2.1 Dictionary study

For this study, a list of 2219 nouns was painstakingly collected from Toporišič's (Reference Toporišič2001) dictionary, which is not publicly available online. The list included all nouns with mobile stress and a random sample of nouns with faithful stress, up to about 300 items for each stressed vowel quality in the nominative singular. The list is available in the online supplementary materials. Nominal stems are mostly short in Slovenian (Jurgec Reference Jurgec2019), and this is true in this dataset as well: monosyllabic 42%, disyllabic 51%, trisyllabic or longer 7%.

Of this list, there are 359 items (16%) with stress on the penultimate syllable of the stem in the nominative. These are presented in Table I. Stem-final stress in the nominative accounts for 83% of the items, and fewer than 1% have antepenultimate stress in the nominative. Table I shows that among words with stress on the stem's penultimate syllable in the nominative, mobile stress is most common with mid lax vowels, fairly common with [a] and very rare with tense vowels.

Table I Dictionary study: mobile stress is most common with [ɛ ɔ].

To quantify the effect of vowel height in the lexicon, we fitted a logistic regression model to the nouns with penultimate stress in the nominative, with mobile vs. faithful stress as the dependent variable, using the glm function in R (R Core Team 2016). Vowels were coded as either marked (mid lax) or unmarked (low or tense), and the unmarked vowels were further divided into low or tense, for a total of three vowel-height categories. Following the recommendation of Gelman & Hill (Reference Gelman and Hill2007: §4.2, §5.5), vowel height was centred using Helmert coding. This coding created two binary predictors: marked vs. unmarked, or more descriptively, Mid lax vs. Low or tense, with values of ⅔ for the mid lax vowels [ɛ ɔ] and ―⅓ for the low or tense vowels [i e a o u], and Low vs. Tense, with values of +½ for low [a], ―½ for tense [i e o u] and zero for mid lax vowels. The model, given in Table II, shows that mobile stress is significantly more common with mid lax vowels than with other vowels, and significantly more common with low [a] than with tense vowels. Masculine gender accounts for the majority of the 359 penultimate stress items (78%), and for an even greater majority of the 72 mobile stress items (83%), leading us to make all nonce words in §3 masculine. To quantify the effect of gender in the lexicon, a binary gender predictor (masculine vs. feminine or neuter) was added to the model in Table II. Gender did not make a significant improvement to the model (χ ²(1) = 0.05, p > 0.1), so this predictor was omitted.

Table II Dictionary-data regression model: mobile stress is significantly more common with mid lax vowels.

A reviewer correctly points out that the lexicon we use overrepresents the mobile stress items, since we collected all of them, but have only a sample of the fixed stress items. To measure the effect of this overrepresentation, we simulated lexicons with increasing numbers of fixed stress items. The estimates (β) of the predictors stay exactly the same, with slightly smaller SE and thus slightly larger z-values and slightly smaller p-values. The only difference in the model is in the intercept, whose estimate decreases with the number of fixed stress items. Therefore, the overrepresentation of mobile stress items does not affect the conclusions we draw from the model.

Of the 72 mobile stress items in Table I, 71 are disyllabic; the one exception is the trisyllabic [peˈtɛlin ~ peteˈlina] ‘rooster’. It would be highly accurate, then, to say that mobile stress shifts from a stem-initial syllable to a stem-peninitial syllable, and this is indeed the analysis we propose. We will claim that productive stress shift is only learned from the 56 items in which stress moves away from an initial mid lax vowel, which compete with the eleven items in which it does not (see Table I). No generalisation is learned from the 16 remaining nouns in which stress shifts from vowels that are either not mid lax or not initial: [ˈtɾebux], [peˈtɛlin] and the 14 items with [a]. These are lexicalised or fossilised, and do not contribute to the productive grammar.

To summarise, the proposed analysis categorises Slovenian stems into three groups: stems with initial stress, which may be trochaic or monosyllabic, and stems with non-initial stress, most of which are iambic disyllables.

