionally raised low vowel is the back round vowel [o] that results from epenthesis of [+round] and implies a violation of MAX(round) and DEP(round). If the raised mid vowel is preceded by a front mid vowel, however, this default strategy implies an additional violation of SHARE(back)_–lo,–hi. This violation was avoided in all previous tableaux since the raised mid vowel was either not preceded by a vowel (cf. (41)) or preceded by a high front vowel (cf. (43)). The tableau (48) now adds the crucial final context of a raised mid vowel preceded by a mid front vowel. Candidate (48a) that undergoes the default strategy of [+round] insertion has two mid vowels with different backness specifications [e] and [o] and violates SHARE(back)_–lo,–hi in addition to MAX(round) and DEP(round). The alternative strategy of spreading [–back] (48b) avoids this violation and becomes optimal.

Since SHARE(back)_–lo,–hi is parasitic on height, no additional backness changes are predicted if an exceptionally raised low vowel is preceded by high vowel. For underlying /misɑ–ijɑ/ (13), for example, only association of the floating [–low] feature and regressive [+ATR] harmony is predicted: SHARE(back)_–lo,–hi is not violated by winning [misolijɑ]. The trigger for the exceptional backness harmony is absolutely parallel to the trigger for regular [+ATR] harmony: It is a SHARE constraint demanding parasitic vowel harmony for vowels that share certain features. As for the [+ATR] harmony, the backness harmony is consequently predicted to be potentially iterative since every pair of adjacent mid vowels with different backness specifications violate the constraints. The weights of constraints in Assamese, however, predict that this change is only possible if the mid vowel that is expected to change is a derived mid vowel — and those only appear in the very restricted context preceding one of the two suffixes that contain an additional floating [–low] feature. The non-iterativity of backness harmony is therefore a simple epiphenomenon following from the pDEE in Assamese.

The crucial difference between an underlying mid vowel (48) and an underlying low vowel that is raised in the context of an exceptional suffix (49) is the fact that the latter has to change either its frontness or its backness anyway to be a well-formed mid vowel of the language. In such a case, the low-weighted constraint demanding a harmonic sequence with respect to backness is crucial and prefers the harmonic sequence.

All the constraint weights of this first possible analysis of Assamese based on a fully specified feature system are summarized again in (50). It is important to keep in mind, however, that these are only some exemplary weights that illustrate the relevant weighting arguments we discussed throughout the preceding subsections. Though the final number of constraints looks rather large, it merely includes many specific versions of general constraint types. There are, for example, only the three general types of faithfulness constraints MAX(F), DEP(F), and *SPREAD (=for L and R respectively) for every relevant feature.

In our analysis, the regressive [+ATR] harmony follows from a SHARE(ATR)_–low constraint and the fact that [+ATR] is the dominant feature from the markedness constraint *[–ATR]. The exceptional raising was analysed as an instance of featural affixation: The exceptional suffixes contain a floating [–low] feature in their lexical representation (cf. section 3).

If a fully specified feature system is assumed, the exceptional backness harmony is a pDEE that is predicted by constraint ganging. The core weighting arguments for the exceptional targets are repeated in (51). First, we established that [o] is in principle the default option for deriving a mid vowel after raising (51a). If this option would result in a newly created disharmonic sequence of mid vowels not agreeing in their backness specification, an [e] is created instead (51b).

These weighting arguments illustrate a gang effect: Whereas MAX(round) and DEP(round) together have a lower weight than MAX(back) and *SPREAD-R_back together, the combination of MAX(round), DEP(round), and SHARE(back)_–lo,–hi overtakes the combination of MAX(back) and *SPREAD-R_back. Consequently, the relevant markedness constraint SHARE(back)_–lo,–hi only has an effect in derived environments and never causes a change for disharmonic sequences of underlyingly mid vowels (51c). This follows since the derived mid vowel has to violate one of the relevant faithfulness constraints anyway: Since a violation of MAX(back) and *SPREAD-R_back or of MAX(round) and DEP(round) is unavoidable in such a context anyway (cf. (41)), the derived mid vowel is free to satisfy the lower-weighted markedness constraint SHARE(back)_–lo,–hi: An emergence of the unmarked effect (McCarthy & Prince 1994; Becker & Flack Potts 2011) in a derived environment results.

