Frequency

Among the multitude of factors that have been reported to play a role in the evolution of preterite morphology, the best known is token frequency: high-frequency verbs tend to weaken less than low-frequency verbs. Bybee (Reference Bybee2006:715) explained this “Conserving effect”: “High-frequency sequences become more entrenched in their morphosyntactic structure and resist restructuring on the basis of productive patterns that might otherwise occur.” This factor has been empirically verified for many Germanic languages, using different methods and on both synchronic and diachronic data (cf., Carroll et al., Reference Carroll, Svare and Salmons2012; Cuskley et al., Reference Cuskley, Pugliese, Castellano, Colaiori, Loreto and Tria2014; De Smet & Van de Velde, Reference De Smet and Van de Velde2019; Lieberman et al., Reference Lieberman, Jean-Baptiste, Jackson, Tang and Nowak2007; Nowak, Reference Nowak2015:166–73; Pijpops et al., Reference Pijpops, Beuls and Van de Velde2015; Strik, Reference Strik2015; Van Santen, Reference Van Santen, Santen and van der Wal1997; Vosters, Reference Vosters, Vogelaer and Seiler2012; inter alia).

A second frequency measure that plays an important role is type frequency, that is, the number of different verbs belonging to a particular ablaut class. A higher type frequency means, again, a lower chance of weakening: the higher the number of verbs in a particular ablaut pattern, the more the pattern is entrenched in memory, so that it is readily available in memory when language users have to form preterites of verbs that fit the pattern, preempting the need to resort to the default weak inflection. This effect, too, is well established in the literature concerning strong and weak verb morphology (e.g., Bybee, Reference Bybee1995; Carroll et al., Reference Carroll, Svare and Salmons2012; Hare & Elman, Reference Hare and Elman1995; Knooihuizen & Strik, Reference Knooihuizen and Strik2014; Nowak, Reference Nowak2015:166–73).

The protective effect of type frequency is basically an analogical effect: a verb is connected with its ablaut neighbors. It has been hypothesized that this effect will, however, be counterbalanced by an analogic pull to other verbs that are phonetically similar but belong to the weak inflection class (Köpcke, Reference Köpcke, Butt and Fuhrhop1999:57). The phonetic similarity may range from full homophony (e.g., scheppen ‘create’–scheppen ‘dig’) to partial resemblance (e.g., scheppen–meppen ‘slap,’ deppen ‘dab,’ reppen ‘mention’).

Another variable related to frequency is the regularity of the verb. While strong verbs are often said to be irregular (Haeseryn, Romijn, Geerts, de Rooij, & van den Toorn, Reference Haeseryn, Romijn, Geerts, de Rooij and van den Toorn1997:84–93; Lieberman et al., Reference Lieberman, Jean-Baptiste, Jackson, Tang and Nowak2007), they show productive regularity, which is evidenced by production experiments with fake verb forms (Knooihuizen & Strik, Reference Knooihuizen and Strik2014; Van de Velde & Kestemont, Reference Van de Velde and Kestemont2015), and by the emergence of strong ablaut patterns in erstwhile weak verbs (cf., Köpcke, Reference Köpcke, Butt and Fuhrhop1999; Nowak, Reference Nowak2015). Moreover, the historical ablaut patterns are systematically related to the consonant skeleton of the root. A number of verbs are, however, genuinely irregular. These change not only their stem vowel when forming the preterite or the past participle. Two examples are houden-hield ‘hold’ and gedijen-gedeeg (now gedijde) ‘flourish.’ The reasons are, once more, the influence of sound laws, ripping up morphological regularity. In houden-hield, the absence of the liquid in the present is due to the sound law that turns it into a glide when it occurs before a back vowel, but not before a front vowel. As neatly fitting an ablaut class makes the verb easier to memorize, these irregularities might be a disadvantage for the strong verbs. However, most irregular forms tend to be very high-frequency forms (low-frequency irregular forms are likely to regularize by analogy) that tend to become autonomous from derivationally related forms (they do not need to be regularly formed on the basis of the present stem [Bybee, Reference Bybee2006:715]) and “often constitute unanalyzed frozen phrases, and so do not contribute to analogical generalization at all” (Ambridge, Kidd, Rowland, & Theakston, Reference Ambridge, Kidd, Rowland and Theakston2015:247). Because, in order to form the weak inflection, we need to know the present tense, it may be the case that weakening is less straightforward for these autonomous irregular verbs than for regular verbs (cf., Köpcke, Reference Köpcke, Butt and Fuhrhop1999:49).

