1.1. The phenomenon and the implications

Turkish is an agglutinative language, with the majority of derivation achieved through suffixation. A rare instance where prefixation is observed in the language is the so-called partial reduplication (or emphatic reduplication) (Demircan, 1987; Lewis, 1967, i.a.). It is the only prefix present in the language except for foreign prefixes in borrowed forms. The partial reduplication is found with modifiers, namely adverbs and adjectives. As shown in (2),² emphatic variants are derived by prefixing a C₁VC₂ syllable. The initial C₁V are identical to the word-initial CV of the base. However, C₂ ends in one of the four consonants: -p, -m, -s, -r (Lewis, 1967), generally referred to as linking consonants (LC).³

On the other hand, if the base is V-initial, the LC surfaces as -p.

Partial reduplication in Turkish has been the subject of many studies, which have aimed to explain the conditions underlying the choice of the LCs (see e.g., Demir, 2018; Demircan, 1987; Dobrovolsky, 1987; Foster, 1969; Hatiboğlu, 1973; Kaufman, 2014; Köylü, 2020; Lewis, 1967; Sofu, 2005; Sofu & Altan, 2008; Taneri, 1990; Wedel, 1999; Yavaş, 1980; Yu, 1998, 1999). This phenomenon is particularly suited for shedding light on the nature of the identity avoidance effect in terms of its locality and feature specificity. It is a type of dissimilation in reduplication with fixed segmentism (Alderete et al. 1999). Crucially, prototypical instances of reduplication with fixed segmentism involve only a single fixed unit (e.g., [a] in Javanese habitual reduplication (Yip 1997:p. 18), or [m] in Turkish, (4), in which an invariant segment appears whatever the features of the base). However, as seen in (2), the number of fixed segments is four in Turkish emphatic reduplication. This allows for many more features to participate in the identity avoidance effect.

Another interesting property of the Turkish emphatic reduplication is that, it differs from most crosslinguistic examples in not just allowing multiple possible fixed segments, but also having variability in which LC would appear for a given stem, as exemplified in (5). As such, although it is not always reported in the literature (though see Müller, 2003; Wedel, 1999; Yu, 1999), there is indeed variability with respect to the choice of the LC with a substantial number of bases.

Moreover, with items that allow multiple LCs, variation exists as to which LC is preferred for a given item across speakers. Even the existence of a varying degree of preference for the choice of a particular linking consonant (and between multiple LCs) differentiates the Turkish phenomenon from the classic reduplication with fixed-segmentism. Crucially, in Turkish partial reduplication, as we will argue in this paper, the choice of the linking consonant is sensitive to the features of the base. In some instances of reduplication with fixed segmentism (e.g., Alderete et al. 1999; McCarthy & Prince, 1993), the identity of that fixed segment is attributed to the emergence of the unmarked, default form. However, as has been noted by Wedel (1999) and Yu (1999), this is not the case in Turkish, and these LCs appear in the output despite clear markedness violations and are not the unmarked segments in the language. The consonants [p, m, s] are not considered the default segments in Turkish, and instead [n, j] are considered to be so (Wedel, 1999). As such, the latter are known as ‘buffer consonants’ in traditional Turkish grammars, appearing in various contexts (e.g., breaking up vowel hiatus as in kedi-y-i ‘cat-[j]-Accusative’ “the cat”).

While the seminal work by Demircan (1987) and subsequent work all identified the identity avoidance effect as a major factor, their accounts often involve heuristically chosen features that operate over the first two consonants in the base.⁴ The current study (re)-examines the factors conditioning the choice of the LC, focusing on the C-initial forms since the V-initial forms are consistently reduplicated with the LC -p. We extend the previous studies by examining the nature of the proximity (locality) and the (dis)similarity (feature specificity) between the consonants in the base and the LC. Starting with the former, we examine both topics in turn, particularly how they have been handled in the Turkish literature on emphatic reduplication as well as their implications for the broader literature beyond Turkish.

1.2. Locality avoidance in Turkish and beyond

With respect to locality, most studies emphasise the importance of C₁ and C₂, yet some of them (implicitly) assume a cut-off after C₂. One of the questions this study addresses is whether the consonants beyond C₂ do not play a role in the choice of the LC. Let us begin by taking a closer look at the classic study on the phenomenon by Demircan (1987), which observes that the selection of the linker is subject to various dissimilation constraints. The primary observations are given in (6), in the format they are succinctly summarized in Yu (1999, p. 5 & p. 18) and with slight modifications.⁵

Note that the observations (ii) and (iii) by Demircan focus on the contrast between the linking consonant and the C₁ and C₂ of the base, while the observations also make reference to the whole base. However, this latter point has been underappreciated in some subsequent studies. For example, based on the minimal pair in (7), Kelepir (1999, 2000) argues that not only C₁ of the base, but also C₂ matters for the choice of the LC.

Similar to Demircan’s (1987) observation (i), Kelepir’s (1999) study also has a constraint/restriction, *Repeat [strident], which makes reference to the whole base (specifically, that rules out the strident linker [s] if there is a strident in the base). Yet, Kelepir’s constraint system centers around the comparison of the LC with respect to C₁ and C₂ of the base.

Wedel (1999, 2000) is another study that argues for the presence of dissimilatory phonological constraints in partial reduplication, yet lacks a constraint of the sort proposed by Demircan and Kelepir, who allowed for the scanning of the whole base to avoid the choice of an LC that shares all or some of the features of the LC. In fact, the author explicitly mentions that there should be a cut-off after C₂. This is reflected in the generalizations and the OT constraints that Wedel proposes. Consider (8):

The assumption that consonants beyond C₂ have no significant effect has led some researchers that do nonce-word studies to only create nonce-words consisting of two-consonant stems. This is most clearly seen in the study of Köylü (2020), whose all 48 nonce-words consists of bases with a maximum of two consonants.

In fact, examining the features labial, coronal, and strident, Yu (1999) found that the strident feature has an effect with respect to C₃, and corroborates Kelepir’s *Repeat [strident] constraint. More generally, this supports the observation that the LC is not restricted to identity avoidance effect with respect to C₁ and C₂. Crucially, Yu did not find a significant effect from segments beyond C₃, which he attributed to the limited number of adjectives and adverbs with more than three consonants in his corpus data.