These gang effects are apparently rather intricate since they involve the combination of two and three constraints respectively. In HG, such an interaction is straightforwardly predicted since all constraint violations contribute to the harmony score for a candidate.

The depiction in (53) visualizes this effect. Whereas an underlying mid vowel cannot become fronted to fulfill the vowel harmony constraint, a derived mid vowel resulting from an original low vowel can indeed undergo fronting. On its path to become a licit mid vowel, this originally low vowel has to change some features anyway, it can hence use this feature change as a free ride to fulfill the vowel harmony demand.

A comparable account with ranked constraints cannot predict the Assamese pattern under the assumptions we made. This is apparent from the weighting arguments in (51). If they were understood as ranking arguments, it is clear that they result in a ranking paradox for MAX(round) and DEP(round) as well as MAX(back) and *SPREAD-R_back. The faithfulness constraints for [±back] must outrank those for [±round] to ensure the [o] is the default vowel for a raised low vowel (53a). Furthermore, to allow backness harmony for derived mid vowels, SHARE(back)_–lo,–hi must outrank MAX(back) and *SPREAD-R_back (53b). But given these ranking arguments, transitivity excludes that any pair of these faithfulness constraints can dominate SHARE(back)_–lo,–hi to block backness harmony for an underived mid vowel. Either MAX(back) and *SPREAD-R_back or MAX(round) and DEP(round) on their own would be sufficient to exclude this unattested vowel harmony, but both these rankings (53c) result in a ranking paradox.

The constraint ganging in (51) involves a markedness (=SHARE(back)_–lo,–hi) and two faithfulness constraints (=MAX(round)+DEP(round)) that together outweigh two other faithfulness constraints (=MAX(back)+*SPREAD-R_back). It has already been argued in Farris-Trimble (2008b) that faithfulness constraints can gang up and predict non-derived environment blocking. The Assamese pattern is interestingly different since the weights of faithfulness and markedness constraints gang up. The relation of the pDEE in Assamese to DEE’s as cumulative effects in general is discussed in more detail in 6.1.

4.2 Contrastive feature specification: Emergence of the Unmarked

In this subsection, we now turn to an alternative possible feature system for Assamese and its consequences for the theoretical account. The account in the last section is based on the fully specified feature system in (54) where all vowels are underlyingly specified for [±high], [±low], [±ATR], [±back], and [±round]. Under the contrastivist hypothesis that the phonology only operates on those features that are necessary to distinguish the phonemes of a language (Dresher et al. 1994; Hall 2007; Dresher 2009), the vowel system in Assamese is represented with fewer underlyingly specified features. Either [±back] or [±round] is sufficient to distinguish the non-low vowels /e, ɛ, i/ and /o, ɔ, u, ʊ/ respectively since there are no unrounded back vowels or round front vowels. Moreover, since there is only a single low vowel, /ɑ/ could only be specified for being low and lack any frontness or [±ATR] specification.¹⁹ One possible contrastive feature system for the Assamese vowels is given in (54).

Interestingly, the account for exceptional fronting in Assamese more or less falls out without any additional assumptions or constraints under this contrastive feature system. This mainly follows from the fact that the only low vowel /ɑ/ is radically underspecified in such a contrastive feature system. That only /ɑ/ undergoes exceptional fronting in case it is raised to a mid vowel then follows as an Emergence of the Unmarked: The faithfulness constraints preserving fronting for other vowels are irrelevant for this underspecified vowel. This account is briefly illustrated below.

Our illustration of the account starts — as in 4.1 — with a closer investigation of the concomitant feature changes that are relevant if a low vowel is raised because a floating [–low] feature needs to be realized (cf. (39)). In such a context, the vowel is also provided with a [+ATR] specification since SHARE(ATR)_–low demands independent ATR-harmony. We end up with a mid vowel that is specified for [–high,–low,+ATR] and has no value for backness. Since backness is contrastive for mid vowels, such a specification is necessary in Assamese, ensured by (55a). Again mirroring the account proposed in 4.1, we assume that the default mid vowel is back [o]. Under the contrastive feature system, this is simply ensured by the relative weightings of *e (55b) and *o (55c). Both have a relatively low weight and only show their effect in this context where a choice between the two mid vowels must be made.