2.2 Real-word survey

To ensure that the dictionary data reflects current usage among Slovenian speakers, an online survey was carried out. Speakers were auditorily presented with nominative–genitive paradigms, and they were asked to judge their acceptability. The results confirmed those of the dictionary study in §2.1.

2.2.3 Procedure

Participants accessed the experiment online, using Experigen (Becker & Levine Reference Becker and Levine2020), with the device and browser of their choice. To keep the survey acceptably short, the server selected a random sample of 38 items for each participant: 16 with [a], 16 with a mid lax vowel, four with a mid tense vowel and two with a high vowel.

After a short written introduction in standard Slovenian, items were presented on the screen, one at a time, following the format in Fig. 1. The nominative form was presented orthographically at the top of the screen, and a button appeared on the next line. When pressed, two words were presented auditorily: the nominative followed by one of the two possible genitives (faithful or mobile), and then three new buttons appeared, labelled v redu ‘ok’, ni v redu ‘not ok’ and ne vem ‘I don't know’. Once one of the three buttons was pressed, the other paradigm was presented auditorily: the same nominative followed by the other genitive. The order of genitives was randomised and counterbalanced. Once both paradigms were evaluated, the following message appeared: Imate v zvezi s to besedo kakšen komentar? (neobvezno) ‘Do you have any comment on this word? (optional)’, before the participant moved to the next screen.

Figure 1 Sample item (zakon ‘law’) used in the real-word survey.

At the end of the experiment, participants were asked to volunteer demographic information, as explained above.

2.2.4 Results

Participants accepted (i.e. selected ‘ok’) one of the two genitives in 83% of the trials, i.e. 83% of the trials were informative. They accepted both genitives in 13% of the trials and neither genitive in 4% of the trials. To simplify the analysis, we collapse the ‘not ok’ responses and the ‘I don't know’ responses; ‘I don't know’ accounted for fewer than 3% of the responses. The raw results are available in the online supplementary materials.

Figure 2 shows mean differences in the acceptability of the two nominative–genitive paradigms by item. The by-item results are reported in Appendix A. The y-axis shows the preference for mobile stress (the difference between acceptance of the mobile and faithful genitives), with higher values indicating a stronger preference for mobile stress. Overall, mobile stress was preferred for mid lax vowels, and faithful stress for all other vowels, with [a] closer to the tense vowels. These results thoroughly confirm the dictionary study in §2.1. Speakers of central dialects accepted mobile stress with mid lax and low vowels slightly more often than speakers of other dialects, and slightly less often with tense vowels.

Figure 2 Difference in acceptability between mobile and faithful stress in real words (92 participants, 163 nouns). Error bars are 95% confidence intervals.

For the statistical analysis of these results, a mixed-effects logistic regression was fitted to the acceptance of the mobile genitive as the dependent variable, using the glmer function from the lme4 package (Bates et al. Reference Bates, Mächler, Bolker and Walker2015) in R. All of the trials were included, regardless of the acceptance of the faithful genitive. Vowel height was Helmert coded exactly as in §2.1. Speaker dialect was coded as a centred binary predictor, with the value +0.47 for central dialects and ―0.53 for other dialects. Following the method described in Barr et al. (Reference Barr, Levy, Scheepers and Tilly2013), an initial fully crossed model was fitted with the two predictors and their interaction, as well as maximal random slopes for participant and item. Since this model did not converge, random slopes were removed one at a time; all random slopes had to be removed before arriving at a model that converged. This model is shown in Table III.

Table III Real-word regression model (92 participants, 163 nouns).

The model confirms that the mobile genitive is significantly more acceptable with mid lax vowels than with other vowels, and with the low vowel [a] more than with the tense vowels.

