Semantic differences between strong and weak verb forms in Dutch

1 Introduction

In Dutch, the past tense of schuilen ‘hide’ can either be realized as schuilde or school. Adding a dental suffix to the stem, as in schuilde, is the strategy of the weak inflection, while changing the stem vowel (ablaut), as in school, is the strategy of the strong inflection^[1]. Usually, there is a clear, lexically determined preference for either the weak or the strong inflection. This lexical preference can change over time. Most frequently, originally strong verbs tend to become weak (e.g., besefte < besief ‘realised’), though sometimes, weak verbs can become strong as well (e.g., vroeg < vraagde ‘asked’). Especially the former change has been studied extensively (see Carroll et al. 2012; De Smet and Van de Velde 2019; Lieberman et al. 2007) and we have quite a clear idea why some verbs tend to become weak and others do not. Put in a nutshell, verbs with a higher frequency, or belonging to a group (e.g., ablaut class or vowel pattern) with a higher frequency tend to weaken less than verbs (or verb groups) with a lower frequency. Sometimes, however, verbs can show both inflections, for example when they are in the process of becoming either strong or weak. In these cases, when both strong and weak forms of the same verb are used frequently, we would like to know what determines the choice for either the strong or the weak inflection.

The usual suspects like token frequency^[2], type frequency (number of verbs per ablaut class) or vowel pattern cannot be the culprits, as they stay more or less constant, at least synchronically (though through time, of course, frequencies might change, or verbs may change their vowel patterns) when looking at individual verbs. A few other possible factors come to mind. Regional or social factors will probably play a role (see for example the morphological Atlas of the Dutch dialects, MAND [Goeman and Taeldeman 1996], De Vriendt [1965] and Taylor [1994] for regional differences), though we also find alternating verbs in texts from the same author. In poetry, rhyme will most likely play a role: to fit the rhyme, authors may opportunistically vary between the strong and the weak inflection (for examples, see De Vriendt 1965; Taylor 1994). Again, however, we can also find both variants of the same verb in prose. This suggests we might need to take a look at the linguistic context of the verbs. More specifically we will look at whether the semantic context might determine if a verb shows up strong or weak.

We start by explaining why we think semantics can play a role in §2. In §3, we make our research question more concrete and we explain precisely which semantic features we will study. §4 describes how we gathered and coded our data, §5 covers the analysis and §6 shows and discusses our results. The discussion will be held in the light of more general trends, like iconicity, exaptation, and Kuryłowicz’s “fourth law of analogy”. §7 concludes this paper.

2 Why do we expect semantics to play a role?

Although strong and weak verb morphology has been studied intensively in many domains of linguistics (historical linguistics, psycholinguistics, dialectology, language acquisition…), the relationship with semantics, aside from papers claiming this relation to be absolutely non-existent (i.a., Kim et al. 1991; Pinker 1999; Pinker and Ullman 2002), has not received much attention. Some studies, however, do show indications of a relationship between strong and weak verb morphology and semantics. In psycholinguistic research, for example, a study by Ramscar (2002) shows that the priming of the semantics associated with either a strong or weak verb facilitates the inflection of the strong or the weak verbs, respectively.

Other research shows that the strong inflection as a whole has a different semantic profile than the weak inflection. Baayen and Moscoso del Prado Martín (2005), for example, find that the strong inflection has a greater semantic density than the weak inflection. More specifically they notice that verbs with a strong inflection tend to show larger synonym sets, tend to have more meanings, tend to take different perfect auxiliaries (which they use as a proxy for telic vs. atelic aspect) and different argument structures, and show different contextual distributional properties. Tabak et al. (2005), too, find differences in the preferred perfect auxiliary and a different number of argument structures for the strong and the weak inflection. They give a functional explanation for the greater semantic density of the strong verbs: “the greater entanglement of irregular^[3] verbs with other parts of the grammar makes it easier to remember irregular forms” (Tabak et al. 2005: 22–23). A different semantic profile for strong and weak verbs also comes forth from the studies of Brodahl Nilsen (2012) on Norwegian and Karlsson and Sahlquist (1974: 77–79) on Swedish (as cited in Strik 2015: 37). Both studies find the strong inflection to be correlated more strongly with momentaneous semantics, which is in line with the findings of both Baayen and Moscoso del Prado Martín (2005) and Tabak et al. (2005) that the strong inflection tends to show telic aspect. Baayen and Moscoso del Prado Martín (2005: 15) suggest that the reason might be that the first verbs we acquire are verbs that are prototypically punctual, have an endpoint and a clear resulting state (Shirai and Anderson 1995). The first verbs we acquire are typically also strong verbs. So, through time, we learn more weak verbs and these verbs would be expected to lie further away from the punctual prototype than the strong verbs acquired earlier.