Differences in weakening between preterites and past participles

A second important observation is that preterites and past participles do not weaken together. Many verbs in contemporary Dutch sport strong participles, but weak preterites (stoten-stootte-gestoten ‘push,’ vouwen-vouwde-gevouwen ‘fold,’ bakken-bakte-gebakken ‘bake,’ malen-maalde-gemalen ‘grind,’ inter alia). Contributing to this asymmetry is Präteritumschwund “preterite loss.” In Germanic, past participles have been grammaticalized into the analytic perfect (ik heb gezongen ‘I sang’), which is in the process of replacing the preterite as a way to express past tense in, among others, (South-)German, Afrikaans, and Dutch (e.g., Abraham & Conradie, Reference Abraham and Conradie2001; Coussé, Reference Coussé2013; Dammel et al., Reference Dammel, Nowak and Schmuck2010; Drinka, Reference Drinka2013). This diachronic trade-off leads to a decrease in the frequency of the preterite and an increase in the frequency of the past participle. Because token frequency is an important determinant in whether a verb weakens or not, we expect that, through time, the ever more frequent past participle is increasingly protected against weakening, while the dwindling preterite is more vulnerable to weakening (as Dammel et al. [Reference Dammel, Nowak and Schmuck2010] showed for synchronic data).Footnote ³

Also contributing to the asymmetry between preterite and past participle weakening is the vowel pattern. Not only do preterite and past participle share semantics (Dammel et al., Reference Dammel, Nowak and Schmuck2010), for some verbs they also share the morphological markers. For the Dutch strong verbs, three different vowel patterns can be discerned (cf., Table 1): an ABA-pattern, where the present and the past participle show a different vowel than the preterite (e.g., lopen-liep-gelopen ‘run’), an ABB-pattern, where the preterite and past participle share a vowel that differs from the vowel in the present (zingen-zong-gezongen ‘sing’), and an ABC-pattern where every vowel change stands on its own (breken-brak-gebroken ‘break’). Different behaviors for these different patterns noticed by Van Haeringen (Reference Van Haeringen1940) have been empirically verified on synchronic (Van de Velde & Kestemont, Reference Van de Velde and Kestemont2015), diachronic (De Smet & Van de Velde, Reference De Smet and Van de Velde2019; Nowak [Reference Nowak2015:168; Reference Nowak, Dammel, Eitelmann and Schmuck2018] for German) and dialect data (Taeldeman, Reference Taeldeman2011). Two main trends can be distinguished: (i) for preterites, verbs showing the same vowel in the preterite and participle (ABB-verbs, e.g., zingen-zong-gezongen ‘sing’) tend to preserve the strong inflection better, while preterites not showing the same vowel in the preterite and participle tend to lose the strong inflection (ABC- and ABA-verbs), especially when the past participle does not show ablaut (in case of an ABA-pattern, for example, the participle gelopen ‘ran’); and (ii) for participles, verbs that share the same vowel for the present and participle (the ABA-pattern) tend to retain the strong inflection.

For the ABB-verbs, the vowel pattern clearly reflects the semantic distinction between the present on the one hand and the past tense (expressed either by preterite or analytic perfect) on the other hand (Nowak, Reference Nowak, Dammel, Eitelmann and Schmuck2018:154). This regular pattern is dominant for both the contemporary strong verbs in Dutch (class I, II, III and some class IV-verbs sport an ABB-pattern) and for the weak verbs (one marker, a dental suffix, is used for both the preterite and past participle), and offers mutual protection for both the preterite and participle. Though preterites might have become less frequent as a result of Präteritumschwund, the vowel of ABB-preterites is still well known, because it is shared with the past participle.

This is not the case for ABA- and ABC-verbs. When preterites of ABA- or ABC-verbs decrease in frequency, the language user cannot fall back on the vowel of the past participle, though it does happen that ABC-preterites take the vowel of the ABC-past participle, turning these verbs into ABB-verbs, for example, breken-brook-gebroken < breken-brak-gebroken ‘break’ or stelen-stool-gestolen < stelen-stal-gestolen ‘steal’Footnote ⁴ (see also Nowak [Reference Nowak, Dammel, Eitelmann and Schmuck2018] for German). This makes ABA- and (to a lesser degree) ABC-preterites more prone to weakening.

