2.1 Morphological characteristics

Slovenian velar palatalization is an alternation which turns velars into postalveolars, before certain suffixes (a). Less frequent is second velar palatalization (here termed velar fronting) which turns velars into alveolars (b). Finally, iotation affects all segments except postalveolars and palatals (c). Iotation is a heterogenous pattern that includes velar (k → ʧ) and coronal palatalization (t → ʧ), lenition (d → j), and l-epenthesis after labials.

The three patterns share the same set of targets, which are velars turning into coronal sounds. At the same time, it is easy to distinguish the three processes. First, they are all morphologically conditioned, which means that any given suffix triggers at most one of the alternations. Second, the segments affected are only partially overlapping. Labials, for instance, show alternations only in iotation. Hence, the processes are clearly distinct.

This paper examines velar palatalization (henceforth, palatalization) rather than the other two patterns. Palatalization is considerably more productive than velar fronting. Furthermore, palatalization is frequent in nominal paradigms, where morpheme boundaries are relatively easy to determine from surface realizations. Iotation, in contrast, is often triggered by affixes that are not realized independently, i.e., as zero affixes, which makes morpheme boundaries harder to identify in a written corpus. In terms of the variation in individual stems, palatalization and iotation appear to be quite similar in Slovenian, so it is likely that the results reported in this paper are representative for both phenomena even though only palatalization will be analyzed.

Velar palatalization in Slovenian is triggered by the suffixes listed in Table 1. Many of these suffixes begin with a front vowel or glide j. There is a strong cross-linguistic tendency for these segments to be likely triggers of palatalization (Bhat 1978; Bateman 2007; Kochetov 2011). Other suffixes begin with a consonant or a back vowel, which are atypical triggers.¹

C-suffixes consist of a single consonant; these suffixes may be preceded by schwa to avoid raising sonority codas or syllabic consonants, typically in the masculine nominative/accusative singular only. The status of Slovenian schwa is controversial. Synchronically speaking, various descriptions have analyzed schwa as underlying (e.g. Toporišič 1976/2000), epenthetic (Jurgec 2007) or morphologically predictable (Bidwell 1969).

An anonymous reviewer suggests that schwa is underlying based on the generalizations from other Slavic languages. Historically, Slovenian schwa has developed from the Common Slavic yer in certain positions, but the synchronic data suggest that the speakers have reanalyzed schwa as epenthetic. The main argument for schwa being epenthetic is that its distribution is to a large extent predictable. Schwa appears in a small set of native roots (e.g. sən ‘dream’ ∼ sna ‘-GEN.SG’, pəs ‘dog’ ∼ psa ‘-GEN.SG’, məgla ‘fog-NOM.SG’ ∼ məgəl ‘-GEN.PL’), in the position before ɾ when not adjacent to a vowel (e.g. bom əɾ ˈdɛʧ ‘I will be red’ ∼ bo ɾ ˈdɛʧ ‘he will be red’), and at the boundary between roots and suffixes (e.g. ˈɾekəw ‘he said’ ∼ ˈɾɛkla ‘she said’, ˈdedək ‘grandpa’ ∼ ˈdetka ‘GEN.SG’), or to break an illicit cluster (ˈpalma ‘palm’ ∼ ˈpaləm ‘-GEN.SG’, ˈkamən ‘rock’ ∼ ˈkamna ‘GEN.SG’). Schwa never appears word-finally and generally avoids stress. Perhaps the most convincing evidence is that schwa does not count as a vowel for purposes of allomorphy. For instance, monosyllabic or shorter inanimate stems ending in ɾ take the allomorph –a in the genitive singular (tɪɾ ‘rail, track’ ∼ ˈtɪɾa ‘-GEN.SG’, mɛɾ ‘mayor’ ∼ ˈmɛɾa ‘-GEN.SG’), whereas disyllabic or longer roots take the allomorph –ja (ˈkuteɾ ‘boat’ ∼ ˈkuteɾja ‘-GEN.SG’; ˈɡɾubeɾ ‘name’ ∼ ˈɡɾubeɾja ‘-GEN.SG’). Schwa, crucially, does not count as syllabic for the purposes of allomorph selection (ˈputəɾ ‘butter’ ∼ ˈputɾa ‘-GEN.SG’, ˈʧɛbəɾ ‘bucket’ ∼ ˈʧɛbəɾa ‘-GEN.SG’), even though the forms with –ja would be well-formed in terms of phonotactics; both unattested forms *ˈputəɾja and *ˈʧɛbəɾja would have schwa in the phonotactically expected positions. Hence, the choice of –a instead of –ja in roots with schwa remains unexplained if we assume that schwa is underlying, but not if it is considered epenthetic. In derivational terms, allomorph selection applies first and blocks schwa epenthesis (along the lines of Hall et al. under review).

In synchronic terms thus both the second and third group of suffixes begin with a consonant and palatalization is unexpected with such suffixes.

Finally, the data in (3) show that not all suffixes beginning with i trigger palatalization.

An anonymous reviewer points out that one way to think of the distinction between triggering and non-triggering suffixes would be to say that there is they are of two classes of suffixes in terms of derivation, which would be consistent with other Slavic languages (for such an analysis of Russian, see Blumenfeld 2002, and for Polish, see Rubach to appear). This may well be the case, but the evidence would have to be particular to palatalization as there does not seem to be any other evidence for these two classes of suffixes. The palatalization-triggering suffixes are all derivational, some of which are nominal whereas other are verbal. In terms of stress, some shift stress, some shift it variably, but others do not. The non-triggering suffixes are all nominal. No inflectional suffix triggers palatalization. Non-triggering suffixes appear to be more recent, appearing in loanwords. In any case, the argument for the two classes of suffixes appears to be tied to palatalization facts alone.

When taken together, the data in this section demonstrate that palatalization is a morphologically conditioned process: only certain suffixes, apparently regardless of their phonological content, trigger palatalization.

