2.1 Variable CSD

CSD is, in simplest terms, the variable deletion of word-final coronal stops following consonants in English (e.g., old versus ol’). This variable has been of enduring interest to sociolinguists, as evidenced by the size of the literature that describes it. One reason for this sustained interest is the grammatical conditioning that the variable exhibits, such that stops in words of different morphological classes are deleted at different rates. The basic pattern is that coronal stops are more commonly deleted at the end of monomorphemes (e.g. mist) than when they comprise an -ed past tense suffix (e.g. missed). This pattern is almost universally replicated across studies of Englishes around the world (Fasold 1972; Labov 1972; Wolfram & Christian 1976; Guy 1980; Patrick 1991; Santa Ana 1991; Lim & Guy 2005; Gut 2007; Hazen 2011; Walker 2012; Hansen Edwards 2016; Baranowski & Turton 2020; cf. Tagliamonte & Temple 2005).

Researchers have posited a range of competing explanations, both formal and functional, for the morphological conditioning of CSD (see, for example, Guy 1996 for discussion). Of particular note for our interests in this paper are approaches that appeal to the different functions of the stops in these word types. An -ed suffix marks past tense information and its absence can give rise to homophony with the present form of the same verb. But stops at the end of monomorphemes contain no particular grammatical information and may even be predictable from the preceding string. Therefore, lower rates of CSD in -ed suffixed forms may have something to do with a functional pressure to preserve informative structure (following e.g. Kiparsky 1972). The functional argument is thematically similar to the argument that reduction phenomena occur when speakers exploit predictable or redundant elements of form. In both, speakers preserve unpredictable and therefore informative structure, with perhaps the difference that the functional CSD story frames the morphological information encoded in the -ed suffix as especially important, rather than simply important because it is an unpredictable ending to a stem. However, despite the attractiveness of a functional story, and even though multiple contextual measures of word and biphone probability are good predictors of reduction in some cases (Jurafsky et al. 1998), CSD has only been found to be sensitive to decontextualized lexical frequency (Jurafsky et al. 2001). Of course, frequency itself is not straightforwardly captured with a single measure, and a central premise of the current study is that there is more than one plausible unit whose frequency matters.

All accounts of the grammatical conditioning of CSD are complicated by other morphological categories whose rates of CSD are intermediate between monomorphemes and regular past forms. These include semiweak verbs where past tense is marked both with a vowel change and a coronal stop (e.g. kept), negative contractions (e.g. didn’t), past forms where the coronal stop already present in the stem is devoiced (e.g. sent), and the irregular past form went. As a further complication, the ‘monomorphemic’ category has traditionally included a number of words that are actually morphologically complex, but where the word-final coronal stop is tautomorphemic with the preceding segment. For example, a superlative form like biggest was labelled as a ‘monomorpheme’, despite the fact that -est is a suffix, because no morphological boundary breaks up the word-final /st/ cluster. In the present study, we follow MacKenzie & Tamminga (2021) in refining this class to include only true monomorphemes (as set out in that paper), and focus on the well-established difference between this class and regular past forms. In this way, we compare forms that are unambiguously morphologically simplex and complex.

While CSD is typically described in terms of a discrete distinction between ‘present’ and ‘absent’ realizations, this perspective has been called into question. Temple (2009; 2014) remarks that the phonetic conditioning and acoustic implementation of CSD resemble that of any other connected speech process. Moreover, articulatory data has revealed that covert tongue tip raising is the norm for inaudible coronal stops (Browman & Goldstein 1990; Purse 2019), in some cases observing CSD-like morphological conditioning on the magnitude of tongue tip raising (Purse & Turk 2016; Purse 2019). These results, of course, raise questions about the proper representation of CSD, but these are beyond the scope of the current paper. Insofar as CSD is the outcome of a gradient phonetic process of lingual reduction, we might expect it to behave like other classical lenition phenomena in being sensitive to lexical frequency, as we discuss in the following subsection. Indeed, this perspective is consistent with a handful of studies finding that, for CSD overall, frequent words have higher rates of deletion than infrequent words (Bybee 2002; Jurafsky et al. 2001; Tamminga 2016, cf. Walker 2012). Other studies, however, report that lexical frequency conditions CSD differently according to morphological class (Myers & Guy 1997; Guy 2019).