To sum up, the role of C₁ and C₂ has been the focus of many previous studies of the Turkish phenomenon.⁶ There is nonetheless evidence from other languages that suggests all the consonants in most dissimilation phenomena could play a role in identity avoidance, being also subject to a distance-based decay effect (Zymet 2014, 2018). This effect states that the likelihood for the application of a phonological process decreases as transparent distance increases. Arabic is the poster child for this kind of effect in identity avoidance (see e.g., Coetzee & Pater, 2008; Frisch & Zawaydeh, 2001; Frisch et al., 2004; McCarthy, 1986; McCarthy & Prince, 1994). For example, OCP-Place is a single gradient constraint that restricts consonant co-occurrence in Arabic based on (i) similarity (see below for the discussion of similarity) and (ii) proximity (Frisch & Zawaydeh, 2001; Frisch et al., 2004). Particularly, the influence of similarity on consonant co-occurrence is affected by distance, as the constraint is weaker for non-adjacent consonants. Accordingly, the fact that many prior studies on Turkish partial reduplication consider only C₁ and C₂, and the null effect of segments beyond C₃ in Yu (1999) might also be due to the distance-based decay effect. The present study indeed uncovers that the identity avoidance in Turkish is subject to such an effect, where it holds for all the segments, with the effect being strongest from C₁ and being weakened as a function of its distance from the target segment.

1.3. (Dis)similarity of features in Turkish and beyond

The (dis)similarity of features plays an even more prominent role in the identity avoidance literature. For example, studies usually aim to probe which specific features identity avoidance constraints are sensitive to whether they are categorical (Bennett, 2013), or gradient, (see e.g., Frisch et al., 2004; Gallagher & Coon, 2009; McCarthy, 1986), or whether features matter to the same degree. If the answer to the latter is no, then what is the extent to which specific features matter as opposed to other features in a given language and cross-linguistically? (see e.g., Bye, 2011; Coetzee & Pater, 2008; Frisch et al., 2004; Gallagher & Coon, 2009; Graff & Jaeger, 2009) It turns out, typologically, not all features participate equally in identity avoidance. Based on an extensive survey of 46 phenomena by Bye (2011), the following phonological dimensions have been found to play a role: The place of articulation, the laryngeal state, the manner of articulation (continuancy, liquid, nasality), vowel height, and suprasegmental properties such as length and tone. However, major class features such as [consonantal], [sonorant], and [approximant] do not play much of a role. Concerning the place feature, [labial] is relatively common, while [coronal] is rare and [dorsal] is unattested. Furthermore, alternations involving laterals and rhotics are relatively common.

When we look at previous studies on Turkish, we observe that they tend to focus on specific (set of) features. For instance, the analysis in Kelepir (1999) is built on avoidance constraints that use specific features. In particular, base consonants contrast with their correspondents in the reduplicant in place and sonorancy (i.e., the features [coronal], [sonorant], [labial], or [continuant]). Demircan (1987) identifies [coronal], [labial], and [nasal] as features of importance, and Yu (1999) adds that [strident] and [approximant] features also play a role in the interaction of segments.

Some remarks are in order regarding these previous studies on Turkish and their connection to the larger literature. Firstly, the choice of the relevant features is usually heuristic. For instance, Wedel (1999, 2000) uses PLOSIVE-αPL to explain why [p] is not selected if C₁ in the base is labial, while Kelepir (2000) uses the exact consonant [b] directly through the constraint, *-pb-. The heuristic nature mostly stems from the fact that researchers posit features, OT-constraints, or perception-related restrictions that can explain the respective datasets. In fact, a lot of the proposed constraints are correlated and refer to overlapping issues, in that a constraint might be implicated or even entailed by another constraint. For example, the rule (i) in (6) encompasses the rule (ii). Secondly, and relatedly, it is unclear whether these features are the only features that matter, or perhaps there are some other features that may also trigger identity avoidance, but have been missed since most studies rely on the researcher’s informed observations rather than a systematic examination of all possible features. Thirdly, cross-linguistic studies have found that features may differ in their strength in identity avoidance. For example, Gallagher and Coon (2009) find that [+strident] and [+ejective] have greater effects than others in Chol, while non-coronal place features were found to play a greater role in Arabic and Muna (Coetzee & Pater, 2008; Pierrehumbert, 1993). As such, various features have been shown to trigger the identity avoidance effects of varying strengths in different languages. Several proposals have been put forth with the aim of capturing the varying strength of the features. For example, Frisch et al. (2004) proposes a similarity metric of natural classes, which are shown to strongly correlate the observed-over-expected ratios (O/E) of consonant pairs co-occurring within Arabic roots, in a way that relies on the feature inventory of a language. However, data from other languages have not been able to replicate the effect of this metric. Coetzee and Pater (2008) failed to establish a correlation using this metric for Muna and Rotuman. Their discussion of place of articulation patterns in Muna and Arabic also demonstrates that the relative strength of place co-occurrence patterns cannot be due to inventory structure alone. Similarly, Graff and Jaeger (2009) found no evidence for such a correlation using this metric by any of the three languages examined (Dutch, Aymara, and Javanese). An alternative approach was proposed by Graff and Jaeger (2009) which did find effects of feature-specific effects in identity avoidance of the three language using a more complex model. This model allows each individual feature from each consonant in the base being weighted freely.

Additionally, previous analyses involved using a specific set of features, but their formulation of these features are often correlated, for example, a constraint is entailed by another. The choice of relevant features, therefore, tends to be heuristic. This is partly due to feature redundancy, a design property of standard versions of feature theory. Multiple sets of relevant features that could play a role in consonant dissimilation. For example, one could use any of following features [nasal], [labial], [sonorant], [voice] to model the dissimilation with /m/. That is not to say redundant feature values do not matter. In fact, Keyser and Stevens (2006), and Stevens and Keyser (1989) have proposed that redundant feature values can enhance the phonetic interpretation of contrastive values (see Clements & Ridouane, 2006 for an overview). On top of feature identity avoidance, previous analyses also make reference to total identity of C₂, such as C₂ cannot be identical to any of the consonants in the base. Again, these two types of factors are correlated. Previous analyses have also not fully addressed the question of whether the effect of total identity can be reduced to the effect of feature identity. These considerations call for revisiting the phenomenon with a different theoretical angle,⁷ as well as new methodological tools that can jointly evaluate identity avoidance of all features and linking consonants, therefore allowing us to tease apart their relative contributions.

