2.1. The puzzle: asymmetrical consonant effects

While this vowel alternation pattern seems to be arbitrary at a first glance, works on Egyptian Arabic as well as on several other dialects report partial predictability based on phonological factors.Footnote ¹ These phonological factors include root consonants and correspondences between the vowels of different forms. I first discuss the root consonants and then discuss the vowel correspondences in §2.2. In Hijazi Arabic, the presence of a guttural (pharyngeal, glottal or uvular) root consonant leads to a preference for [a], and a pharyngealised alveolar leads to a preference for [u] both in lexicon trends and in nonce-word experiment results (Ahyad Reference Ahyad2019; Ahyad & Becker Reference Ahyad and Becker2020). Similar gradient effects of gutturals have been reported in the lexicon of MSA (McCarthy Reference McCarthy1991), and pharyngealised alveolar effects are reported for Egyptian Arabic (Abdel-Massih et al. Reference Abdel-Massih, Abdel-Malek and Badawi1979). Similar effects are shown to be categorical in Palestinian Arabic (Herzallah Reference Herzallah1990).

These frequently cited consonant–vowel co-occurrence effects are phonetically motivated. Guttural consonants involve retraction of the tongue root, which creates an affinity for low vowels like [a] over high vowels like [i, u] (McCarthy Reference McCarthy and Keating1994). Pharyngealised alveolars have the effect of lowering the F2 of surrounding vowels (Norlin Reference Norlin1987; Laufer & Baer Reference Laufer and Baer1988), explaining the preference for [u] over [i, a].

A curious commonality of these findings is the asymmetry between the two tense/aspect forms, where consonant effects on vowel choice are typically reported only in the imperfective and not in the perfective (cf. Blanc Reference Blanc1964, where the opposite pattern is reported in Muslim Baghdadi Arabic). This asymmetry is most likely not due to differences in the segmental contexts, since the alternating vowels are embedded between the second and third consonants in the underlined portion of both forms, CCVC vs. CVCVC. (Potential effects that uniquely come from the first consonant will be discussed in §3.) It is possible that the morphological structure of the wazn I verb paradigm provides some explanations for this asymmetry, but no such accounts have been proposed because the paradigm structure itself is contended, as I discuss below.

2.2. Morphological structure of the paradigm: surface correspondence vs. morphosyntax

I discuss here two ways the morphological structure of the wazn I verb paradigm has been investigated in the literature. One line of work has focussed on surface-based mappings between vowels in the imperfective and perfective forms, but this approach has not led to any definitive results. It is possible to account for the vowel alternation in MSA with great accuracy if the imperfective form is taken as the base that predicts the perfective form (McOmber Reference McOmber and Eid1995). However, Guerssel & Lowenstamm (Reference Guerssel, Lowenstamm, Lecarme, Lowenstamm and Shlonsky1996) have also shown that greater accuracy can be achieved by analysing the perfective form as the base from which imperfective form is predicted, though their analysis involves positing an abstract underlying vowel for a subclass of verbs. In either case, it is not clear if such surface correspondences are even a part of speakers’ grammars. Ahyad (Reference Ahyad2019) found that [i] in perfective forms predicts [a] in imperfective forms 90% of the time in Hijazi Arabic, but speakers completely fail to generalise this trend in wug tests. Whether speakers generalise the opposite imperfective-to-perfective pattern has not been tested before now.

Data from morphosyntactic studies on the imperfective and perfective verb forms in Arabic offers new insight into this issue. While the perfective and imperfective in Arabic are often treated as two different tense/aspect forms, syntactic and acquisition studies give evidence that the imperfective should actually be understood as an infinitive form. Thus, the imperfective occupies a smaller morphosyntactic domain, which is embedded within the morphosyntactic domain occupied by the perfective. Crucially, this suggests that the perfective is more likely to be derived from the imperfective than the other way around.

Benmamoun (Reference Benmamoun1999) shows that in Arabic dialects like MSA and Egyptian Arabic, the perfective form of the verb is always used in past tense clauses (2), whereas the imperfective form occurs in a wide range of contexts, including present tense, as in (3a); future tense, as in (3b); after modals and auxiliaries, as in (3c) and (3d); and in infinitival clauses, as in (3e).

While the perfective verb stem is always specified for past tense, Benmamoun notes that in the sentences containing an imperfective verb, the imperfective verb stem itself is never specified for tense; tense is always conveyed by some other particle (e.g., ha- for future). This suggests that the ‘imperfective’ in Arabic is actually an infinitival verb or present participle. Acquisition studies offer additional support: Aljenaie (Reference Aljenaie2010) found that Kuwaiti children in the age range of 1;8–3;1 use the bare imperfective stem as a non-finite form (see also Omar Reference Omar1973).