Not only are ABA-preterites more susceptible to change, ABA-participles also seem to be overall more conservative. We hypothesize that this scenario might work on the observation that ABA-verb participles are midway between strong and weak. Regarding ablaut vowel, their participles are weak, and it is only their nasal suffix that testifies to their allegiance to the strong strategy. As ablaut is the most salient indicator of Germanic past tenses, language users may be under the impression that the participle is weak already, seeing no need to level the form further. Support for this second scenario can be found in the less common weak-to-strong evolution. Weak verbs that have become strong by analogy of a class with an ABA-pattern, namely jagen-joeg-gejaagd ‘hunt’ or vragen-vroeg-gevraagd ‘ask,’ kept their original dental suffixes in their participles: the vowel pattern already fits, so the suffix does not need to change. However, one could argue that, if the participle is conceived of as weak already, language users might as well shift to the weak inflection entirely. Indeed, in German, precisely these past participles of ABA-verbs tend to weaken more (Nowak, Reference Nowak2015:167). A second explanation could be that the strong participles are protected by the supporting rhyme in present plural and infinitive forms. This explanation fits better with the observation that German ABA-past participles tend to weaken more: though these could also receive protection from present plural and infinitive, in German, only the first and third person present plural are formed with -en (the second person plural is formed with -t, for example, ihr lest ‘you read’) and the infinitive is used almost 7% less often than in Dutch.Footnote ⁵

The effect of ambiguous forms

As mentioned, similar forms might influence the strong-to-weak evolution. If a strong verb is homophonous with a weak verb (cf., Nowak [Reference Nowak, Schmid and Ziegler2011] for German examples), this could cause confusion for language users who might use weak forms for the originally strong verb (cf., [a]) and strong forms for the weak verb (cf., [b]), as in the case of the originally strong verb scheppen-schiep-geschapen ‘create,’ which is homophonous with the originally weak verb scheppen-schepte-geschept ‘dig, draw from.’

• (a) want die looze Feeks wist de St. Huiberts Heeren zo aardig te betoveren, dat de Leeuwen, in Bokken, de Tygers, in Zwynen, en de Luipaarts, in Bandrekels, herschept wierden (Van Olmen [Reference Van Olmen2019]: observation from 1724)

  ‘because the false Witch managed to bewitch the Gentlemen of St. Huibert so well, that the Lions were recreated into Bucks, the Tigers into Swines, the Leopards into Chain Dogs.’

• (b) zoo grooten behagen schiep zy in haar kuisch voornemen. (Van Olmen [Reference Van Olmen2019]: observation from 1659)

  ‘so much delight did she draw from her chaste intention.’

Homophony is not only found between the present forms of two verbs but can also be seen at work in the preterite. In some cases the strong preterite plural can be homophonous with the present of another verb, for example, roken (smell-pret.pl or smoke-pres.pl) or the stem of the preterite can be homophonous with a noun or adjective, for example, rook (smell-pret.pl or smoke-noun). Previous research suggests that homophones can strengthen the memory trace of a word (Jescheniak & Levelt [Reference Jescheniak and Levelt1994], though other studies contradict these results, e.g., Gahl [Reference Gahl2008]) and we also see cases where superficial resemblance can generate contamination effects in semantics, morphology, and syntax (see Dauzat, [Reference Dauzat1927], Pijpops, De Smet, & Van de Velde [Reference Pijpops, De Smet and Van de Velde2018], and Van de Velde, De Smet, & Ghesquière [Reference Van de Velde, De Smet and Ghesquière2013] on paronymic attraction).

Nouns can influence the inflection of the strong verbs in yet another way (Fertig, Reference Fertig2009). In some cases, a strong verb can be reanalyzed as a zero derivation of a noun. As derivations almost always take the weak inflection (Bailey, Reference Bailey1997:6–29; Pinker & Ullman, Reference Pinker and Ullman2002:459; though see Ramscar, Reference Ramscar2002), this might cause the verb to weaken. For example, it might have been the case that schellen-schal-geschollen ‘sound’ became weak, because it was reanalyzed as a derivation of the noun schel ‘bell.’