This possible interaction between frequency and morphological structure, and its implications for theories of frequency and variation in form, is a central topic that we probe in this paper. In the following two subsections, we review first the literature on how frequency influences the phonetics and phonology of word forms, and second the literature on how frequency relates to morphological representation and processing.

2.2 Frequency and variation in form

Frequency, specifically whole-word frequency, is associated with variation in phonetic and phonological form in many cases. In general, frequent whole-words tend to be pronounced faster, and in more lenited or reduced forms, than infrequent whole-words. This is relevant insofar as we conceive of CSD as an example of lenition, and we generally expect phonetic reduction and lenition to be intimately related to duration (Lindblom 1963). However, in laboratory studies, evidence for the precise details of the relationship between duration and frequency is somewhat mixed. Wright (1979) claims that rare words are spoken as much as 24% slower than common words, but some subsequent studies have failed to replicate these effects between matching segments (Damian 2003; Mousikou et al. 2015; cf. Kawamoto 1999). Laboratory studies are also not entirely aligned in terms of how phonetic reduction and lenition are sensitive to frequency. In one articulatory study, Lin et al. (2014) find that tongue tip activity is generally reduced in highly frequent words, but Tomaschek et al. (2013; 2014) find that the magnitude of vowel gestures is highly sensitive to segmental context and may only be compressed for frequent words with phonologically short vowels. In contrast with these laboratory studies, corpus studies on more spontaneous speech reliably find that frequent whole-words are produced with shorter durations (Aylett & Turk 2004; Gahl 2008), and with more centralised vowels (Munson & Solomon 2004) than infrequent whole-words.

Beyond gradient phonetic properties like duration, there exist a number of variables where the apparent rate of discrete variants² is correlated with lexical frequency. This is particularly well exemplified by work on varieties of Spanish. Highly frequent Spanish whole-words are more likely to exhibit intervocalic /d/ deletion (Bybee 2002; Diaz-Campós & Gradoville 2011), /r/ deletion (Diaz-Campós & Carmen 2008), vowel coalescence (Alba 2006), /s/ lenition and deletion (Brown & Cacoullos 2003; Brown 2009; File-Muriel 2009), and less likely to feature /ʒ/ devoicing than infrequent whole-words. In English, too, schwa deletion (Hooper 1976), yod retention (Phillips 1981; 1984), and alveolar word-final -in’ for the ING variable (Tamminga 2016; Forrest 2017), have all been found to be more common in frequent whole-words than infrequent whole-words. In a more general approach that is not limited to specific sociolinguistic variables, Turnbull (2018) compares phonological and phonetic transcriptions in corpora of English and Japanese and computes the segment deletions necessary between the underlying and surface forms. He finds that whole-word frequency (among other predictability measures) conditions the rate of segment deletion, and concludes that these patterns mirror those of phonetic reduction.

Outlying results notwithstanding, it seems generally true that frequent words are more susceptible to compression and ‘weakening’ of their pronunciations. Explanations for this kind of reduction phenomenon fall into three main theoretical camps (Clopper & Turnbull 2018), two of which link production effects to robust results that frequent words are recognised more quickly and accurately in perception experiments (e.g. Howes 1957; Savin 1963; Taft & Hambly 1986; Connine 1990; Dupoux & Mehler 1990). (1) ‘Listener-oriented’ accounts (e.g Lindblom 1990; Aylett & Turk 2004) explain production effects in terms of word predictability, and the optimization of the speech signal in order to maximize communicative efficacy while minimizing effort. In other words, speakers use tacit knowledge that frequent words are easier to perceive and attenuate the articulatory effort spent on them. (2) For ‘talker-oriented’ accounts (e.g. Baese-Berk & Goldrick 2009), frequency effects arise as part of the cognitive mechanisms of speech production. Just as in perception, infrequent word forms have a higher threshold for activation during production, and properties of timing and magnitude of activation during retrieval are passed on to properties of timing and articulation in the phonetic implementation. (3) Finally, there are ‘passive’ perspectives, in which word frequency directly shapes the mental representation of words, rather than creating on-line production pressures. A notable example of this kind of perspective is Exemplar Theory (Pierrehumbert 2002; Bybee 2002), in which a persistent leniting bias affects all words, but high frequency words—which are encountered most often—most quickly accumulate exemplars with compressed and ‘weakened’ pronunciations. While the frequency measures we discuss in more detail in §3.2 are correlated with each other (e.g. a word with a high stem frequency is likely to also have a high whole-word frequency), each one is likely more indicative of one of these theoretical mechanisms being at play than the others. For example, an effect of stem frequency is more likely to be indicative of a talker-oriented account than a listener-oriented or a passive account. This is discussed in more detail below.