Dialectal differences were small and insignificant. Speakers of central dialects have a vowel system that is similar to the vowel system of standard Slovenian, and have quality-driven stress shift in their non-standard dialect. Speakers of non-central dialects have different vowel systems, which are either smaller or do not correspond to the standard system; furthermore, some non-central dialects lack stress shift. Despite all of this, the absence of a significant effect suggests that speakers were largely accessing their knowledge of standard Slovenian, the prestigious dialect of the media and education, regardless of their native dialect. For the relatively young participants in the study, this is to be expected, given the increase in fluency in standard Slovenian since the 1950s. Mastery of standard Slovenian is also high among older speakers; in Jurgec (Reference Jurgec2019) even the oldest, most fluent speakers of local dialects without mobile stress appeared to have easy access to the mobile stress patterns of standard Slovenian.

2.3 Laxness alternations under stress

The same tense–lax alternation that accompanies stress shift is also observed in nouns with fixed stem-final stress. In these nouns, exemplified in (5), a mid lax vowel may be either retained faithfully, as in (5a), or tensed, as in (5b). Some nouns display variability, e.g. [ˈʃkɔf ~ ˈʃkɔfa, ˈʃkofa] ‘bishop’. This type of vowel alternation suggests that mid lax vowels are generally tolerated better in the nominative singular, but they are ideally eliminated in the rest of the paradigm, even under stress.

1. (5)

   

The prevalence of laxness alternations under stress is summarised in Table IV, using the dictionary data from §2.1. The alternation affects the majority of polysyllabic words, but most monosyllables are spared. We attribute this protection of monosyllables to faithfulness to the initial syllable (Beckman Reference Beckman1997, Reference Beckman1998, Casali Reference Casali1998, Barnes Reference Barnes2006, Becker Reference Becker2009, Steriade Reference Steriade, Hanson and Inkelas2009, Becker et al. Reference Becker, Ketrez and Nevins2011, Jesney Reference Jesney2011, Becker et al. Reference Becker, Nevins and Levine2012, Becker & Gouskova Reference Becker and Gouskova2016, Becker et al. Reference Becker, Jurgec, Kehrein, Köhnlein, Boersma and van Oostendorp2017). In monosyllabic stems such as [ˈsʋɛt ~ ˈsʋeta], the [ɛ ~ e] alternation is in the initial syllable, and therefore violates initial syllable faithfulness. In polysyllables such as [uˈspɛx ~ uˈspexa], the same [ɛ ~ e] alternation does not occur in the initial syllable, and initial syllable faithfulness is not violated.

Table IV Mid tense-lax alternations under stress; monosyllables tend to be protected.

The nominative singular is also more tolerant of marked vowels when low vowels are considered. Of the two low vowels of Slovenian, [a] and [ʌ], the unmarked lax [a] is freely distributed. Tense [ʌ], however, is strictly limited to the stressed final vowel of unsuffixed masculine nominative singular nouns and adjectives. In the rest of its paradigm, [a] is found without exception, e.g. [ˈbɾʌt ~ ˈbɾata] ‘brother’ (Jurgec Reference Jurgec2011).

To summarise, vowel quality drives two different types of alternations: stress alternations that allow marked vowels to reduce inconspicuously by shifting stress away from an initial syllable, and reduction in situ when stress is not initial. The marked vowels [ɛ ɔ ʌ] are mostly tolerated in the nominative singular, and are mostly eliminated in the rest of the paradigm.

2.4 Summary

Our examination of the lexicon revealed an asymmetry: stress shift is significantly more acceptable when it moves stress away from the mid lax vowels [ɛ ɔ] than from other vowels. The effect was found in the dictionary study (§2.1), and confirmed in the questionnaire with native speakers (§2.2). The argument that stress shift is driven by marked vowels is supported by the observation that when they are non-initial, marked vowels are eliminated even under stress (§2.3).

In existing words, shifting stress away from the low [a] is significantly more common than it is from the tense vowels [i e o u]; we will see in §3 below that this effect is not extended to nonce words.

