2.1 A low-frequency pattern: agent nouns in -ist-a/-yst-a

The suffix -ist-a/-yst-a [ista/ɨsta] is used to form agent nouns in Polish. It is of Latin/Greek origin and entered Polish at the turn of Old and Middle Polish, that is, around the 15^th century (Długosz-Kurczabowa & Dubisz 2006: 369). The suffix is productive and no longer restricted to Latinate words. In spite of its long history in Polish, the suffix remains a low-frequency pattern. The quantitative data to back this claim will be given in Section 2.3. We start out with some facts about the distribution and consonant mutations triggered by the two variants of the suffix: -ist-a and -yst-a. I begin with examples illustrating the usage of the variant -ist-a. The data are drawn from Gussmann (2007: 157–161). For the purposes of this analysis, a consonant mutation is defined as a featural change in a consonant that results in a phonologically distinct (i.e. contrastive) segment. There is convincing evidence that palatalization of labials, dentals and velars before [i], [pʲ bʲ fʲ vʲ mʲ tʲ dʲ sʲ zʲ c ɟ ç], is a non-contrastive, coarticulatory effect of the following vowel. First, labials, dentals and velars do not contrast for palatality preconsonantally and word finally. Second, Święciński’s (2014) acoustic study shows that [pʲ bʲ fʲ vʲ mʲ tʲ dʲ sʲ zʲ c ɟ ç] appear exclusively before the vowel [i]; before other vowels palatalization is manifested on a distinct segment, the palatal glide [j], e.g. piasek [pjasɛk] ‘sand’, diabeł [djabɛw] ‘devil’ and kiosk [kjɔsk] ‘kiosk’. In light of this, [pʲ bʲ fʲ vʲ mʲ tʲ dʲ sʲ zʲ c ɟ ç] should not be regarded as mutated variants of [p b m f v t d s z k g x]. Such coarticulatory palatalization is omitted from the transcription in this paper, as it does not represent an instance of mutation in the relevant sense.⁴

The items in (2) show that the variant -ist-a is used after labials, (a), palatalized labials, (b), coronals [s z t d], (c), and velars, (e).⁵ There are no cases of -ist-a after base-final palatalized coronals except after [j], shown in (d). The front glide is variably elided before -ist-a. Coronals (except [l]) show mutations before the suffix, while velars and labials (except [w]) do not. The patterning of base-final [l] in (c) and [w] in (a) requires an explanation. The coronal [l] does not alternate because it does not have a distinct mutated variant before [i]. The alternation [w ~ l] in (a) is a reflex of a historical alternation between a velarized /ɫ/ and a palatalized /lʲ/. It is omitted from further consideration for several reasons. First, it arose relatively recently due to the context-free change /ɫ/ > /w/. More time is necessary for paradigm uniformity pressures to take an effect on this alternation. Second, there are no words that end in […w-ɨst-a] or […w-ist-a]. In contrast, there is a large number of words in […l-ist-a]: 135 (Bańko et al. 2003). There are two sources of such words: words that have a base in […l], e.g. motocykl-ist-a ‘motorcyclist’ (< motocykl), and words that have no independently existing base and are not fully decomposable (i.e. they contain bound roots), e.g. popul-ist-a ‘populist’. The existing words in […l-ista] likely form an attractive bias for words in -ist-a with base-final […w], perpetuating the [w ~ l] alternation. There is no comparable attractive bias from […w-ɨsta] or […w-ista], as these patterns are not attested.⁶ The [w ~ l] alternation, thus, persists for lack of an alternative and for that reason it is not compatible with the other alternations.

The other variant, -yst-a [ɨst-a], is selected after base-final [r]. The rhotic alternates with the fricative /ʐ/ in the derived form, as illustrated in (3a). A context-free diachronic change /rʲ/ → /ʒʲ/ → /ʐ/ gave rise to this alternation in modern Polish. As exemplified in (3b), the variant -yst-a [ɨsta] also appears after alveolar and postalveolar (retroflex) affricates [ts dz tʂ dʐ] and the postalveolar (retroflex) fricatives [ʂ] and [ʐ], the latter fricative showing a similar behavior to the /ʐ/ ← /rʲ/.

In addition to the words with coronals [t d r] that show mutations before -ist-a, illustrated in and (2) and (3), there are also those which do not evidence consonant mutations and choose the other shape of the suffix, as exemplified in (4). The distribution of the suffix alternants in (4) diverges from the distribution shown in (2) and (3).

A number of words with the suffix -ist-a vacillate in current usage. Their bases end in [t d r]. This different behavior of words in [t d r] in (2) and (3), on the one hand, and (4) and (5), on the other, will be linked to differences in their token frequency in Section 3.3.2.

As the summary of the distribution of the suffix -ist-a/-yst-a in (6) shows, the -ist-a variant appears after labials and velars, while -yst-a occurs after [ts dz tʂ dʐ ʂ ʐ]. Base-final coronals [t d r] show a more complex pattern: both allomorphs are attested. Based on these distributional facts, it is assumed that -ist-a is the basic variant of the suffix, while -yst-a, whose usage is restricted to retroflexes and alveolar affricates (categorically) and [t d r] (variably), is a positional variant of the suffix. In usage-based models, it is assumed that relations between morphological units, including identity relations, are established on the basis of phonological and semantic similarity. The suffixes -ist-a and -yst-a are both phonologically and semantically similar. A rationale for this assumption is provided in Section 2.5, where the properties of schemas are fleshed out.

(6)

We now turn to evidence indicating that consonant mutations before front vowels are historically and synchronically motivated in Polish. As for the historical motivation, there is compelling evidence that CV coarticulation preceded and induced the emergence of distinctive palatalization in Slavic languages. Jakobson (1929/1962: 71ff.) and Andersen (1978: 11–15) argue that in Common Slavic palatalized consonants arose due to coarticulation with the following front vowel. The developments of the reconstructed Common Slavic *věra ‘faith’ and *klětŭka ‘cage’ in Polish are used as illustrative examples (ĕ traditionally stands for yat’, a long open front vowel). At the stage of [vʲɛra] and [klʲɛtka], the palatality of the consonant was interpreted as coarticulatory and attributed to the following front vowel. A series of changes in the quality of the vowel, depicted in (7), was partly responsible for the phonologization of a palatalized consonant.⁷ When the vowel became back, as in [vʲara] and [klʲatka], the palatality could no longer be attributed to the vowel and was phonologized on the consonant, which gave rise to the distinctively palatalized /vʲ/ in /vʲara/ and /lʲ/ in /klʲatka/ (Andersen 1978: 11–15). A similar mechanism was likely involved in the development of mutations before -ist-a/-yst-a. Thus, phonetic reconstruction and comparative evidence indicate that palatalization is historically motivated before front vowels in Polish.

Synchronic alternations in Modern Polish also provide motivation for mutations before front vowels. Suffix-initial front vowels commonly trigger consonant mutations (Rubach 1984). In (7) mutations before the adjectival suffix -ist-y are shown. Two types are directly relevant for the discussion at hand: Coronal Mutation and Velar Mutation. Coronal Mutation is applicable to coronals [t d s z n r] and results in alveolopalatals and a postalveolar [tɕ dʑ ɕ ʑ ɲ ʐ], as illustrated in (8a). Velar Mutation applies to velars [k g x] and results in postalveolars [tʂ ʐ dʐ ʂ], as illustrated in (8b).⁸ In addition to -ist-y, there is a host of other morphological patterns that generate mutations before coronals and velars (Rubach 1984; Gussmann 2007). Both types of mutation appear before /i/ and /ɨ/, as shown in (8), as well as before [ɛ] (e.g. u[d]-o ‘thigh’ – u[d ʑ]-ec ‘haunch’, kro[k] ‘step’ – kro[tʂ]-ek DIM). The mutations result in a change of the featural specification of the consonant. The latter aspect will be fleshed out in Section 2.6.

