English comparative constructions at different levels of schematicity: what is the role of adjective-specific variability?

2.1 Competition at multiple levels of schematicity

Building on the growing body of empirical research on lexically specific constructions, several studies have investigated lexical specificity in linguistic alternations. Perek (2014) uses collostructional analysis (Stefanowitsch and Gries 2003) to examine the English at-alternation, demonstrating that the construction expresses distinct meanings depending on the semantic class of the verb. Examples (1) and (2) below, taken from Pijpops et al. (2021: 490), effectively illustrate this contrast. In (1), with the “cutting” verb chip, the at-variant conveys a repeated action in which contact with the object is made. In contrast, when the variant involves a “striking” verb such as slap, as in (2), it conveys an attempted action where contact with the object is not necessarily made.

Pijpops et al. (2021) further develop the approach introduced by Perek (2014), employing collostructional analysis at multiple levels of schematicity and utilizing distributional semantics to investigate the Dutch naar-alternation. Crucially, their findings do not support systematic variation in the naar-alternation without lexical specifications, possibly due to verb-specific naar-alternations that exhibit distinct meaning differences. In other words, since lower-level alternations exhibit distinct distributions that are independent of each other, this dialutes clear semantic distinction at the most schematic level.

Both Perek (2014) and Pijpops et al. (2021) highlight the importance of taking into account the multilevel structure of the constructional network in alternation research. While systematic variation can emerge at any level of schematicity, lexemes that fill the open slots of constructions can cluster based on the meaning expressed, giving rise to lower-level generalizations. Alternations at lower levels tend to capture more tangible meaning differences, which arguably align more closely with the distinctions to which speakers are attuned. The alternations at the schematic level in these studies (English at-alternation and Dutch naar-alternation without any lexical specifications) may emerge as conglomerates of their lower-level alternations, without directly addressing a functional need of the speaker.

Most studies on alternations focus on cases where speakers of a language can choose between two (or more) variants (see Pijpops 2020 for a discussion on the notion of alternations). However, many lexical elements that occupy the slots of alternating constructions exhibit a bias toward one variant over the other. A method that can be used to reveal lexical bias in alternations is distinctive collexeme analysis (Gries and Stefanowitsch 2004). For instance, in the case of the English dative alternation, the verb tell exhibits a substantial preference for the ditransitive construction over the to-dative (128_DITR:2_TO), whereas send (64_DITR:113_TO) has a more balanced distribution (Gries and Stefanowitsch 2004: 106). In the dative alternation, lexical bias is so pervasive that the verbs can actually be used to predict the correct variant with well above chance accuracy (Gries and Stefanowitsch 2004: 119). The strong lexical bias in the dative alternation is corroborated by many other studies (e.g., Lehmann and Schneider 2012; Röthlisberger et al. 2017).

There are various ways to interpret lexical bias in alternations. Gries and Stefanowitsch (2004) argue that distinctive collexemes represent semantic differences between the two variants of the alternation. For example, considering that the central meaning of the ditransitive construction involves the successful transfer of an object from an agent to a recipient (Goldberg 1995), the preference of tell for the ditransitive may be explained as a metaphorical extension of communication as transfer (Gries and Stefanowitsch 2004: 106) However, quantitative corpus-based studies generally suggest that the determinants of alternations are inherently probabilistic, encompassing, for instance, phonological, syntactic, semantic, and discourse-level factors (e.g., Bresnan et al. 2007; Mondorf 2009; Röthlisberger et al. 2017). Can these conditioning factors account for the strong bias of tell in the ditransitive? An even more extreme example is the verb explain, which almost invariably occurs with the to-dative construction. For verbs like explain, detailed memory traces of their usage in the prepositional dative, where their usage in the ditransitive might have been expected, may preempt speakers from using the ditransitive construction for the verb (Boyd and Goldberg 2011; Goldberg 2011). In other words, for highly biased lexical elements, the probabilistic constraints of the parent schemas may be less relevant to speakers when learning to use the conventional variant.