We attribute the special treatment of mid lax vowels to positional faithfulness to the initial stressed syllable: stress moves away from marked vowels in order that they can be eliminated in a less prominent position, an example of the Beckman-Noyer ranking.

3.1 Participants

Native speakers of Slovenian were recruited online via social networks and by word of mouth, as in §2.2 above. Data was analysed from the 145 participants who completed the survey and self-identified as being 18 or older, living in Slovenia and speaking Slovenian natively; other data was discarded.

Of these 145 participants, 90 self-identified as female and 36 as male; 19 did not say. The average reported age was 31 (range 18–66, median 29). We categorised 79 participants as speakers of central dialects (Ljubljana, Gorenjska, Dolenjska), and 67 as speakers of other dialects (Primorska, Štajerska, Koroška, Prekmurje or unreported).

3.2 Materials

Nonce-word stimuli consisted of nominative–genitive masculine nominal paradigms, with ten paradigms per stressed vowel, giving a total of 70 paradigms, where each paradigm contained one nominative, one faithful genitive and one mobile genitive, e.g. [ˈbidip ~ ˈbidipa, biˈdipa]. All stems were disyllabic CVCVC.

For mid vowels, identical consonants were used for tense and lax vowels, e.g. tense [ˈʒolup ~ ˈʒolupa, ʒoˈlupa] and lax [ˈʒɔlup ~ ˈʒɔlupa, ʒoˈlupa]. The mobile stress genitive is identical in both paradigms, since the contrast between tense and lax mid vowels is completely neutralised, as mentioned in §2 above. The stimuli (provided in Appendix B) were recorded as in §2.2, using the same speaker and methods.

3.3 Procedure

The experiment was conducted online using Experigen. The server made a random selection of 20 paradigms for each participant: four paradigms with high vowels, twelve with mid vowels (six tense, six lax) and four with the low vowel [a]. Each item had a unique consonantal make-up, so that no participant was asked to judge both tense [ˈʒolup] and lax [ˈʒɔlup], for example.

Each trial started with the frame sentence shown in Fig. 3, with blank slots for the nominative and the genitive respectively: To je moj ＿.‘This is my ＿.’ Ali nimaš svojega ＿? ‘Don't you have your own ＿?’. A button appeared on the next line; when pressed, one nominative–genitive paradigm was played, either faithful or mobile, and then two buttons appeared, v redu ‘ok’ and ni v redu ‘not ok’. Once pressed, a second sound button appeared for the other paradigm, and then two more judgement buttons. The participants were free to listen to the materials as many times as they wanted. Once they had judged both paradigms, the next stimulus was played.

Figure 3 Web interface for the nonce-word survey.

4.1 Strong faithfulness in the masculine nominative singular

In the masculine nominative singular, stress is contrastive, and marked vowels are tolerated, which we attribute to strong faithfulness to underlying stress and vowel qualities. In the rest of the paradigm, weaker faithfulness allows the elimination of mid lax vowels and low tense vowels. For example, the low tense [ʌ] surfaces in the nominative singular [ˈbɾʌt], due to Ident[ATR] ⪢ *LowTense, but the genitive [ˈbɾata] is the result of the opposite ranking. What singles out the masculine nominative singular? Language-internally, as argued by Jurgec & Bjorkman (Reference Jurgec and Bjorkman2018), it is the only part of the paradigm that allows bare roots, or morphologically simple forms. Cross-linguistically, it represents the confluence of unmarked number, unmarked gender and unmarked case; see e.g. McFadden (Reference McFadden and Grohmann2009, Reference McFadden2018) for arguments that the nominative is syntactically unmarked.

Further support for the unique status of the zero-ending masculine nominative singular comes from frequency. While evidence from Slovenian is not available, data from Russian (Slioussar & Samojlova Reference Slioussar and Samojlova2015) is entirely consistent with the proposal: masculine is the most common gender (47%), nominative the most common case (30%), singular the most common number (78%) and zero the most common ending (24%).