In theories where morphological structure is syntactic in nature, such as Distributed Morphology (Halle & Marantz Reference Halle, Marantz, Hale and Keyser1993), word formation happens bottom-up in the syntax. Under this framework, the Arabic imperfective would be used as the input to derive the perfective. The perfective form, always specified for past tense, must have an underlying structure that includes the tense head T and its projection. The imperfective form is never specified for tense, so it cannot contain T, meaning that it has simpler morphosyntactic structure than the perfective form. (4) shows a sketch of the underlying structure of Arabic wazn I verbs, adapted from Tucker (Reference Tucker, LaCara, Thompson and Tucker2011):

The functional head v combines with a consonantal root to form a verb, which is then selected by a Voice head. This structure then gets selected by an Asp head, which I treat as the position where the imperfective vowel resides. Whereas previous approaches either treat the imperfective vowel as the Voice head (Arad Reference Arad2005 for Hebrew) or a composite Tense/Aspect/Voice (TAV) head (Tucker Reference Tucker, LaCara, Thompson and Tucker2011 for MSA), this modification was made because vowel change does not signal active/passive alternations in Egyptian Arabic. Furthermore, since the imperfective verb always combines with other elements that carry tense information, its structure should only include AspP (boxed), not T. On the other hand, the perfective verb is always inflected for past tense. Therefore, the perfective vowel is treated as a T head specified for [past], and the structure of the perfective verb will minimally include the T head.

Following such an analysis, the vowel alternation in the perfective and imperfective are allomorphic processes limited by locality constraints, which allow only adjacent overt morphemes to affect each other in phonologically conditioned alternations (Embick Reference Embick2010; Marantz Reference Marantz, Marantz and Matushansky2013). Evidence of this locality constraint has been found in case studies involving both concatenative morphology (Northen Paiute, Toosarvandani Reference Toosarvandani2016; Nez Perce, Deal & Wolf Reference Deal, Wolf, Gribanova and Shih2017; Greek, Paparounas Reference Paparounas2021) and non-concatenative morphology (Hebrew, Kastner Reference Kastner2016, Reference Kastner2019). Imperfective vowel choice is conditioned by the root consonants because there is no overt intervening structure between the imperfective vowel and the root. Perfective vowel choice is not conditioned by the root consonants because there is overt intervening structure, namely the imperfective vowel. The asymmetry between imperfective and perfective verb forms in whether vowels are predictable from adjacent consonants is therefore an instance of morphologically constrained phonotactics.

In the rest of the article, I present results first from a lexicon study, then from a wug test investigating predictors of the vowel alternation in Egyptian Arabic wazn I verbs. Consistent with previous studies, I find that consonants play a role in the choice of vowel for imperfective but not perfective forms; these effects are successfully learned by speakers and generalised onto novel forms. Additionally, I identify vowel correspondence effects in the lexicon and show that speakers selectively generalise them in ways that align with the morphosyntactic structure of the paradigm proposed above. I argue that the vowel alternation in Egyptian Arabic is a case where non-concatenative morphological domains constrain phonotactics, just as has been found for concatenative morphological domains. However, non-concatenative morphological domains are unique in that they are by default abstract and not readily evident on the surface.

3.1. Data collection

I compiled a corpus of 330 wazn I verbs in Egyptian Arabic (available on OSF at https://osf.io/tdr6x/?view_only=4aec9be4f2204849ac81932103fe3dd9). Only words with the prosodic shape CVCVC in the perfective (traditionally known as ‘sound verbs’) were included, as verbs with other prosodic shapes have different vowel alternation patterns. The words were extracted from the online dictionary Lisaan Masry (Green Reference Green2007) and checked with a native speaker.

Since the focus of this study is on the colloquial Arabic dialect spoken in Cairo, it is important to acknowledge that nearly all speakers of Egyptian Arabic also use MSA to some degree (Ferguson Reference Ferguson1959; Eid Reference Eid and Owens2007). To control for the differences in the morphophonology of these two varieties of Arabic, verbs that had clear features of MSA (e.g., having [a] in the imperfective agreement prefix, as in [ja-sbat] ‘he proves’, as opposed to [i] or [u] in the prefix as in Egyptian Arabic) were excluded.

The breakdown of the corpus by vowel alternation pattern is shown in Table 2. The perfective vowel is either [a] or [i], and the vowels in the two syllables (CVCVC) are always identical. The single imperfective stem vowel (-CCVC) can be [a], [i], or [u]. Note that imperfective stems always occur with an agreement prefix. The prefix vowel is generally [i] but optionally undergoes harmony when the stem vowel is [u]. All combinations of vowels in the two forms are attested, though the perfective [i] – imperfective [u] combination is very rare.