To understand lexical bias in alternations, it is useful to remind ourselves of the multi-level nature of the constructional network. At the schematic level, distinct lexical items can indeed reflect meaning differences between alternating constructions. Accordingly, the verb bring is strongly distinctive for the prepositional dative, as it conveys the meaning of caused motion, rather than successful transfer. As Perek (2014) and Pijpops et al. (2021) have demonstrated, they may be meso-schematic alternations that emerge from clusters of lexemes. Even further down the vertical line of the constructional network, since combinations of schematic constructions and individual lexical items can themselves become entrenched and conventionalized, their entrenchment may override the conditioning factors that govern the alternation between the schematic variants. For the verb explain, speakers simply discard the fact that it is semantically compatible with the ditransitive, and they use it exclusively in the prepositional dative. Through progressive conventionalization, one variant may thus ultimately drive out the other (De Smet et al. 2018). This means that lexical bias in alternations can result either from the compatibility between the lexical meaning of a verb and the constructional meaning of an abstract schema, or it can result from the conventionalization of a lexically specific construction (Diessel 2016). As Goldberg (2011: 151) notes in reference to alternations, “While much of language is semantically and historically motivated, there remain pockets of idiosyncrasy that speakers must learn.” One such idiosyncrasy is skewness in distribution, which is often not entirely predictable from the parent schemas.

In morphology, situations where multiple variants fulfill the same grammatical function are referred to as overabundance. Thornton (2019) provides a comprehensive overview of this phenomenon, presenting examples of overabundant cells across a diverse range of languages. Similar to syntactic variation, most cases of overabundance are conditioned by a variety of language-internal factors (such as syntactic or semantic variables) and language-external factors (such as social or stylistic variables). However, there is also a significant area where the choice appears to be “free”, with no apparent factors determining the selection. If two morphological variants are truly in free variation for a given lexeme, Thornton (2019) predicts that they should occur with roughly equal frequencies. Corroborating this prediction, Bermel and Knittl (2012) report on frequency proportions of competing variants in Czech nominal declension (see Tables 4 and 5 in Bermel and Knittl 2012). In the paradigm of “masculine hard inanimate nouns”, genitive singular forms exhibit variation between the endings – u and – a, while locative singular forms vary between the endings – u and – ě. They note that, although only small groups of nouns in the paradigm show variation in the first place, within these groups, many nouns exhibit roughly equal frequencies for both variants. Therefore, there are instances where one variant does not appear to be driving out the competing variant in their dataset.

In the case of English comparatives from Hilpert’s (2008) data, drawn from the British National Corpus, adjectives such as angry (angrier/more angry), choosy (choosier/more choosy), and risky (riskier/more risky) exhibit a relatively balanced distribution, although most adjectives tend to favor one variant over the other. The conditioning factors influencing the choice of English comparative forms have been extensively studied in the literature, which will be reviewed in the next section.

2.2 The English comparative alternation

English has two ways of expressing the comparative degree: by adding the suffix – er or by placing the degree marker more before the adjective to form the periphrastic variant. The strongest determinant of the alternation is the length of the adjective (e.g., Hilpert 2008); monosyllabic adjectives overwhelmingly take the morphological form, while adjectives with three or more syllables almost invariably take the periphrastic form. Disyllabic adjectives generally exhibit more robust variation, though some of them, such as happy, are extremely skewed (Hilpert 2008, 414–415). However, previous studies have uncovered a complex set of factors influencing the alternation, with no condition rising to the level of a categorical rule. Other factors include the final elements of adjectives and syntactic patterns, such as attributive or predicative usage and the presence of following complements. Mondorf (2003, 2009) argues that these factors can be attributed to considerations in processing effort. Specifically, speakers tend to prefer the more explicit periphrastic variant in cognitively complex environments (Rohdenburg 1996), such as when followed by a complement, as it offers a processing advantage.

From the extensive body of literature on the English comparatives, Hilpert (2008) and Mondorf (2009) provide comprehensive overviews of the conditioning factors. This section focuses on studies that identify idiosyncratic patterns of alternation specific to individual adjectives or clusters of adjectives, aligning with the goals of the present study.