2.2 Slovenian palatalization in the Slavic context

Slovenian resembles other Slavic languages in that it exhibits several palatalization-like processes. Common Slavic distinguished three kinds of palatalization and iotation (Shevelov 1964). Determining which of the three processes applied depended on the phonological triggers. Some front vowels triggered first palatalization, whereas others triggered the second or third palatalization. Iotation was triggered by the palatal approximant /j/. The palatalizations also differed in targets. While palatalization affected only velars, iotation affected most consonants. The resulting palatalized segments differ across Slavic languages.

To put the Slovenian facts in the broader Slavic context, let us consider palatalization-like processes in the most closely related language, Bosnian/Croatian/Serbian (henceforth, BCS). Both Slovenian and BCS evidently exhibit all three patterns, albeit with some significant differences. To start with, BCS has a larger set of coronals, which include apical and laminal postalveolars; all Slovenian postalveolars are apical. Velar palatalization in BCS involves the same targets, velars, that turn into postalveolars (Browne 1993; Morén 2006). However, unlike Slovenian, the suffixes that trigger palatalization in BCS all begin with a front vowel. In Slovenian, this is not the case; some suffixes begin with a back vowel or consonant. As such, only the BCS velar palatalization has a clear phonological trigger. Velar fronting in BCS is triggered by some i-initial suffixes, and the resulting consonants are alveolar. In this sense, BCS and Slovenian are identical. However, while BCS shows velar fronting with several suffixes, such as the masculine nominative, dative/instrumental/locative plural and imperfective, Slovenian velar fronting is productive only in the imperative; all other examples are limited to a small set of exceptional nouns (e.g. otɾɔk ‘child’ ∼ otɾɔts-i/coll. otɾɔk-i ‘-NOM.PL’). Iotation appears to be the most similar across the two languages, with the largest set of targets. The difference between the two languages is in the resulting iotized segments; BCS has a set of palatal sonorants, which Slovenian lacks. In short, while there are significant differences between the two languages, the overall facts are strikingly similar. Crucially for the discussion that follows, in both languages palatalization-like processes are morphologically conditioned.

When focusing on velar palatalization, what the two languages have in common is that the process is no longer completely phonological. Not all suffixes that begin with a front vowel trigger palatalization, as seen for Slovenian in (3). However, only Slovenian has suffixes that start with other segments, such as back vowels or consonants in Table 1. Because Morén (2006) posits that the trigger of palatalization is phonological, a front vowel i, his analysis cannot be straightforwardly extended to Slovenian. Another way of analyzing the Slovenian pattern would be to say that the triggering segment must be sufficiently abstract, such as a floating feature bundle (along the lines of Morén’s analysis of velar fronting) or a ghost segment (Szpira 1992). Such a segment can be found at the beginning of the palatalizing suffixes. The Slavic literature refers to these segments as yers, to mirror the Common Slavic schwa-like segments which have been deleted in late Common Slavic. For instance, Havlík’s Law states that most yers are deleted, except for strings of schwas in adjacent syllables where every even yer (counting from the right) is retained. Yer segments that are subject to deletion thus cannot be realized independently, but can palatalize the preceding root. Thus, yers could be thought of as underlying on the synchronic level. For such an analysis of Slovenian see Morén & Jurgec (2007).

The focus of this paper is variation in palatalization. As such, the ghost segment trigger could not account for the full set of facts. As we will see, the suffixes starting with i/j are better palatalizers than other palatalizing suffixes.² Furthermore, palatalization is less frequent in roots containing another postalveolar. Whatever the ultimate account of the triggering segment is, it will have to capture these tendencies. Because the morphological facts about palatalization are well established in the literature, this paper will not attempt a different analysis. I will instead assume that the set of triggering suffixes is lexically marked and the palatalization triggering constraint refers only to those suffixes. In short, the constraint PALATALIZATION is lexically indexed to specific suffixes. Lexical indexation is a common and well-established approach to capturing such morphological exceptionality (Pater 2000; 2007; 2009; Inkelas & Zoll 2007; Gouskova 2007; Flack 2007; Jurgec 2010). This makes sense since the focus of this paper is on the phonological factors that condition the variability of palatalization, not morphology.

2.3 Phonological characteristics

The literature already addresses some phonological characteristics of palatalization in Slovenian. In particular, only velars {k ɡ x} undergo palatalization and map to postalveolars {ʧ ʒ ʃ}. Nothing can be said about the triggering segments on the synchronic level, even though historically these triggers were all front vowels. Thus, the focus of this paper will be the phonological factors that affect palatalization. The main contribution of this paper is in demonstrating that variation in Slovenian velar palatalization is tied to purely phonological factors, such as the initial segment of the suffix, the phonotactic restrictions at the morpheme boundary and long-distance interactions between consonants in the whole word.

3.1 Data collection

To determine the scope of variation, I first searched for words ending in palatalizing suffixes in two dictionaries of Standard Slovenian: The Dictionary of Standard Slovenian (Bajec 2000) and The Slovenian Orthographic Dictionary (Toporišič 2001). The Dictionary of Standard Slovenian contains about 110,000 words and was published between 1970 and 1991. The Slovenian Orthographic Dictionary contains about 130,000 words and was produced in the 1990s. The searchable versions of these dictionaries are available online at http://bos.zrc-sazu.si. The dictionaries were produced using pre-digital techniques. Words with only a handful of records or words that have multiple overt derivational suffixes are less likely to be included (see Bajec 2000).