2.3 Morphology and frequency

We now turn to a brief examination of the relationship between frequency and morphological structure, with reference to both sociolinguistic and psycholinguistic results that highlight possible frequency–morphology interactions. As we have mentioned, there is already some reason to believe that frequency and morphological structure interact in how they condition CSD itself. Myers & Guy (1997) report, based on data from two Philadelphian speakers, that there is a robust effect of whole-word frequency among monomorphemes, but no such effect among -ed suffixed words. Similarly, Bayley (2014) finds a small effect of whole-word frequency that is limited to monomorphemes in San Antonio Chicano English CSD. Interactions between lexical frequency and grammatically-defined conditioning contexts in sociolinguistics have also been reported for morphosyntactic variables. Erker & Guy (2012) find that lexical frequency has an ‘amplification’ effect on the grammatical conditions influencing null subjects in Spanish: effects of verb regularity, verb semantics, subject person/number, and utterance tense/mood/aspect are small or nonexistent among low frequency verbs, but very significant among high frequency verbs. An interesting question we return to in our discussion in §5 is whether reported frequency/grammatical context interactions are the same kind of effect for CSD and null subjects.

The relevance of morphological structure for word processing has led to more widely recognized interactions in this domain. There is some evidence that morphologically complex words are generally recognised faster than monomorphemic words of equal length and frequency (Fiorentino & Poeppel 2007), but highly frequent complex words are disadvantaged if the suffix is also highly frequent (Balling & Baayen 2008). Moreover, it has been suggested that the frequency-bearing unit most appropriate to capture variance in word recognition latencies depends on the morphological complexity of the word (Vannest et al. 2011). Morphologically complex words are recognised at speeds that vary according to lemma—or “base”—frequency, while monomorphemic words’ recognition speeds are best accounted for with whole-word—or “surface”—frequency.

In addition to basic frequency/morphology interactions in behavioral reaction times, there is also a growing body of work making inferences about what level of representation is active at a given point in the timecourse of spoken word recognition based on what kind of frequency measure correlates best with neural activity during processing. Specifically, a number of MEG studies find neurological activity to be most strongly correlated with measures of morphological structure, including lemma frequency and the transition probability between stem and suffix, at around 170ms (Solomyak & Marantz 2009; 2010; Zweig & Pylkkänen 2009; Lewis et al. 2011; Fruchter et al. 2013) and again at around 350ms (Solomyak & Marantz 2009) following exposure to visual word stimuli. These results are taken as evidence for word recognition making reference to smaller morphological units, since these frequency measures associated with activation levels reflect the frequency of those sub-word units. While this literature has typically discussed these sub-lexical units in terms of decomposition (see also Embick et al. forthcoming on the nature of decomposition), we do not believe it is necessary to endorse a particular view on whether morphologically complex words are decomposed per se in order to draw similar inferences about the relevance of sub-lexical structure for variation in morphologically complex words. Even models that posit whole-word episodic storage in the lexicon allow for morphological relationships to emerge from patterns of phonetic and semantic overlap (Bybee 2002); these relationships may in principle influence variable outcomes.