Returning to the formation of agent nouns in -ist-a/-yst-a and the behavior of the base-final consonant, three facts require explanation in the context of Coronal and Velar Mutations. First, in considering the words in -ist-a/-yst-a, labials and velars do not show any significant consonant mutations before the suffix, while coronals (variably) do. This stands in contrast to the data exemplified in (8) showing that mutations apply to both coronals and velars in other morphophonological patterns.⁹ The failure of velars to undergo mutations before -ist-a/-yst-a is in this light surprising and will be addressed in Section 3.1. Second, why do some words with base-final [t d r] show mutations before -ist-a/-yst-a, while others do not? Looking at the data illustrating Coronal Mutation, it seems safe to assume that mutations of coronals before front vowels are historically and phonetically motivated in Polish. There is evidence that palatalization before -ist-a in words with base-final [t d r] used to be a fully regular process. Words of this type without palatalization appeared later (Rubach 1984: 65–68). Therefore, the lack of mutations in this context for some words with base-final [t d r] before -ist-a/-yst-a must be seen as a later development, whose emergence is in need of explanation. This issue is tackled in Section 3.2.3. The third fact that requires explanation is the gradient elimination of mutations for base-final [t d r] and their preservation for base-final [s z n]. The gradient elimination of mutations for base-final [t d r] is dealt with in Section 3.2.3. In the summary in (9), the effects of -ist-a/-yst-a on the base-final consonant are juxtaposed with the effects of other morphological patterns, both inflectional and derivational (such as the adjectival suffix -ist-y). For compactness, in (9) and below I use -ist-a to refer to both alternants of the suffix: -ist-a and -yst-a, unless the choice of the alternant is somehow relevant.

(9)

2.2 A high-frequency pattern: diminutives in -ek

Let us take a closer look at the patterning of the diminutive suffix -ek in relation to different base-final consonants (Czaplicki 2013a; 2014a).

Labials and coronals in (10a) and (10b) fail to mutate. In the case of palatalized labials and coronals in (10c), the palatal element disappears before -ek, which can be described in traditional terms as depalatalization.¹⁰ Velars in (10d) mutate and appear as their postalveolar (retroflex) reflexes before the [ɛ] of the suffix.

There is inconsistency in the applicability of consonant mutations before the suffixes -ek and -ist-a (the latter discussed in the previous section). In contrast to the behavior of the -ist-a suffix, which triggers mutations of coronals but not of velars, the diminutive suffix -ek results in mutations of velars but not of coronals. There is ample evidence that the high front /i/ is typologically more likely to trigger mutations than the mid front /ɛ/ (Bateman 2007; Rubach 2007). In fact, phonological conditioning alone predicts that -ist-a/-yst-a should be responsible for higher rates of mutations than -ek, as mutations are more likely to occur in the context of /i/ than in the context of /ɛ/. In a similar vein, Czaplicki’s (2019) quantitative analysis of the effects of 27 suffixes in Polish has shown that /i/ is far more likely to trigger mutations than /ɛ/. The fact that the suffix -ek triggers palatalization before velars, but depalatalization before labials and coronals has been interpreted by Czaplicki (2019) as evidence for the irrelevance of phonological naturalness in the conditioning of consonant mutations. What is relevant for our purposes is that words in -ek do not exhibit variability, as opposed to words in -ist-a. There are no variants without mutations for base-final velars before -ek. Mutated coronals before -ek are equally uncommon.¹¹ It is shown that an element that is crucial in an explanatory analysis of morphophonological patterns is their frequency. Specifically, it will be shown that type frequency determines the stability of a pattern.

2.3 Frequency

There is growing evidence suggesting that frequency plays an important role in morphophonology (Mańczak 1980; Bybee 2001; Ellis 2002; Albright & Hayes 2003; Baayen et al. 2003; Dąbrowska 2008; Czaplicki 2013a; 2013b; 2014a; 2014b). As already mentioned, the generalizability of a pattern has been shown to crucially depend on the number of stored words that exhibit the pattern (gang effects). More specifically, when two (or more) patterns are available in a particular context, the pattern with a higher frequency is the one most likely to become generalized to novel words. In other words, pattern extension deploys the most robust of the several patterns used in a particular morphological context. An explanation along these lines is available for the change of classes of the English verb help, which in Old English belonged to Strong Verbs, but now forms its past tense by means of the more robust -ed suffixation. Frequency enters into interactions with other factors. Dawdy-Hesterberg & Pierrehumbert (2014) have found that, while the generalization of the various patterns of Arabic broken plurals in large part depends on prosodic templates, gang effects are important predictors as well.

Another factor that is often implicated in language change and pattern extension is token frequency (Bybee 2001). High-frequency words are predicted to resist pattern extension, as evidenced by the preservation of such suppletive forms in English as go – went. As already mentioned, the relative token frequency of the derivative and its base plays a critical role in dual-route models of lexical access. This issue becomes relevant in Section 3.2, where the strength of lexical representations is considered.

Anttila (2006) discusses an interesting case of assibilation in Finnish. The bimoraic verbs which meet the structural description of assibilation (context of a following /i/) show three types of behavior that can be directly linked to their frequency: the most frequent verbs exhibit assibilation, the least frequent verbs fail to undergo assibilation and verbs of medium frequency show variation. Czaplicki (2016) argues that these regularities can be insightfully explained by reference to frequency, on the one hand, and avoidance of mutations between the base and the derivative (output-output correspondence, formulated below), on the other.

In an attempt to explain the different behavior of words in -ist-a and -ek with respect to mutations, I make reference to the frequency of the two patterns in the lexicon. It is argued that for robust patterns (i.e. those showing a high type frequency) identity pressures are overridden. Let us compare the type frequencies of the words in -ist-a and -ek using dictionary and corpus data. Table 1 presents the counts of words in -ist-a and -ek drawn from a reverse dictionary of Polish (Indeks a tergo do Uniwersalnego słownika języka polskiego pod redakcją Stanisława Dubisza) (Bańko et al. 2003).

The number of dictionary entries in -ek is nearly twice as high as the number of dictionary entries in -ist-a. There is good reason to believe that the difference in robustness between the two patterns is even greater. First, the list of words with the suffix -ek is not exhaustive, as it does not include recent borrowings and many diminutives whose semantics is fully predictable. Grzegorczykowa & Puzynina (1999: 425) mention that diminutives in -ek and -ik constitute an open class and that most nouns can form diminutives using one of these suffixes.¹² Second, a considerable number of words in -ist-a (but definitely not all) belongs to the learned stratum of vocabulary and their usage is often restricted to formal and technical registers. Therefore, the factor that is missing from Table 1 is token frequency, i.e. the frequency with which each of the words is used in the discourse.

In analyzing the frequency of the two suffixes, I use data extracted from the corpus plTenTen: Corpus of the Polish Web, available in Sketchengine, which is made up of texts collected from the internet in 2012 and comprises more than 7.7 billion words. Table 2 provides the type and total token frequency of words in -ist-a and -ek in the corpus (accessed 12 October 2020). Words below the frequency of 50 have not been considered, as they turned out to be mostly proper names and spelling errors.¹³

The type frequency of words in -ek in the corpus is 7.6 times higher than the type frequency of words in -ist-a. Predictably, the difference between the number of words in -ist-a and -ek in the corpus (i.e. their type frequency) is greater than for the data taken from the dictionary shown in Table 1. Table 2 also shows the total token frequency, which is the sum of the token frequencies (the frequency with which a given word is used) of all the words in -ist-a and -ek in the corpus. The total token frequency of words in -ek is 9.2 times higher than the total token frequency of words in -ist-a.

The token frequency data are not normally distributed, so they have been log-scaled as in Figure 1. The histograms in Figure 1 show that words in -ek outnumber words in -ist-a for all token frequency ranges. A non-parametric Mann-Whitney test run on the data reveals that the token frequency of words in -ek (Mdn = 459, M = 7433.76) is significantly lower than the token frequency of words in -ist-a (Mdn = 884, M = 6189.56), U = 390640.5, z = 3.49, p < .001. This result is likely due to the highly positively skewed distribution for words in -ek (skewness = 29.44). As can be seen from the histogram on the left in Figure 1, low-frequency words in -ek outnumber comparable words of medium and high frequency. It follows that type frequency is a better measure of the strength of a pattern than token frequency.

Figure 1

Histogram of token frequency by suffix.

To sum up, a close analysis of type and token frequency confirms the difference in the robustness of the two patterns: the pattern -ek is significantly more robust than the pattern -ist-a in the grammar. In addition, low-frequency words in -ek are overrepresented relative to words in -ek of medium and high frequency.