I searched for stems ending in velars and followed by palatalizing suffixes. Palatalization is reliably indicated in the orthography; variable words are counted as two separate entries. I made the following exclusions. First, I excluded null suffixes, which would require a full morphological analysis of a large part of the sources. This rules out two null suffixes ‘ABSTRACT’ and ‘VERBAL’ (Table 1).³ Second, I only included those words in which the relevant suffixes were the rightmost derivational suffixes within a word. The more morphologically complex the word is, the less likely it is to be included in the sources, regardless of the frequency. As such, morphologically very complex words are not adequately represented in the sources. Third, I excluded verbs as they failed to display any variation. The findings regarding the adjectival and nominal suffixes turned out to be entirely consistent with the verbal suffixes (e.g. i-initial suffixes display little variation). Finally, I only included the suffixes that appeared in the dictionary in at least 20 words. Two palatalizing suffixes, –tʋ(o) and –iʃʧ(e) ‘ABSTRACT’ in Table 1, failed to meet this threshold.

The words collected from the dictionary sources were then checked for frequencies in the text corpus Gigafida (Logar-Berginc et al. 2012). This text corpus contains 1.2 billion tokens from a variety of written sources, published between 1990–2011 (with very few sources dated before 1997). The corpus is partially morphologically parsed (stem-inflection boundary only), but not lexically disambiguated. Each item was checked for accuracy and homophonous items were manually discarded. For very frequent suffixes, an estimate of valid tokens was made on the basis of the first 200 hits.

The subsequent descriptive and inferential statistics was performed in R (R Core Team 2013).

3.2 Palatalization by suffix

The search turned out 9 suffixes with more than 20 pairs of word-types in the dictionaries. A type-pair consists of two words, one with palatalization and the other without (e.g. nɔɡitsa ∼ nɔʒitsa ‘leg-DIMINUTIVE’). Some dictionary words did not appear in the corpus at all. These type-pairs were excluded from analysis. As a result, one suffix, –nat, had only 17 type-pairs.

All in all, 612 type-pairs had at least one token in the corpus. There were 5.7 million tokens collected, which represents 0.5% of all tokens in the corpus, leading to the conclusion that the palatalizing suffixes are quite frequent. Of the 612 type-pairs, 447 or 73% exhibit palatalization in the majority of tokens. This represents 4.9 million tokens or an 86% share. While this suggests that very frequent suffixes display more palatalization, this is not the case statistically. The Pearson product-moment correlation coefficient is 0.072, indicating that type frequency plays little role in palatalization. Frequency of individual types will not be discussed further.

Individual suffixes differ in frequency from one another, ranging from 17 to 169 types, shown in Table 2. The most common suffix is –n which stands out with 169 types and almost 3.9 million tokens (over two thirds of all tokens). At the other extreme, –nat has a tenth of the types and less than one hundredth of the tokens.

The suffixes also vary in terms of how frequently they trigger palatalization. On the one hand, all 20 type-pairs with –iʧ trigger palatalization more frequently than not, and only 31 tokens lack palatalization. On the other hand, only 32 out of 86 (37%) type-pairs with –ts display palatalization in the majority of tokens.

For each type-pair, I calculated a palatalization ratio which is the share of palatalized tokens among all tokens for that type-pair. The higher the palatalization ratio, the more frequent the palatalization is within the type-pair. A palatalization ratio of 1.00 means that all tokens in the corpus were palatalized, whereas the ratio of 0.00 indicates a complete lack of palatalization.

The palatalization ratios across type-pairs and suffixes differ considerably (Table 3). The mean value for all data is 0.72, varying from 0.36 for –ts and –owje to 0.99 for –je. The median value for all data is 1.00, which means that the majority of types show palatalization. In fact, the median value of palatalization ratios for most suffixes is 1.00, with the exception of –owje (0.00), –je (0.02) and –itsa (0.27).

The Shapiro-Wilk test reveals that palatalization ratios are not normally distributed for any of the suffixes (p < 0.0001 for all suffixes). To get a better idea how the palatalization ratios are distributed for each suffix, I will visualize the data using the package Beanplot in R (Kampstra 2008). The function draws one bean per group of data. A bean consists of (i) a one-dimensional scatter plot, (ii) its distribution as a density shape, and (iii) a thick mean/median line for the distribution. The dashed line represents the overall mean/median (across all beans). Showing both mean and median appeared confusing, so I chose only the more informative measure of central tendency for each figure.

Figure 1 presents palatalization ratios by suffix, with mean values for each suffix. Each short horizontal dash represents the palatalization ratio of one type-pair. (The overall frequency is not represented here, so a rare type-pair is treated the same way as a very frequent type-pair.) The ranges with more type-pairs have a broader density shape. For example, –ts has mostly types with low palatalization ratios, which is indicated by a broader density shape and more dashes at the lower end. The same suffix also has quite a few type-pairs with palatalization ratios around 1.00, which is marked with a broader density shape at the top of the bean. Very few type-pairs have palatalization ratios around 0.5, hence the density shape is thin in that range. In short, the bean of the suffix –ts resembles the number 8, with many type-pairs at both extremes, but very few in the middle. The distribution is bimodal.

Figure 1

Palatalization ratios by suffix, with means. Vertical beans represent the density distributions, with short lines corresponding to stems. The longer lines (green) mark mean values for each suffix. The dotted line indicates the mean value across all suffixes.

Three other suffixes exhibit bimodal distributions: –itsa, –nat, and –owje. Combined with –ts, these suffixes have the lowest mean values. The remaining suffixes have mean palatalization ratios close to 1.00, with very few outliers.

These results indicate significant differences among the suffixes. Many suffixes almost invariantly trigger palatalization in all stems, and the remaining suffixes exhibit less palatalization. Only a small share of stems have ratios around 0.50, suggesting that the variation is not random. In the following sections, I will show that much of the variation can be attributed to phonological factors.

3.3 Palatalization by stem

The results so far suggest that individual suffixes differ significantly in their likelihood to trigger palatalization. This section looks at the overall results in stems.