Table 2 Vowel alternation in wazn I verbs from Egyptian Arabic (Table 1) with counts added (a total of 330 verbs).

3.2. Modelling

Logistic regression models with vowel and consonant predictors were fitted using the nnet package (Venables & Ripley Reference Venables and Ripley2002) in R (R Core Team 2021). Models predicting perfective vowel choices had two levels of the dependent variable ([a] or [i]), and models predicting imperfective vowel choices had three ([a], [i] or [u]), which called for multinomial logistic regression. Vowel predictors based on the vowels in the form other than the one that the model predicts were used to investigate whether vowel-to-vowel correspondence plays a role in the paradigm. The consonant predictors assess whether a consonant with a particular place of articulation is present in the word; they were used to investigate whether there is avoidance of more marked consonant–vowel combinations. Place of articulation was chosen to be the main consonant property investigated because consonant effects on vowel alternations in other dialects generally involve place features (e.g., Ahyad & Becker Reference Ahyad and Becker2020). The place of articulation classes are listed in Table 3. Words containing two glide consonants, {j, w}, were excluded from this study because they generally have different prosodic shapes and vowel patterns.

Table 3 Consonant natural classes by place of articulation.

Pharyngealised alveolar consonants, also known as emphatics, are characterised by a secondary constriction in the upper pharynx (Ghazeli Reference Ghazeli1977; Laufer & Baer Reference Laufer and Baer1988). They exert a lowering and backing effect known as emphasis spreading on all vowels within the same phonological word, as well as affecting formant transitions surrounding the consonants (Lehn Reference Lehn1963; Norlin Reference Norlin1987; Watson Reference Watson2002). The status of the tap /r/ in Egyptian Arabic is debated. Specifically, it has been argued that there is a contrast between a non-emphatic /r/ and an emphatic one (Youssef Reference Youssef2019), which causes emphasis spreading like other pharyngealised alveolars (Younes Reference Younes, Eid, Cantarino and Walters1994; Watson Reference Watson2002). My corpus based on the dictionary did not include such a contrast, and I grouped /r/ with pharyngealised alveolars.Footnote ²

In the imperfective forms (-C₁C₂VC₃), the second and the third root consonants are directly adjacent to the vowel, so one might expect stronger effects for them compared to the first consonant. To test this, I tried running models with positional consonant predictors, which specify whether a given natural class is present in each of the three positions in the consonantal root. However, because of the relatively small sample size, the models with positional predictors overfitted (as indicated by extremely large coefficient values and poor accuracy in cross-validation) and thus were not informative. I will return to this issue of consonant position when discussing the results below.

Besides quantitatively evaluating the role of vowels and consonants in predicting the vowel alternation, the models were designed with an additional purpose, namely to model different paradigm organisations. Numerous case studies (e.g., Albright Reference Albright, Downing, Alan Hall and Raffelsiefen2005; Kuo Reference Kuo2020) have shown that speakers preferentially select a paradigm member as the base to derive other forms in ways that maximise predictability. Specifically, I evaluate two possibilities that both instantiate such unidirectional paradigm mappings but in opposite directions. The first is that speakers use the imperfective forms as the bases to derive the perfective forms. This possibility is instantiated by a model that predicts perfective vowel choices based on all and only the phonological predictors that are available in the imperfective forms, namely the root consonants and the imperfective vowels (§3.3). The second possibility is that speakers use the perfective forms as the bases to derive the imperfective forms. The model that instantiates this possibility uses root consonants and perfective vowels to predict imperfective vowel choices (§3.4). The models are then compared on the basis of predictive accuracy and explanatory power (§3.5).

3.3. Predicting the perfective

Vowel-to-vowel correspondences play a role in predicting the perfective vowel. As shown in Table 4 and plotted in Figure 1, while the overall distribution of [a]- and [i]-perfectives is even, the breakdown of perfective vowel choice for each imperfective vowel differs. There is a strong preference for [a]-perfectives (91%) when the imperfective vowel is [u]. When the imperfective vowel is [i] or [a], there are moderate preferences for [i]-perfectives (69% and 56%, respectively).

Table 4 Perfective and imperfective vowel frequencies.

Figure 1 Breakdown of perfective vowel by imperfective vowel.

On the other hand, perfective vowel choice does not seem to be influenced by consonant place of articulation, as shown in Table 5 and Figure 2. Note that each verb in the corpus is represented three times in Table 5, once for each consonant in the root. Having a root consonant of a particular place of articulation tends not to bias the perfective vowel distribution away from the overall distribution. While pharyngealised alveolars, palatals and uvulars show a preference for [a], the effects are quite small.

Table 5 Effects of consonant natural classes on perfective vowel distribution.