Idiosyncratic patterns within the alternation have been sporadically noted in previous studies. For example, it has been reported that adjectives ending with the sound /i/ do not show a consistent tendency (Mondorf 2009; Chua 2018), despite findings suggesting that this group leans toward the morphological variant (Kytö and Romaine 1997). The variation can be partially explained by separating adjectives ending in /li/ from the broader group, as they have been found to favor the periphrastic variant (Hilpert 2008; Lindquist 1998). However, even within the /li/ group, the tendency appears to vary (Watanabe and Iyeiri 2020). For example, in Hilpert’s (2008) data, costly (82_MORE:19_-ER), deadly (30:9), and likely (3,724:17) prefer the periphrastic form, whereas lively (49:76), lovely (7:42), and ugly (4:41) do not (Hilpert 2008: 414–415). It is possible that the significant effect of the /li/ ending toward the periphrastic form observed in Hilpert (2008) was driven by highly frequent adjectives like likely. The fixed-effects-only model used in that study treats the final segments of adjectives as unique to each observation, allowing highly frequent adjectives to disproportionately influence the results. This potential confound is addressed in the present study by incorporating individual adjectives as random effects and treating adjective-dependent variables as second-level predictors (see Section 3.2 for details).

With regard to the degree of bias in comparative forms, D’Arcy (2014) observes that individual adjectives do not exhibit robust variation in casual speech. In the Origin of New Zealand English (ONZE) corpus, which comprises 1.5 million words of casual speech, only a few disyllabic adjectives alternate between the comparative variants. While the author attributes this lack of variation to the style and “simplified grammar” characteristic of casual conversation (D’Arcy 2014: 238), partitioning toward either variant within the English comparative alternation is also observed, albeit to a lesser degree, in other registers, as noted in the previous section with reference to the British National Corpus. Although it is standard practice in variationist research to omit lexical elements that exhibit near-categorical distributions (Pijpops 2020), the existence of such partitioning itself raises theoretically interesting questions for the alternation. This is particularly relevant given the proposed view of the present study, which posits that competition occurs at multiple levels of schematicity.

What remains as a challenge is a systematic examination of variation and idiosyncrasies within the comparative alternation. This requires a method that can quantify the influence of conditioning factors and the idiosyncrasies of individual adjectives simultaneously, while accounting for one another. A mixed-effects regression model, incorporating the nested structure of the dataset, offers a robust solution. The next section provides an overview of the dataset and explains the basics of this regression model and its structure for the present study.

3.1 Data and annotation

The dataset from Hilpert (2008) that is re-analyzed in the present study is based on comparative usage in the British National Corpus, which is a corpus of 100 million words that contains both written and spoken data.^[2] Morphological comparatives are marked with a dedicated part-of-speech tag, which facilitates their exhaustive retrieval. For the periphrastic comparatives, a regular expression was used to retrieve instances of more to the left of an adjective. False positives such as We publish more fine books than ever before were manually identified and discarded. The dataset only includes alternating adjectives, which yields a total of 247 adjective types. The representation of morphological and periphrastic comparatives in the data set is highly uneven. Morphological comparatives (71,622 tokens) vastly outnumber periphrastic comparatives (8,256 tokens). Each token is annotated for eleven variables (cf. Table 1 below). Besides the number of syllables and morphemes, the dataset holds information on the final phonological segment of the adjective (e.g., final /i/). Tokens are further annotated for the presence or absence of adjective-final stress (e.g., robúst) and word-initial stress of its right-side collocate (e.g., stronger númbers). With regard to syntactic contexts, the dataset offers annotation on attributive and predicative uses, tokens with to-infinitive complementation, a following than, and the presence of premodification. The dataset also includes frequency information. Besides the frequency of each adjective type in its positive form, the ratio of positive to comparative forms, i.e., the relative frequency of comparatives, is determined for each adjective type. For the present study, minor inconsistencies in the data were corrected.^[3] A table with all adjective types and their frequencies in the morphological and periphrastic comparative is offered in Appendix 1.