The two suffixes are useful in assessing the role of frequency, as they are compatible in the relevant aspects of distribution and phonological behavior. First, they are both derivational. Second, they are both fully productive and readily extended to new words. Third, neither of them is restricted to loanwords.¹⁴ Finally, they both begin with front vowels and trigger mutations. In fact, as mentioned in Section 2.2, phonological conditioning alone predicts that -ist-a should be responsible for higher rates of mutations than -ek, as /i/ is more likely to trigger mutations than /ɛ/ both in Polish and more generally (Bateman 2007; Rubach 2007; Czaplicki 2019). It is claimed that the differences in the phonological and morphological behavior of the two constructions are derivable from the differences in their frequency.

2.4 Lexical storage of allomorphs

Lexical storage of allomorphs is not a new idea. There is ample evidence from usage-based research that morphologically complex words are stored whole (McQueen & Cutler 1998; Bybee 2001; Baayen et al. 2003; Hay 2003; see also Section 3.2.2). Similarly, certain analyses representative of generative phonology demonstrate that specific alternations cannot be derived using purely phonological operations and the relevant allomorphs need to be listed. In phonologically conditioned suppletion two phonologically dissimilar alternants need to be stored, however, their distribution is phonologically conditioned (Carstairs 1988; 1990). Anderson (2008) discusses a pattern from Surmiran and argues that vowel reduction, once a phonologically conditioned process, has become opaque. A solution proposed by Anderson involves reference to two listed alternants of the stem whose distribution is governed by prosodic considerations (i.e. stress). The distribution of alternants may also be regulated by phonologically neutral considerations. Paster (2006) and Embick (2010) use subcategorization frames which make reference to phonological and lexical information. However, allomorph selection does not as a rule result in phonologically optimized structures. In Kaititj the ergative suffix appears as [-ŋ] after disyllabic stems and as [-l] after trisyllabic stems. Although the formula that captures the generalization makes reference to phonological vocabulary, in this case the syllable, this instance of allomorph selection does not in any way improve phonological well-formedness (Paster 2006). The existence of such patterns shows two things: allomorph selection need not be phonologically optimizing and some alternants, including different shapes of a stem, have to be listed (i.e. stored).

2.5 Morphophonological schemas and cophonologies

In order to shed light on the differences in the assumptions of traditional generative and usage-base approaches, it is vital to compare the relevant aspects of rules (Current generative analyses generally employ constraints instead of rules, but certain properties of rules remain implicit.) and schemas (usage-based approaches). Schemas in contradistinction to rules emerge from the lexicon, that is, from the stored representations of words and phrases. As a consequence, they have “no existence independent of the lexical units from which they emerge” (Bybee 2001: 27). Rules, on the other hand, exist independently of the stored items and form part of a module that is separate from the lexicon. The productivity of a schema is a function of the number of participant items (gang effects). Put differently, the more words comply with a given schema, the more productive the schema is predicted to be.¹⁵ In this view, productivity of a schema is gradient and probabilistic, which is a consequence of the close connection between schemas and stored words. While in early generative models rules did not show a direct relationship with the number of words they apply to, more recently, modeling non-categorical effects, including the effects of frequency, has been facilitated by the use of gradient and probabilistic OT constraints (e.g. MaxEnt, Hayes & Wilson 2008).

Since the publication of The Sound Pattern of English (Chomsky & Halle 1968), the well-formedness of rules has typically been associated with the notion of markedness or phonological naturalness. Chomsky & Halle (1968) observe that their model largely overpredicts the types of processes that occur in natural languages and propose to constrain the set of possible rules by appealing to phonological markedness (Chapter 9). Rules that lead to the reduction in markedness are preferred over those that do not. More recently, Hayes and Steriade (2004: 1) claimed that markedness constraints are gleaned from phonetic knowledge, the latter being somewhat vaguely defined as “the speakers’ partial understanding of the physical conditions under which speech is produced and perceived”. In this view, final devoicing of obstruents is possible, while final voicing is predicted to be impossible, as it would lead to more marked structures (Kiparsky 2006). The position that synchronic universals (i.e. markedness) constrain diachronic change and shape linguistic patterns is advocated in, for instance, de Lacy (2002; 2006), Kiparsky (2006; 2008) and de Lacy & Kingston (2013).

On the other hand, there is accumulating evidence that undermines the role of markedness as an active bias in synchronic grammars. Processes that result in more marked structures are pervasive (Bach & Harms 1972; Anderson 1981; Blevins 2004; Hale & Reiss 2008; Czaplicki 2013a; 2014a; 2019). Examples of such processes include final voicing in Lezgian (Blevins 2006) and unnatural patterns of consonant epenthesis in various languages (Blevins 2008). Typological asymmetries can be explained by extragrammatical factors (e.g. common trajectories of phonetically based sound change) (Ohala 1983; Blevins 2004). Insofar as schemas are based on stored representations, they are language specific and not necessarily dependent on naturalness, understood as a universal learning bias. Thus, schemas are a priori markedness-free. However, it should be noted that markedness constraints are not logically incompatible with schemas. In principle, the emergence of markedness constraints is independent of schemas. Yet, one of the predictions of usage-based models is that language-specific considerations should override markedness constraints when a conflict between the two pressures arises (a preference for morphological conditioning over phonological conditioning, see below). The role of markedness in schema-based approaches certainly deserves more attention.

Rules refer to distinctive features (Chomsky & Halle 1968). If segments are referred to, this is done as a shorthand for featural specification that underlies a particular segment. Schemas are not similarly restricted in their vocabulary.

As mentioned in (12), schemas can refer to various organizational units, such as segment, syllable and feature, as long as these units emerge from stored representations. The claim cited in (12) points to yet another important difference between rules and schemas. Rules are preferably stated using phonological vocabulary (this is a requirement of modularity, Scheer 2012). In contrast, schemas require that different types of information – phonological, syntactic, morphological and semantic – be simultaneously accessible. This is the fundamental property of Parallel Architecture, a theory of grammar developed by Ray Jackendoff (cf. Jackendoff 2002). Schemas are formed on the basis of phonological and semantic similarity between stored words. Morphological structure emerges from these identity relations. On this view, -ist-a and -yst-a are predicted to be identified as alternants of a single suffix because of their phonological and semantic similarity as well as their near complementary distribution.

Schemas contain information about morphological structure (e.g. English past tense formation using -ed) (Bybee 2001: 23–24). In fact, Bybee (2001: 97–100) argues that segmental alternations display a preference for morphological conditioning over phonological conditioning. She also claims that “once morphological conditioning becomes dominant, it follows that phonological principles, such as patterning based on natural classes, will no longer be applied in the same way as for phonetically conditioned processes” (2001: 105). Put differently, markedness considerations are less important for morphologized patterns than for patterns that are fully phonetically transparent. Two phonologically similar contexts can give rise to different segmental alternations, as long as the morphological conditioning is different (i.e. two different morphemes). For example, Velar Mutation applies before the suffix -ek but not before -ist-a (see Sections 2.1–2.2), even though the context of a following front vowel is present in both cases.

Schemas are compatible with approaches that assume the existence of multiple cophonologies within one grammar (Booij 2010; Inkelas 2014; Booij & Audring 2017; Czaplicki 2020), where each morphological construction is associated with its own phonological subgrammar. In other words, each morphological construction has its own phonological properties. Representative of early approaches that assume the existence of multiple cophonologies within one grammar is Itô & Mester’s (1995) Core-Periphery Model of the lexicon, which identifies the core area of the lexicon, governed by a maximum set of markedness constraints (markedness dominates faithfulness). Their markedness constraints are syllable- and segment-related and penalize, for example, voiced obstruent geminates and non-geminate [p]. Structures that occupy less and less central areas of the lexicon show increasingly more violations of markedness constraints (faithfulness dominates markedness). By extension, the Core-Periphery Model predicts that the less nativized (more peripheral) a structure, the more faithful it should be to its input. Crucially, the model predicts that there are multiple layers within the lexicon and each layer is differentiated from the others in that it has its own specific constraint ranking (phonological grammar). In Japanese four layers are distinguished: Yamato, Sino-Japanese, Assimilated Foreign and Unassimilated Foreign. An important claim of Itô & Mester (1995) is that the differences between the core area of the grammar and more peripheral areas relate to the reranking of faithfulness constraints. The ranking of markedness constraints remains constant for the whole grammar. For example, the hypothetical input /paka/ is realized differently depending on the layer of the grammar in which it is generated. If it is processed in the Sino-Japanese stratum it surfaces as [haka], due to the ranking of No-P above Faith. In the Assimilated Foreign stratum, the output is [paka], due to the reranking of Faith above No-P.¹⁶ Morphological conditioning can also be handled by indexing constraints to specific morphological constructions (Itô & Mester 1999).