Furthermore, as pointed out by Strik (2015: 37), we can also find semantic pairs of causatives and the verbs they are derived from with the causative showing the weak inflection and the original verb showing the strong inflection^[4]. Examples are leiden – leidde (‘lead’) vs. lijden – leed (‘suffer’, but originally ‘go’ as well), zetten – zette (‘set’) vs. zitten – zat (‘sit’), leggen – legde (‘lay’) vs. liggen – lag (‘lie’) and so on. More in general, we also know that the original weak verbs were mostly derivatives (except for later loans that entered our language as weak verbs) (Bailey 1997: 6–29), while the strong verbs usually were not. This original distribution, with the strong verbs perhaps belonging more to our basic, core vocabulary and the weak verbs having more secondary, derived meanings, might also have contributed to the differences in their semantic profiles.

Zooming in on differences in inflection within individual verbs, we can find a few instances where a verb has undergone semantic differentiation (cf. also Bolinger (1968a) for English examples and Nowak (2011) for German examples)^[5]. Sometimes, meanings drift so far apart that the verb practically splits up in two homonyms, e.g., (1)^[6], while in other cases only a (temporary) trend can be distinguished (especially [2] and [3]), and still other examples are somewhere in between, e.g., (4)–(10). Karlsson and Sahlquist (1974) specify that the correlation they found in Swedish of momentaneous aspect with the strong inflection becomes most clear for verbs that show both strong and weak forms as a result of semantic differentiation.

1. plegen – placht ‘used to’ vs. pleegde ‘committed’

2. jagen – joeg ‘chased’ and all other meanings vs. jaagde ‘hunted’ (Haeseryn et al. 1997: 86)

3. buigen – boog ‘bent’ vs. buigde ‘bowed’ (Verwijs et al. 1885)

4. zinnen – zon ‘thought about’ vs. zinde ‘pleased’

5. pluizen – ploos ‘picked’ vs. pluisde ‘produced fluff’

6. stijven – steef ‘starched’ vs. stijfde ‘strenghtened’

7. spinnen – spon ‘spinned’ vs. spinde ‘purred’

8. schrikken – schrok ‘scared’ vs. schrikte ‘cooled down suddenly’

9. scheren – schoor ‘shaved’ vs. scheerde ‘moved quickly past’

10. snuiten – snoot ‘blowed [his nose]’ vs. snuitte ‘took away a sharp angle’

Aspectual differences between the strong and weak inflection also occur within the latter class of verbs: in English the weak class has regular forms in –ed and irregular forms in –t (dreamed vs. dreamt, burned vs. burnt) (see also footnote 1). Quirk (1970) and Levin (2009) find that –ed past tenses tend to be used more often in durative contexts, while –t past tenses can be found more in punctual contexts. All of this suggests that semantic factors might indeed (partly) determine when strong and when weak verb forms are found.

3 Research questions

We want to add to the research into semantic factors by specifically looking at possible meaning differences between strong and weak variants of vacillating verbs. This means we will study the influence of semantics on an individual token level, instead of on a verb level (as in Baayen and Moscoso del Prado Martín 2005). We first need to select which semantic features we want to take into consideration. All kinds of semantic differences are imaginable: from concreteness or animacy, to whether or not the verb denotes a bodily process. However, it would definitely not be feasible, or even useful, to code for all these semantic factors. We want to select semantic features that are not inherent to the verb, in contrast to the features that were studied by Baayen and Moscoso del Prado Martín (2005) and Tabak et al. (2005), because those factors remain constant when looking at individual verbs and would only be able to explain the variation between verbs, not within verbs. Rather, we will look at semantic features that are context-dependent.