3.2 Mixed-effects logistic regression

The present study employs logistic mixed-effects models as the primary method of investigation. Logistic regression is especially well-suited for research on linguistic alternations, as it predicts a binary outcome on the basis of multiple continuous and categorical variables. The mixed-effects model incorporates random effects alongside fixed predictors, accounting for dependencies within random groups when estimating fixed effects. For a detailed and practical introduction to applying mixed-effects logistic regression models in corpus-based alternation research, Schäfer (2020) offers a highly informative resource.

Mixed-effects logistic regression models have gained significant traction in corpus-based studies on alternations in recent years. Most studies, however, include random effects primarily to control for group-dependent variability, often without delving deeply into them (a notable exception being Röthlisberger et al. 2017). As Winter (2019, Section 15.9) points out, random effects can provide valuable insights in linguistic research, such as uncovering individual differences or lexical specificity. In this study, individual adjectives receive their own intercepts in the estimation of the fixed effects, offering insights into the extent to which specific adjectives deviate from the overall probabilistic tendencies.

Another advantage of mixed models, or more specifically multilevel models, is their ability to account for the nested structure within a dataset. Table 1 presents a list of fixed effects included in Hilpert’s (2008: 407) model. The predictors are grouped into two categories. First-level predictors, such as whether the comparative form is used attributively or predicatively, vary at the level of individual observations. Second-level predictors are group-level variables; they vary across random groups (in this case, individual adjectives) but remain constant within them.^[4] For instance, each adjective has a fixed number of syllables, which does not change across its observations. Without treating these variables as second-level predictors, highly frequent adjectives overrepresent the data points for these variables. To illustrate, in Hilpert’s (2008) model, simple contributes 1,175 data points for the syllable count, while hazy contributes only 3. By designating these variables as second-level predictors, each adjective contributes only one data point for these variables, ensuring fair representation among adjectives.

4.1 Mixed-effects logistic regression

All analyses presented in this section were conducted in R, version 4.4.2 (R Core Team 2024). Mixed-effects logistic regression models were implemented using the lme4 package (Bates et al. 2015).^[5] To ensure model convergence, the bobyqa optimizer was used. The call for our final model is presented in (3). This model incorporates all the fixed effects listed in Table 1, with their respective levels specified. The frequency of positive form was log-transformed to base e, since frequency effects in language generally accrue logarithmically. Additionally, the comparative/positive ratio was logit-transformed to create an unbounded continuous scale, enabling a more interpretable coefficient estimate.^[6] The model did not encounter multicollinearity issues (Zuur et al. 2010). Generalized variance inflation factors (GVIF) for the fixed effects were calculated, with the highest corrected GVIF^1/2df being 2.26 for the final stress (FINSTR). The random effects include 247 adjectives, each assigned its own intercept to account for group-level variability.

4.1.2 Random effects

In the mixed-model presented in (3), each adjective has its own intercept for the regression model. Importantly, this approach differs from fitting separate regression lines for each adjective (i.e., the “no-pooling approach”; Gelman and Hill 2007: Ch. 12). In a mixed-effects model, the random groups are assumed to originate from a shared population and are therefore expected to share certain characteristics (Bates 2010: Section 3.4). As a result, the intercepts are “shrunk” or drawn toward the global intercept (Gelman and Hill 2007; Winter 2019: Section 15.9). This shrinkage effect is more pronounced for groups with fewer observations, as there is less evidence to suggest that these groups deviate significantly from the overall tendency (Gelman and Hill 2007; Schäfer 2020).

Random effects can be analyzed by focusing on those that exhibit significant deviations from the overall tendency. Figure 1 displays the top 10 adjectives with the largest positive intercepts, while Figure 2 displays those with the largest negative intercepts. In other words, these figures illustrate adjectives that show either a preference for the periphrastic variant (Figure 1) or the morphological variant (Figure 2) to a degree that is unexpected based on the fixed effects alone. It is important to interpret the rankings with caution due to the shrinkage effect. The figures include 95 % prediction intervals for the intercepts, which are generally inversely related to the number of observations.^[8]

Figure 1:

Top 10 adjectives with largest positive intercepts (idiosyncratic biases toward the periphrastic variant).