Several stems appear with multiple suffixes that trigger palatalization. Of these stems, twelve appear with four or more suffixes, shown in Table 4. Individual stems show high degrees of similarity amongst varying suffixes: if a stem has a high palatalization ratio with one suffix, the same stem will have a high palatalization ratio with another suffix. For example, dlak ‘hair’ has palatalization ratios ranging from 1.00–0.83. Some roots show more variation. For instance, boɡ ‘god’ has low palatalization ratios with two suffixes (–ts = 0.12, –k = 0.08), but high palatalization ratios with the other three (–iʧ = 1.00, –tsa = 0.92, –ina = 1.00). It is likely that these differences are conditioned by the phonological properties of the suffixes rather than the stem itself.⁴

We can also group stems by the stem-final velar. Not all of them are equally common, which mirrors the general trends in the lexicon (Table 5).

Figure 2 presents palatalization ratios by stem-final consonant, with mean values for each consonant. The beans are horizontal, to make the distinction between stems and suffixes apparent. Each short vertical dash represents the palatalization ratio of one type-pair (henceforth, stem). The differences between the three groups of stems do not appear to be clearly distinct, with the mean only slightly higher for the stops when compared to the fricative.

Figure 2

Palatalization ratios by stem-final consonant, with means. Horizontal beans represent the density distributions, with short lines corresponding to stems. The longer lines (green) mark mean values for each suffix. The dotted line indicates the mean value across all words.

To test whether the effects of suffixes and stem-final consonants on palatalization ratios were significant, I fit the data into a linear mixed-effects model with the target velar and suffix as fixed effects, and with random intercept and slopes for stem, using the lmer package in R. Both fixed factors were dummy-coded, with target ɡ and suffix –ts serving as the reference levels.

The results for fixed effects are presented in Table 6. Significant predictors have t-values above 2 or below –2. When compared to the stem-final ɡ, only x is significantly different, while k is not. Most suffixes have rather high t-values when compared to the suffix –ts, which suggests that they affect the palatalization ratios significantly.

3.4 Interim summary

The data reviewed so far suggest that palatalization may be less variable than assumed at the outset of this paper. It may be, for example, that palatalization is triggered in most instances of a particular stem or suffix, but a particular phonotactic restriction may block palatalization in some combinations. These phonological restrictions will be the focus of the rest of the paper. Another source of variation is lexical, namely that each suffix has its own likelihood to cause palatalization, independently of its phonological shape (Pater 2009; Gouskova & Linzen 2015). While the ultimate analysis will need to account for this second source of variation, lexical variation will not be the focus of the analysis in this paper. The vast majority of variation will be shown to be phonological.

4.1 Triggers

Languages with phonological palatalization reveal two asymmetries regarding the triggers. First, high vowels are better triggers than non-high vowels (Chen 1973; Bhat 1978; Bateman 2007; 2011; Kochetov 2011). In many languages only high vowels trigger palatalization, and in languages in which non-high vowels trigger palatalization, high vowels are also triggers.⁵ Second, front vowels are better triggers than back vowels. That is, if a back vowel triggers palatalization in a particular language, front vowels will also be triggers.

As we have seen in section 2, Slovenian palatalization is not purely phonological; there is no phonological property common to the triggering suffixes. Recall that the palatalization triggering suffixes fall into three groups (Table 1). One group of suffixes starts with typical phonological triggers (–itsa, iʧa –ina, –je), while the rest do not. When pooled together, the suffixes with front vocoids do not differ from other suffixes. However, this is due to the effect of one single suffix, –itsa, which is much more variable than all other suffixes. There are phonological reasons why this suffix behaves differently than other i-initial suffixes, and this issue will be addressed in section 5.1. Once –itsa is excluded, front suffixes appear to be much stronger triggers than other suffixes, as shown in Figure 3.

Figure 3

Palatalization ratios by suffix-initial segment, with means. Horizontal beans represent the density distributions, with short lines corresponding to stems. The longer lines (green) mark mean values for each suffix. The dotted line indicates the mean value across all words.

While most suffixes have many stems with low palatalization ratios, front suffixes almost invariantly trigger palatalization in all stems. The few exceptions are shown in Table 7. Only four such words exist, with two having palatalization ratios below 0.90. The word with the lowest palatalization ratio is a predictable exception, as it contains another postalveolar—see section 5.2 for further discussion. In contrast, suffixes beginning with a segment other than a front vowel or glide have many words with low palatalization ratios. This asymmetry between the two sets of suffixes suggests that Slovenian mirrors the cross-linguistic tendencies—despite the fact that velar palatalization in Slovenian is morphologically conditioned.

4.2 Targets

Cross-linguistically, palatalization is more likely to affect velar consonants, followed by coronals and labials (Chen 1973; Bhat 1978; Bateman 2007; 2011; Kochetov 2011). This preference is immediately clear in Slovenian, since palatalization affects only velars. Other palatalization-like patterns in Slovenian affect a larger set of consonants. For instance, iotation affects all consonants (2), yet the output of palatalization is different for different places of articulation, with velars and most coronals becoming postalveolars (telet-a ‘calves’ ∼ teleʧ-ji ‘calf-ADJ’) while labials exhibit l-epenthesis (ɡlob-ɔk ‘deep’ ∼ ɡlobl-ji ‘deeper’), thus preserving the underlying place of articulation.

Other target asymmetries are less clear. There is a cross-linguistic restriction on palatalization of rhotics (Walsh Dickey 1997; Hall 2000; Hall & Hamann 2010), but other natural classes do not seem to display any clear trends. For instance, obstruents and sonorants (or stops and fricatives) may display different behavior, but there is no uniformity.

Stem-final consonants display significant differences across all suffixes, with the fricative x showing less palatalization than the stops {k, ɡ}, as seen in Figure 2. This tendency remains even after we account for all other interactions; phonotactically neutral n-suffixes mirror the general tendencies, as shown in Figure 4.