Various theoretical frameworks allow the masculine nominative singular to be determined by a different constraint ranking, such as cophonologies (Inkelas et al. Reference Inkelas, Orhan Orgun and Zoll1996, Anttila Reference Anttila2002, Inkelas & Zoll Reference Inkelas and Zoll2007) or indexed constraints (Itô & Mester Reference Ito, Mester, Beckman, Dickey and Urbanczyk1995, Pater Reference Pater2000, Reference Pater and Parker2009a, Reference Paterb, Flack Reference Flack2007, Gouskova Reference Gouskova2007, Becker Reference Becker2009, Jurgec Reference Jurgec2010). Jurgec & Bjorkman (Reference Jurgec and Bjorkman2018) propose that constraints can be specific to unaffixed words, modifying the lexical indexation schema. The current proposal is compatible with any of these possibilities.

The marked vowels [ɛ ɔ ʌ] are further limited to native nouns; they do not appear in established loanwords (Jurgec Reference Jurgec2010). This again suggests the regulation of these vowels along lexical lines.

4.2 Protection of existing items

Leaving the nominative singular behind, we focus on the more interesting grammar of the rest of the paradigm, exemplified throughout by the genitive singular. Our goal is to present a grammar that is learned from real words of the language, and correctly derives all of the real words, but also makes gradient predictions for nonce words, capturing the results of our nonce-word task. Zuraw's (Reference Zuraw2000) UseListed framework was designed to capture such effects: existing words are memorised and protected by a UseListed constraint, and novel words are derived productively by a grammar that is trained on the words of the lexicon (see Moore-Cantwell & Pater Reference Moore-Cantwell and Pater2016 and Zymet Reference Zymet2018 for more recent work on the grammar of irregular lexicons). To account for the gradience in our data, we switch from categorical Optimality Theory to a gradient constraint-based approach. Gradience in Zuraw (Reference Zuraw2000) was modelled by the Gradual Learning Algorithm (Boersma & Hayes Reference Boersma and Hayes2001), which was later shown in Pater (Reference Pater2008) to lack a solid mathematical foundation. We opt for the more established Maximum Entropy framework (MaxEnt; Goldwater & Johnson Reference Goldwater, Johnson, Spenader, Eriksson and Dahl2003, Wilson Reference Wilson2006). For a broader review of gradient constraint-based approaches, see Hayes (Reference Hayes2017).

The grammar has two context-free markedness constraints, *MidLax and *LowTense, and several faithfulness constraints, to be listed below. Crucially, there is no constraint penalising the unmarked vowels [i e a o u], and therefore nothing compels the alternation when stress shifts away from these vowels. Canonically, stress shifts away from an initial mid lax vowel, but there are also 16 lexical items that have stress shift under different conditions: stress shift away from [e] in [ˈtɾebux ~ tɾeˈbuxa] and stress shift away from a non-initial vowel in [peˈtɛlin ~ peteˈlina], and also 14 items with stress shift away from [a], e.g. [ˈnazoɾ ~ naˈzoɾa]. Like all nouns, regardless of their phonology, these 16 nouns are memorised and protected by UseListed, as seen in (6), but they make no contribution to any general productive pattern. Candidates (6a) and (6b) are generated productively from the concatenation of the nominative singular base with the genitive suffix, and compete with the next two candidates, which are derived directly from the memorised genitive [tɾeˈbuxa]. The use of the memorised form is favoured by UseListed, and faithfulness allows it to surface unchanged.

1. (6)

   

In our analysis, then, [ˈtɾebux], [peˈtɛlin] and the 14 items with [a] are lexicalised, or fossilised. Productive stress shift is only learned from the remaining 56 items in which stress shifts away from an initial mid lax vowel, which compete with eleven items in which it does not (see Table I above). Words with canonical stress shift, e.g. [ˈjɛzik ~ jeˈzika], are also listed, and their genitives are therefore protected by UseListed, but they also provide evidence for stress shift motivated by *MidLax.