Figure 2:

Top 10 adjectives with largest negative intercepts (idiosyncratic biases toward the morphological variant).

We can now examine the adjectives shown in the figures and explore how their idiosyncrasies manifest themselves. Starting with Figure 1, the adjective that stands out is likely. The idiosyncrasies of this adjective have also been noted in previous research (e.g., Mondorf 2009: 41; Watanabe and Iyeiri 2020: 79). The average estimate for all comparative tokens of this adjective, without considering the random intercept (+7.98), is −2.68, indicating that the fixed effects alone predict a tendency toward the morphological variant. This is partly because a majority of disyllabic adjectives favor the morphological variant, and because final /li/ has no effect on the alternation in the current model. Thus, likely’s substantial bias toward the periphrastic form (3724:17) is remarkably unexpected.

Figure 1 also includes several monosyllabic adjectives, such as real (109:4), dead (19:4), and apt (36:13), whose preference for the periphrastic variant is unexpected based on the fixed effects. For dead and real, their infrequent use in comparisons, reflected in their very low comparative/positive ratios (0.002 and 0.005, respectively), partially explains their preference for the periphrastic variant. In the case of apt, the final consonant cluster contributes to this preference (Sweet 1968 [1891]: 326). Nonetheless, since monosyllabic adjectives overwhelmingly favor the morphological variant, these monosyllabic adjectives require large positive adjustments to their intercepts to account for their unexpected behavior.

Turning to Figure 2, the majority of these adjectives are disyllabic, as the preference of monosyllabic adjectives for the morphological variant can largely be explained by the fixed effect (i.e., syllable count). However, the almost categorical distribution of the monosyllabic adjective hard (1:1745) requires a substantial adjustment to its intercept, particularly given that it ends with a consonant cluster, which favors the periphrastic form.^[9] The disyllabic adjectives in Figure 2 – silly, curly, ugly, quiet, noble, and lovely – substantially favor the morphological variant (see Appendix 1). Notably, half of the adjectives in the figure (silly, curly, ugly, lovely, and unruly) end with the final /li/ sound, despite the other -ly adjectives such as costly, deadly, or likely, which exhibit the opposite tendency. Additionally, trisyllabic adjectives containing the prefix un- such as untidy, unhappy, and unruly are described as “exceptions” by Quirk et al. (1985: 462). Their occurrence in the morphological form is particularly unexpected given their syllable counts.

Figures 1 and 2 indicate that the current model effectively accounts for lexically specific variability by incorporating random effects. The degree to which random effects explain the variance can be assessed using pseudo R² for logistic regression. Nakagawa and Schielzeth (2013) introduced a method to distinguish between marginal R² (variance explained solely by fixed effects) and conditional R² (variance explained by both fixed and random effects). In this case, the marginal R² was 0.49, while the conditional R² was 0.68, highlighting that a considerable proportion of variance is attributed to random effects.

4.1.3 Model performance

We are now examining whether the mixed-effects model offers a better fit for the dataset than the fixed-effects model. For this purpose, we independently constructed a fixed-effects model using the same predictors in (3), since the present study involved minor corrections to the annotations and a non-linear transformation of the frequency predictors. The prediction accuracy of the fixed-effects model was 96.1 %.^[10] Prediction accuracy was calculated by dividing the number of correct predictions by the total number of predictions, using a probability threshold of 0.5 to determine the predicted outcome. Note that 89.7 % of the tokens in the dataset are in the morphological form, meaning the model must exceed this baseline to demonstrate meaningful accuracy. By comparison, the mixed-effects model achieves an improved prediction accuracy of 97.6 %.

Additionally, we are interested in comparing the performances between these models at the aggregated level for individual adjectives. Specifically, we aim to determine how each model predicts the probability of selecting the periphrastic variant for each adjective. Figure 3 presents a density plot showing the predicted proportions of choosing the periphrastic form for the alternating 247 adjectives, based on the fixed-effects model and the mixed-effects model. The actual proportion is represented by the (largely overplotted) red density curve. A density plot illustrates the concentration of data points (in this case, adjectives) across the dataset. For instance, we observe that in both the fixed-effects and mixed-effects models, adjectives are concentrated toward low proportions of the periphrastic form, which corresponds to high proportions of the morphological form. Additionally, we can observe that the curves slightly rise toward the high proportions of the periphrastic variant, creating a highly asymmetrical U-shape. This indicates that adjectives are generally skewed toward either variant.