Figure 2 Effects of consonant natural classes on perfective vowel distribution.

Since the two vowels in a perfective form (C₁V₁C₂V₂C₃) are identical in Egyptian Arabic, it is possible that only one of the vowels is represented and the other derived by copying. If this is the case, we may expect the consonant effects on vowel distribution to be positionally specific. For instance, if only V₁ is represented, it is possible that consonant effects are only found with the adjacent C₁ and C₂. Table 6 presents the vowel distributions in the presence of pharyngeals and pharyngealised alveolars at each consonant position. There is no clear evidence for any positional effect. For example, perfectives with a pharyngeal as the first root consonant have [a] 44% of the time. This number is 46% if the pharyngeal is the second consonant and 51% if it’s the third.

Table 6 Effects of consonant natural classes by position in perfectives: pharyngeals and pharyngealised alveolars.

These two observations (that perfective vowel choice is influenced by vowel-to-vowel correspondences but not by consonant–vowel co-occurrences) are borne out by the regression model, shown in Table 7. Positive coefficients indicate a preference for [i]-perfectives, whereas negative coefficients indicate a preference for [a]-perfectives. Note that there are two imperfective vowel predictors, where the effect of [i] and [u] are each compared to that of [a].

Table 7 Imperfective-to-perfective model. Residual deviance: 378.35; pseudo- $R^2$ : McFadden 0.173, CoxSnell 0.213, Nagelkerke 0.284; cross-validation accuracy: 0.639. Significant factors are bolded.

Table 8 Effects of consonant natural classes on imperfective vowel distribution.

There are only two significant predictors in the above model, both of which are vowel predictors. First, imperfective [u] results in a large preference for perfective [a]. Second, there is a moderate preference for perfective [i] when there is imperfective [i]. None of the consonant predictors are significant.

3.4. Predicting the imperfective

Imperfective vowel choices also show sensitivity to perfective vowels. Notably, there is a preference for imperfective [u] when the perfective vowel is [a]. This trend is less drastic than the same preference in the opposite direction (see Table 4 and Figure 1). Perfective [i] results in a slight preference for imperfective [i] and dispreference for imperfective [u].

Additionally, there are noticeable consonant effects on the imperfective vowel distributions, as shown in Table 8 and Figure 3. Compared to the overall vowel distribution, the presence of pharyngeal and glottal root consonants is associated with more [a]-imperfectives (70% and 60%, respectively); labials and plain alveolars with more [i]-imperfectives (37% for both); and pharyngealised alveolars with more [u]-imperfectives (36%).

Figure 3 Effects of consonant natural classes on imperfective vowel distribution.

A multinomial regression model predicting imperfective vowels is shown in Table 9. The model shows pairwise comparisons for [i] vs. [a] and [u] vs. [a], where positive coefficients indicate a preference for [i]- or [u]-imperfectives over [a]. Note that there is just one perfective vowel predictor, where the effect of [a] is compared to that of [i].

Table 9 Perfective-to-imperfective model. Residual deviance: 515.22; pseudo- $R^2$ : McFadden 0.253, CoxSnell 0.411, Nagelkerke 0.469; cross-validation accuracy: 0.606. Significant factors are bolded.

Consonant predictors make a substantial contribution in this model, consistent with the various consonantal effects observed above. Pharyngeals and glottals strongly favour imperfective [a] over [i, u], and uvulars strongly favour [a] over [u]. Pharyngealised alveolars strongly favour [a] over [i] but exhibit no preference between [a] and [u]. As discussed in §2, these effects are consistent with previous findings and have phonetic motivations.

A gross inspection of vowel distribution based on the presence of consonants in specific positions finds striking positional effects for pharyngeals. Table 10 shows that the preference for imperfective [a] is very strong when the imperfective vowel is immediately adjacent to a pharyngeal (as C₂ or C₃) but is absent when the pharyngeal (as C₁) is separated from the vowel by an intervening consonant. Interestingly, strong effects of consonant position are not found in verbs with pharyngealised alveolars, possibly because pharyngealised alveolars can influence vowel qualities across the entire phonological word, not just adjacent vowels (e.g., Watson Reference Watson2002).

Table 10 Effects of consonant natural classes by position in imperfectives: pharyngeals and pharyngealised alveolars.

3.5. Comparing the models

The two models’ goodness-of-fit was compared in order to assess which direction of paradigm mapping yields greater predictability. The first method I used compares the accuracy of the two models with k-fold cross-validation ( $k=5$ ). The dataset was randomly divided into five parts, and each model was run on four of the parts and tested on the other. This process was repeated for all five parts, and the average model accuracy from all the trials was calculated by comparing the model predictions on the testing data in each run with the corpus. The imperfective-to-perfective model had a higher average accuracy (0.639) than the perfective-to-imperfective model (0.606), but the difference is very small.