To sum up, schemas are dependent on stored representations, their productivity (strength) depends on the number of words they derive (type frequency), they are a priori markedness-free, they can be stated at various degrees of generality (e.g. segment, feature, syllable) and they include morphological information. Supporting evidence for these properties of schemas can be found in, for example, Bybee (2001), Ellis (2002) (frequency effects), Booij & Audring (2017), Czaplicki (2013a; 2019; 2020) (morphological conditioning) and Blevins (2004) (markedness-free generalizations).

Schemas can be product- or source-oriented. An example of a product-oriented schema specifying English Past Tense in walked, begged and wanted was given in (1) and is repeated in (13) (Bybee 2001: 126). Source-oriented schemas mention input (base) as well as output. Several source-oriented schemas representing Polish consonant mutations are exemplified in (14).

The schemas in (14) express the formation of diminutives, (a), and instrumentals, (b), from nouns whose stems end in various consonants. The former pattern was illustrated in Section 2.2, the latter can be applied to kro[k] ‘step’, pro[g]-u ‘threshold’ GEN SG and du[x] ‘ghost’. Source-oriented schemas are preferable for describing consonant mutations, as the applicability and type of mutation in the derivative crucially depend on the final consonant in the base. In (13a) base-final velars appear as their mutated alternants before -ek. However, at the same time base-final [t] fails to mutate and the cluster [ɕtɕ] depalatalizes to [st] in the same context. So we could not refer to a product-oriented schema requiring that a consonant appear in its palatalized (or mutated, or retroflex) form before -ek. Rather, the output of the concatenation of a suffix depends on a particular base-final consonant. Before the instrumental suffix -em, illustrated in (14b), velar plosives mutate but the fricative remains unchanged. Such arbitrary suffix- and consonant-specific alternations, which abound in Polish, would be difficult to conceptualize as product-oriented schemas (see Becker & Gouskova 2016 for more evidence that source-oriented schemas are necessary).

Schemas defined along these lines constitute the core of the proposed analysis. Type frequency determines the strength of linguistic patterns, in the sense that the more frequent the pattern, the more likely it is to be extended to novel words. In the formalization of the analysis, type frequency finds reflection in the ranking of constraints representing morphophonological schemas. Constraints representing more frequent schemas are ranked higher than constraints corresponding to less frequent schemas. In the case at hand, schemas representing -ek will be ranked higher than schemas pertaining to -ist-a, as illustrated in (15). Schema-constraints are interspersed with phonological constraints, as will be demonstrated in Section 3.2.3.

2.6 Representing consonant mutations in Polish

In traditional terms, palatalization is a type of consonant mutation that is caused by a following front vowel or a palatal glide. Palatalization has its diachronic roots in the phonologization of coarticulatory effects between a consonant and the following front vowel or a palatal glide (Jakobson 1929/1962; Bateman 2007; Kochetov 2011). In phonological analyses, palatalization has been represented as agreement in certain features between a consonant and a following vowel. Several feature theories have been in use, for example, the Halle-Sagey model (Sagey 1986) and the Clements-Hume model (Clements & Hume 1995). In the Halle-Sagey model palatalization involves the agreement in the feature [–back] and in the Clements-Hume model the process is represented as agreement in the features [coronal, –anterior]. More recent approaches cast in Optimality Theory concentrate on defining feature classes, i.e. features that function together in phonology (see, for instance, Padgett 2002 and Halle 2005). In the present analysis front vowels are represented as [+coronal, –back] and, consequently, palatalization imposes agreement in the features [+coronal, –back] on adjacent consonants and vowels. The assumption that the feature [±back] can refer to both [±coronal] and [±dorsal] is in line with Padgett (2002) and Halle (2005) in the sense that feature sets (classes of segments) are defined on a language- and alternation-particular basis.

Coronal Mutation, e.g. [t] ~ [tɕ], can be viewed as a change of the feature [±back], alongside the change of [±anterior]. With relevance to this analysis, Coronal Mutation does not affect the major place of articulation ([±coronal]).

Velar Mutation, e.g. [k] ~ [tʂ], on the other hand, involves a change of the major place of articulation ([+dorsal, –coronal] > [–dorsal, +coronal]), with the feature [±back] remaining unaffected. The assumption that the output of Velar Mutation, postalveolars, is [+back] gets support from their phonetics and distribution. Hamann (2002) has found that Polish postalveolars meet the criteria of retroflex consonants and retroflexion is incompatible with palatalization. In addition, postalveolars never appear before the high front vowel /i/ (except in recent borrowings). Based on these facts, Velar Mutation is an instance of coronalization.

The analysis at hand focuses on one important aspect of these mutations: Velar Mutation results in a change of the major place of articulation (from dorsal to coronal), while Coronal Mutation generates no similar change (the coronal place remains). This difference will be central.

A family of output-output faithfulness (paradigm uniformity, PU) constraints will be relevant for the present purposes (Kenstowicz 1996; Benua 1997; Steriade 2000). Such constraints have an important function of improving the transparency of morphological relationships between words and, thus, may facilitate lexical access. Dressler (2003: 464) refers to this pressure as “morphotactic transparency” and adds that “the most natural forms are those where there is no opacifying obstruction to ease of perception”. Research on morphological processing suggests that transparent phonology aids morphological decomposition (Frauenfelder & Schreuder 1992: 173). Output-output faithfulness is violated whenever a consonant undergoes a mutation. Both Velar Mutation and Coronal Mutation incur a violation of output-output faithfulness.

IDENTPl, an output-output faithfulness constraint given in (16), is used to represent the difference between the effects of Velar Mutation and Coronal Mutation.

IDENTPl is violated when Velar Mutation applies (–coronal, +dorsal > +coronal, –dorsal). When Coronal Mutation occurs, the constraint is respected (+coronal, –dorsal > +coronal, –dorsal). This analysis relies on output-output faithfulness constraints, as opposed to input-output faithfulness constraints. Identity relations between a derivative and its base are evaluated.

3.1 Historical change: why coronals underwent mutations but velars did not

This section aims to account for the emergence of a general pattern, in which coronals show mutations before -ist-a, while velars do not. Thus, the proposed analysis is historical and refers to the time when the suffix -ist-a became productive in the language. At that time, there were no morphophonological schemas related to the suffix -ist-a that would be sufficiently entrenched in the lexicon to determine the output. As mentioned in Section 2.1, there is historical evidence that Coronal Mutation initially applied across the board and was phonetically conditioned. Therefore, the evaluation of candidates proceeds according to phonological constraints. The modern-day complexities surrounding coronals are addressed in Section 3.2, where it will be argued that currently morphophonological schemas play a greater role than at the inception of the -ist-a pattern.

In explaining why coronals underwent mutations before -ist-a and velars did not, I make use of the observation made in Section 2.6 that Coronal Mutation does not change the major place of articulation, while Velar Mutation does. Thus, IDENTPl is violated by Velar Mutation but not by Coronal Mutation.

In addition to IDENTPl, a faithfulness constraint that will be relevant is MAXV_[±back], formulated in (17a). Given that [ista] is the principal variant of the suffix, while its allomorph [ɨsta] is positionally restricted (see Sections 2.1–2.2), MAXV_[±back] is violated when [ɨsta] is used. MAXV_[±back], rather than IDENTV_[±back], is used because central vowels, including [ɨ], lack place features (Clements and Hume 1995). In consequence, the change of [i] to [ɨ] involves the deletion of [–back], which is penalized by MAXV_[±back]. IDENT_[±anterior] in (17b) enforces faithfulness to [±anterior]. The constraints regulating consonant mutations in Polish are formulated in (17c, d) (based on Rubach 2007). This analysis rests on the assumption that all features are binary.¹⁷

The AGREE constraints are used here as shorthands for restrictions on specific consonant + vowel sequences. AGREE_[±coron] effectively bans velars before front vowels, e.g. [ki gi xi]. Following Rubach (2007), the formulation of AGREE_[±coron] in (17c) ensures that the constraint is moot for the vowel [ɨ], as central vowels lack place features. AGREECOR_[±back] in (17d) bans non-palatal coronal consonants before front vowels, e.g. [ʂi ʐi tʂi dʐi] and [ti di tsi dzi]. In line with the evidence mentioned in Section 2.1, the latter constraint requires phonological palatalization; presence of coarticulatory effects of the following vowel, e.g. [tʲi], is not sufficient to satisfy it. The formulation of AGREECOR_[±back] illustrates the already-mentioned assumption that feature sets are established on the basis of the phonological behavior of segments in a particular language and possibly in a particular alternation, rather than on the basis of a universal hierarchy of features. In the proposed analysis the feature [±back] can co-occur with both [±dorsal] and [±coronal], which in essence is reminiscent of the claims of Padgett (2002) and Halle (2005).