Accordingly, besides probing the locality effect, this study also aims to better understand the inventory of features that may be involved in triggering the identity avoidance effect in Turkish emphatic reduplication, and if certain features have a greater effect in this phenomenon in relation to other features. In so doing, we also aim to elucidate the question of how the similarity between the consonants in the base and the LC should be specified, for instance, at the level of the total identity of the consonants, or at the level of the individual features? As will be discussed later in Section 2, this study adopts the methodological approach by Graff and Jaeger (2009) given its ability to help us rigorously examine the nature of both locality and feature specificity, and establish whether locality-sensitive feature-specific identity avoidance constraints are part of the grammar.

To address the questions that revolve around locality and feature-specificity, in this study we conducted an acceptability rating judgement experiment (which has rarely been used for the study of the partial reduplication phenomenon) as opposed to researcher’s intuitions or experimentally obtained forced-choice task responses. Among other things, our findings provide support for the view that speakers’ grammars have active identity avoidance constraints that operate on specific features (e.g., Gallagher & Coon, 2009) and the strength of their effects is a function of the distance between similar consonants (e.g., Pierrehumbert, 1993; Zymet, 2014, 2018). These are in line with previous findings on Arabic (Frisch et al., 2004), Dutch, Aymara, and Javanese (Graff & Jaeger, 2009) which highlighted a gradient feature-similarity-based restriction that is also subject to locality (see also Suzuki, 1998).

In terms of the nature of feature specificity, our results reveal that in the Turkish emphatic reduplication process, the similarity between the consonants in the base and the LCs operates at the segmental level (total identity) as well as the level of individual phonological features (partial identity). Furthermore, we found that not all features participate equally by allowing individual features to be free parameters. Our study also confirms the importance of only some of the features employed in the previous studies, such as [strident], [labial], and [nasal], but not others, such as [coronal], [sonorant], or [continuant]. The important features like [strident] and [labial] and the unimportant features such as [coronal] and [sonorant] are more in line with the cross-linguistic tendencies (see e.g., Bye, 2011; Pierrehumbert, 1993).

In terms of the effect of locality, the study demonstrates that not only the first two consonants, but all the consonants in the base form contribute to the identity avoidance effect (Zymet, 2014), with the strength of the effect decreasing further into the base. Another locality-related factor that studies on assimilation (and dissimilation) have noted is the role of syllable position (Rose & Walker, 2004) or syllable-role specific correspondence (Bennett, 2012). This approach states that matching syllable roles might contribute to segments’ (dis)similarity. Given this aspect has not been investigated in the context of Turkish emphatic reduplication, our study also examines whether syllable position is a significant factor. Strikingly, it turns out the identity avoidance effect in Turkish is also sensitive to the syllable position. In this regard, it lends support to the view that syllable position might play a role in contributing to segments’ (dis)similarity (Bennett, 2012; Rose & Walker, 2004; Suzuki, 1998). In particular, the effect would decay with the distance from the linking consonant and would be enhanced if the consonant in the base matches the constituency of the linking consonant (coda vs onset). For instance, in the case of Turkish, the linking consonant is always a coda.⁸ When a base form has the word shape C₁VC₂V, a distance-based decay effect was found, as hypothesised, with C₂ being less influential than C₁. However, the word shape C₁VC₂ exhibits the opposite pattern with C₁ being less influential than C₂. This surprising effect can be attributed to that the LC and the C₂ in C₁VC₂ are both codas (see Section 3.2.2 for the complete result). Accordingly, the Turkish phenomenon is particularly insightful in this regard too: It showcases the interplay of distance-based decay effect and syllable position effect, which in most studies are treated separately (due to the data involved). Moreover, unlike other examples that illustrate the syllable position effect, the presence of this effect in Turkish is not immediately clear both due to the just-mentioned interplay with distance-based decay effect and also because the dissimilation is not categorical (cf. Bennett, 2013), but much more gradient.

This study also replicates some of the previous Turkish-specific findings. For example, the preference hierarchy regarding the choice of the LC ([p] > [s] > [m] > [r]) still broadly holds. Methodologically we demonstrate that the precise nature of the identity avoidance effect can be revealed using hierarchical regression and statistical model comparisons (Graff & Jaeger, 2009; Zymet, 2019).

The rest of the paper is organised as follows: Section 2 outlines the details of our study. We first introduce the logic of this study before we present the materials (Section 2.1), followed by the methodological details of the experiment (Section 2.2). Section 2.3 introduces the variables of interest. Section 2.4 outlines the modelling procedure. In this section, we will introduce the model specification and evaluation as well as the modelling details of Study I on feature specificity (Section 2.4.2) and Study II on locality (Section 2.4.3). Section 3 presents our results for the two studies. Section 4 contextualises our findings of the Turkish case into the broader literature, focusing on feature specificity and locality, particularly the interaction between distance-based decay and the syllable position. This section also discusses issues such as factors beyond identity avoidance and representation of speakers’ knowledge. Section 5 summarises and concludes the paper.

2.1. Materials

The experimental items were 162 real base forms taken from previous studies to enhance the comparability and the replicability of this work. Most of the items were taken from the classic study by Hatiboğlu (1973) because many of the later studies also examined a subset of these items, and the rest were sampled from the other studies.⁹ The 162 items were then evenly divided into five lists (three lists have 33 items; one list has 32 items; and one list with 31 items) with each list containing roughly the same distribution of dominant linking consonants as well as variable items.¹⁰

2.3. Variables

This section describes the fixed effect variables and the random effect variables we included in our analyses. Variables that we excluded can be found in Appendix B.

2.4. Modelling procedure

The 162 items consist of 27 vowel-initial items and 135 consonant-initial items. Vowel-initial items were not analysed since they have an overwhelming preference for the LC [p] (Demircan, 1987; Kelepir, 1999; Sofu, 2005; Sofu & Altan, 2008, i.a.). This preference for [p] is supported by the descriptive statistics of the ratings of the 27 items. The mean and median ratings (on a scale of 1 to 7) are 6.402 and 7 for [p], 1.502 and 1 for [m], 1.533 and 1 for [s], and 1.197 and 1 for [r]. The 135 consonant-initial items were divided up into four groups based on the number of consonants they contain in the base and their word shapes as shown in Table 3. The five-consonant items were filtered out because there were only six of them and they might not provide enough statistical power for the analyses.