Among the studies that do apply this strategy of comparing frequency measures to explore the role of morphological structure in production, one interesting result that has emerged is evidence of ‘paradigmatic enhancement’ effects. As well as the basic effect whereby frequent items are realized (and recognized) faster as a result of their predictability or ease of retrieval, some words with a high frequency compared to morphologically related words within the same paradigm are reinforced and pronounced with less phonetic reduction. An intuitive way to conceptualize this idea is in terms of speaker confidence, such that speakers are reassured that they are ‘correct’ in selecting the most relatively frequent form and do not hold back in production (Kuperman et al. 2007). Originally, paradigmatic enhancement was proposed to explain effects in ‘pockets of uncertainty’ between functionally equivalent forms that directly compete for use in the same position, like Dutch compound linking morphemes (Kuperman et al. 2007) and variable Russian agreement suffixes (Cohen 2015). This ‘pocket of uncertainty’ aligns fairly closely with mainstream variationist conceptions of the linguistic variable, and indeed we see parallel results in the effect of variant frequency on variant duration in French variable schwa (Bürki et al. 2011). More recently, however, research on paradigmatic probability has been extended to explain variation in pronunciation across paradigmatically related words that are not in direct competition, with evidence for both the enhancement (Schuppler et al. 2012; Tucker et al. 2019; Tomaschek et al. 2019; 2021; Bell et al. 2021) and reduction (Hanique & Ernestus 2011; Smith et al. 2012; Ben Hedia & Plag 2017; Plag & Ben Hedia 2017) of more relatively frequent forms. The present study represents, among other things, a contribution to this literature that may help reconcile these seemingly contradictory results.

4.1 Modeling both monomorphemes and past tense

In order to probe which frequency measure best captures variance in CSD, we compared a series of logistic regression models predicting CSD outcomes. The baseline model does not contain any frequency measures but does include the fixed effects for speech rate, grammatical class, and following segmental context, plus a random intercept for speaker. The subsequent models add all possible combinations of the three frequency measures to this baseline model. We use likelihood ratio tests to assess whether each additional level of complexity (i.e. each additional frequency measure) was warranted as a significant improvement over the nested smaller models.¹⁰

In addition to likelihood ratio tests, each model’s fixed AIC (Akaike Information Criterion), BIC (Bayesian Information Criterion) and log-likelihood statistics were recorded. While the log-likelihood is inevitably improved by adding additional complexity to a model, the AIC and BIC penalize model complexity at the same time as evaluating a model’s ability to account for variance. This is especially true of the BIC, whose penalty for additional complexity is proportional to the number of observations, and frequently disagrees with the AIC in favour of a simpler model. Together, these information criteria provide the clearest evaluation of these models, indicating in particular where multiple frequency measures do not account for a sufficient amount of variance to justify their inclusion. Figure 2 shows the degree to which models with various combinations of frequency measures reduce the AIC and BIC, compared to a baseline model with no frequency measures.

Figure 2

Information criteria reduction from baseline comparing models of full dataset (triangles = most reduced).

Including each of the three frequency measures, individually, yields information criteria statistics that are somewhat reduced compared to the baseline model. This result is reinforced by significant likelihood ratio tests (p < .001) in each case. However, the reduction in both AIC and BIC that is attained from the addition of conditional frequency far outstrips that of the other measures. In fact, the addition of conditional frequency provides a large reductions in both AIC and BIC regardless of any other frequency measures already included in a model. The model comparison also suggests that the combination of stem frequency and whole-word frequency in a single model is a significant improvement over just one of these measures. However, we cannot rule out the possibility that this is an artifact of the strong correlation between these measures causing inflation of their estimated effects. In addition, neither stem nor whole-word frequency significantly improves any model that already includes an effect of conditional frequency. This is demonstrated by likelihood ratio tests (p > .05), and the fact that these measures do not account for enough additional variance to counteract the penalty for model complexity that occurs in either the AIC or the BIC.

The initial model comparison results point to a need to reconsider how frequency is accounted for in linguistic variation. In particular, the success of conditional frequency over other measures in terms of accounting for variance suggests that the interplay between word frequency and morphological structure within the lexicon is important and underexplored. Morphological structure is particularly relevant for a variable like coronal stop deletion, since it has repeatedly been reported that coronal stops at the end of monomorphemes are more likely to be deleted than coronal stops that constitute -ed suffixes (Guy 1980; 1991). This basic difference is controlled for with the main effect of grammatical category in each of the models in Figure 2. However, the effect of morphology may be more complicated still, as Figure 3 shows. In the top-left panel of the figure, we replicate previous findings that only monomorphemes are sensitive to whole-word frequency, and not regular past forms (Guy 2019). This result strengthens our confidence in the interaction between frequency and morphological category for CSD, because compared to previous reports it is based on a significantly larger dataset with more narrowly defined morphological categories. Unsurprisingly, this interaction also holds for the closely-related stem frequency in the top-right panel. In addition to replicating previous reports for CSD, we note that these results resemble the “amplification” effect described by Erker & Guy (2012), in which the effect of grammatical class is stronger at high frequencies, and may not exist at all between low frequency words. However, in all of the CSD studies, including ours, the slope of the line for regular past forms does not significantly deviate from 0, suggesting that any apparent amplification of the morphological effect does not affect the morphological categories evenly, but rather is driven by differences between high- and low-frequency monomorphemes.¹¹