In (18), an evaluation of a word in -ist-a with a base-final velar is shown. The high-ranked IDENTPl mandates that a candidate with no change of major place be selected (candidate a). Candidates (b) and (c) are eliminated due to a change from dorsal to coronal. Some remarks are in order about the AGREE constraints. Candidate (a) violates AGREE_[±coron] because the consonant does not agree in [±coronal] with the following vowel. Candidate (b) violates AGREECOR_[±back] because a postalveolar [ʐ] is [+back], while the following vowel is [–back]. AGREE_[±coron] is respected because both the consonant and the vowel are [+coronal]. Candidate (c) fares well on AGREE_[±back], as both the consonant and the vowel are [+back]. It also vacuously satisfies AGREE_[±coron], because, as mentioned above, [ɨ], being a central vowel, is not specified for place features. Candidate (d) incurs a fatal violation of MAXV_[±back], because the backness of the relevant vowel has been modified. This evaluation demonstrates that AGREE_[±coron] must be ranked below IDENTPl and MAXV_[±back].

(18)	Evaluation of a derivative in -ist-a of [tʂɔwg-u] ‘tank’

The evaluation in (19) focuses on words in -ist-a with base-final coronals. The faithful candidate shown in (a) fails to respect agreement in backness (a violation of AGREECOR_[±back]) and loses to candidate (b), which satisfies AGREECOR_[±back] by employing an alveolopalatal before a front vowel. IDENTPl and AGREE_[±coron], the latter not shown, are respected by all the candidates. Surely, the change of [t] to [tɕ] in the winning candidate cannot be of no consequence for faithfulness. The constraint that is violated here is the low-ranked IDENT_[±anterior]. The tableau also confirms the relevance of MAXV_[±back]. Candidate (c) is not optimal because the quality of the vowel has been changed.

(19)	Evaluation of a derivative in -ist-a of [flɛt] ‘flute’

It has been shown that mutations of velars were avoided, as they involve a change of major place, the latter being detrimental to base recognition. In contrast, mutations of coronals were tolerated, because they do not result in a comparable modification. The ranking responsible for the emergence of the pattern -ist-a is given in (20).

3.2 Present-day developments: why (some) mutations are being eliminated

In this section, we address the issue of the gradual elimination of consonant alternations from -ist-a words with base-final coronals [t d r]. It is worth noting that alternations with base-final coronals [s z n] show a weaker tendency towards elimination. I refer to an interplay of the type frequency of a pattern with the requirement that the base be transparent in the derivative (output-output faithfulness). In contrast to the analysis in Section 3.1, where phonological constraints determined the output, I assume that, the analysis in this section must refer to morphophonological patterns because at the present stage some patterns are more entrenched than others and this must be reflected in the grammar. This assumption is based on the finding that the frequency of a pattern in the lexicon (its type frequency) determines its productivity. As words showing a particular pattern accumulate, they form a gang. When the gang reaches a certain threshold, a schema emerges and the size of the gang determines the productivity of the schema (gang effects; see Section 2.3). In Section 3.2.1–3 we look at data drawn from a dictionary and a corpus, and Section 3.2.4 examines experimental data probing native speaker intuition.

3.2.1 Dictionary entries

Table 3 provides the counts of dictionary entries with the variants -ist-a and -yst-a for each base-final consonant drawn from the reverse dictionary of Polish (Bańko et al. 2003). Labials, velars and the coronals [s z n] show categorical behavior: labials and velars do not mutate, while the coronals [s z n] mutate. The coronals [t d r] show two competing patterns. Looking at the type frequency of the mutated and non-mutated variants for each derivative with base-final [t d r], it appears that the lexical entries without mutations are actually more numerous. A corpus-based analysis of words with base-final [t d r] yields similar results. Based on a search of The National Corpus of Polish, words showing stable alternating patterns exhibit an average frequency of M = 391.17, while words without alternations and fluctuating words have an average frequency of M = 38.13. The difference is significant at p < .05. An in-depth analysis of the corpus-based data is postponed until the next section, where the concept of listedness is elucidated.

Table 3

Number of alternating and non-alternating patterns for words in -ist-a/-yst-a.

BASE-FINAL CONS.	MUTATIONS	NO MUTATIONS	% OF NO MUTATIONS
[p]		13	100
[b]		10	100
[m]		37	100
[f]		11	100
[v]		30	100
[t]	20	33	62.3
[d]	18	21	53.8
[r]	28	51	64.6
[n]	109		0
[s]	19		0
[z]	2		0
[ts]		6	100
[ʐ]		11	100
[ʂ]		6	100
[dʐ]		1	100
[k]		7	100
[g]		11	100
[x]		5	100

It is argued that PU pressures are responsible for the elimination of morphophonological alternations between the base and the derivative. The schemas in Table 4 represent some of the attested morphophonological patterns based on dictionary entries. They are divided into two categories according to the behavior of the base-final consonant: the schemas in the first column show alternations (mutations), unlike the schemas in the second column.

Table 4

Selected alternating and non-alternating schemas.

	ALTERNATING SCHEMAS	NON-ALTERNATING SCHEMAS
a.		[…p] ↔ [[[…p]ist]a]
b.		[…m] ↔ [[[…m]ist]a]
c.		[…f] ↔ [[[…f]ist]a]
d.	[…t] ↔ [[[…tɕ]ist]a]	[…t] ↔ [[[…t]ɨst]a]
e.	[…d] ↔ [[[…dʑ]ist]a]	[…d] ↔ [[[…d]ɨst]a]
f.	[…r] ↔ [[[…ʐ]ɨst]a]	[…r] ↔ [[[…r]ɨst]a]
g.	[…s] ↔ [[[…ɕ]ist]a]
h.	[…z] ↔ [[[…ʑ]ist]a]
i.	[…n] ↔ [[[…ɲ]ist]a]
j.		[…ts] ↔ [[[…ts]ɨst]a]
k.		[…dʐ] ↔ [[[…dʐ]ɨst]a]
l.		[…ʐ] ↔ [[[…ʐ]ɨst]a]
m.		[…k] ↔ [[[…k]ist]a]
n.		[…x] ↔ [[[…x]ist]a]

For the base-final consonants [t d r] in (d–f), both the alternating and the non-alternating schemas are present. In fact, in accordance with the quantitative data in Table 3, the non-alternating patterns seem to be more frequent in the lexicon than the alternating patterns. Given that the frequency of a pattern determines its strength, the strength of the non-alternating schemas for [t d r] is higher than the strength of the corresponding alternating schemas in modern usage and is reflected in the relative ranking of the schema-constraints in (21) (“>>” indicates dominance).

(21)	a.	[…t] ↔ [[[…t]ɨst]a] >> […t] ↔ [[[…tɕ]ist]a]
	b.	[…d] ↔ [[[…d]ɨst]a] >> […d] ↔ [[[…dʑ]ist]a]
	c.	[…r] ↔ [[[…r]ɨst]a] >> […r] ↔ [[[…ʐ]ɨst]a]

Returning to Table 4, the base-final consonants [s z n] in (g–i) show only alternating patterns, while for the base-final consonants [p b f] in (a–c) and [ts dʐ ʐ k x] in (j–n) only non-alternating patterns can be identified. The patterns in the first column impinge on PU because they introduce segmental alternations in the derivatives. In other words, the transparency of the base is diminished.

The issue that needs to be tackled first is the selectivity of the bias against alternations within coronals.