Figure 3:

Density plot of predicted proportions of the periphrastic variant per adjective (n = 247).

The close alignment between the mixed-effects model’s estimates and the actual proportions is expected, as this model incorporates the proportions of each adjective to calculate the random effects. What can be seen in the graph is that the fixed-effects model tends to overestimate the skew in proportions for both variants. In other words, the fixed-effects model predicts that adjectives are strongly biased toward one variant over the other, whereas in reality, the bias, particularly for the periphrastic form, is more nuanced.

Why does the fixed-effects model behave this way? It is important to reemphasize that the fixed-effects model does not ‘know’ that the data originates from 247 adjectives with varying frequencies (see Section 3.2). As a result, the model tends to generalize data from highly frequent adjectives to less frequent ones, since the former contribute the majority of data points. Therefore, the extreme skew predicted by the fixed-effects model reflects the tendency of highly frequent adjectives to exhibit a stronger bias toward one variant over the other.

Note that the fixed-effects model includes frequency variables, namely the logarithm of positive frequency and the logit transformation of comparative/positive ratio. However, these variables are not treated as adjective-dependent in the model’s structure, as all variables are handled as observation-level predictors. If these frequency variables are excluded from the fixed-effects model, the skews become much more extreme. A corresponding figure is included in Appendix 2. This suggests that frequency plays a pivotal role in determining the strength of bias (skewness) in the present dataset. The following section discusses this issue in more detail.

4.2 Token frequency and strength of bias

In light of the results from Section 4.1.3, we systematically investigate the effects of frequency on the strength of bias for individual adjectives in the comparative alternation. To this end, the following mixed-effects model (4) was constructed. The purpose of this model is to examine the effect of frequency while controlling for other adjective-dependent factors that may also impact skewness.

This model predicts the probability of selecting the preferred variant for an adjective, whether it is the periphrastic or morphological form. Consequently, the predicted probability will always exceed 0.5. The frequency variable reflects the comparative forms (periphrastic + morphological) rather than the positive form, aligning with prior studies like Mondorf (2009). The frequency values were log-transformed to base e and subsequently centered. The interaction term with “Preferred Variant Type” (a factor variable with two levels: morphological and periphrastic) was included because frequency effects may influence skewness for one variant but not for the other. All second-level predictors from model (3) were included, except for the positive frequency and the comparative-to-positive ratio. The model did not encounter multicollinearity issues, with the highest corrected GVIF^1/2df being 2.71 for the final stress (FINSTR).

Adjectives with no preference (i.e., a completely balanced distribution of variants) were excluded (N = 9). Additionally, adjectives with token frequencies less than six were removed (N = 11). This exclusion was necessary because of the unbalanced distribution of variants in the data set. The mean proportion of the preferred variant among adjectives was 0.84. Adjectives with fewer than or equal to six comparative tokens could not achieve this proportion (5/6 = 0.833) and were therefore excluded. As a result, only 227 adjectives were included for this analysis.

Table 3 presents the coefficient table for model (4). The effects of frequency are divided into two rows: one for when the preferred variant is morphological (the top row after the intercept) and the other for when it is periphrastic (the interaction term). In both cases, frequency is positively associated with the degree of bias (morphological: β = 0.88, p < 0.001; periphrastic: β = −0.40, p < 0.001). The coefficient for the latter case should be interpreted as a “slope adjustment term” (Winter 2019: 138), indicating that while the effect is weaker when the periphrastic variant is preferred, it remains positive (β = 0.88–0.40 = 0.48). Note that the preceding model (3) and previous research (e.g., Cheung and Zhang 2016; Hilpert 2008; Mondorf 2009), with the exception of D’Arcy (2014: 237), suggest positive effects of frequency (in positive or comparative forms) on choosing the morphological variant. Mondorf (2009) attributes this effect to the ease of cognitive processing resulting from the higher entrenchment of frequent forms, which she argues biases speakers toward using the morphological variant. While this explanation is certainly plausible and aligns with the results of model (3), it appears that once a preference is established for an adjective, repeated usage reinforces that bias regardless of the direction.