I then assessed the two models relative to chance-level performance and found that the model predicting the imperfective is superior. The reason is that the random baseline accuracy for this model is 0.33, since there are three imperfective vowel choices, whereas the random baseline accuracy of the model predicting the perfective (between two alternatives) is 0.5. This intuition is also supported by pseudo- $R^2$ measures. Since the two models predict different dependent variables, they cannot be directly compared with AIC or likelihood measures, and pseudo- $R^2$ measures are appropriate. All three pseudo- $R^2$ measures (McFadden, Cox & Snell and Nagelkerke) calculated were higher for the perfective-to-imperfective model, which suggests that it is superior in terms of accounting for the variability in the data.

3.6. Summary

Modelling results based on lexicon data showed that both perfective and imperfective vowels are partially predictable from phonological factors. A crucial difference between the models lies in the types of relevant phonological predictors. The imperfective vowel can be predicted primarily by consonants in ways governed by the phonotactic well-formedness of consonant–vowel sequences, with a minor role played by correspondences to the perfective vowel. The perfective vowel can be predicted based solely on the imperfective vowel. Furthermore, model comparison metrics showed that the model predicting the imperfectives from the perfectives had higher explanatory power, even though the opposite model had a slightly higher accuracy. The next section examines whether these predictive trends are generalised by native speakers in a wug test.

4.1. Methods

4.1.2. Materials

A list of nonce verbs in both perfective (C₁VC₂VC₃) and imperfective (-C₁C₂VC₃) forms was compiled. The list contained 50 distinct roots, each occurring with the two perfective vowels and three imperfective vowels, yielding a total of 250 nonce verbs. All possible vowel choices in both forms were included. Only a subset of consonants was included to make it easier to observe and test the effects of consonant class. The first two root consonants were always drawn from labials {b, f, m}, alveolar obstruents {t, d, s, z}, and alveolar sonorants {l, n}. These consonants showed a preference for imperfective [i] but did not affect the vowel distribution as strongly as other consonants, making them ideal control consonants. In keeping with OCP co-occurrence restrictions (McCarthy Reference McCarthy and Keating1994), a root never contained two segments drawn from the same class. In half of the roots, the third consonant was drawn from the control set as well, while for the other half, the third consonant was a pharyngeal {ħ, ʕ}. Pharyngeals only occurred as the third consonant because their effect on imperfective vowel was observed most strongly in this position (see Table 10).

All nonce verbs that could be created from a pair of roots differing by C₃ (l-b-ħ and l-b-z) are illustrated in Table 11. Each root occurs in both forms, with all possible vowel choices. Each participant would only hear an individual root in one aspect, with one vowel choice for that aspect, pseudorandomised via Latin square. Each participant heard 50 stimuli in a randomised order, 20 of which were presented to them in the perfective (10 [a], 10 [i]) and 30 of which were presented in the imperfective (10 [a], 10 [i], 10 [u]).

Table 11 Example stimuli with the roots l-b-ħ and l-b-z.

The nonce words were checked with a native speaker to ensure that they were well-formed and not existing words. The words were recorded in the two frame sentences in (5) by a female native speaker. The two frame sentences contain adverbial phrases disambiguating the verb form that should occur (perfective in (5a), imperfective in (5b)). Both forms agree with a third person masculine singular subject: the imperfective has the indicative prefix [bi-] and the agreement prefix [ji-], while the corresponding perfective has no overt agreement marking.

• 

Transitivity has been claimed to affect perfective vowel choice in Egyptian Arabic (Abdel-Massih et al. Reference Abdel-Massih, Abdel-Malek and Badawi1979). The lexicon compiled in this study also showed gradient effects of transitivity in Egyptian Arabic. Transitive verbs were more likely to have perfective [a] (58%) than intransitives (34%). Opposite trends were found with imperfective [a] (40% in transitives, 69% in intransitives). This was controlled for by keeping all frame sentences intransitive. Additionally, while semantic effects on vowel choice have been reported for MSA (McCarthy Reference McCarthy and Keating1994), Egyptian Arabic does not have similar patterns, and semantic effects are generally not expected to affect nonce-word experiment results.

4.3. Summary

The experimental results show that Egyptian Arabic speakers generalise some predictive trends when deciding on vowel choices in novel verb forms, but crucially do not generalise all trends that can be found in the lexicon. Specifically, when predicting imperfective vowels, consonant effects were generalised, but perfective vowel effects were not, even though predictive trends of both types were found in the lexicon. When predicting perfective vowels, imperfective vowel effects in the lexicon were generalised. Root consonants did not help predict perfective vowels in the lexicon and were also not used by speakers.