(22)	a.		Alternating patterns	Non-alternating patterns
		[t]	[t ~ tɕ]	[t ~ t]
		[d]	[d ~ dʑ]	[d ~ d]
		[r]	[r ~ ʐ]	[r ~ r]
	b.
		[s]	[s ~ ɕ]
		[z]	[z ~ ʑ]
		[n]	[n ~ ɲ]

As for the base-final coronals [t d r] in (22a), mutations can be avoided thanks to the patterns on the right, which have emerged recently and are affecting more and more words (evidence for the latter claim is given below). Looking at (22b) it is clear that for the coronals [s z n] the older alternating patterns have not been replaced.

How do we account for the asymmetry in the treatment of the two groups of coronals in (22)? Crucial to the analysis is the fact that, while all of the patterns in (22) satisfy IDENTPl, the alternating patterns in (22a) incur a violation of IDENT_[±strid], a faithfulness constraint defined in (23).

(23)	IDENT[±strid]O-O
	Corresponding consonants in the stem of the base and the output have an identical value for the feature [±strident].

IDENT_[±strid] is violated by the alternating patterns in (22a) because [t d r] are specified as [-strident] and when they mutate in -ist-a words they become [tɕ dʑ ʐ], which are [+strident]. In contrast, the specification of the [s z n] in (22b) remains the same in this respect: both [s z] and [ɕ ʑ] are [+strident] and both [n] and [ɲ] are [-strident]. IDENT_[±strid] is respected.

Similarity can be measured in terms of features, as shown in (24). The contrast [s] vs. [ɕ] involves a difference in one feature, while the contrasts [t] vs. [tɕ] and [k] vs. [tʂ] each involve a change of two features, the latter mutation crucially involves a change in major place.¹⁸ As a result, mutations of [t d r] and [k g x] involve more contrast (i.e. they result in less similar segments) between the base and the derivative than mutations of [s z n]. This observation can be used to explain the relative acceptability of the mutations for [s z n] and their avoidance for [t d r] (gradient) and [k g x] (categorical).

(24)

3.2.2 Listedness

This analysis makes crucial reference to the availability of a lexical representation of a particular word. During lexical retrieval, the lexical representations of some morphologically complex words are stored and available. The mental representations of other words are not available and the words need to be processed on-line from their component parts, e.g. base + affix. I refer to a dual-route model of lexical access (McQueen and Cutler 1998; Hay 2003; Plag 2012), which proposes that the availability of a mental representation of a word depends on the word’s token frequency. A complex word may be accessed via the whole-word route or the decomposed route. The choice between the two ways of access is determined by the relative frequency of the derivative and the base, as well as the phonotactic constraints of the language. If the derivative is more frequent than the base, the derivative is more likely to be accessed via the whole-word route. If, on the other hand, the base is more frequent than the derivative, we expect the derivative to be accessed via the decomposed route. The latter type of relationship occurs more commonly in general and has been identified in all the cases of -ist-a words discussed here. I extend the model and propose that some words in -ist-a are accessed via the whole-word route and others via the decomposed route. The choice depends on their absolute frequency. In other words, frequency impacts the morphological decomposability of the derivative. The second factor that has been found to determine the decomposability of words is phonological: the phonotactic probability of segmental sequences (Plag 2012). Plag’s (2012) phonotactic probability in some crucial ways parallels the phonological restrictions (formulated in terms of features) on consonant-vowel sequences, e.g. *[ʂi ʐi tʂi dʐi], enforced by AGREE constraints in Section 3.1.

The constraint USELISTED promotes the selection of a stored lexical representation as the input. If such a form is unavailable (for instance, as for rare and novel words), the constraint is moot, as all the potential outputs violate it (Zuraw 2000). For example, the word [altɕ-ist-a] can be accessed via the listed form /altɕ-ist-a/ or via its component morphemes /alt/ + /ist-a/. USELISTED promotes the former type of access, whenever available.

(25)	USELISTED
	The input portion of a candidate must be a single lexical entry.

The strength of a word’s lexical entry has been found to depend on its frequency of use (token frequency) (Hay 2003). Therefore, it is necessary to gauge the strength of the lexical representations of words in -ist-a. In (26) I give the frequencies of the relevant bases with stem-final [t d r] and their derivatives in -ist-a. The words were compiled on the basis of the data extracted from the reverse dictionary (Bańko et al. 2003) and later used to investigate token frequency in a corpus. The frequencies of the lemmas were extracted from the balanced National Corpus of Polish (12 January 2018), which contains 250 million words from various sources from 1988 to 2010, including literature, newspapers, journals, conversations and internet texts. The Pelcra search engine (Pęzik 2012) was used. The items in (a) illustrate stable alternating patterns. The items in (b) show the only four vacillating words found in the corpus. The two values stand for the number of extracted alternating/non-alternating lemmas. Stable non-alternating patterns are exemplified in (c). Figure 2 shows boxplots for frequency of all the words in -ist-a with and without mutations identified in the corpus.

Figure 2

Boxplots for frequency of words in -ist-a with and without mutations on a log scale with base 10.

(26)	Token frequencies of representative derivatives in -ist-a and their bases

All the derivatives extracted from the corpus show a lower frequency than their corresponding bases. An interesting tendency is discernible in the boxplots in Figure 2. The derivatives with mutations, such as those in (26a), on average exhibit a higher frequency than the derivatives without mutations, exemplified in (26b, c) (with the notable exception of stypen[d]-yst-a, marked as an outlier). This agrees with the predictions of usage-based models. Well-established words are expected to be more stable, hence more resistant to PU pressures, than are rare or novel words, because of their strong mental representations. Rare and novel words, on the other hand, are more susceptible to the influence of PU pressures because their representations are weak or unavailable. Thus, we expect a difference in the frequency of words with and without mutations.

In analyzing the frequency of all the words with base-final [t d r] available in the corpus, two things deserve a mention. First, the overall number of the words is not very high: 47 with mutations and 24 without mutations. Second, among the words with mutations the degree of variance of frequency is very high. For example, there are 11 words in this group with the frequency below 10. At the same time, 6 words in this group exceed the frequency of 1000 (two of them exceed 3000). The mean frequency of all the derivatives showing stable alternating patterns, (26a), is M = 391.17, SE = 123.94 and the median is Mdn = 92. In the group of vacillating and non-alternating words, (26b) and (c), the degree of variance is not that high. Only one word, stypen[d]-yst-a (visible as an outlier in Figure 2), shows a frequency higher than 100; the frequency of most of the remaining words is well below 100. Crucially, the frequencies of the words that vacillate, all of them shown in (26b), are low. In fact, propagan[dʑ]-ist-a/propagan[d]-yst-a is the only one among them whose frequency exceeds 100. Note, however, that the two variants of this word show very similar frequencies, which might explain their persistence. In the remaining cases in (b), one of the variants is significantly more frequent than the other (even though the overall frequency does not exceed 10). The mean frequency of the vacillating and non-alternating derivatives (excluding stypen[d]-yst-a) is M = 38.13, SE = 6.21 and the median is Mdn = 33. A non-parametric Mann-Whitney test was run on the data. On average, the words with mutations are more frequent than the words without mutations or vacillating words, U = 356.00, z = –2.308, p < .05.¹⁹

Finally, how can we explain the unusual behavior of stypen[d]-yst-a? Being of recent origin, the word was formed when the non-alternating pattern was more frequent than the corresponding alternating pattern. It began to be used very frequently and its representation stabilized (without mutations). Words like stypen[d]ysta show that high-frequency words can represent two patterns in modern usage: an alternating one when the word was formed earlier and a non-alternating one when it is of a more recent origin.

Some of the important predictions of usage-based models are upheld. Words of the highest frequency (i.e. above 1000 in the analyzed data) are phonologically stable. We would not expect a word like ren[tɕ]-ist-a, whose frequency is the highest among the analyzed words (3780), to appear in current usage in its corresponding form without a mutation, i.e. as *ren[t]-yst-a. Its strong memory trace prevents such an outcome and USELISTED promotes the selection of its stored representation as input. In contrast, words showing a relatively low value of frequency are susceptible to PU pressures, for example, bonapar[t]-yst-a with the frequency of 29. In addition, low-frequency words show more variation than high-frequency words, e.g. al[tɕ]-ist-a/al[t]-yst-a (4/1). A less expected discovery is a substantial number of words with mutations whose frequency is low – 11 of them have a frequency below 10, for example, kontraban[dʑ]-ist-a (7). Their persistence may be due to the gradual and probabilistic nature of change (lexical diffusion). What is more, some of the words with mutations (e.g. kontraban[dʑ]-ist-a ‘contrabandist’) are relatively old and predictably comply with the phonological requirements which were regular at the time of their formation. Their low frequency today may well reflect changes in the society. In low-frequency vacillating words such as al[tɕ]-ist-a ~ al[t]-yst-a the variant with a mutation is most likely to be older than the variant without a mutation. Another reason for the persistence of low-frequency words with mutations might have to do with the fact that the corpus represents mostly written language, which means that the data reflect conservative usage and may be an imperfect representation of spoken language. In order to address this issue, in Section 3.2.4, we will examine data elicited in an experiment involving native speakers of Polish.