Although we just adduced reasons to hypothesize that homophonous forms can attract a verb to the weak inflection by an analogical pull at the formal level, in other cases language users avoid homonymy (Gilliéron & Roques, Reference Gilliéron and Roques1912). Though homonymy avoidance has met skepticism in historical linguistic literature (e.g., Lass, Reference Lass1997:355–61), its effects are supported by several observations in the history of Dutch verbal morphology and elsewhere (Baerman, Reference Baerman2011; De Clerck & Vanopstal, Reference De Clerck, Vanopstal and Collins2015:364; De Vogelaer & Coussé, Reference De Vogelaer and Coussé2011). When presents and preterites might formally merge due to sound change, language users appear to maintain the distinction. For example, by the end of the Middle Ages, central Dutch dialect areas and the burgeoning standard language underwent a massive apocope of the schwa in the Auslaut “word final position.” However, the apocope was suspended in weak preterites where the dental preterite (either in its voiceless allomorph environment or by Auslautverhärtung “final obstruent devoicing”) would have merged with the third person singular ending -t of the present (Van Loon, Reference Van Loon2014:261): werk-t (work-prs.3sg) would be indistinguishable from werk-t(e) (work-pret.sg) with apocope. It is revealing that the schwa apocope did take place in verbs where there is no risk of confusion, as in the irregular weak forms like koch-t(e) (buy-pret.sg) versus koop-t (buy-pret.3sg). Another example is the rise of the /i:/-preterite in many originally class III verbs. In some Middle Dutch dialects /ɑ/ before liquids changed to /ɛ/, which led to a temporary merger of the first person singular of the present and preterite (e.g., help: help:pres.1sg or help:pret.sg). To sidestep this ambiguity, these verbs adopted the /i:/-preterite of ablaut class VII (Van Bree, Reference Van Bree1987:212). Ambiguity avoidance might also play a role in the strong-to-weak evolution. In case the stem of a verb ends in a dental, the weak preterite will be homophonous with the present plural (e.g., zitten sit:pres.pl or sit:pret.pl).

Finally, two phonologically similar verbs can come to be seen as one lexical item by language users. In case one of the verbs was strong and the other one weak, the inflection of the resulting verb can be either strong or weak (see Fertig [Reference Fertig2009], who shows that this is by no means a rare phenomenon; see Nowak [Reference Nowak, Schmid and Ziegler2011] for German). For instance, in Dutch, a weak verb dwalon ‘wander’ has merged with a strong verb dwelan ‘wander.’ The resulting verb dwalen took the weak inflection. These mergers can skew our results, so they have to be taken into account.

Response variable

Every attestation was coded as strong or weak.

Fixed effects

• – Century. Century was coded as a scaled and centered numerical variable.

• – Token frequency. To calculate the token frequency, we split our data into two subperiods (800–1500 and 1500–2000).Footnote ⁹ Every preterite received a value of how often its verb stem was attested in the preterite in this period, divided by the sum total of all observations (both preterites and participles) in that period. The calculation was also applied to the past participles. This relative frequency was subsequently log-transformed.Footnote ¹⁰

• – Type frequency. For the type frequency, each verb stem was assigned an ablaut class on the basis of its attestations in the dataset, though in case the class could not be deduced from our dataset (e.g., because only the preterite or only the participle appears), the class was assigned based on reference grammars. When a verb shows forms belonging to more than one ablaut class, the most frequent ablaut pattern was chosen. When the verb shows only weak forms in a specific century, it was assigned to the class it last belonged to before it weakened. This variable was coded on a century level (in order to be able to take into account verbs that change ablaut class, e.g., wasschen-woesch ‘wash’ used to belong to class VI until the thirteenth century when it changed to class VII, wiesch), although the ninth through thirteenth century were aggregated, because of the data sparsity in the first four centuries. We then counted the number of verb stems (of which at least one attestation was strong) in each class in each century. This value was divided by the total number of verb stems in that century. The end result was log-transformed.

• – Weak type frequency. To test whether the phonological similarity with weak verbs plays a role, the reference dictionaries were used (because the originally weak verbs are not included in our dataset). For each verb stem in our dataset, the number of weak verb stems with the same rhyme was counted. Different weak homonyms were not taken into account (i.e., the weak homonyms were counted as one type). This value was log-transformed.Footnote ¹¹

• – Regularity. Regularity was coded for each verb stem per period in the dictionaries.

• – Principal part. Each token was coded as a preterite or past participle.