While vowel-to-vowel correspondences in both paradigm directions were found in the lexicon, speakers only made use of imperfective-to-perfective correspondences. This suggests that speakers derive perfective forms based on imperfective forms and not the other way around; this is discussed in more detail in §5.

Notably, the asymmetry of consonant–vowel phonotactics was faithfully learned by speakers. Whereas consonants played no role in perfective vowel choice, pharyngeals strongly predicted [a] in the imperfective, and labials and plain alveolars strongly predicted [i]. Since the overlapping consonantal contexts for the alternating vowels in the perfective and imperfective forms mean that this pattern has no clear phonological explanation, the explanation must come from the morphological structure of the wazn I verb paradigm.

5.1. A morphosyntactic account

First, I show that a morphosyntactic analysis of the Egyptian Arabic verbal paradigm aligns well with the lexicon and wug-test results and provides an explanation for the asymmetrical effects of consonant-vowel co-occurrences. The verbal morphosyntactic structure in (4) is repeated in (7):

I follow previous morphosyntactic analyses of Arabic verbal structure such as Tucker (Reference Tucker, LaCara, Thompson and Tucker2011) in positing separate morphemic status for the consonantal root, the imperfective vowel and the perfective vowel. The imperfective vowel is treated as an Asp head, and the perfective vowel as a T(ense) head, consistent with the proposal that the imperfective is a tenseless, infinitival form, whereas the perfective is inflected for past tense (Benmamoun Reference Benmamoun1999). According to the structure in (7), the consonantal root first combines with the phonologically null functional head v, and then combines with the imperfective vowel to form the imperfective. Prosodic constraints have been shown to be successful at deriving the correct surface linearisation, with the consonantal and vocalic morphemes interleaved (Tucker Reference Tucker, LaCara, Thompson and Tucker2011); I abstract away from them here.

The predictions of this analysis align perfectly with the wug test results. Since the perfective is derived from the imperfective and not vice versa, we expect that speakers should generalise lexicon trends based on the imperfective vowel when deriving perfective forms but do not expect generalisation in the opposite direction. These predictions are borne out.

Likewise, the absence of consonant–vowel interactions in the perfective form follows from independently proposed syntactic locality constraints, which disallow allomorph selection between any two elements that are separated by other overt material in the morphosyntactic structure (Embick Reference Embick2010). Since the consonantal root merges with the imperfective vowel first, it is possible for phonological interactions between consonants and vowels to influence the choice of imperfective vowel. Further, since the perfective vowel is structurally adjacent to the imperfective vowel but not to the consonantal root, it follows that only imperfective vowel predictors contribute to predicting perfective vowel choice.

The vowel choice in both the imperfectives and the perfectives is understood as suppletive allomorphy in this analysis. Specifically, the consonant–vowel co-occurrence restrictions are phonological conditioning factors for allomorphic selection of the imperfective vowel, which is the Asp morpheme in (7). There are several different ways to implement such phonological conditioning factors. I briefly discuss two possible implementations here. One possibility is that such information is encoded as subcategorisation constraints, such that allomorphs of affixes select stems based on the stems’ underlying phonological properties (Paster Reference Paster2005). An illustration of one such subcategorisation rule is shown in (8a). This rule states that the Asp morpheme should be realised as [a] if it precedes a root morpheme that contains a pharyngeal. Another possibility is that selectional restrictions of allomorphs are learned as phonotactic grammars over the subsets of the lexicon which take the various allomorphs (Gouskova et al. Reference Gouskova, Newlin-Łukowicz and Kasyanenko2015; Becker & Gouskova Reference Becker and Gouskova2016). Phonotactic properties of the stems themselves and those of the affixed output forms can both play a role. Example rules are shown in (8b), following Gouskova et al. (Reference Gouskova, Newlin-Łukowicz and Kasyanenko2015). Specifically, the [a] allomorph of the Asp morpheme may be indexed to a sublexicon of consonantal roots that contain pharyngeal consonants, as in (8b-i), or to a sublexicon of imperfective forms that contain sequences of low vowel and pharyngeal, as in (8b-ii).

Similarly, the selection of perfective vowels is treated as suppletive allomorphy of the past tense morpheme. However, phonological conditioning of this morpheme’s realisation can only refer to the identity of the imperfective vowel, which is local in the morphosyntactic structure, and not to the consonantal root.

There are two other factors that could potentially condition the perfective vowel allomorphy, though neither was supported by the data presented here. One is that [i] may be emerging as the default perfective vowel in Egyptian Arabic. This observation was made by my native speaker consultant based on the fact that a few words borrowed from MSA have undergone a change in the perfective vowel from [a] to [i] after becoming common in the colloquial variety. The second factor is transitivity: there is a slight preference in the lexicon for intransitive verbs to have [i] in the perfective (§4.1.2). However, in the wug test, [a] was predominantly the preferred choice for perfective vowel, even though the frame sentences force an intransitive interpretation of the verb.