How these weights predict the default realization of a raised low vowel as [o] is shown in tableau (56). Only realization of [–low] and spreading of [+ATR] fatally violates HAVE(BK)_–lo,–hi (56a) (the underspecified mid vowel is notated as [O] for reasons of simplicity). The choice between inserting [–back] (56b) and [+back] (56c) is then made in favor of the latter due to the fact that *e has a higher weight than *o. For completeness, DEP(back) is also given which is violated in both candidates that insert [±back] (56b) +(56c). As in all other autosegmental structures, the [±high] feature is omitted to make the depictions more readable.

The trigger for fronting is taken to be absolutely identical to the one assumed in 4.1: SHARE(back)_–lo,–hi demands that adjacent mid vowels should be associated to the same [±back] feature. As in the account in 4.1, leftward spreading of [±back] is taken to be excluded by a very high weight for *SPREAD-L_back (=100). Backness harmony is hence always progressive in Assamese and we won’t consider candidates in the following that spread the feature regressively.

The first crucial weighting argument ensures that underlying mid vowels never undergo fronting, shown in (59). The weight of MAX(back) is higher than the weight of SHARE(back)_–lo,–hi and the faithful candidate (59a) wins over the harmonic candidate (59b).²⁰

In the context of an underlyingly low vowel that is raised to a mid vowel and therefore potentially subject to SHARE(back)_–lo,–hi, however, MAX(back) is irrelevant simply because this vowel never had an underlying [±back] specification. It is an automatic consequence of the contrastive feature system that all low vowels are maximally underspecified and escape the scope of certain faithfulness constraints. Backness harmony is an Emergence of the Unmarked Effect (McCarthy & Prince 1994; Becker & Flack Potts 2011) under this interpretation: SHARE(back)_–lo,–hi has no effect for all the vowels which are subject to the general faithfulness constraints but suddenly shows its effect if certain vowels are exempt from faithfulness. This is shown in tableau (61).

As was established in tableau (59), we expect a back mid vowel [o] as the default after a floating [–low] is realized on a low vowel: The vowel needs a specification for [±back] and the lower weight of *o prefers the insertion of [+back]. If a mid front vowel precedes the underlyingly low vowel, however, this strategy induces an additional violation of SHARE(back)_–lo,–hi since the two mid vowels are disharmonic. Crucially, the harmonic candidate (61b) can avoid this violation of SHARE(back)_–lo,–hi without inducing an additional violation of MAX(back) since the vowel undergoing backness harmony never had an underlying [±back] specification. In addition, spreading even avoids a violation of DEP(back) since no epenthetic feature needs to be inserted. Backness harmony consequently emerges as optimal for an underlyingly underspecified originally low vowel.

4.3 Exceptional undergoers: Summary

The preceding two subsections showed two possible analyses for the exceptional undergoers of backness harmony in Assamese. Under the assumption of a fully specified feature system, exceptional undergoers are analysed as a pDEE that follows from constraint ganging in HG (4.1), and under a contrastive feature system, the exceptional undergoers fall out as a simple Emergence of the Unmarked Effect arising from underspecification (4.2). Importantly, the account for the exceptional triggers of raising in section 3.3 is completely independent of either one of these feature systems. Only height features and [±ATR] are relevant for this account and the specification differences which distinguish the two feature systems in 4.1 and 4.2 are not relevant. Under both feature systems, we assume that exceptional raising is due to a floating feature in the representation of certain morphemes. These reanalyses are summarized in (63).

We crucially do not aim to make an argument for either one of these feature systems that would rely on general claims about contrastivity or underspecification (Archangeli 1988; Itô et al. 1995; Dresher 2009). The important conclusion for us is that there are two reasonable different accounts for the Assamese exceptional backness harmony that avoid the assumption of lexically indexed constraints (cf. 5) and are based on purely phonological mechanisms. In contrast to the alternative account that is discussed in the following section, the analysis proposed here is hence compatible with a strict modularity between phonology and morphology. Our interim conclusion is therefore that the assumption of lexically indexed constraints is completely unnecessary for an account of vowel harmony in Assamese.