• – Vowel pattern. The coding for the vowel pattern was largely done the same way as the coding for ablaut class, that is, based on our own data, or when necessary, based on reference grammars, for each verb stem and for each century. Usually singular and plural share the same vowel, which makes the generalization to an ABA, ABB, or ABC pattern rather straightforward. However, in some cases, this is not true. Class I and II still show a different vowel in preterite singular and preterite plural more or less until the Middle Ages (/ê/ versus /ē/ for class I and /ô/ versus /ō/ for class II). This difference was leveled by regular sound changes (now reed-reden ‘drove’ and boog-bogen ‘bent’). Class III shows a different vowel in singular and plural largely until the sixteenth century (e.g., drank-dronken ‘drank’). This difference disappeared through analogical leveling in the direction of the plural (Van Bree, Reference Van Bree1987:210). Class IV and V still show a different vowel, i.a. /ɑ–a/ (e.g., nam-namen ‘took’ and gaf-gaven ‘gave’); however, these vowels are morphologically connected for language users, as they are vowel pairs that occur in lengthening in open syllables in Dutch nouns, for example, dak ‘roof’ and plural daken (Van Bree, Reference Van Bree1987:232). When there is an actual difference between singular and plural, as for class I, II, and III in certain centuries, this cannot be summarized in a simple three-letter combination. Therefore, the more frequent singular was chosen to represent the second letter, for example, drinken-drank-dronken-gedronken is summarized as ABC with the preterite singular drank representing the second letter (cf., De Smet & Van de Velde, Reference De Smet and Van de Velde2019).

• – Homophony preterite noun, homophony preterite verb, homophony weak verb. For all the verb stems in the dataset, the dictionaries were checked to see whether a homophone existed for each period. Because there are obviously fewer words attested in the earlier dictionaries than actually existed, material from the later dictionaries was extrapolated, except when it was mentioned specifically that the word only entered the Dutch language later. For each verb stem, it was checked whether its stem was homophonous with a weak verb (e.g., scheppen strong ‘create’ or weak ‘dig’), whether its preterite plural was homophonous with another verb (either strong or weak, e.g., roken smell:pret.pl or smoke:pres.pl) and whether its preterite (singular or plural) was homophonous with an adjective or noun (e.g., rook smoke:pret.sg or smoke:noun). Hapaxes were ignored.

• – Possible reanalysis. To account for possible reanalysis, we checked for each period whether a semantically related noun with the same present stem as the verb stem existed (e.g., verdriet ‘grief’–verdrieten ‘to have grief’).

• – Merger. The difference between the merger variable and the homophony with a weak verb variable is that, in the case of merger, the two verbs will be conceived of as the same lexeme by language users. Because this is hard to measure, we could not code for actual cases of merger, but only for plausible cases of merger. This variable partly consists of the homophony with a weak verb variable but also captures other plausible scenarios for merger. We consider merger a plausible scenario for verbs that are homonymous with a weak verb, verbs that have a causative derivation (cf., Fertig, Reference Fertig2009), verbs that have a homophone and verbs that only differ in the length of their vowel (e.g., frequentatives, drupen-druppen ‘drip’) or show e instead of a or i instead of e (these vowels behave as allophones or are orthographically interchangeable between regions and time periods in earlier stages of Dutch, for example, strong schellen – weak schallen ‘sound’), or vice versa. For the fully homophonous verbs, no extra semantic constraints were applied, and we assumed that the exact formal resemblance was already enough for language users to regard both verbs as one lexical item. In the cases without full homophony, as in the causative or frequentative derivations, or in the case of small vowel differences, we added the constraint that there needs to be a plausible semantic connection. We did not differentiate on the level of this lexical merger variable: we treated it as a binary distinction ‘possible merger’ versus ‘no merger,’ and kept this distinction constant for all periods and all different verbs (because it is hard to see when a verb has actually merged and because its consequences can be long-term).

• – Dental stem. This was coded for each verb stem per period in the dictionaries.

Control variables

• – Adjectival participle. To control for adjectival use of past participles in our dataset, we added for each lemma whether its strong participle was lexicalized as an adjective, again according to the historical dictionaries.

• – Rhyme. We added whether the text in which the observations occur has end rhyme or not, as the strong form may be preserved longer to fit the rhyme.Footnote ¹² Note that the rhyme variable is not measured at the level of the individual observations but at the level of the text. This means that observations from rhyming texts that do not occur in rhyme position are nevertheless also considered as undergoing a conservation effect due to rhyme. The reason is that it was not feasible to manually check each individual observation. At the same time, this variable also controls for the genre of the text, as poetry often contains more archaic language than prose.