Novel items presented to speakers in their nominative singular form lack a memorised genitive, regardless of whether the genitive has faithful or mobile stress, and are therefore all equally penalised by UseListed. This is exemplified by the nonce item [ˈbagem] in (7), whose genitive does not exist and therefore must be derived productively. UseListed penalises all of the candidates equally, and the decision is handed down to the grammar, in this case favouring the fully faithful genitive.

1. (7)

   

With unmarked vowels such as [a] and [e], stress shift is entirely unmotivated; the shifted candidate is harmonically bounded by the faithful candidate. This grammar is unable to reproduce the difference in the lexicon between [e] (one alternating item) and [a] (14 alternating items). This prediction is borne out in the treatment of nonce words, as seen in Fig. 4 above, where stress shift is equally disfavoured in nonce words with [a] and [e]; indeed, there is a non-significant trend in the opposite direction.

The stress-shifted genitive [baˈgema] is harmonically bounded in (7), which guarantees its absolute ungrammaticality in a categorical grammar. In our MaxEnt grammar, the predicted well-formedness of stress shift from an unmarked vowel is low (6%), and is the same for all five vowels.

4.3 Stress shift eliminates mid lax vowels

Mid lax vowels in forms other than the masculine nominative singular are reduced in two ways: when the stressed mid lax vowel is non-initial, it is tensed in situ in most words. In nouns with initial stress, positional faithfulness blocks the alternation in monosyllables, whereas in trochaic stems, the same protection of the initial stressed position causes stress to move away, to allow tensing in the less prominent unstressed initial position.

We use a family of Ident[ATR] constraints to model these effects: first, general Ident[ATR] penalises all differences in [ATR] between the nominative singular and the rest of the paradigm. Next, the positional constraints Ident[ATR]/σ₁ and Ident[ATR]/ˈσ penalise changes in word-initial and stressed positions respectively (Beckman Reference Beckman1997, Reference Beckman1998). Finally, the most specific Ident[ATR]/ˈσ₁ penalises changes in stressed initial position, and can be construed either as an atomic constraint or as the conjunction of the previous two (see Tessier Reference Tessier2007: §2.6.3–§2.6.5 on choosing between positional faithfulness constraints, and Shih Reference Shih2017 for further discussion of constraint conjunction in MaxEnt grammars). Stress shift is penalised by the general constraint Ident[stress]. The weights of the constraints were obtained by training the grammar on the lexicon from §2.1 using the MaxEnt Grammar Tool (Hayes & Wilson Reference Hayes and Wilson2008). Following Zuraw (Reference Zuraw2000), the training did not include the constraint UseListed.

We present the application of this grammar in (8)–(10), starting with a mid lax vowel in the less prominent non-initial position, followed by mid lax vowels in the initial stressed position.

Faithfully retaining a mid lax vowel is penalised by *MidLax, as seen in (8a). Tensing in stressed position violates both the general Ident[ATR] and the specific Ident[ATR]/ˈσ. To keep the tableaux manageable, we stipulate that stress shift to suffixes and stress shift to the left are blocked. The prediction in (8) is an expected well-formedness score of 61%, i.e. tensing a non-initial mid lax vowel is fairly acceptable.

1. (8)

   

In monosyllabic stems, the stressed syllable is word-initial, and thus tensing it incurs an additional violation of Ident[ATR]/ˈσ₁ relative to polysyllabic stems, as seen in (9).

1. (9)

   

This added violation of Ident[ATR]/ˈσ₁ reduces the predicted well-formedness of tensing from 61% in polysyllables to 21% in monosyllables; cf. the difference in the lexicon of 70–82% and 35–44% respectively (see Table IV in §2.3). The numbers do not match perfectly, because the proposed constraints are general, and apply both to tensing under stress and to tensing with stress shift. The same positional faithfulness constraints do double duty, maintaining mid lax vowels in monosyllables while preventing the maintenance of the same vowels in trochees, as seen in (10).