3.2.3 Analysis

We begin with an analysis of words with base-final [s z n], which show mutations. IDENTPl is not shown, as it is moot for coronals. In the evaluation of bas-ist-a [baɕ-ist-a] ‘bass player’ (a recent word), the listed form is unavailable, as indicated below the Input in (27). The mutation in the winning candidate causes a violation of the low-ranked IDENT_[±anterior], the latter requires faithfulness to the feature [±anterior]. The evaluation highlights two important issues. First, novel words do not have strong lexical representations, therefore, USELISTED is violated by all the candidates, and the relevant schema-constraint determines the output. Second, once morphophonological schemas become entrenched (i.e. patterns are morphologized), markedness constraints regulating mutations (e.g. AGREECOR_[±back]) take a back seat. This is due to the preference for the morphological over phonological conditioning of patterns, discussed in Section 2.5. The entrenchment of a pattern is a function of the size of the gang it represents (gang effects, see Section 2.3). In OT formalism, the morphologization of a pattern may be viewed as the emergence and promotion of the relevant schema-constraint in response to the increasing number of words that observe the pattern.

(27)	Evaluation of a derivative in -ist-a of [bas].

As regards derivatives with base-final [t d r], the token frequency values in the corpus of the representative words al[tɕ]-ist-a, bonapar[t]-yst-a and fle[tɕ]-ist-a are 4, 29 and 96, respectively. The appearance of al[tɕ]-ist-a alongside al[t]-yst-a is an indication that the word is variably accessed via either the whole-word route or the decomposed route. bonapar[t]-yst-a is mainly retrieved via the decomposed route. As the mental records of both words are weak (due to their low frequency), transparent bases facilitate their retrieval. Finally, fle[t] ~ fle[tɕ]-ist-a does not show fluctuations because the derivative, having its own mental trace (due to a considerably higher frequency than al[tɕ]-ist-a and bonapar[t]-yst-a), is mainly accessed via the whole-word route.

In the evaluation of the derivative from Bonapar[t]e the listed form is unavailable due to a low token frequency and, therefore, USELISTED (not shown) is violated by all the candidates. IDENT_[±strid] and IDENT_[±anterior] represent PU pressures.

(28)	Evaluation of a derivative in -ist-a of [bɔnapartɛ]

Currently, the non-alternating schema for base-final […t] has a higher type frequency, hence higher strength, than the alternating schema. This finds reflection in the ranking of the two schemas in tableau (28). Candidate (b) is selected because it satisfies both IDENT_[±strid] and the dominant schema. Candidate (a) uses neither of the schemas available in the lexicon for the formation of words in -ist-a and hence violates the respective schema-constraints. Candidate (c) is eliminated due to a violation of IDENT_[±strid].

To derive an earlier state of affairs when the alternating pattern prevailed in the lexicon, the ranking of the two schema-constraints must be reversed in (28). This would account for the expansion of non-alternating patterns at the stage when the alternating patterns were in fact more common. With IDENT_[±strid] ranked high, the candidate with a transparent base, (b), is selected regardless of the ranking of the two schema-constraints with respect to each other.

In the case of fle[t] ~ fle[tɕ]-ist-a, the derivative fle[tɕ]-ist-a has a relatively high token frequency, which entails that the listed form is available. The morphologically complex word is independently stored and hence is accessed via the whole-word route. The impact of PU pressures as well as schemas is outweighed by the impact of the high-ranked USELISTED.

(29)	Evaluation of a derivative in -ist-a of [flɛt]

Vacillating forms such as al[tɕ]-ist-a ~ al[t]-yst-a also indicate that the impact of lexical strength, established on the basis of token frequency, may reduce the influence of PU pressures. The token frequency of al[tɕ]-ist-a is low. Together with the attested variation, this is interpreted as evidence that the word may be alternately accessed via the whole-word route when the listed form is available, or the decomposed route when the listed form is unavailable. The availability of the listed form is determined probabilistically on the basis of token frequency.²⁰ When the word is accessed via the whole-word route, faithfulness constraints are outranked, and USELISTED effectively selects the output. The tableaux in (30) illustrate the variability of the output. When a stored derivative is available, as shown in the first tableau, candidate (c) is selected (with a mutation), as it is the only one that satisfies USELISTED. On the other hand, when the stored representation is unavailable and USELISTED is violated by all the candidates (the second tableau), the word is accessed via the decomposed route and derived on-line from its component parts. In such a situation, candidate (a) fails the evaluation because it does not respect either of the available schemas. Candidate (c) loses to candidate (b) because it contains a modification of stridency specification. To summarize, the variation between al[t]-yst-a and al[tɕ]-ist-a can be explained in a dual-route model of lexical access by assuming that the former variant is derived on-line from its component parts, while the latter is retrieved from memory via the whole-word route. The outcome depends on the probabilistic availability of the memorized representation.

(30)	Evaluation of a derivative in -ist-a of [alt]

Such vacillating forms suggest that the three pressures, i.e. PU, type frequency, and token frequency, are important in the language and that the actual usage reflects their combined effects. The differences between the impact of type and token frequency will be elaborated in Section 5. The attestation of al[tɕ]-ist-a alongside al[t]-yst-a may be explained by the different method of retrieval of the word. When it is accessed via the whole-word route, the stored representation is used, i.e. al[tɕ]-ist-a. When the word is accessed via the decomposed route, mutations in the base are detrimental to word recognition and PU becomes relevant, yielding al[t]-yst-a. In the case of bonapar[t]-yst-a, as the derivative with a mutation is not listed in the mental lexicon, a transparent base serves to facilitate word recognition. Finally, in the case of well-established words, such as fle[t] ~ fle[tɕ]-ist-a, the impact of PU constraints is mitigated by their strong lexical representations every time they are accessed. The ranking of the relevant constraints established in the course of the analysis is given in (31).

(31)	USELISTED, IDENTPl >> IDENT[±strid], […t] ↔ [[[…t]ɨst]a], […s] ↔ [[[…ɕ]ist]a] >> […t] ↔ [[[…tɕ]ist]a], IDENT[±anter]

3.2.4 Experiment

In order to offer more evidence for the continuous role of PU pressures in the grammar, I use the findings of a nonce-word experiment conducted on 61 participants, speakers of Polish, aged 19–23. The participants were asked to fill out a written questionnaire probing their intuitions regarding the formation of words in -ist-a/-yst-a. The main area of interest was the formation of -ist-a/-yst-a words for base-final coronals [t d r] and [s z n]. In the first part of the questionnaire, the participants were given examples of -ist-a/-yst-a words along with their meaning. At this stage, real words were used and only those whose bases ended in a non-coronal. In the second part, the participants were given a list of two-syllable nonce words and asked to form words in -ist-a/-yst-a. Non-coronals were used as fillers, as we do not expect significant variation here. In order to get a more varied sample of the relevant patterns, twice as many words with base-final [t d r] and [s z n] were elicited than with the remaining consonants, that is, each participant had to form two words with base-final [t d r s z n] and one word with base-final [p b m f v ʂ ʐ tʂ dʐ k g x]. The words comply with the Polish phonotactic constraints. The base-final syllable is of the structure CVC(C), where the V is [a], [ɔ], [u], [ɛ] or [i]. Two versions of the questionnaire, differing in the choice of nonce words, were assigned randomly to the participants, each containing 24 stimuli. 30 participants completed Questionnaire A and 31 participants completed Questionnaire B. The list of words used in the questionnaires is provided in the Appendix. An example of a task (translated) is given in (32). The respondents were instructed to provide the word that sounds the most natural to them and that there are no bad answers. They were asked to provide the first word that comes to their mind. The responses of one participant were excluded, as they were largely incongruous.

(32)	klunar is a new scientific discipline.
	A person who studies or represents klunar is called ______________.