6.1 HG and Derived Environment Effects

The existence of pDEE’s is notoriously challenging for theoretical accounts of phonology. In rule-based phonology, for example, they only follow if rules can be explicitly marked for applying only in a derived environment. This is exactly the claim made in Lexical Phonology where cyclic rules only apply in morphologically or phonologically derived contexts (=Strict Cycle Condition; Kiparsky 1982). In Assamese, the rule triggering progressive backness harmony has been taken to be cyclic to correctly exclude backness harmony for non-derived mid vowels.

For parallel OT based only on standard markedness and faithfulness constraints (McCarthy & Prince 1995), pDEE’s are a serious problem and are inherently impossible to predict. If a high-ranked markedness constraint *X excludes a surface structure X in a language and a change into another surface structure Y is predicted, this repair is expected irrespective of whether X was underlyingly present or is derived in the phonology. No standard markedness- or faithfulness constraint can restrict the change to the latter context (Lubowicz 2002; Burzio 2011). Two proposals that argue for an extension of a basic OT system to account for pDEE are the proposal of comparative markedness (McCarthy 2002; 2003a) and of constraint conjunction (Smolensky 1993; 2006). The former theory assumes that all markedness constraints exist in two versions and one explicitly bans a ‘new’ marked configuration that was not yet present in the input. For Assamese, only a newly created sequence of mid vowels disharmonic for their backness specification would be penalized. The mechanism of constraint conjunction, on the other hand, creates new constraints via conjoining two existing constraints into a new one that is only violated if both source constraints are violated in a single locus. pDEE then follow under conjoining a markedness and a faithfulness constraint basically predicting that a marked configuration can be worse if it arises after changing an underlying specification (Lubowicz 2002; 2003). In Assamese, a high-ranked conjunction of IDENT(high) and SHARE(back)_{–high,–low} would then exclude a sequence of disharmonic mid vowels if the offending second vowel is also unfaithful to its [±high] specification (cf. 6.2 for more details).

The account proposed here models the same intuition as such an account based on Local Constraint Conjunction (=LCC). Both predict the pDEE from cumulativity: Whereas a single violation of a markedness constraint M is tolerated in order to satisfy a faithfulness constraint F₁, the conjoined effect of the markedness constraint M and another faithfulness constraint F₂ overrides F₁ and a process applies. In section 6.2, we further justify the choice of HG over LCC to model cumulativity.

The HG account of a pDEE in Assamese presented here can easily be generalized to other pDEE’s (Kiparsky 1973; Lubowicz 2002; Burzio 2011) as well. An example is diphthongization in Slovak that only affects derived long mid vowels (/čel+μ/→ [čiel] ‘forehead’; Lubowicz 2002: 10) whereas underlyingly long vowels are realized faithfully (/dceːr+a/ → [dceːra] ‘daughter’; Lubowicz 2002: 10). In an account absolutely parallel to our proposal for Assamese, the combination of having a marked long vowel and adding an association line between a mora and a vowel to lengthen it is too much and diphthongization is optimal. Only being a long vowel is therefore not a sufficient trigger for the repair and underlying long vowels surface faithfully. Only the combination of being a marked vowel and undergoing the unfaithful operation of vowel lengthening results in a worse harmony score than a candidate undergoing another repair operation avoiding the marked structure.

On the other hand, constraint ganging is by no means a general account of all pDEE types. It is crucially restricted to a certain sub-class where applying one phonological process to one target element induces additional faithfulness or markedness violations which could be avoided by changing it to another element altogether by applying a second process.²³ This second process, however, is blocked in isolation on another target element. Apart from the Slovak example, the famous case of Polish spirantization (Rubach 1984) and palatalization in Kinyamwezi (Kula 2008) are further examples of pDEE’s that can fall out as gang effects in HG. The present account is not generalizable, however, to other famous cases of pDEE’s like Finnish assibilation (Kiparsky 1973; 1993) or Makassarese /ʔ/-epenthesis (Aronoff et al. 1987; Basri et al. 1997; McCarthy 2002). In both these examples, a process derives the context for a second process, it does not derive the target of the process. Both these patterns have, however, received a reanalysis or reinterpretation of the data (Hammond (1992) and Jukes (2006) respectively). In general, convincing examples of pDEE’s seem to be rather rare which is in line with the claim that there should be no special mechanism in phonological theory that explicitly predicts pDEE’s since such a mechanism would apparently overgenerate. Conversely, the relative rarity of pDEE’s strengthens our claim for HG that predicts a subset of pDEE’s as epiphenomena. Future research needs to reveal the real typology of pDEE’s and the question whether it can fall out completely as the subset of pDEE’s predicted as ganging in HG.