1. (10)

   

In the genitive of nonce [ˈbɛdiʃ], the fully faithful candidate in (10a) incurs only a violation of *MidLax, just like (9a). Tensing under stress in (10b) incurs violations of positional faithfulness, just like (9b). However, the best candidate, (10d), involves stress shift, allowing the elimination of the marked vowel in an unstressed position. Stress shift is favoured because it avoids penalties from positional faithfulness constraints. There is no violation of faithfulness to the stressed initial position in (10d), because the initial syllable is not stressed in the output, and prominent positions are defined relative to the output in positional faithfulness theory (Beckman Reference Beckman1997, Reference Beckman1998).

Plain initial syllable faithfulness is violated in both (9b) and (10d), which is why Ident[ATR]/σ₁ was assigned a weight of zero and left out of all the tableaux.

The grammar presented here exemplifies a straightforward prediction of positional faithfulness: the stressed prominent position moves in order to avoid an unfaithful mapping in the prominent position. The prediction is borne out in Slovenian, as seen in (10) above, showing that the Beckman-Noyer ranking is not problematic.

The tableau in (10) is a version for Slovenian of Jesney's (Reference Jesney2011) tableau (6b), with one difference: in Jesney's account, moving the prominent position incurs a violation of the markedness constraint Trochee, whereas here it incurs a violation of the faithfulness constraint Ident[stress].

Jesney also considers inputs in which all of the underlying vowels are marked, e.g. /ˈbɛdɛʃ/ (her tableau (7)). In our analysis of Slovenian, a hypothetical stem with only mid lax vowels strongly prefers faithful stress, e.g. /ˈbɛdɛʃ-a/ → [ˈbɛdeʃa], because shifting stress to [beˈdɛʃa] adds a violation of Ident[stress], without removing a violation of *MidLax. Recall that shift to a suffix vowel is blocked by fiat. In this type of grammar, stress is optimally moved away from a mid lax vowel, but if there is no eligible unmarked vowel to move to, one mid lax vowel surfaces faithfully and all of the other mid lax vowels reduce to tense.

Given a nominative singular with a peninitial mid vowel such as [ˈbɛdeʃ], a Slovenian speaker could potentially posit an underlying representation such as /ˈbɛdɛʃ/ rather than the more surface-true /ˈbɛdeʃ/, and therefore indirectly allow the quality of the peninitial vowel to influence the position of stress in the genitive. However, in the nonce-word experiment, preference for stress shift was stronger with a peninitial mid vowel, not weaker (14% preference for stress shift with a non-mid peninitial vowel vs. 22% preference for stress shift with a mid peninitial vowel). This result is incompatible with the idea that speakers assume two lax vowels in the underlying representation to any detectable extent.

Jesney (Reference Jesney2011) proposes a method for excluding Slovenian-like patterns in the framework of Harmonic Serialism (McCarthy Reference McCarthy2010, Reference McCarthy, McCarthy and Pater2016). In this theory, prominent positions are defined relative to the input to the previous stage of the derivation, and therefore stress shift in /ˈbɛdiʃ-a/ → [beˈdiʃa] is blocked, because the reduction of the mid lax vowel still takes place in the prominent position. It seems likely, however that Slovenian could be analysed in Harmonic Serialism with a different set of underlying representations and constraints, for example by replacing underlying stress and faithfulness to stress with lexically specific alignment. We leave this task for future work.

4.4 Possible alternatives to positional faithfulness

We have provided an adequate analysis of Slovenian stress shift: stress moves away from a marked vowel to allow it to reduce inconspicuously, and this alternation is governed by the same positional faithfulness constraints that protect marked vowels in monosyllables from being eliminated under stress. Because this analysis relies on a type of constraint interaction that has been considered problematic, we must ask whether there is a workable analysis without positional faithfulness, of course within plausible limits on the admissible types of constraints.