There is also a significant effect for the preferred variant (β = −0.51, p = 0.011). Since the morphological variant serves as the reference level, the negative coefficient indicates that the degree of bias is stronger for the morphological variant.

For all second-level predictors except for final stress, no significant effects were observed. This is because their effects, if present, are unidirectional. For instance, final consonant clusters predict a bias toward the periphrastic variant. However, the current model predicts biases toward both variants, which counteracts and diminishes the impact of unidirectional effects. The significant effect of final stress (β = 1.12, p = 0.001) may appear puzzling, as it biases adjectives toward the periphrastic form (see Table 2). However, this can be attributed to an interaction between final stress and syllable count. Since monosyllabic adjectives automatically received positive values for final stress (Hilpert 2008: 400), the effects of this variable operate in two ways: one favoring the periphrastic variant (from the stress effect) and the other favoring the morphological variant (from monosyllabic adjectives). In the previous mixed-effects model (3), this dual effect was controlled by the significant effect of syllable count. In the current model, however, with no significant effect of syllable count, the dual effect persists. When the interaction term between final stress and syllable count is included, the main effect of final stress is no longer present (β = 1.05, p = 0.206), demonstrating that it is the interaction that explains the observed puzzling effect.

Figure 4 presents a scatterplot illustrating the relationship between the log frequency of comparative forms and the proportion of the periphrastic form. Several representative data points are labeled with the names of the corresponding adjectives, including gentle and remote, whose positions somewhat diverge from other adjectives. The plot highlights the bidirectional nature of frequency effects, showing a positive relationship between frequency and the strength of bias (skewness) for both directions of the bias. However, it should be kept in mind the dataset contains fewer adjectives that favor the periphrastic form (N = 49) compared to those favoring the morphological form (N = 178).

Figure 4:

The relationship between log frequency and proportion of the periphrastic variant of 227 alternating adjectives in the BNC.

Additionally, it is noteworthy that the relationship between frequency and skewness exhibits a heteroskedastic pattern: there is substantial variability in the proportion among low-frequency adjectives, while high-frequency adjectives exhibit much less variability. For instance, on the low-frequency end, adjectives such as sticky and risky are relatively evenly distributed, whereas gross and obscure display considerable skewness. On the high-frequency end, adjectives are generally much less variable and skewed toward their preferred variant. This is evident from the near absence of data points in the middle-right section of the plot.

The robust variation reflects the probabilistic nature of the variation; adjectives may be skewed toward one variant or remain more balanced depending on the conditioning factors associated with the adjectives or the contexts in which they are used. Conversely, on the high-frequency end, most adjectives show near-categorical distributions, including happy, likely, and great. The low variability within this group suggests that these adjectives escape probabilistic constraints and become almost exclusively associated with their preferred variant.^[11]

It is worth comparing this result to that of Bermel and Knittl (2012), who report a substantial number of alternating nominal endings with a relatively balanced distribution in Czech (see Section 2.1). They find that noun groups exhibiting variation in the genitive singular and locative singular are, on average, highly frequent compared to groups that show no variation within the same grammatical paradigm (Bermel and Knittl 2012, Table 2). However, they do not provide frequency distributions within these variable groups, leaving it uncertain whether token frequency influences skewness among nouns capable of forming alternating endings.

To summarize Section 4, the results demonstrate that individual adjectives in the BNC exhibit robust idiosyncrasies within the comparative alternation. While most fixed effects remain significant factors in the alternation, the strength of their effects is substantially qualified in the mixed-effects model. A closer examination reveals that individual adjectives show preferences that diverge from the general tendencies of the conditioning factors. The mixed-effects model can further account for the strength of bias in individual adjectives, which positively correlates with the token frequency of adjectives in comparative forms.