A first potentially relevant factor that comes to mind is metaphor. This feature was chosen on the basis of the idea that verb tokens expressing a metaphorical meaning might be less associated with their original, prototypical inflection because their metaphorical meanings as well lie further away from the original, prototypical, literal meaning of the verbs. Because our dataset consists of both originally strong and originally weak verbs, we can test whether the differences we find in meaning (e.g., metaphorical versus literal) are associated with differences in innovativeness or with differences in inflection. If the difference in meaning is indeed a case of innovativeness versus conservativeness, we would see different results for originally strong and originally weak verbs, as in the group of originally strong verbs, the weak forms are more innovative, while in the group of the originally weak verbs, the strong forms are the newer forms. However, if the differences in meaning are inherent to the different inflections, we would expect the same results for originally strong and originally weak verbs.

We will also investigate whether this effect plays out the same way for preterites and past participles, because we know that past participles are in general more conservative in the process of weakening in Dutch. This is most likely an effect of ‘Präteritumschwund’, where preterites are being replaced by perfects (using past participles) to express past tense, which increases the frequency of the past participles, making them more protected from weakening (Dammel et al. 2010). Furthermore, past participles can also lexicalize as adjectives and then fossilize in a certain inflection, while the verb itself might still change inflection (e.g., lexicalized strong past participle verbolgen ‘angry’ < verbelgen ‘to become angry’ vs. non-lexicalized weakened preterite belgde ‘became angry’), which is not the case for preterites. For these reasons, we expect to see differences between past participles and preterites with regard to the effect of metaphor.

A second factor that has already been studied at the token level (instead of the type level as in Baayen and Moscoso del Prado Martín 2005 and Tabak et al. 2005), though not for strong versus weak, but for English –ed vs. –t past tenses, is aspect (Quirk 1970; Levin 2009) and more specifically durative versus punctual aspect, which indeed has to be assessed based on the context of the verb. Compare (11) and (12), where (11) shows durative aspect and (12) punctual aspect. Quirk (1970) and Levin (2009) both show that the longer –ed forms tend to appear more often in cases with durative aspect, while the shorter –t forms tend to appear more often in cases with punctual aspect.

1. It burned for three days (Levin 2009: 65).

2. It was brilliant, except I burnt someone’s leg with a firework (Levin 2009: 65).

Bolinger (1968b: 110) and Rohdenburg (2003: 277) suggest that iconicity may be at work here. If we assume this might be going on in Dutch as well, we need to compare the phonetic size of strong and weak verbs. For the preterites, we can observe that the weak variants are phonetically longer, because of the dental suffix (compare duikte versus dook ‘dived’)^[7]. For the past participles on the other hand, we see that the strong variants are longer (the nasal suffix of the strong verbs is phonetically longer than the dental suffix of the weak verbs, compare geduikt vs. gedoken ‘dived’). This means that if iconicity is at work in Dutch verb morphology as well, we would expect past participles with a durative meaning to be strong more often and past participles with a punctual meaning to be weak more often, while we would expect preterites with a durative meaning to be weak more often and preterites with a punctual meaning to be strong more often.

It could also be the case that, in line with previous research on the type level, the strong inflection overall tends to be more associated with punctual aspect, while the weak inflection tends to be more associated with durative aspect. However, if that effect occurs, we still need to make sure that it really is the strong inflection being associated with punctuality and the weak inflection being associated with durativity and not innovative forms being associated with one aspect and conservative forms associated with the other aspect, like we suggested might be the case for metaphor. Again, if this is rather a case of innovative versus conservative meanings, we will expect a different effect for originally strong and originally weak verbs, while we would expect the same effect when punctuality is really connected with the strong inflection. All of this can be summarized in the following research questions:

1. Does metaphor play a role in the choice for either the strong or the weak variant?

   1. Does metaphor determine whether we find strong or weak forms or whether we find conservative or innovative verb forms?