3.1. Study I: Feature specificity

In the following sections, we present the results from the model comparisons of different levels of feature specificity in Section 3.1.1. Given the model comparisons, the selected best models were then evaluated further individually for each of the three item groups. To enhance the interpretability of each of the predictors in the best models, all pairwise associations between the response and each of the predictors were computed and can be found in Appendix D. Finally, we report the detailed model evaluations of each of the predictors in Sections 3.1.2, 3.1.3, and 3.1.3 for the two-consonant, three-consonant, and four-consonant groups, respectively. The regression tables of all of the models can be found on the osf.io repository (see Section 5: Data accessibility statement).

3.2. Study II: Locality: Distance-based decay and syllable role

In the following sections, we examine which consonants play a role in the identity avoidance effect and whether a distance-based decay effect can be found through a series of model comparisons in Section 3.2.1. We then repeat the same analyses with the five most frequent word shapes in our dataset in Section 3.2.2 to examine the role of syllable position in the sense of Bennett (2012) and Rose and Walker (2004). The regression tables of all of the models can be found on the osf.io repository (see Section 5: Data accessibility statement).

3.2.2. Syllable position

The method of evaluating variable importance of each consonant remains the same as section 3.2.1. The difference is that in Section 3.2.1, the data was divided up by the number of consonants the items contain, while in the current section, the data was divided up by the five word shapes to enable an examination of the syllable position. These five word shapes are C₁VC₂V, C₁VC₂, C₁VC₂C₃V, C₁VC₂VC₃, and C₁VC₂C₃VC₄. The word shapes with the same number of consonants were compared. C₁VC₂ and C₁VC₂V were compared because C₂ is a coda consonant in C₁VC₂ but an onset consonant in C₁VC₂V. Similarly, C₁VC₂C₃V and C₁VC₂VC₃ were compared because C₂ and C₃ are syllabified different across the two shapes. Finally, C₁VC₂C₃VC₄ was examined because it could reveal the combined effect of syllable position and the distance decay effect, since it contains two sets of onsets (C₁ and C₃) and codas (C₂ and C₄) which differ only in their distance from the linking consonant.

AIC is being used only to compare models with the same set of data. For example, a fully-specified model was fitted over the subset of the data with the C₁VC₂V items. To evaluate the importance of C₁, we fitted a model on the same subset of the data but without any identity variables of C₁ (i.e., the model regression structure has the identity variables of C₂ and not of C₁). AIC was computed for both of these models (the initial model with both identity variables of C₁ and C₂, and the new model with only identity variables of C₂). The two AIC values were used to compare these two models.

The word shapes with the same number of consonants were compared and visualised. Figure 1 visualises the variable importance of each consonant in C₁VC₂V and C₁VC₂ base forms. C₁VC₂V shows a distance-based decay effect with C₂ being less important than C₁, while the C₁VC₂ shows the opposite pattern with C₁ being less important than C₂. These differences indicate that the divergence observed in the two-consonant group (as described in Section 3.2.1) was driven by C₁VC₂ which has slightly more base forms than C₁VC₂V (23 C₁VC₂ base forms compared to 19 C₁VC₂V base forms). The primary difference between C₁VC₂ and C₁VC₂V is that C₂ is a coda consonant in C₁VC₂ but an onset consonant in C₁VC₂V. The extra vowel/syllable in C₁VC₂V is also a possible cause of the divergence. To further evaluate these observations, we turn to the two word shapes in the three-consonant group.

Figure 1

Variable importance of each consonant in two-consonant base forms by word shape: C₁VC₂V and C₁VC₂.

Figure 2 visualises the variable importance of each consonant in C₁VC₂VC₃ and C₁VC₂C₃V base forms. C₁VC₂VC₃ and C₁VC₂C₃V both have three consonants and two syllables but they differ in terms of the constituency of their C₂ and C₃. The C₂ is an onset in C₁VC₂VC₃, but a coda in C₁VC₂C₃V. The C₃ is a coda in C₁VC₂VC₃, but an onset in C₁VC₂C₃V. C₁VC₂VC₃ shows a distance-based decay effect with a decrease in importance from C₁ to C₃. However, C₁VC₂C₃V shows a different pattern. A general distance-based decay can still be seen with C₁ being more important than C₃, but C₂ diverges from the pattern being more important than both C₁ and C₃. C₁VC₂C₃V matches the distance-based decay pattern observed in the three-consonant group (as described in Section 3.2.1). This is again not surprising since C₁VC₂VC₃ is the dominant word shape in the three-consonant group with 37 base forms, while C₁VC₂C₃V has 14 base forms.

Figure 2

Variable importance of each consonant in three-consonant base forms by word shape: C₁VC₂VC₃ and C₁VC₂C₃V.

One explanation of the patterns of these four word shapes is to consider two effects operating in tandem. The first effect is the distance-based decay effect which predicts a decrease in importance as distance increases. The second effect is that the syllable position effect which predicts coda consonants to be more important than onset consonants. The pattern of C₁VC₂C₃V can be explained if we consider both effects together. C₁ is more important than C₃ because of the distance-based decay effect, while the divergence of C₂ being more important than C₁ is because C₂ is a coda consonant. The C₂ of C₁VC₂VC₃ conforms to the distance-based decay effect and did not diverge because it is not a coda. Similarly, the divergence of C₂ from the distance-based decay effect in C₁VC₂ but not in C₁VC₂V can be explained since the C₂ in C₁VC₂ is a coda but an onset in C₂VC₂V. To further evaluate the two hypothesised effects, we turn to the four-consonant group.

Figure 3 visualises the variable importance of each consonant in C₁VC₂C₃VC₄ base forms. C₁ and C₃ are both onsets, while C₂ and C₄ are both codas. A distance-based decay effect can be observed with the two onsets and the two codas; C₁ is more important than C₃, and C₂ is more important than C₄. The syllable position effect can also be observed with C₂ and C₄ (codas) being more important than C₁ and C₃ (onsets).

Figure 3

Variable importance of each consonant in four-consonant base forms by word shape: C₁VC₂C₃VC₄.