We maintain the assumption that the grammar only shifts stress away from mid lax vowels, and consider an alternative analysis that dispenses with positional faithfulness and instead introduces positional markedness. To motivate stress shift, we use a stress-window constraint that we dub StemFinal (‘stress must be stem-final’) or a similar positional markedness constraint. This constraint must be dominated by the general faithfulness constraint Ident[stress], as shown in (11), to prevent stress shift away from the unmarked [i e a o u].

1. (11)

   

The problem with this grammar is that without positional faithfulness, *MidLax can only cause tensing, never shifting, as shown in (12). Without positional faithfulness, the intended winner [beˈdiʃa] loses, or, in a stochastic grammar, [beˈdiʃa] would be predicted to be less well-formed than [ˈbediʃa], contrary to fact.

1. (12)

   

Similarly, a positional markedness constraint that penalises initial mid lax vowels (*InitialMidLax) will not help, for the same reason: it would penalise the fully faithful candidate (12a), but leave tensing in situ, as in (12b), as the winner.

Even less useful would be a positional markedness constraint against mid lax vowels in unstressed syllables, *MidLax_unstressed; this constraint would not be violated at all in (12) or in the tableaux below, and would do no work. We thus reject the possibility of an analysis that eliminates positional faithfulness and retains only markedness and general faithfulness.

A second alternative to positional faithfulness would be a conjunction of positional markedness and general faithfulness, in this case, Ident[ATR] and StemFinal (Smolensky Reference Smolensky1997, Reference Smolensky, Smolensky and Legendre2006, Łubowicz Reference Łubowicz2005), as in (13). The conjoined constraint penalises tensing under non-stem-final stress, correctly allowing [beˈdiʃa] to be the most well-formed, followed by [ˈbɛdiʃa].

1. (13)

   

This analysis works for nouns with non-stem-final stress, and extends correctly to polysyllables with stem-final stress, as shown in (14). Tensing in situ is predicted by the ranking *MidLax ⪢ Ident[ATR].

1. (14)

   

Without positional faithfulness, however, nothing blocks tensing in monosyllabic stems, as seen in (15). To block tensing in monosyllables, an additional constraint would have to outrank *MidLax and prevent tensing, while still allowing the tensing in trochees in (13), for example a constraint against lax vowels in initial syllables. Such a constraint has no typological support, since languages that allow lax vowels (or any other marked structure) in all syllables but the initial one are unknown.

1. (15)

   

The beauty of positional faithfulness lies precisely in its ability to connect the behaviour of monosyllabic stems with the opposite behaviour of trochaic stems; abandoning this family of constraints leads to a loss of generality. We hope to have shown that any alternative to positional faithfulness is rather unappealing, combining lack of generality with unattractive constraints.

To reiterate, the current proposal depends on the types of constraints that the analyst allows: general (context-free) markedness, general faithfulness, positional markedness and positional faithfulness. Positional faithfulness, we argue, is necessary for an analysis that captures the entire range of phenomena in Slovenian.

4.5 Local summary

The proposed analysis consists of several components: the unsuffixed masculine nominative singular is assumed to be the base of the derivation, and to enjoy stronger resistance to markedness pressures. In the rest of the nominal paradigm, a markedness constraint against mid lax vowels allows the grammar to capture the elimination of these vowels, either via stress shift or in situ. The analysis leaves behind those few lexical items in which stress shifts away from other vowels, or away from a non-initial syllable; those items are listed and retrieved from the lexicon correctly, but they do not contribute to the productive grammar.

A positional faithfulness constraint that penalises tensing in the initial stressed position does double duty: it prevents tensing in situ in monosyllabic stems, and drives stress shift, with concomitant tensing in trochaic stems. While there is an alternative analysis that relies on markedness–faithfulness conjunction, it cannot capture the connection between stress shift and tensing in situ.