Figure 4

Palatalization ratios for the two suffixes without clear local interactions (-n ‘DIMINUTIVE’ and –nat ‘ADJECTIVE’) by stem-final consonant, with means. Vertical beans represent the density distributions, with short lines corresponding to stems. The longer lines (green) mark mean values for each suffix. The dotted line indicates the mean value across all words.

For these two suffixes, there are no clear phonotactic grounds that could explain the differences between the stem-final consonants. That is, these suffixes do not contain overt vowels that would increase the likelihood of palatalization or have clear local interactions between the consonants at the stem-suffix boundary. This general tendency is language-specific.

In other words, not all segments are equally likely to undergo palatalization. Two tentative explanations are possible. On the one hand, it could be that some velars are more marked in stem-final position compared to other velars. The problem is that there is no independent evidence for such markedness. On the other hand, it could be that some segments are more resistant to palatalization than other segments. In the language of Optimality Theory, some segments violate more faithfulness constraints than other segments.

A particular caveat concerning Slovenian velar palatalization is an asymmetrical obstruent inventory (Table 8). Slovenian has no postalveolar stops and velar affricates; both postalveolar and velar fricatives are attested.

This asymmetry leads to different mappings between inputs and outputs in palatalization (4). While the velar fricative maps to the postalveolar fricative while preserving voicing, stops cannot map to stops. Instead, the voiceless stop maps to the voiceless affricate. Voiced velars do not have the same option, as there are no voiced postalveolar affricates. Instead, voiced velars map to voiced postalveolar fricatives.

If we make use of an SPE-style representational model, the mapping x → ʃ will incur the fewest violations of faithfulness constraints, as the two segments are more similar than the other pairs. This fact alone suggests that palatalization of x is more likely than of the other two segments, which is contrary to the data (Figure 4). This, perhaps surprising mapping of the voiced stop ɡ to the fricative ʒ is also found in other Slavic languages. These data have been used to support an analysis based on hierarchical contrastive models, such as in Radišić (2009). Under this approach, ɡ is not specified for the feature [continuant]. These small differences among the stem-final consonants are left for further research.

4.3 Triggers and targets

The final piece of the data concerns cases in which palatalization depends on segments on both sides of the morpheme boundary. I will discuss one particularly clear example, but the generalizations extend to other similar cases.

Consider the suffix –k ‘DIMINUTIVE’. Figure 5 shows that this suffix mirrors the general tendency in the lexicon, although at a more extreme level. The stops show all but invariant palatalization, whereas a sizable share of x-final stems has low palatalization ratios.

Figure 5

Palatalization ratios for –k ‘DIMINUTIVE’ by stem-final consonant, with means. Horizontal beans represent the density distributions, with short lines corresponding to stems. The longer lines (green) mark mean values for each group of stems. The dotted line indicates the mean value across all words.

There is a single outlying ɡ-final stem with a palatalization ratio other than 1.00: /boɡ/ at 0.08. When this stem is followed by the suffix –k, the resulting non-palatalized word contains a sequence of two identical consonants (the stops are both voiceless and released), which is exceedingly rare in Slovenian and never tautomorphemic. However, this stem is exceptional in other ways and limited to a single idiom bok-k-ow kot ‘corner with a cross’. As such, its phonological exceptionality is not surprising.

Identical adjacent consonants are very rare in Slovenian, with the only instances limited to the boundaries between root and derivational affixes. There are several active processes that resolve such sequences. For instance, the clitic k ‘to (directional)’ dissimilates to x when followed by a word starting with a velar stop (5).

Why is the restriction on identical adjacent consonants relevant to palatalization? When stems ending in velar stops {k ɡ} are followed by k-initial suffixes, the resulting non-palatalized word contains a highly marked sequence. Palatalization turns velar stops into postalveolars, which avoids sequences of identical segments. For example, *otɾɔk-k-a ‘child-DIM-NOM.DUAL’ contains an illicit sequence which is resolved by palatalization: otɾɔʧ-k-a. In this sense, palatalization is facilitated by the phonotactic restrictions in Slovenian, resulting in more palatalization for stops. No facilitation exists for the fricative x, as strings xk are well-formed in Slovenian.

An anonymous reviewer notes that Slovenian is not alone in this restriction. The Common Slavic diminutive *-ъk triggered velar palatalization, whereas the suffix *-ъk ‘similar to stem’, which has a back yer, would not. However, *-ъk would become *-ъk after velars, and thus triggers palatalization.

Similar interactions between adjacent segments are found with the suffixes –ts and –owje. In a nutshell, the likelihood of palatalization depends on the underlying voicing in the case of –ts and whether the resulting target is an affricate in the case of –owje. Like in the examples above, these patterns can be supported by the phonotactics of Slovenian.

4.4 Interim summary

Slovenian palatalization displays several interactions between segments at the stem-suffix boundary. First, suffixes containing front vowels and glides are more likely to trigger palatalization than other suffixes, matching the cross-linguistic tendencies. Second, palatalization is less likely with the fricative x compared to the stops {k, ɡ}. Finally, palatalization can be conditioned by the phonotactic restrictions applying at the morpheme boundary.

These asymmetries can be formalized in both rule-and constraint-based frameworks. In a rule-based approach, adjacency is specified in the environment of the rule. In a constraint-based approach, constraints refer to strings of segments. Because the data contain a level of inherent variation that cannot be attributed to lexical factors, I will be using a specific constraint-based model (section 6).