4.1. Feature specificity

The model comparisons in Study I revealed two key findings. The first finding concerns the importance of total identity and partial identity. While previous studies have evaluated both types of identity in their formal analyses, it is entirely conceivable that total identity might not be needed after partial identity has been taken into account. Partial identity is a function of individual featural identity. If two segments are totally identical, both total identity and partial identity would be able to capture the same degree of similarity. Given that the consonants in the base forms might not be one of the four linking consonants, partial identity is expected to be much more important factor than total identity. In other words, partial identity can in theory capture what total identity can but not vice-versa. Our result suggests that partial identity is more important than total identity across all three item groups but total identity still contributes above and beyond partial identity.

The second finding concerns the nature of partial identity. Recall that cross-linguistic studies have found that features that participate in identity avoidance processes are not weighed equally, in that they may differ in the extent to which they influence the phenomenon in question. Gallagher and Coon (2009) found that features such as [+strident] and [+ejective] trigger greater OCP effects than others in Chol. Moreover, non-coronal place features have been shown to trigger OCP effects of varying strengths (see e.g., Frisch et al., 2004; Pierrehumbert 1993). In light of this background, we addressed the question of whether or not the individual features contribute equally in Turkish. It was found that the models with individual featural identity variables consistently outperformed the models with sum featural identity variables, therefore partial identity operates on an individual featural level. This finding echoes many of the formal analyses from previous studies which formulated a number of identity-avoidance constraints using specific features. In particular, Demircan (1987) identified coronal, labial, and nasal as features of importance and Kelepir (2000)’s analyses involve strident, labial, continuant, and sonorant.

To determine whether these six features are of particular importance in Turkish partial reduplication in general, we inspect the model summary of each of the three item groups in Tables 4, 5, and 6.²² Strident appears to be the most prevalent feature since it was significant in all of the consonant positions across all three models. Labial was similarly prevalent since it was only insignificant once (C₄ of the four-consonant model). Nasal was insignificant twice (C₂ of the two-consonant model and the four-consonant model). Continuant was insignificant three times (C₁ of the two-consonant model, and C₁ and C₄ of the four-consonant model) and it was in the opposite direction of identity avoidance twice (C₃ of the three-consonant and the four-consonant model). Sonorant was insignificant three times (C₁ of the three-consonant model, and the four-consonant model, and C₂ of the three-consonant model) and it was in the opposite direction in C₃ of the four-consonant model. Coronal was insignificant four times (C₁ of the three-consonant model and the four-consonant model, C₂ of the two-consonant model and C₄ of the four-consonant model) and it was in the opposite direction in C₃ of the four-consonant model. In terms of the size of the coefficients, strident and labial were in generally higher than the other features. The features, continuant, sonorant, and coronal, were not as consistent in terms of their statistical significance and the direction of their effects.

Overall, these models suggest that strident and labial were the most prevalent features. This finding is in line both with two of the previous studies on Turkish (i.e., Kelepir, 1999; Yu, 1999), which emphasised the importance of [strident], as well as cross-linguistic studies that found a similar effect (e.g., Bye, 2011; Gallagher & Coon, 2009). The feature [labial] has also been found to be influential as opposed to some other features, such as [nasal], which again are in line with the cross-linguistic typology, in that non-coronal place features were argued to play a larger role in the OCP literature (e.g., Bye, 2011; Pierrehumbert, 1993). Another finding of our study concerns the status of [coronal]: While both Demircan, 1987 and Kelepir, 1999 argue for the importance of coronal feature in their systems, our results found no such effect. In this regard, our finding is more consistent with the cross-linguistic generalization that coronal is not an influential feature (e.g., Bye 2011; Coetzee & Pater, 2008; Pierrehumbert, 1993).

While our findings are in line with cross-linguistic typology, it is important to remind ourselves that these cross-linguistic tendencies might be triggered by language-specific phenomena. In the case of Turkish, the alternating segments being [p, m, s, r] largely determine which features could potentially participate in identity avoidance. As such, although [strident] is influential in both Chol and Turkish, this importance is due to different factors.

We also believe that it is unsurprising that strident, labial, and nasal are particularly prevalent. Strident and nasal decrease the preference for [s] and [m], respectively, while labial decreases the preference for [p] and [m]. The three features together influence the preference for three of the four linking consonants except for [r]. The preference for [r] can be captured by using a markedness constraint since, as we have discussed earlier, it is the least preferred linking consonant even when other fixed and random effects were taken into account. In contrast to strident, for example, continuant was not a prevalent feature in our study, contrary to Kelepir’s analyses. Similarly, the nasal feature is more important than the sonorant feature, also used in the previous literature. These findings, we believe, further confirm the heuristic side of the inventory of features employed by many previous studies. Additionally, we speculate that this state of affairs could be explained by resorting subset-superset relation between the features in question. In particular, the feature sonorant encompasses a larger set of consonants as opposed to nasals; likewise, continuant picks out a larger set of consonant than strident. The results indicate that the feature that is more specific applies first or is more influential, in a way that makes the superset feature redundant.

Our results also reveal that the identity effects of a given feature for a given consonant can differ across different stem types. For example, the continuant identity effect of C₁ is not significant in bi-consonantal and quadri-consonantal stems, but it is significant in tri-consonantal stems.

These apparent inconsistencies call for some post-hoc speculative remarks. While it is plausible to assume that the same feature in the same position will exert the same strength across the board for stems of different lengths, we believe this does not need to be the case. The strength of the effect of a feature may very well be not determined once and for all, rather it may be contextually determined, which might include the number of other consonants and the syllable roles in the stem. For example, while a feature F in C₁ might be strong in a two-consonant stem, the same feature F might have a weaker role in a three-consonant stem if a competing feature is found in C₃ position. In statistical terms, additional consonants could introduce additional identity avoidance effects which could influence the significance of other consonants (e.g., C₁) by taking up variance in the regression model. As an analogy, in phonological theories like OT, constraints are not strictly-ranked, but rather are weighted, where multiple violations of lower-priority constraints are able to overcome the violation of a higher-priority constraint (the “gang effects”) (Pater, 2016). As such, we should not necessarily interpret the mismatches/inconsistencies about the effect of a particular feature as a negative or as a shortcoming of the model, but rather as indicators of the contextual factors that give rise to the apparent mismatch.

Relatedly, what can our statistical models, which were fitted separately over different word shapes, tell us about the Turkish speakers’ grammar? The models were trained on Turkish speakers’ responses to real words which as discussed in Section 2.2, are confined to certain feature combinations. Therefore, our models would allow us to predict how a Turkish speaker would choose C₂ when emphatically reduplicating a novel existing or nonce stem of given length with a C with a given feature specification in a particular position, as long as the stems have comparable feature combinations as the real words that we examined.