As mentioned above, the phonological conditioning factors for both the imperfective and perfective vowel allomorphy are stochastic and exceptionful. As a result, allomorphy rules that refer to idiosyncratically indexed sets of lexical items are needed in addition to the phonologically based rules.

I illustrate how this analysis plays out with the following example. To form the imperfective form of ‘return’, [-rgaʕ], the consonantal root r-g-ʕ first selects [a] as the imperfective vowel, as opposed to [i] or [u], motivated by the avoidance of a marked phonotactic structure (namely, a high vowel next to a pharyngeal consonant). This happens below the AspP level.Footnote ³ The imperfective form can then be used as the input to form the perfective form [rigiʕ]. Because of locality restrictions, phonological information about the lowest embedded morpheme, the consonantal root, is no longer accessible. As the probabilistic correspondences between imperfective and perfective vowels found in this study do not include one that favours perfective [i] when the imperfective has [a], selection of [i] as the perfective vowel is based on a lexical indexation rule that is not phonologically conditioned.

5.2. Surface correspondence accounts

Having shown that an analysis which assumes both non-concatenative morphemes and morphosyntactic structure successfully account for how Egyptian Arabic speakers represent the wazn I verb paradigm, I now explore the extent to which these representations are necessary. To this end, I discuss how morphological theories based on surface correspondence within paradigms may explain the same data.

One idea that emerges from surface correspondence theories is that learners approach morphological paradigms by selecting one paradigm member as the base and using it to derive other members of the paradigm. Crucially, speakers choose a base that maximises informativity (Albright Reference Albright2002), that is, one that can allow them to derive other forms most reliably. This kind of base choice can be independent of morphosyntax. For example, in Yiddish nouns, plurals have been shown to be more informative bases than singulars, even though singulars have simpler morphosyntactic structure, and patterns of paradigm levelling are consistent with the hypothesis that plurals are the base (Albright Reference Albright, Bachrach and Nevins2008).

There is, however, a mismatch between what would be the most informative base choice and what Egyptian Arabic speakers opted to use. §3.5 compared two regression models which instantiate the two paradigmatic directions. The model that used the perfective as the base to predict the imperfective resulted in more improvement from chance, as shown by pseudo- $R^2$ metrics. In other words, there is more predictability to be gained from learning the trends that predict the imperfectives from the perfectives. This was the opposite of what speakers actually did in the wug test.

Additionally, vowel correspondence trends in the paradigm are redundant in that they contribute significantly to vowel predictability in both the perfective and the imperfective, and such redundancy in theory maximises predictability if speakers can use all of the trends (Bochner Reference Bochner1993). However, speakers selectively generalised trends in one paradigmatic direction (imperfective to perfective) but not the other.

We must therefore explain why Egyptian Arabic speakers represent the wazn I verb paradigm in a way that does not use paradigmatic relations most informatively. An analysis in which morphosyntax plays a direct role has been explored in §5.1. Another alternative explanation may be found in acquisition. Imperfectives may be preferentially chosen as the base because they are more common than perfectives in the early input and early production of Arabic-acquiring children (Aljenaie Reference Aljenaie2010). However, previous work on correspondences within paradigms has not made clear claims about how the acquisition timeline affects base selection, and more work on the early acquisition of Arabic verbs is necessary to address these questions.

A surface-based account where speakers select the imperfectives as the base to derive the perfectives has two implications that differ from the morphosyntax based analysis. First, an anonymous reviewer points out that analysing the imperfective as the derivational base has the consequence that the consonant–vowel co-occurrence restrictions active in the imperfective may be understood as phonotactic restrictions learned from stems (Anderson Reference Anderson1992). In contrast, under the morphosyntactic analysis, these restrictions are allomorphic selection criteria specific to the morphological context of the imperfective. Since this article focuses solely on the wazn I verb paradigm, I do not have the data to determine whether these consonant–vowel restrictions are morphologically specific or are part of more general phonotactic knowledge. Further work, such as lexicon studies on other morphologically simple words in Egyptian Arabic or patterns of loanword adaptation, would be informative on this issue.