Figure 3

Observed CSD outcomes according to each frequency measure and morphological class.

Compared to the results for whole-word and stem frequency, the results for conditional frequency are striking. Here, not only is there an effect for both monomorphemes and regular past forms, but the lines are almost parallel. This helps to explain why conditional frequency was so highly favored by the model comparison for the combined data in Figure 2 with sum-coded morphological categories. Furthermore, recall from Figure 1 that the relationship between conditional frequency and the other measures was weaker than the relationship between whole-word and stem frequency; the robust conditional frequency effect observed here accounts for a portion of the variance in CSD outcomes that is virtually untapped by controlling for just whole-word or stem frequency.

The differences in the effects of the frequency measures between morphological categories is not captured by the regression models we have been discussing, because they do not include any interaction terms targeting the non-independence of frequency and grammatical category. As a result, the best models we’ve presented so far, which combine data from regular past and monomorphemic words (sum-coded), will compromise between the two. In other words, a frequency measure that might be best for one group of words will be penalized if it is inappropriate for another. This raises questions about the performance of frequency measures within morphological categories, which are not addressed by the models we have presented so far. Therefore, in the following subsections we divide the data by morphological class and test the different frequency predictors within each word type.

4.2 Modeling only monomorphemes

We begin by adopting the same method of model comparison as described for the full dataset, implemented over a subset of the data containing only monomorphemes. Once again, all models include fixed effects for speech rate and following segmental context, and a random intercept for speaker, but since all the words are monomorphemic no morphological category predictor is included.

In Figure 4, we can see that the picture for monomorphemes alone is very similar. As in the models for the full dataset, all the frequency measures significantly improve model fit over the baseline when they are added individually, but conditional frequency outperforms the other measures and improves every model to which it is added. These results are reinforced by likelihood ratio tests (p < .001). In this case, the addition of stem frequency provides a slightly more obvious reduction in AIC and BIC than was observed for the full dataset, but it is still the smallest in magnitude out of the three frequency measures. Once again, we see that the combination of stem and whole-word frequency outperforms either measure on its own, but this is very likely an artifact of the especially strong multicollinearity between these measures for monomorphemes.

Figure 4

Information criteria reduction from baseline comparing models of monomorpheme subset (triangles = most reduced).

In terms of the models that best reduce the information criteria, the results for the monomorpheme models are slightly less straightforward than for the full dataset in that the AIC and BIC disagree. Once again, the BIC is lowest for the model with just conditional frequency in addition to the baseline effects. However, the AIC is lower in the models containing at least one other frequency measure in addition to conditional frequency, and lowest in the model with all three measures. This suggests the other measures do capture enough variance in monomorphemes to outperform the relatively small penalty for additional model complexity that is applied in the computation of AIC. This seems especially true for stem frequency, which significantly improves the fit of every model it is added to according to likelihood ratio tests (p < .05). This includes all models with conditional frequency and/or whole-word frequency already present. In contrast, likelihood ratio tests do not show whole-word frequency to significantly improve models with conditional frequency already present. This is likely due, in large part, to the complete absence of a correlation between conditional and stem frequency for monomorphemes, such that they do not compete to account for the same variance.

4.3 Modeling only past tense

Just like for monomorphemes, we conducted the same method of model comparison for the regular past forms alone. Again, all models include a fixed effect of speech rate and following segmental context, a random intercepts for speaker. According to Figure 3, only conditional frequency appears to have the expected frequency effect for this group, but model comparison allows us to observe the interplay between the different measures when they are included in different combinations. The AIC and BIC values for each of the regular past models are plotted in Figure 5.