   2. Do preterites and past participles behave the same way with regards to metaphor?

2. Does aspect play a role in the choice for either the strong or the weak variant?

   1. Does aspect determine whether we find strong or weak forms or whether we find conservative or innovative verb forms?

   2. Do preterites and past participles behave the same way with regards to aspect?

4 Data collection

To answer these questions, we gathered data from an existing dataset (De Smet and Van de Velde forthcoming) that contains about a quarter of a million attestations of all Dutch verbs that were strong at one point in time from the 9th until the 20th century. These data were gathered from a collection of corpora, namely the Corpus Old Dutch (Pijnenburg et al. 2012), the Corpus Gysseling (), the Corpus Van Reenen Mulder (Van Reenen and Mulder 1993), the Corpus Middle Dutch (Kuiper 2017), a corpus assembled by Daniel Van Olmen (2019) and the Letters as Loot corpus (Rutten and van der Wal 2014). From this dataset, we selected all verbs (both originally strong and originally weak verbs) that showed more than 25% and less than 75% weak preterites or more than 25% and less than 75% weak past participles in a given century (with n > 10 for that century). Of the verbs that met these conditions, we collected all attestations from each century in which they showed variation from the 14th until the 20th century^[8]. We did this in order to also capture the beginning or end of the period of variation. We decided to only zoom in on the 14th until the 20th century, because earlier centuries do not yet show enough variation. We also only included the principal parts of the verb (in this case preterites or past participles) that showed variation in the period from 1300 to 2000. Stoten ‘push’ for example, has a preterite that shows both strong and weak verb forms, while the past participle only shows strong forms. We therefore did not include past participles of stoten^[9]. Finally, we also excluded graven ‘dig’, belgen ‘be angry’, scheppen ‘create’, leggen ‘lay’, rinnen ‘run’, bederven ‘spoil’ and zinnen ‘think’. Graven was excluded because all weak forms were used by only one author, belgen was excluded because all strong forms were actually instances of the adjective verbolgen ‘enraged’ (in which the strong inflection has fossilized, while the actual verb forms do not show strong forms any longer). The other verbs were excluded because the variance was clearly due to the verb being the outcome of a merger of two different, though phonologically (and sometimes semantically) close, lexical items that become perceived as one lexical item^[10] (cf. Fertig 2009) or because the variance was due to ambiguity with a weak (or strong verb, in case of an originally weak verb, e.g., leggen ‘lay’ – liggen ‘lie’, scheppen ‘create’ – scheppen ‘shovel’), which might also be visible in the semantic distribution of the weak and strong forms, leaving less room for the semantics we want to study in this paper to play a role. This left us with 26 verbs of which 15 are originally strong verbs (barsten ‘burst’, bergen ‘hide’, buigen ‘bend’, kerven ‘carve’, lachen ‘laugh’, plegen, ‘commit’, raden ‘guess’, rouwen ‘mourn’, spugen ‘spit’, stoten ‘push’, (be)tamen ‘befit’, tijgen ‘accuse’, waaien ‘blow’, waken ‘wake’, wegen ‘weigh’) and 11 are originally weak verbs^[11] (eisen ‘demand’, jagen ‘hunt’, prijzen ‘praise’, schenden ‘violate’, schenken ‘pour’, schrikken ‘scare’, schuilen ‘hide’, stijven ‘make stiff’, treffen ‘hit’, waaien ‘blow’, wijzen ‘point’, zenden ‘send’). All these verbs make up a total of 125 different verb lemmas and 7597 verb tokens.

These 7597 tokens were manually coded for metaphor and aspect. We decided to interpret metaphor very broadly: only attestations that could definitely not be taken in their literal sense, were coded as metaphorical. Verbs in bridging contexts as well were coded as literal meanings, see example (13).

1. […] een soort van wilden orkaan, die door de wereld van studie en betamelijke levensvreugd heen woei […] (corpus Van Olmen 2019: observation from 1879)

   […] a kind of wild hurricane, that blew through the world of study and befitting life joy

For aspect we had to decide whether the starting point and the end point of the action denoted by the verb in a particular context coincide, or whether they are separated in time (minutes or months, depending on the action). 915 observations were coded by a second rater. For these observations an interrater agreement was calculated. Here, Cohen’s kappa was 0.833 (p < 0.01), indicating a high reliability.