If our set of real words are generally representative of the reduplicable items in the Turkish lexicon, then two theoretical stances are conceivable in terms of what aspects of speakers’ grammar our models are capturing. On the one hand, one could take a position with a usage-based lexicon-based approach (e.g., Baayen et al., 2019; Bybee, 2003, 2006, 2010; Chuang & Baayen, 2021). Under this view, Turkish speakers’ grammar is confined to the attested feature combinations in their mental lexicon. Therefore, there are limits on what a Turkish speaker can learn and thus predict from the stems they know, and therefore on what we, as the modelers, can learn and thus predict from our models of what a Turkish speaker can learn from their lexicons. On the other hand, one could take a different position which allows for Turkish speakers’ grammar to not be solely based on their mental lexicon, but rather also under the influence of learning strategies, such as analogical learning (e.g., Arndt-Lappe, 2014; Nosofsky, 1986; Skousen et al., 2002) and discriminative learning (e.g., Baayen et al., 2011). Further research is needed to tease these two stances apart by examining nonce words that contains unattested feature combinations.^23,24

4.2. Locality: Distance-based decay and syllable role

The analyses of locality revealed a number of findings, including the domain over which this phenomenon operates, whether its effect is weighed equally across the domain or other factors such as syllable position effects are at play.

The first finding concerns the rightward boundary up to which the identity avoidance effect applies. Recall that most previous studies emphasise the importance of C₁ and C₂ (e.g., Kelepir, 1999; Wedel, 1999), with some of them (explicitly) assuming a cut-off after C₂ (Wedel, 1999). At the same time, Yu (1999) found that the strident feature has an effect with respect to C₃, which suggests that the LC is not restricted to identity avoidance effect only with respect to C₁ and C₂. Our results reveal that Yu (1999)’s conclusion was on the right track, but goes further to include all the consonants in the base, and that the effects of the other consonants in the base were not particularly strong or insignificant compared to C₁ and C₂.

The second finding shows that despite all the consonants contributing to the identity avoidance effect, the effect was not uniform but linear, with the strength of the effect decreasing further into the base. This suggests all the consonants play a role in identity avoidance but they are subject to a distance-based decay effect (Zymet, 2014, 2018), which states that the likelihood for the application of a phonological process decreases as transparent distance increases.

The third finding is that the distance-based decay effect is not completely linear, and is shown to be sensitive to the word shape. Specifically, some consonants diverged from the linear order, with C₃ of the four-consonant base forms is more important C₄, and C₂ of the two-consonant base forms is more important than C₁. We found that these divergences can be fully explained, once the constituency of a consonant is taken into account. We postulated that on top of the distance-based decay effect, the syllable position effect is also present with a coda consonant exhibiting a stronger identity avoidance effect than an onset consonant. When considering both effects in tandem, we were able to capture the patterns of importance across a number of frequent word shapes in our dataset. To further illustrate the syllable position effect, we can consider a minimal pair from our dataset – sık ‘tight’ and sıkı ‘frequent’. In both base forms, the C₂ /k/ should disprefer the linking consonant [p]. Given the syllable position effect, C₂ /k/ in sık should disprefer [p] more than that in sıkı. This is indeed confirmed by our rating study. sıp + sık has a mean rating of 3.7 which is less acceptable than sıp + sıkı which has a mean rating of 5.3. The coda effect was observed in a recent nonce word study. Köylü (2020) asked 14 native speakers of Turkish to reduplicate 48 nonce-words of four word shapes (VCV, CVC, CVCV, and VCCV). It was found that the linking consonant was never identical to C₁ or C₂ with the CVC nonce-words. However, the linking consonant was never identical to C₁ but sometimes identical to C₂ with the CVCV nonce-words. In other words, C₂ did not always affect the linking consonant in terms of total identity in CVCV but not in CVC. This can be explained by how C₂ is a coda in CVC which has a stronger effect on the linking consonant than when it is an onset in CVCV. Why should a coda consonant outweigh an onset consonant? We speculate that it is due to the fact that the linking consonant itself is also a coda consonant for consonant-initial base forms. Since the identity avoidance effect is a function of the similarity between a consonant in the base and the linking consonant, the similarity of the two consonants would be stronger if they have the same type of constituency. This finding lends support to the view that syllable position might play a role in contributing to segments’ (dis)similarity (Bennett, 2012; Rose & Walker, 2004).²⁵ In particular, the effect would decay with the distance from the linking consonant and would be enhanced if the consonant in the base matches the constituency of the linking consonant (coda vs onset). By uncovering the presence of this effect in the Turkish emphatic reduplication, our study also adds to the typology of syllable position effects. As mentioned earlier in the paper, Bennett (2013) notes that while harmony/assimilation is predicted between consonants with matching syllable roles, dissimilation is predicted for consonants with mismatching syllable roles (but not those with matching syllable roles). Turkish constitutes an example in which dissimilation favors matching syllable roles, with the effect of syllable role observed in non-categorical patterns with multiple features.

All in all, Turkish serves as a fruitful testing ground in showing that the identity avoidance effect holds for all the segments, with the effect being strongest from C₁ and being weakened as a function of its distance from the target segment. Moreover, our study reveals that in addition to the distance, the syllable position also plays a role in the application of the partial reduplication. Given our findings, one could formulate the grammar by deriving the weights of the distance-decay function (e.g., the decay parameter of the decay function which describes the shape of the decay), the weights of syllable position (e.g., the weight of coda), and their possible interactions (McMullin & Burness, 2021; Zymet, 2014). This, however, will likely require a more well-balanced dataset with nonce words in terms of the representation of particular features, word shapes and the number of consonants in the base. This is left for future studies.

4.3. Beyond identity avoidance

The current study examined the Turkish partial reduplication phenomenon as a phonological operation with a focus on the identity avoidance effect. Our models were able to explain a sizable portion of the variance of our naturalness judgements.²⁶ The identity avoidance factors captured around 25% to 35% of the variance (two-consonant group: 26.96%, three-consonant group: 33.67%, and four-consonant group: 28.36%) and, together with the random effects which captured the idiosyncracies of lexical items, the models were able to explain around 50% to 70% of the total variance (two-consonant group: 55.73%, three-consonant group: 55.75%, and four-consonant group: 70.98%). Our study therefore provided ample evidence that phonological factors play a major role in the process behind Turkish partial reduplication. However, there is still variance to be explained by factors beyond phonology, such as the lexicon, morphology, and semantics.