Figure 5

Information criteria reduction from baseline comparing models of complex form subset (triangles = most reduced).

Unsurprisingly, conditional frequency once again introduces a large reduction in both the AIC and BIC of every model it is added to, as well as a significant improvement in terms of likelihood ratio tests (p < .001). Unlike for the full and monomorpheme datasets, not all of the frequency measures improve the baseline model when they are added individually. The addition of whole-word frequency does not account for enough variance to overcome the penalty for model complexity in either the AIC or BIC, and does not significantly improve model fit according to a likelihood ratio test (p>.1). Stem frequency, on the other hand, does marginally reduce the AIC and significantly improve model fit according to a likelihood ratio test (p < .05), but the magnitude of its improvement is still less than the penalty applied by the BIC for introducing additional complexity to the baseline model. Once again, the combination of both whole-word and stem frequency apparently reduces both the AIC and BIC by a fair amount compared to the baseline model. Even though the correlation between whole-word and stem frequency is weaker for regular past forms than for monomorphemes, it is still strong enough that this effect is likely to be an artifact of multicollinearity, especially given how poorly both whole-word and stem frequency perform individually.

Like for the monomorpheme models, the AIC and BIC disagree as to the optimal model for regular past forms. For the third time, the model with conditional frequency alone is favored by the BIC, and additional frequency measures are penalized for unnecessary complexity. However, this time, the AIC is minimized in the model with both conditional and whole-word frequency. This is despite the fact that whole-word frequency performed poorest when it was added to the baseline model individually, and the fact that it is more strongly correlated with conditional frequency than stem frequency is, for the regular past forms.

5.1 Interaction between whole-word/stem frequency and morphology

Whole-word frequency and stem (or ‘base’ or ‘lemma’) frequency are the measures of frequency most commonly incorporated into studies in contemporary sociolinguistics and psycholinguistics. For our data, it turns out that these measures are very highly correlated, and correspondingly predict extremely similar patterns of CSD across different subsets of the data. On the assumption that these frequency measures would also correlate this strongly throughout the lexicon (not just for our subset of CSD words), we offer the methodological recommendation that whole-word frequency, which is considerably more straightforward to implement than stem frequency, will be at least as effective as stem frequency for capturing frequency-related variance in other linguistic variables. In other words, for researchers who simply want to incorporate a reasonable frequency control into studies that are primarily aimed at investigating other phenomena, it will not be worth the effort to operationalize a stem frequency measure. With regard to the specific pattern found using these measures, we observe a main effect of whole-word and stem frequency on CSD outcomes for the monomorphemes—coronal stops are more likely to be deleted at the end of frequent monomorphemes than infrequent monomorphemes—but not for regular past forms. An equivalent interaction between morphological category and whole-word frequency has also been reported for other CSD datasets (Myers & Guy 1997; Bayley 2014), but never at such a large scale, or corroborated with the same finding for stem frequency. In our data, as in these previous studies, the effect of frequency within monomorphemes is similar to that which has been observed for a number variable lenition and reduction phenomena, specifically that highly frequent and therefore highly predictable and/or highly practised words are pronounced with more reduced and lenited forms. However, additional explanation is required for why the same effect is not straightforwardly found for whole-word or stem frequency among regular past forms.

A potential avenue for explanation comes from Erker & Guy (2012), who report a similar interaction between whole-word frequency and grammatical category in the rate of subject personal pronoun omission in Spanish. In their data, the effects of verb regularity, verb semantics, subject person/number, and utterance tense/mood/aspect are small or nonexistent among low (whole-word) frequency verbs, but large among high frequency verbs. Thus, whole-word frequency is taken to ‘amplify’ the effect of these grammatical categories. The proposed reason for this is that speakers need a certain amount of experience with a word in order for the effects of its grammatical category to be learned and reproduced, either as emergent from the particular contexts in which words of that category appear or as a more abstract property that entails a particular rate of some variant. This aligns with a ‘passive’ perspective on frequency effects in that it is the mental representation of words that is implicated, rather than any on-line mechanism. Our results are also consistent with this idea of amplification: among high frequency words, the rate of CSD is far higher in monomorphemes than in regular past forms, but there is very little difference between low frequency monomorphemes and regular past forms in terms of rates of CSD.