The results of the experiment are given in Table 5. Predictably, the results are not as straightforward as the ones given for dictionary entries in Table 3. The category “other” subsumes all the cases in which the participants failed to form a derivative in -ist-a/-yst-a using one of the attested patterns (e.g. the expected derivatives of miarte[z] are miarte[ʑ]-ist-a and miarte[z]-yst-a, yet, 5 participants chose miarte[ʐ]-yst-a instead). For labials and retroflexes mutations were not elicited. Of particular interest are consonants which appear in both non-alternating (no mutation) and alternating (mutation) patterns in Table 5, i.e. velars and coronals (except retroflexes). While the former appear in alternating patterns sporadically, in the case of the latter, the variation reaches a noticeable level.

Table 5

Number of alternating and non-alternating patterns in -ist-a words for each base-final consonant in the nonce-word experiment.

BASE-FINAL CONS.	MUTATION	NO MUTATION	% OF NO MUTATION	OTHER
t	10	105	91.3	5
d	8	107	93	5
r	19	96	83.5	5
s	66	47	41.6	7
z	45	62	57.9	13
n	113	7	5.8	0
p		60	100	0
b		60	100	0
m		60	100	0
f		59	100	1
v		60	100	0
ʂ		58	100	2
ʐ		59	100	1
tʂ		60	100	0
dʐ		59	100	1
k	2	58	96.7	0
g	3	56	94.9	1
x	1	59	98.3	0

Table 5 provides grounds for several interesting observations. Mutations of velars are marginal and may be the result of an indirect influence of Velar Mutation (discussed in Section 2.1) on this pattern. Mutations occur more frequently in the case of coronals (except retroflexes). While non-alternating patterns occur in more than 80% of the cases for [t d r], for [s z] the rates of non-alternating patterns do not exceed 60%. The coronal nasal is the most susceptible to mutations; the rate of non-alternating patterns for [n] does not reach 6%.

A conditional inference tree analysis using the party package in R (Hothorn et al. 2006) has been run on the data, with outcome (mutation/no mutation) as the dependent variable and consonant as the predictor (18 levels). The category “other” has been omitted, as it is not particularly revealing. A conditional inference tree provides estimates of the likelihood of the value of the response variable (mutation/no mutation) on the basis of a series of binary questions about the values of predictor variable (consonant). This method has been chosen, as it is designed for binary dependent variables and deals well with unbalanced data and small or even zero cell counts (Tagliamonte & Baayen 2012).

The first partitioning in the decision tree in Figure 3 shows that mutations of [s z n] are overwhelmingly more common than mutations of the remaining consonants (p < .001). Similarly, mutations of [t d r] are more common than mutations of all the other consonants on the right-hand side of the tree (p < .001). In addition, [n] is significantly more susceptible to mutations than both [s] and [z] (p < .001); the latter two are also different from each other (p = .016). The likelihood of mutations for [r] is higher than for [t d] (p = .045). In a similar way, [k g] are different from all the other consonants. Crucially, mutations of [s z n] are significantly more common than mutations of [t d r], and the latter are more common than mutations of the remaining consonants.²¹ These findings are compatible with the OT analysis provided in the previous sections.

Figure 3

Conditional inference tree for outcome (mutation/no mutation) as the response variable and consonant as a predictor. Mutation is depicted as light gray, No mutation as dark gray.

Figure 3 shows a highly significant difference between the likelihood of mutations for [s z] and [n] in the experimental results. Why is the tendency to eliminate mutations stronger for both [s] and [z] than for [n]: 41.6% and 57.9 vs. 5.8%? The answer might have to do with cue robustness and perceptibility. Wright (2004: 43–44) reviews the activity of auditory nerve fibers and suggests that consonants receive an auditory boost in CV sequences. Crucially, the boost is particularly true of stops, fricatives and affricates but less so of nasals. Thus, the auditory difference between [n] and [ɲ] is smaller than the distance between [s] and [ɕ] (and [z] and [ʑ]). It follows that a higher rate of mutations for [n] than for [s z] can be explained by appealing to cue perceptibility. Steriade’s (2008) approach using P-maps and contrast-based constraints offers a formal solution. PU constraints require that the relevant parts of derivatives be maximally similar to their bases in order to aid word recognition. Acceptability of mutations of the base in the derivative is in fact gradient, with mutations that diverge the most from their base correspondents being less acceptable than mutations of consonants that render them more similar to their correspondents. In other words, mutations that introduce less perceptual contrast between the derivative and the base are preferred. An analysis employing P-maps appeals to the ranking of the contrast-based constraints ▵ (s~ɕ) >> ▵ (n~ɲ) and their interaction with phonotactic constraints and morphophonological schemas. This ranking derives from the fact that the alternation (s~ɕ) universally involves more perceptual contrast than (n~ɲ) and, therefore, is more likely to be avoided, all else being equal. Perceptibility also explains the distribution of palatal coronals in Polish: [s] is possible before [i] in recent borrowings ([sinus] sinus ‘sine’) and is thus potentially contrastive with [ɕ] ([ɕivɨ] siwy ‘gray’), while [n] is impossible before [i] across the board and neutralizes to [ɲ] in this context ([uɲifɔrm] uniform ‘uniform’).

The rates in Table 5 can be directly compared with the results of the investigation of dictionary entries in Table 3. Table 6 shows the rates of words without mutations in the two relevant groupings of consonants, that is, [t d r] and [s z n], and juxtaposes the results obtained in the experiment with those taken from the dictionary. A cursory look at Table 6 warrants a claim that the non-alternating patterns (no mutations) are more common in the experimental data than in the dictionary data for [t d r s z n].

Table 6

Rates of words without mutations in the experiment and the dictionary.

% OF WORDS WITH NO MUTATIONS	t	d	r	s	z	n
experiment	91.3	93	83.5	41.6	57.9	5.8
dictionary	62.3	53.8	64.6	0	0	0

Figure 4 offers a conditional inference tree with outcome (mutation/no mutation) as the dependent variable, and consonant ([t d r s z n]) and source (experiment/dictionary) as predictor variables. Table 6 and the decision tree in Figure 4 show two things: first, the experimental results are not as categorical as the dictionary data and, second, the alternating patterns are significantly less commonly used by the participants of the experiment than in the dictionary (p < .001 for [t d r s z] and p = .021 for [n]).²²

Figure 4

Conditional inference tree for outcome (mutation/no mutation) as the dependent variable and consonant ([t d r s z n]) and source (dictionary/experiment) as predictors. Mutation is depicted as light gray, No mutation as dark gray.

Admittedly, the comparison across the two conditions (dictionary/experiment) should be approached with caution because of the different methods of data collection. Dictionaries offer prescriptive data, often ignoring variation found in actual language use. Data extracted from a corpus provide a better match for experimental data because the method of collecting data for a corpus is more compatible with the method of obtaining data from respondents in experimental conditions.²³ Table 7 compares the number of mutations of base-final [t d r] in two conditions: experiment vs. corpus (The National Corpus of Polish). A logistic regression run on the data suggests that the condition variable (experiment/corpus) has a significant influence on the outcome variable (mutation/no mutation). The coefficient on the condition variable has a Wald statistic equal to 83.581, which is significant at the .001 level (df = 1). The overall model is significant at the .001 level according to the Model chi-square statistic (𝜒²(1) = 92.607).

Table 7

Number of words in base-final [t d r] with and without mutations in the experiment and in the corpus.

	MUTATION	NO MUTATION
experiment	37	308
corpus	47	24

There are two possible reasons for the different rates of paradigm uniformity effects in the experiment vs. in the corpus. First, while novel and rare words in the corpus correspond to nonce words in the experiment, well-established words in the corpus do not have analogues among the experimental stimuli. Given that novel words are more susceptible to paradigm uniformity pressures than established words, the rates of mutations are expected to be lower for the experimental results. Second, it is likely that the data indicate an on-going change that might lead to the gradual elimination of mutations from the -ist-a pattern. The fact that there is a difference between the dictionary and corpus data, on the one hand, and the experimental data, on the other, might suggest that PU pressures continuously affect mental grammars and are responsible for a gradual change in morphophonological patterns. In particular, the promotion of IDENT_[±anterior] above schema-constraints is held accountable for the reduction in the number of mutations for base-final [s z n] in the experimental results, as compared with the dictionary and corpus data.