As also noted by an anonymous reviewer, our analysis captures a large portion of the patterns, yet certain minimal contrasts between reduplicated forms such as köp-kötü ‘very bad’ vs. kas-katı ‘very hard’ still are not fully accounted for. As shown in the Appendix, different linking consonants are selected for the same k-t sequence of consonants in the base, and in the same syllabic positions (as argued in Yu 1999, we take it that vowels do not play a role in the reduplication process). Previous analyses have also failed in providing a satisfactory explanation for such minimal contrasts. For example, Demircan (1987) brings up the intuition that speakers might avoid a reduplicant that resembles an existing root and confirms it in a corpus study by Kılıç and Bozşahin (2013), which demonstrated that root-level lexical statistics inversely correlate with the preference of a linking consonant. Speakers’ preference therefore might depend on the knowledge of distributional statistics at the morpholexical level. Speakers might not prefer to use a reduplicated form for the emphatic meaning, but instead prefer to use the word çok ‘very’ with the base form. However, in the case of kas-katı, it is still unclear why s is selected over p since both kas ‘muscle’ and kap ‘container’ are meaningful words in Turkish. In fact, the only linking consonant that does not result in a form resembling another Turkish root is m as kam is not a Turkish word, whereas kar ‘snow’ also is. Therefore, we acknowledge the presence of factors beyond phonological considerations that play a role in the choice of an LC.²⁷

In a series of corpus and experimental studies by Kaufman (2014), it was observed that the preference to reduplicate a base form depends on its semantic class and the semantic class of existing base forms that are frequently reduplicated. A low rating of a reduplicated form might not be due to phonological factors but rather due to the participant’s dispreference to reduplicate the base form. Incorporating morpholexical statistics and semantics can therefore provide a more complete picture of the Turkish partial reduplication phenomenon and we leave this for future research.

In light of the phonological generalizations uncovered by our experimental study, in the next section we discuss what speakers’ knowledge of this pattern looks like (i.e., what is the potential representation speakers have in mind when using this phenomenon?)

4.4. Representation of speakers’ knowledge

With respect to the representation of speakers’ knowledge of the Turkish partial reduplication phenomenon, one point of investigation that researchers have focused on is whether the choice of a particular linking consonant in partial reduplication is simply a matter of lexicalization, a term which has been interpreted in more than one way in the literature. One interpretation of this approach is whether the choice of the LC is random/arbitrary, or follows a set of generalizations or rules. If the conclusion is that the choice of the LC is arbitrary, then the phenomenon is considered to be lexicalised. Although earlier studies assumed the choice of the LC to be arbitrary, thus lexicalised (e.g., Foster, 1969; Lewis, 1967; Yavaş, 1980), a number of studies have argued that the choice is not lexicalised, and is indeed conditioned by various rules. These studies include Hatiboğlu (1973), Demircan (1987), Dobrovolsky (1987), Taneri (1990), Wedel (1999), Yu (1999), Kelepir (2000), Sofu and Altan (2008), Kaufman (2014). While these works vary considerably in their implementations of the observations, they converge on the view that the choice of the LC is not arbitrary or lexicalised, and that it is subject to several dissimilation constraints motivated by the OCP, similar to the analyses given for dissimilation processes in other, unrelated languages.

Some other studies have used nonce-words as a diagnostic as to whether Turkish partial reduplication is lexicalised or productive. Under the (often implicit) assumption that real words that participate in this phenomenon obey various rules or generalizations, studies in this line of research probe whether nonce-words are subject to the same generalizations. It turns out the results from these studies are far from clear, and have found varying, and sometimes conflicting results. For example, while Sofu (2005) concludes that speakers seem to extend at least some of the rules to nonce-words, two more recent studies, Demir (2018) and Köylü (2020), reach opposing conclusions as to the status of lexicalization and productivity of the reduplication patterns. While Demir (2018) interprets her results in favour of a lexicalization approach, Köylü (2020) argues that speakers do extend the same strategies they use for real words to nonce-words.

On the side of studies that investigate whether the Turkish emphatic reduplication is arbitrary or rule-governed, Wedel (1999) for example, argues that ‘native speakers do abstract some productive phonological generalization from the emphatic forms that exist’, as such speakers ‘have access to a uniform, constraint-based schema’ in using this phenomenon. This is at least the case for the LCs [p, s, m], while the LC [r] may indeed have lost its productivity since it is not used with novel forms and is the least utilised LC as confirmed by other studies including Yu (1999) and this current study.²⁸ Along the same lines, Yu (1998) also notes that ‘modern speakers of Turkish have some grammatical knowledge of emphatic reduplication’ on the basis of the results that speakers do not blindly choose a particular LC (e.g., [p], across the board) and concludes:

Proposals that claim the emphatic construction is unproductive and that all reduplicative closers must be lexically listed with the base form must be taken with great precaution, if not rejected altogether. The experiments reported here clearly suggest that native speakers of Turkish still retain some grammatical knowledge of the selectional restriction of the closer in the emphatic reduplication construction. (Yu, 1998, p. 39)

In this regard, we also add that if partial reduplication was simply a matter of lexicalization, it would be surprising that for a major number of items, speakers exhibit variation, in the sense that more than one linking consonant is permitted. Crucially, in these variable cases, the permissible linking consonants are not identical across items, and even for items that permit multiple identical LCs, the relative acceptability differs from one to another and across speakers. All these suggest that speakers are not blindly memorizing a specific LC for each base form, but instead, as Wedel (1999) and Yu (1999) already suggested, observe certain phonological generalizations, which our study aims to make precise.

Let us now turn to the other interpretation of lexicalization, which focuses on whether nonce-words parallel the behavior of real words. For example, a study by Demir (2018) examines the choice of LC with adults via a comparison of 10 real- and 34 nonce-words using an open-set response task. In one experiment, nonce-words are chosen to have a real word counterpart with identical consonants and word shapes, but different vowels. Demir finds that real words follow the expected observations noted by previous studies, as such using the four common LCs, [p,m,s,r]. On the other hand, in the case of nonce-words, speakers resort to strategies that are not available for real words. For example, they might copy the CVC from the base, as opposed to the CV plus LC strategy, or they might omit a linker altogether. Moreover, nonce-words which were chosen to analogise the real words do not show the same linking consonant as their supposed real word counterparts. Demir (2018) interprets these divergent results between real- and nonce-words to mean that there is not a pattern or set of generalizations for emphatic reduplication of real words, which one would expect to be extended to nonce-words. Therefore, reduplication of real words must be lexicalised.