5.1 Suffixes

Two suffixes stand out from the rest: –ts and –itsa (see Figure 1). These suffixes display the highest amount of variation, with many words having palatalization ratios between 0.2 and 0.8, whereas most of other suffixes have very few words in that range. Why is this the case? One possible answer is that variability is a lexical property of these two suffixes. The alternative would be that the reason is phonological, related to the special status of ts in Slovenian. Both affricates in Slovenian are quite restricted phonotactically, such that complex onsets and codas with at least one affricate are rarer than expected by chance. Once we consider heterosyllabic sequences, there is a variety of restrictions. Alveolar sibilants, including ts, cannot appear immediately before postalveolars even across word boundaries, assimilating in place to the following postalveolar (e.g. kɔnəts leta ‘end of year’∼ kɔnəʧ ʃole ‘end of school’). There is evidence that these restrictions extend across vowels. Slovenian has variant sibilant harmony, which makes anterior sibilants posterior when followed by another posterior sibilant within a word (Jurgec 2011): sliʃi → ʃliʃi ‘hears’. Across a single vowel, two alveolar affricates (29 words in Toporišič 2001) or an alveolar affricate plus a postalveolar (7) are extremely rare. In contrast, velars can freely co-occur with both affricates.

These restrictions also explain why the suffix –itsa behaves differently than other i-initial suffixes, which all invariantly trigger palatalization (section 4.1). That is to say, ts can freely occur with velars but not postalveolars.

5.2 Stems

Section 4.1 revealed that i-initial suffixes are better triggers of palatalization than other suffixes, while section 5.1 discussed why ts inhibits palatalization. Because of these two competing tendencies, the suffix –itsa displays a great deal of variation. In this section, I examine other variables that influence the likelihood of palatalization with this suffix.

I focus on different consonants appearing in the stems. Two groups of consonants are large enough to be representative and display unique behavior—velars and postalveolars. The velar group are stems that contained at least two velars: in addition to the final velar in the stem, there is another non-adjacent velar. The postalveolar group contains an underlying postalveolar that is not adjacent to the final velar of the stem.

Consider first stems with multiple velars, for example knjig-itsa ∼ knjiʒ-itsa ‘booklet’. The velars do not need to match in terms of voicing. The non-final velar could appear anywhere in the root, except as one of the last two segments, ruling out any local interactions with the stem-final segment.

Figure 6 presents a consolidated beanplot. For each stem-final consonant, the bean is split in half. The lower part presents the stems without non-final velars, whereas the upper part presents stems with such velars. Palatalization is less likely in stems with multiple velar stops. The corpus does not contain enough velar fricatives to generate the beanplot or make a reliable generalization.

Figure 6

Palatalization ratios in stems with multiple velars for the suffix –itsa ‘DIMINUTIVE’ by stem-final consonant, with median values. Horizontal beans represent the density distributions, with short lines corresponding to stems. Each bean is split in half, with the upper half representing stems with multiple velars and the lower half representing all other stems. The longer (red) lines mark median values for each group of stems. The two dotted lines indicate the median value across all words, with the lower median value corresponding to stems with multiple velars, and the upper median value corresponding to stems with a single, final velar.

A much stronger generalization can be drawn from the second case, namely stems with another postalveolar, which are shown in Figure 7. In this case, the asymmetry between stems with no postalveolars and the ones with a postalveolar (at least two segments from the right edge of the stem) is striking. Palatalization does not apply in stems with another postalveolar, without exception.

Figure 7

Palatalization ratios in stems with a non-final postaleolars for the suffix-itsa ‘DIMINUTIVE’ by stem-final consonant, with means. Horizontal beans represent the density distributions, with short lines corresponding to stems. Each bean is split in half, with the upper half representing stems with a non-final postalveolar and the lower half representing all other stems. The longer (green) lines mark mean values for each group of stems. The two dotted lines indicate the mean value across all words, with the lower mean value corresponding to stems with non-final postalvelars, and the upper mean value corresponding to stems without postalveolars.

Such convincing long-distance blocking of palatalization is surprising, as these sort of long-distance interactions are cross-linguistically rare. In what follows I provide further examples of such patterns. In fact, the Slovenian data provides evidence to draw a parallel between some well-known sound patterns.

5.3 Implications for the typology of Derived Environment Effects

To understand the significance of long-distance interactions in Slovenian palatalization, I now discuss the pattern in broader perspective of morphologically conditioned sound patterns.

Slovenian velar palatalization can be characterized as a Derived Environment Effect (DEE; Kiparsky 1973; Mascaró 1978; Wolf 2008). DEEs are alternations that appear in morphologically complex forms, typically at the morpheme boundary, but not within morphemes. To illustrate, consider a subset of the Slovenian pattern in (6).

To keep this illustration brief, I will simplify the pattern by taking the sequence ki as representative for velar palatalization as a whole. The sequence ki is possible within morphemes (6-a). At the morpheme boundary, ki is generally not possible, resulting in palatalization (b).

Slovenian palatalization is a local DEE, applying at the morpheme boundary. Long-distance DEEs are much rarer, but a few cases have been reported (Kurisu 2008; Jurgec 2014; Gouskova & Linzen 2015). In Tagalog, for instance, f is possible in some underived words (7-a). When prefixed, initial f changes to p (b), mirroring the local pattern observed in Slovenian. Suffixes—perhaps surprisingly—have the same effect (c). In fact, f appearing anywhere within the root alternates with p in derived words.

DEEs do not need to invariantly apply, but can be blocked by other phonological processes. This blocking is typically local. An example comes from German (8). Sequences tj assibilate to tsj at the morpheme boundaries (a), but not if the sequence is preceded by a sibilant (b), even though sts is possible morpheme internally (c).

The Slovenian palatalization data present a previously unnoticed case of local DEEs that are blocked at a distance. The generalization is that postalveolars and velars can freely cooccur within roots, and that palatalization can create additional postalveolars. However, palatalization is categorically blocked when the root contains another postalveolar (and to a lesser degree, another velar). More data is provided in (9).

This interpretation of the data fills the gap in the typology of DEEs and their blocking (Table 9). DEEs can be local and long-distance. Local DEEs can be blocked locally or long-distance. Slovenian palatalization is a locally triggered DEE that is blocked long-distance. Blocking in long-distance DEEs has not been studied yet. All in all, the Slovenian data fill the typological gap in DEE blocking.