Second, the morphosyntactic analysis makes a strong prediction about the asymmetrical presence of the consonant–vowel restrictions: given the simpler morphosyntactic structure of the imperfective, the absence of these restrictions in the perfective follows directly from syntactic locality constraints. This is different from approaches that allow different phonological grammars for different levels of morphological constructions or different lexical classes, such as Stratal OT (Bermúdez-Otero Reference Bermúdez-Otero1999; Kiparsky Reference Kiparsky2000), Cophonology (Anttila Reference Anttila, Hinskens, Hout and Leo Wetzels1997; Inkelas Reference Inkelas, Booij and Marle1998), and Cophonologies by Phase (Sande et al. Reference Sande, Jenks and Inkelas2020). Under these frameworks, this morphologically constrained phonotactic pattern would be modelled by demoting constraints against marked consonant–vowel sequences when computing the perfective form as opposed to the imperfective form. However, there is no a priori reason that the markedness constraints should be demoted rather than promoted when evaluating the perfective. Existing work on other Arabic dialects generally reports that the consonant–vowel co-occurrence restrictions are found solely in the imperfectives, as in Egyptian Arabic, with the exception of Baghdadi Arabic (Blanc Reference Blanc1964). Comparative work and detailed investigation on dialects like Baghdadi is needed to decide between these approaches.

5.3. Discussion

While morphological domain effects on phonotactic patterns are a common phenomenon typologically, they have been argued to be challenging to learn. Specifically, without knowledge of morphological domains, licit occurrences of a phonotactic sequence across morphological boundaries would become noise, making it more difficult for a learner to correctly reject these sequences within the relevant domain. Precisely for this reason, Gallagher et al. (Reference Gallagher, Gouskova and Rios2019) show that a computational phonotactic learner performs better when trained on data with morphological segmentation than when trained on data without this information. Perhaps paradoxically, however, such patterns have also been argued to aid in early morphological learning, since the sequences that violate phonotactic restrictions occur exclusively at morphological boundaries and thus can serve as cues for discovering these boundaries (Trubetzkoy Reference Trubetzkoy1939). The usefulness of these cues in segmentation has been demonstrated using computational models employing some version of a heuristic whereby word or morpheme boundaries are inserted to break up low-probability sequences. They have been shown to outperform models that only employ other distributional cues for segmentation, such as transitional probability (Daland Reference Daland2009; Adriaans & Kager Reference Adriaans and Kager2010, etc.). Morphological domain effects on phonotactics thus provide a good testing ground for the relationship between phonotactic and morphological learning.

Previous discussions of such patterns have focused on concatenative morphology. A crucial feature of non-concatenative morphology is that morphological domains are difficult to distinguish at the surface level. This article was only able to find evidence supporting the complicated morphological structure through a detailed study of the predictors for the vowel alternation pattern in a lexicon study and a behavioural experiment. Furthermore, there are contending accounts of the nature of the morphological knowledge needed to describe this pattern. The data presented in this article are compatible with two distinct analyses that differ in the representations they assume and offer distinct accounts of the nature of the morphological domain effect. In turn, they have different implications for learning.

The first analysis presented assumes the representation of abstract morphosyntactic structure. One potential acquisition trajectory for this account is that children have to be able to associate verb forms with their underlying morphosyntactic structure before learning that this structure plays a role in phonotactics. Prior to having the morphosyntactic knowledge, discovery of phonotactic patterns is harder because cross-morphemic occurrences of the restricted sequences introduce noise. This mirrors the problem proposed for patterns in concatenative morphology. On the other hand, morphologically constrained phonotactics could potentially also aid in the discovery of morphosyntactic structure. Locality constraints may be crucial, in that the absence of phonological effects on allomorphy serves as a cue for intervening morphosyntactic structure. However, since morphological domain effects on phonotactics are all restricted to specific allomorphic conditioning factors in such accounts, it is still likely that they are learned late because of the relatively small amount of data available.

The second assumption of the morphosyntactic account is non-concatenative morphemes, specifically the consonantal root and vocalic morphemes (McCarthy Reference McCarthy1979). In particular, the distinct functions that these morphemes have been argued to carry lend themselves well to highly decompositional frameworks like Distributed Morphology (e.g., Arad Reference Arad2005). The involvement of the consonantal root in forming the imperfective form and its inaccessibility – due to syntactic locality constraints – in subsequent morphological processes are crucial for this analysis to account for the absence of consonant effects in predicting the perfective vowel in Egyptian Arabic. This article, however, has not explored any alternative account which dissociates the role of non-concatenative morphemes and that of morphosyntactic structure.

The account based on paradigmatic correspondences requires neither of the two above assumptions. However, the challenge for this account is to explain why learners choose imperfective-to-perfective over other possible paradigm organisations. As discussed above, the criterion of maximising informativity prefers the perfective as the base instead, so some alternative explanation is called for. While the acquisition literature on Arabic suggests that the imperfective is learned earlier (Aljenaie Reference Aljenaie2010), the extent to which acquisition order determines base selection has not been explored extensively in the literature.