On the other hand, a deficiency of the amplification story is that, at least for CSD, grammatical categories are treated more or less like arbitrary labels for words. In reality, monomorphemes and regular past forms differ in terms of morphological complexity, which may explain what we observe in terms of sensitivity (or lack thereof) to measures of frequency. Morphological complexity has two relevant properties as pertains to frequency. The first is that of informativity: while coronal stops at the end of monomorphemes are often highly predictable and contain no additional disambiguating information about the word, coronal stops at the end of regular past forms constitute a suffix that marks past tense. Moreover, when this suffix is deleted, regular past forms are always homophonous with a present or infinitival form of the verb. These are some of the primary concerns of linguists who ascribe a ‘functional’ motivation to grammatical patterns of CSD, arguing that deletion is avoided in cases where it would eliminate important past tense information (e.g. Kiparsky 1972).

The second relevant property of morphological complexity is that it entails pieces (whether independently-represented or emergent from shared phonology and semantics) being shared across words. That is, not only does CSD target an informative suffix when it applies in regular past forms, it targets the same suffix identity for every lexical item in the grammatical category. Given that we are asking about the frequency of different linguistic units, we are forced to consider whether the relevant frequency measure for this kind of word might not be of the whole-word or the stem, but the -ed suffix itself. Of course, a raw measure of this kind would amount to a single (high) frequency value, and would not be particularly useful for explaining the basic effect of grammatical category, never mind differences between words within a single category. Therefore, instead of considering the frequency of a suffix overall, it may be more fruitful to consider the frequency of a suffix (or the absence of a suffix) under certain conditions. This is what is achieved by our conditional frequency measure.

5.2 Main effect of conditional frequency

What we have called ‘conditional frequency’ is the proportion of instances of a stem that are realized as a certain whole-word. Unlike for whole-word and stem frequency, we find strong effects of conditional frequency on predicting CSD outcomes in all of our regression models. For regular past forms, conditional frequency corresponds to the decontextualized probability of the -ed suffix as an ending for a given stem. -ed suffixed forms that are common, relative to other reflexes of the same stem, are less likely to retain the coronal stop that marks this suffix. We can consider this result in light of the functionalist framing that is sometimes used to describe the main effect of grammatical category, which states that coronal stops are less likely to be deleted when they encode important grammatical information, i.e. an -ed suffix. As previously mentioned, the functional analysis has a great deal in common with listener-oriented perspectives of reduction/lenition phenomena (Lindblom 1990; Aylett & Turk 2004; Jurafsky et al. 2001), which also describe the preservation of unpredictable and therefore informative structure, but the functional account makes specific reference to the grammatical information encoded by an -ed suffix rather than more general properties of word or segment probability. Our findings show that even a functionally important coronal stop, representing the past tense suffix, is frequently deleted when that suffix is a highly paradigmatically frequent—and therefore highly predictable—ending to the stem.

While -ed suffixed words lend themselves to an intuitive interpretation of the conditional frequency measure, it may be more difficult to conceptualize a similar effect in monomorphemes. If the effect of conditional frequency is to be explained in terms of how predictable a suffix is given a stem, why would we see the same effect for no suffix at all? Indeed, as well as the frequency of an -ed suffix given a stem, conditional frequency in regular past forms is also equivalent to the frequency of an underlying coronal stop in this context. The conditional frequency of kicked is both the rate at which is used in the past or passive and the rate at which /kik/ is followed by an underlying word-final /t/ with no intervening word boundary. In other words, both morphological and phonological levels of representation are captured with the same measure. Conversely, for the monomorphemes in this study, all or most of the words that are morphologically related to them also have underlying representations that contain the relevant coronal stop. The conditional frequency of act does not capture the rate at which a /t/ appears with the stem , because that /t/ is part of the reflex of the stem itself and therefore also occurs in acts, acting and actor. Instead, conditional frequency in monomorphemes only corresponds to the rate at which the stem occurs with a coronal stop in word-final position, as opposed to being followed by additional phonological material within the word. If we are to reconcile this with the predictability view that works well for -ed suffixed forms, we can think of the conditional frequency effect in monomorphemes in terms of edge marking. Stems that do not commonly appear with a word-final coronal stop, relative to other possible forms where the stem is combined with various suffixes, are more likely to retain this coronal stop due to hyperarticulation at the word edge. In other words, stem-conditionally predictable word endings promote deletion, just as stem-conditionally predictable suffixes favor deletion.