• – Contemporary Dutch. In order to control for the unwanted effect that verbs that have ceased to exist would show much less weakening than expected by the model, because they did not have the chance to weaken before they disappeared (e.g., dwaan-dwoeg ‘wash’), we added a variable indicating whether the verb still exists in contemporary Dutch. This was determined by a search in a reference corpus for contemporary Dutch (SoNaR, Oostdijk, Reynaert, Hoste, & van den Heuvel [Reference Oostdijk, Reynaert, Hoste and van den Heuvel2013]).

• – Earlier weakening. Our dataset suffers from an autocorrelation problem: the different observations in our dataset are not fully independent from each other. If a verb was weak in the previous centuries (e.g., geschieden ‘happen’ weakened quite early on), it should not come as a surprise that it will be weak in consecutive centuries. Nonindependence increases the risk of type I errors. Statistical models accounting for autocorrelation have not been fully integrated in historical linguistics yet (see Van de Velde & Petré, Reference Van de Velde, Petré, Knight and Adolphs2020), and full independence is rare in statistical studies (Salsburg, Reference Salsburg2001:144). We followed a practical solution to manage this problem (see also Baayen & Milin, Reference Baayen and Milin2010) by adding a variable that represents the amount of weakening in the previous century, separately for participles and preterites. In case the participle or preterite was not attested, we resorted to the nearest previous century where the participle or preterite was attested. If the verb was not attested before, we added a default value of 0% weakening. This solution ensures that the model will fit the probability of the weak form of an observation under multivariate control, which now includes a measurement of overall rate of weakening. If it is indeed the case that the fitted probability is largely determined by the verb's prior weakening, the model will promote the new variable in the overall regression model, leaving less room for the other factors to enhance the goodness-of-fit. In other words, with this extra control variable, the model will incur a penalty for the variables of interest in periods with high weakening.

Random effects

• – Source. First, it was important to add a random effect for the source in which the attestation was found; as there were many short text files, we merged all files with five or fewer observations in one level. This random effect has 1434 levels.

• – Verb stem – verb lemma. Second, a random effect was added for each verb. This is not straightforward: it is not immediately clear whether we should distinguish verbs on the basis of their stem or on the basis of their lemma. We decided to consider both by nesting the verb lemma in the verb stem. Verb lemma consists of 1961 and verb stem of 237 levels. A random slope was added for the verb stem and lemma with century, because the effect of century may differ for each verb. Some verbs never weaken through time, while others weaken quite early and still other verbs only sport weak forms in the twentieth century. This random slope can either be correlated with the random effect for the verb stem and verb lemma or not. We set out by including the correlation and tested whether this enhanced the model quality.

Interaction effects

As to interaction effects, Century was entered in an interaction with Vowel pattern. Principal part was entered in a two-way interaction with (allowing for the effect of the variables of interest to play out differently for preterites or past participles)Footnote ¹⁴:

• – Vowel pattern (especially for ABA-verbs, as we expect different results for preterites and participles here)

• – Dental stem (weak forms with a dental in the stem will mainly cause ambiguity in the preterite)

• – Adjectival participle

• – Century

If the interactions are significant, this is an indication that preterites and participles are not acting in unison, whereas nonsignificant effects could be an indication to the contrary, implying a paradigmatic connection between the two principal parts, if the main effects are significant at least. However, absence of significance should not be promoted to significance of absence.

Model selection

For our model selection, we first checked whether our random structure was necessary by using likelihood ratio tests to compare nested models and by using the AIC (Aikaike Information Criterion), which imposes a penalty for model complexity (Gries, Reference Gries2015:112–15; Zuur, Ieno, Walker, Saveliev, & Smith, Reference Zuur, Ieno, Walker, Saveliev and M. Smith2009:121–2). All random effects made our model significantly better, although the random slope did not need to be correlated with the random effect for verb lemma nested in verb stem. Then, we checked whether our interactions were necessary, by deleting them one by one from our model and comparing against the previous model. Finally, we checked whether the main effects that were not part of an interaction could be added to make the model significantly better by again using the same method.Footnote ¹⁵ We checked our final model for multicollinearity and found reassuringly low VIF-scores. Our final model attained a C-value (area under the ROC curve) of 0.999 (indicating a very good fit), a marginal R² of 0.246, and a conditional R² of 0.933, though of course this very good fit is partly due to our control variable Earlier weakening.