However, this conclusion might be a bit too hasty. As we just noted, there are other nonce-word studies carried out on Turkish partial reduplication that arrive at different conclusions or interpretations. For example, Sofu (2005) examines the choice of LC with both adults and children using 38 nonce-words. She finds that adults and children conform to the expected patterns in their use of the classic LCs, [p,m,s,r], while children use [p] more than adults. Moreover, children use linking consonants that are different from [p,m,s,r] much more often than adults (e.g., [t,n,f]). This study shows that both adults and children do extend the patterns observed in real words to nonce-words as well.²⁹ Similarly, a more recent study by Köylü (2020) also concludes that Turkish native speakers extend the reduplication strategies they employ in real words to nonce-words. These conflicting results call for a careful investigation regarding the causes behind them. Here we speculate on a few potential issues.

There might be various methodological or linguistic factors that lead to this divergence in nonce-word studies. For example, all the previous studies simply present the nonce-word test items by themselves out of a context without assigning them any meaning. This is not a trivial choice in light of the fact that reduplication is applicable only to gradable modifiers. It is likely that these out-of-context nonce-words were not interpreted as such by the participants, and therefore participants resorted to strategies that differ from real words, whose meaning and property of being gradable they are aware of. The restriction regarding the category and property this reduplication requires could also be a factor as to why nonce-word studies might not be replicating or reaching lower scores.

This last point also relates to another concept, productivity, which usually comes up in the discussion of lexicalization. On the point of productivity, it is worth highlighting that it should be approached with caution. This is because emphatic reduplication, as just noted above, applies specifically to a subset of adjectives (i.e., gradable adjectives) and adverbs (which usually are built on adjectives in Turkish) and not to absolute adjectives or modifiers in general. As such, we can make sense of why it is found in a relatively small number of items in the language, due to its nature. Therefore, any statement about productivity should take this important aspect into consideration, and its potential role especially on studies investigating nonce-words.

Moreover, recall that Demir (2018) created certain nonce-words in anticipation of analogy with real words. For example, the nonce-word mava was created by Demir on the assumption that the participants would analogise it to the existing adjective mavi ‘blue’ while they are attempting to reduplicate mava. This is not an innocuous assumption, however. First, analogy does not rely on just a single word, but a number of words (e.g., the Generalised Neighbourhood Model by Bailey and Hahn (2001) considers all the words in the lexicon weighted by lexical frequency and form similarity to the target (nonce-)words). As such this assumption overlooks the complicated aspect of how analogy works. Secondly, some words might have a large number of noun neighbors, which might make the reduplication harder. For example, in the context of the nonce-word boyuz from Demir’s study, our native intuition (and those of our consultants) analogises boyuz to boyoz, which is an existing word that refers to a food item. As such, although the experimenter might have a certain real word in mind while designing the nonce-words, participants might have completely different real word analogies. Similarly, with mava, native speakers we have consulted brought up the nouns hava ‘air’, or tava ‘pan’ as the first words that come to their minds, both of which are nouns, rather than the adjective mavi that Demir had in mind as a control item. This is significant since the control items that Demir has in mind may not be the items participants are supposedly analogizing to, as such might explain part of the results. A further related note is that analogy may not be solely based on phonological properties, but might be due to semantic resemblance participants establish with real words they might think of.

As just discussed in Section 4.3, there are also other factors beyond identity avoidance, such as lexical (Kılıç & Bozşahin, 2013) and semantic factors (Kaufman, 2014), that play a role but were not consistently considered in these nonce-word studies.

In light of these considerations, the results of our study are in support of the view that partial reduplication is subject to various active phonological rules, particularly speakers have access to locality and feature-based conditions (or generalizations) that they are applying to items that are potentially intensifiable. In this regard, it corroborates the findings/intuitions raised by studies such as Yu (1998), Wedel (1999), and accords with Demircan (1987) and Kelepir (1999) who also conclude that speakers obey various phonological rules when they do partial reduplication. In particular, in this study we have uncovered that the phonological rules that are exhibited by the real words are much more graded than previously thought. The identity avoidance effect is both locality-sensitive (distance-based decay effect, and syllable position effect) and feature-sensitive (individual features with different weights).

With that said, it is important to keep in mind that as Frisch et al. (2004) argues, lexical information, including lexical idiosyncrasies, and rules are not mutually exclusive, in that there does not need to be a categorical choice between the two interpretations. It is possible to have a phenomenon which respects various constraints (e.g., locality- or feature-specificity, or phonotactic ones). As such, it is not necessarily the case that presence of phonological rules conflicts with or rules out the presence of lexical information. The Turkish emphatic reduplication may very well be an example of this sort. Wedel’s (1999) conclusion might also be in line with this interpretation, in that while certain linking consonants [p, s, m] still actively participate in the identity avoidance effect, [r] might have fallen on the memorization side of it.

Having better understood the nature of the phonological grammar that speakers might have, this naturally leads to the second question about lexicalization, (i.e., whether and to what extent these rules are being extended to nonce-words). Concerning nonce-words, we can bring in the insights raised by Becker et al. (2011) who found, on the basis of another phenomenon in Turkish, that nonce-words are subject to only some of the rules exhibited by real words (see also Harris, Neasom & Tang, 2016; Hayes & White, 2013). As such, there is no a-priori reason to expect that nonce-words fully conform to the same generalizations as real words, or reflect all of the generalizations/rules found for real words. Moreover, as discussed above, any study that aims to investigate nonce-words in the Turkish emphatic reduplication must also address a number of methodological and linguistic factors such as framing the nonce-words in a context that signifies its meaning and property of being gradable, as well as controlling for lexical and semantic factors. This can be done by using a combination of careful stimuli design and statistical modelling (Redington & Chater, 1996; Tang & Baer-Henney, 2023). Once these factors are controlled for, we expect (following Becker et al., 2011) that at least some of the real word generalisations would apply to nonce-words (some of which have already been argued to be the case, e.g., Köylü, 2020; Sofu, 2005).