6.2 An alternative: Local Constraint Conjunction

Our account of exceptional backness harmony in Assamese as a pDEE in 4.1 is formally based on a ganging between faithfulness and markedness constraints (cf. type i. and j. in Farris-Trimble 2008a: 6). It is not a new claim that the cumulativity of markedness and faithfulness constraints can predict Derived Environment Effects. As was already mentioned in 6.1, the proposal in Lubowicz (2002) argues explicitly that the mechanism of LCC predicts phonological and morphological Derived Environment Effects. LCC is an approach to cumulative effects in OT brought forwards by Smolensky (1995), similar to Harmonic Grammar. By conjoining two lower ranked constraints *A and *B, a higher ranked constraint *C (=*A&_D*B) is created. This higher ranked constraint is violated iff A and B are violated in a certain local domain D.

For illustration, an LCC account for the exceptional backness harmony in Assamese is given in (71) and (72). The first crucial difference to the HG account in 4.1 is that all constraints are strictly ranked and not weighted. The second important difference is the existence of the constraint DEP(round)&_VSHARE(back)_–lo,–hi: The constraint that results from locally conjoining the backness harmony constraint on mid vowels and the constraint against epenthetic [round] features for the domain of a vowel.²⁴ Whenever a vowel now induces a violation of both constraints, DEP(round)&_VSHARE(back)_–lo,–hi is also violated.

In (71), a non-derived context of two mid vowels that are disharmonic for [±back] is optimized. The general constraint SHARE(back)_–lo,–hi is violated in the candidate in which all [back] values surface faithfully (71a). Backness harmony as in (71b) avoids this violation but since MAX(back) is ranked higher than SHARE(back)_–lo,–hi, this option is sub-optimal. Only the need to be harmonic for backness is therefore not a good enough reason to violate MAX(back).

In (72), on the other hand, a front mid vowel precedes an underlyingly low back vowel that is raised to a mid vowel (ensured by independent mechanisms/constraints; cf. 3). Candidate (73a) that raises the low vowel to /o/ under epenthesis of [+round] not only violates general SHARE(back)_–lo,–hi and DEP(round) but also the conjoined constraint SHARE(back) &_VDEP(round). In this case, the backness harmony candidate (73b) is optimal. Crucially, each individual violation of SHARE(back)_–lo,–hi or DEP(round) would not be fatal for a candidate since the individual constraints are ranked lower than MAX(back). The locally conjoined constraint SHARE(back)_–lo,–hi &_VDEP(round), however, is ranked higher than this faithfulness constraint. Absolutely parallel to the constraint ganging in HG, only the combination of violating two lower ranked constraints outweighs the violation of a higher ranked constraint.

If we compare this account with the HG account of the pDEE in 4.1, it is apparent that the LCC account is apparently simpler. Whereas the HG account has rather complex ganging arguments involving two and three constraints (cf. (51)), the LCC account only involves a constraint that conjoins two other constraints. The assumption of recursive constraint conjunction (e.g. Itô & Mester 2003) that would mirror a complex ganging effect involving three constraints is hence not necessary. As we already discussed at the end of section 4.1, this is mainly due to the fact that all constraints that are violated by a candidate contribute to a potential ganging effect.

It is a matter of ongoing debate that ganging in HG and LCC can in principle predict the same type of effect but still differ in their concrete predictions. One main argument against HG is often the missing locality restriction. On the other hand, Pater (2009; 2016) argue that HG is more restrictive than LCC and can avoid overgeneration problems. Walker (2017) makes the converse argument in favour of HG in showing that HG can predict some attested patterns that are at least problematic to capture with LCC. The local domain in Constraint Conjunction has to be stipulated for each conjunction and can — in principle — be broadened to derive certain non-local patterns. Pater (2016) argues that corelevance restrictions in HG, on the other hand, are inherent. Another central argument often cited in favor of HG is the fact that there is no learning algorithm for local conjunction (Pater 2009).