5.4 Similar patterns in other languages

The Slovenian data fit the cross-linguistics typology of DEE blocking. A related question is whether there are any other similar patterns.

Slovenian palatalization involves creating a postalveolar, but not when there is another postalveolar in the stem. This characterization of DEE blocking resembles dissimilation. While dissimilation often applies to adjacent segments, many cases involve some larger domain. In Tashlhiyt Berber, for instance, a prefix m- dissimilates to n- as long as the following root contains a labial, even if root-final (El Medlaoui 1995; Odden 1994; Alderete 1997). The Slovenian case is unusual, because there is no overt dissimilation within stems. In fact, there are many instances of postalveolars within the stem (9). The crucial generalization is that sequences of postalveolars cannot be created by palatalization, whereas there are no restrictions applying to underlying postalveolars.

Unlike Tashlhiyt Berber, Slovenian distinguishes between underlying and morphologically derived segments of a particular kind. Another such language is Japanese.⁷ Japanese Rendaku involves voicing of root-initial obstruents in compounds (10). Non-initial stems in certain kinds of compounds normally show voicing (a). However, voicing does not apply in roots with another voiced obstruent anywhere in the stem (b). This blocking pattern is known as Lyman’s Law.

Itô & Mester (1998, 2003) consider Rendaku voicing a case of DEE. Lyman’s Law is a case of DEE blocking at a distance: underlying and derived voiced obstruents are well-formed, but not their combination. This matches the Slovenian data, except that different morphological constituents and phonological features are involved. A similar process is Korean compound tensification (Zuraw 2011; Ito 2012).

However, Japanese is not identical to Slovenian DEE blocking. First, Japanese involves laryngeal features, whereas Slovenian involves place features. Second, Japanese generally allows at most one voiced obstruent per root in native (Yamato) words. Slovenian, on the other hand, permits multiple underlying postalveolars within a word, either tauto- or heteromorphemic (9). It is only that a postalveolar cannot be derived by palatalization if the stem already contains another postalveolar. In this sense, Slovenian is a case of DEE blocking without a general restriction on multiple underlying postalveolars. To the best of my knowledge, no other language exhibits such a pattern.

6.1 Constraints

I propose several constraints to capture the empirical generalizations about palatalization in Slovenian. I start with the local constraints (these data are reviewed in section 4), followed by long-distance consonant co-occurrence constraints (section 5).

As we have seen in section 4.1, suffixes beginning on front vowels and the glide j are better triggers than other suffixes. This asymmetry can be captured by splitting the markedness constraint preferring palatalization, PAL, into several constraints. Rubach (2003) proposes that each palatalization constraint penalizes subsets of segments with respect to backness. For instance, the constraint against the sequence of a back consonant followed by a high front vowel is universally ranked higher than the constraint against a back consonant and a mid or low vowel: PAL-i >> PAL-e >> PAL-æ. Since the Slovenian suffixes here fall into two groups, only two constraints are required (11). I propose a specific constraint against front vowel sequences (a) and a more general constraint which prefers palatalization before any suffix. These two constraints are in a stringency relation: a front vowel suffix can violate both constraints, whereas other suffixes can only violate PAL. If PAL-i has a non-zero weight, this would confirm an asymmetry between high vowel and other suffixes.

It is worth noting that PAL is satisfied regardless of what representational assumptions we make about the ultimate source of palatalization. One option is that the source is a ghost or similar underlying segment (Halle 1959; Szpira 1992; Gussmann 2007). In autosegmental terms, the trigger could be a floating feature. Regardless of which of these options we take, PAL constraints will be satisfied when palatalization applies. In addition, the following surface segment will also have an effect, as evident by the corpus data.

Palatalization violates several IDENT constraints. Recall that the mappings in palatalization are asymmetrical. In particular, the voiceless velar stop k becomes the postalveolar affricate ʧ, whereas the voiced velar stop ɡ becomes the postalveolar fricative ʒ, as shown in (4). Because the featural representations of the triggers have been controversial in the literature (Morén 2006; Radišić 2009), I will model the data using a single IDENT constraint.⁸

The final local constraint penalizes adjacent identical consonants (12).

The second set of constraints will account for long-distance interactions in the data. As we have seen, these interactions resemble consonant harmony and dissimilation, as there are restrictions on what kind of consonants can cooccur within a stem or word. One standard account of consonant harmony and dissimilation is Agreement-by-Correspondence (Rose & Walker 2004; Hansson 2001; Bennett 2013/2015). According to this approach, consonants can be in correspondence with one another. Correspondence is a formal relation between segments, which is referred to by constraints. The constraints preferring consonant harmony—that is, similarity between consonants—are of two kinds: CORR(ESPONDENCE) constraints make sure that specific consonants are in correspondence, whereas CC-IDENT constraints require the corresponding consonants to be similar. The harmonized candidates have agreeing consonants in correspondence, but non-harmonized candidates can have (disagreeing) consonants in correspondence or not. Thus, there are two ways of representing identical, non-harmonized forms. Given that there is no way to tell these apart in the corpus, one could stipulate that both are equally likely. However, this arbitrary decision would significantly affect the weights of CORR and CC-IDENT.

Hansson (2014) proposed an alternative model, Agreement-by-Projection (ABP). In this approach, markedness constraints penalize sequences of segments of a particular kind on some tier (13). An example of this constraint is *[+anterior][–anterior]_[+coronal] which penalizes anterior coronals followed by posterior coronals, as long as no other coronal intervenes.

While Hansson’s proposal is relatively recent, it has several crucial advantages over Agreement-by-Correspondence. The main advantage is formal simplicity. ABP does not require any abstract representational relationships, such as correspondence; phonetically identical candidates are not represented in more than one way. The new projection constraints (13) achieve the effect of two families of constraints—CORR and CC-IDENT—at the same time. Other advantages are empirical. A sequence of three segments with [F] violate the constraint in (13) at most once, whereas CC-IDENT[F] could be violated three times (once for each segment-pair). The additional violation marks create pathologies (Hansson 2007). For these reasons, I will adopt Hansson’s ABP. The proposed constraints penalize co-occurrence of two specific consonants within a domain; I leave the formal implementation in terms of features and tiers for further research.