Our results, that high conditional frequency corresponds to a high rate of coronal stop deletion, conflict with some recent findings of ‘paradigmatic enhancement’ effects. This is the class of results where the most common reflexes of a particular word or morpheme are found to be phonetically reinforced rather than reduced. These effects are framed from both speaker-oriented and passive perspectives. They are commonly interpreted in terms of speakers articulating common reflexes of a morpheme with increased confidence, suggesting an on-line pressure to reduce in cases where the speaker is unconfident. At the same time, speaker confidence itself has been explained as the result of extensive motor practice, allowing these words to be executed with enhanced kinematic skill (Tomaschek et al., 2018), suggesting an evolution of the specific representation or implementation associated with a word that is not generated on-line. However, comparison between paradigmatic enhancement findings and our own results is not straightforward. As we have already discussed in this section, the conditional frequency measure captures different facts about the coronal stops in monomorphemes versus -ed suffixed forms. Our results for monomorphemes lend themselves to a comparison with findings regarding the pronunciation of stem vowels with various suffixes (Tucker et al. 2019; Tomaschek et al. 2021), and perhaps even more pertinently with those concerning the pronunciation of consonants in the component pieces of compound nouns (Bell et al. 2021). All three of these studies find evidence of reinforcement when a stem or word is followed by a common ending; the present study finds the opposite. Instead, we show that monomorphemes whose stems typically occur in that form, with no additional suffix, are more likely to exhibit CSD than monomorphemes whose stems are more commonly suffixed.

In the case of -ed suffixed forms, our results can be more straightforwardly compared to research on the pronunciation of affixes themselves in terms of their relationship to a given stem (Kuperman et al. 2007; Hanique & Ernestus 2011; Smith et al. 2012; Schuppler et al. 2012; Cohen 2015; Ben Hedia & Plag 2017; Plag & Ben Hedia 2017; Tomaschek et al. 2019). In these studies and in terms of the regular past forms in ours, the frequency of the affix itself, as attached to a given stem, is what is compared to the frequency of the same word/stem with other affixes or with no affix at all. While some studies of this type look at functionally equivalent affixes in direct competition (Kuperman et al. 2007; Cohen 2015), our past tense form study aligns with the many others comparing the frequency of one affix as an ending to a stem to the frequency of the whole paradigm (Hanique & Ernestus 2011; Cohen 2015; Tomaschek et al. 2019). Like these studies, using a different suffix in place of -ed will no longer denote the past tense. This means the conditional frequency of an -ed suffixed form does not capture the same ‘pocket of uncertainty’, where a language user could use more than one form to convey the same thing, that was considered to be so important in many early paradigmatic enhancement results. Correspondingly, our results indicate more reduction when -ed is a frequent ending to a stem, aligning in particular with Hanique & Ernestus’s (2011) result of greater reduction and deletion of word-final /t/ in Dutch irregular past verb forms when it is frequent within the paradigm, as opposed to Schuppler et al.’s (2012) result of greater word-final /t/ retention in Dutch 3SG present verb forms when this form is more frequently used than the 1SG form of the same verb. However, this pocket of uncertainty is surely to be found at every site of a sociolinguistic variable. And certainly, Bürki et al. (2011) find a comparable enhancement effect such that French variable schwa (e.g. fenêtre [f(ǝ)nɛtr] ‘window’) is longer in words that appear relatively more frequently with schwa compared to without it. In other words, even though we find no evidence of paradigmatic enhancement effects in this study, we might predict that future studies would find such effects corresponding to variant frequency, e.g. enhancement that is negatively correlated with the rate of CSD for a given word, such that more commonly retained stops have a reinforced pronunciation when they are retained. What our present results for conditional frequency do point towards is an effect of suffix predictability, and a corresponding effect of word edge predictability, that are correlated with CSD rates. These effects are ultimately different reflexes of the same, listener-oriented, goal to signal that the listener should not expect another suffix.