Section 5 reviewed three types of long-distance interactions found in the data. The strongest restriction was that non-stem-final postalveolars block palatalization of the final velar, which is just another way of saying that a postalveolar cannot be followed by another postalveolar within the stem (14-a). Any two postalveolars within a stem will violate the constraint *Š…Č, regardless of their distance. I assume that a derived postalveolar is counted as being a part of the stem, even though the place features of such a postalveolar originate from the suffix. Next, velars cannot be followed by postalveolars within the stem (14-b). The final restriction concerns suffixes containing ts; such suffixes have lower palatalization rates when compared to other similar suffixes. The constraint that mirrors this effect is *Š…C (14–c). Crucially, the suffix is included in the domain of the constraint.

The constraints in (14) can be satisfied by changing one or the other of the two consonants, and only one of the options is attested in the data. To distinguish between the two repairs, we need another constraint. As it turns out, this constraint needs to mirror the difference between changing a velar into a postaveolar (attested) and vice versa (not attested). There is a variety of ways to define this constraint, either by splitting up IDENT(place) into two more specific constraints or by introducing a MAX constraint, such as MAX[+posterior] or MAX[+coronal] (Morén 1999/2001). I will take the latter. Other alternatives include positional licensing referring to the stem-final position (Zoll 1998; Crosswhite 2001; Walker 2001, 2011).

In addition to the constraints just mentioned, I also included MAX (≡ No deletion) for some instances where palatalization resulted in deletion of a penultimate consonant at the end of the stem; such instances were rare, but attested. Finally, I also considered several other long-distance constraints. Among them were the constraints against two affricates and two velars; none of these constraints had weights larger than 0.0.

6.2 Weights

The corpus frequencies and violation profiles were fed into the MaxEnt grammar tool (Hayes & Wilson 2008), which assigns weights that fit the data best. In addition to the type-pairs mentioned throughout this paper, I also included those unattested candidates (corpus frequency = 0) in which postalveolars become velars. The model learned the weights shown in Table 10.

The highest weighted constraints are the ones that are violated by the least attested (and most unattested) forms—MAX[+cor] and *C_iC_i. The weight of PALATALIZATION is 1.6, while PALATALIZATION-i is weighted at 3.4, suggesting that i-suffixes are much more likely to trigger palatalization than other suffixes. Of the long-distance constraints, *Š…Č has the highest weight, indicating that multiple postalveolars within a stem are strongly dispreferred; the effect is much lower for velars followed by postalvolars, and the constraint *K…Č has a marginal weight of 0.2. The weight of the constraint *Š…Č is also rather high at 3.4, indicating that suffixes containing ts are considerably less likely to trigger palatalization than other suffixes. The faithfulness constraints have a null weight. This suggests that palatalization generally obtains in the data. Furthermore, the absence of palatalization is not due to IDENT but other markedness constraints, particularly long-distance consonant co-occurrence constraints.

Let us look at these effects more closely by examining individual tableaux. MaxEnt tableaux differ from classic OT tableaux in that constraints are weighted, with the weights decreasing from left to right. Each candidate comes with frequency in the corpus (indicated by the percentage next to the output). The sum of violations can be found in the penultimate column, indicated by ‘ℋ’, and the probability estimate is in the final column (‘p’). In a perfect model, the corpus frequencies and predicted frequencies match.

First, consider a stem without non-final velars or postalveolars (15). When followed by a front suffix palatalization invariably obtains, which is due to PAL-i. In fact, the weight of this constraint is so high that—absent any alveolar affricates within the suffix—all stems show palatalization, as in candidate (b).

The same stem may be followed by another kind of suffix (16). The corpus frequencies for the palatalized candidate (b) are lower, which is attributed to the fact that Pal-i is not violated by neither candidates. The predicted frequencies are somewhat different than the ones found in the corpus, even though the general tendency is clear. It is reasonable to assume that some variation is lexical and could not be accounted for in this purely phonological MaxEnt grammar. I leave the lexical factors for further research.

Sequences of identical consonants are extremely rare in Slovenian, and in the present corpus. The constraint *C_iC_i has a high weight, which effectively facilitates palatalization (17). The most frequent candidate (b) does not violate any constraints with non-null weights. Note that schwa was not considered in these cases, as its distribution is opaque (see section 2.1 for further information).

The remaining tableaux illustrate the long-distance interactions. As we have already seen, forms without a velar or postalveolar in non-final positions generally show palatalization (18). This is particularly so for i-initial suffixes, even though the suffix may contain a ts, which somewhat lowers the probability of palatalization.

The effect of long-distance constraints can be seen in forms with another velar or postalveolar within the stem (19). In this case, the most frequent candidate (a) lacks palatalization. The palatalized candidate (b) violates *Š…Č. The model predicts that these forms are rare, but attested. What is not attested is depalatalization of non-final segments (c), which would violate MAX[+cor].

The model does not match the data perfectly, and this is related to the fact that sequences of postalveolars are well-formed if underlying. Furthermore, –itsa may not be entirely representative for all other other suffixes, particularly because stems with a non-final postalveolar are infrequent. I also ran an alternative model in which all palatalization constraints were specific to each suffix. In this case, the weights were as follows: MAX[+cor] 16.0, *Š…Č 5.2, *K…Č 2.6, PALATALIZATION_-itsa 2.3, and IDENT 0.0. This grammar correctly predicts blocking in stems with another postalveolar, that is a 0% probability for candidate (19-b).

This concludes the MaxEnt analysis of variable palatalization in Slovenian, which captures both local and long-distance asymmetries by relying on purely phonological constraints.