5 Analysis

We ran two different analyses, one for originally strong and one for originally weak verbs. The dataset with originally strong verbs consists of 3526 instances of 15 different verb stems. The dataset with originally weak verbs consists of 4071 instances of 11 different verb stems. The reason why we decided to split up the data is that the control variables in this model are based on our understanding of their influence on the strong-to-weak evolution (cfr. i.a., De Smet and Van de Velde 2019, forthcoming). However, we do not yet have the same understanding of their influence on the weak-to-strong evolution and we do not know whether to expect effects in the same direction. This might also skew the effects of our semantic factors. We could solve this by adding an interaction effect to each variable with the original inflection, but this would entail three-way-interactions, as the model will already consist of some interactions (e.g., between aspect and principal part). We prefer not to use three-way-interactions, because they are visually very hard to interpret, but also computationally heavy for the model.

Because our research questions probe into more general trends and because some verbs hardly show any variation with regards to aspect or metaphor (e.g., barsten ‘burst’ almost always shows a punctual meaning), we decided to combine all verbs into two models. This means we needed to take into account that some verbs are more likely to show weak forms than others. Therefore, we first of all worked with generalized linear mixed models by adding a random effect for the verb lemma nested in the verb stem, as the data showed that some lemmas of the same verb stem can be more or less vulnerable to change than other lemmas (e.g., bewegen ‘move’ shows 22% weakening in the entire corpus mentioned above while wegen ‘weigh’ only shows 0.6% weakening). A random effect for the source was added as well^[12]. Secondly, we also added a number of control variables which recent research has shown to play a role (e.g., De Smet and Van de Velde 2019; forthcoming). In order not to overcomplicate the models, we only focused on the most important ones. We added a control variable for token frequency, because verbs with a high token frequency tend to weaken less than verbs with a low token frequency (cfr. Bybee 1985; Carroll et al. 2012; De Smet and Van de Velde 2019; Lieberman et al. 2007). We also added a control variable for type frequency, which represents how many members a certain ablaut class has (see De Smet and Van de Velde forthcoming). Again, verbs belonging to ablaut classes with a high type frequency tend to weaken less, while verbs belonging to ablaut classes with a low type frequency tend to weaken more. The vowel pattern of the verbs was added as well. Three different patterns can be distinguished: ABA^[13], where infinitive and past participle share the same vowel, e.g., raden – ried – geraden ‘guess – guessed – guessed’, ABB, where preterite and past participle share the same vowel, e.g., zingen – zong – gezongen ‘sing – sang – sung’ and ABC, where no vowels are shared, e.g., stelen – stal – gestolen ‘steal – stole – stolen’. Previous research has shown that preterites of ABA-verbs tend to weaken substantially more, while past participles of ABA-verbs are very conservative (De Smet and Van de Velde forthcoming; Van Haeringen 1940). To account for this interaction with principal part, we included vowel pattern and principal part as interaction terms in the model. For the model with all originally weak verbs, we had to alter the vowel pattern variable slightly, because there were no past participles in this model of verbs with an ABA-pattern, which means this variable could not enter in an interaction with principal part. Therefore, we joined all ABC- and ABA-verbs, so that we could at least distinguish between verbs with a shared vowel for preterite and past participle on the one hand (ABB-verbs) and verbs that do not share a vowel in these principal parts (ABC- and ABA-verbs). Finally, we also added whether or not the verb is a complex verb. Though it is not yet clear to what extent this plays a significant role in the distribution of the strong and the weak inflection, it might be important on this level of the variance, especially in combination with aspect, because prefixes tend to be involved in the expression of aspectual distinctions (see De Vooys 1967: 249–255).

Other, non-verb-related, control variables are rhyme and century. As noted above, authors might choose to inflect the verb with a dental suffix or with ablaut to fit the rhyme. This variable was coded at the text level, which means that we distinguished between texts with rhyme and prose, but we did not look at whether the verb form actually occurs in rhyme position in each specific attestation. We also added a numeric variable for century, because we know that (at least) the weakening of the originally strong preterites increases through time (Carroll et al. 2012; De Smet and Van de Velde 2019; Lieberman et al. 2007). Because this is less clear for the past participles (which become more conservative through time, as an effect of the Präteritumschwund, see section Research questions), we again added an interaction with century and principal part.

Because we are dealing with an evolution through time, we need to be aware of non-independence (the observations are not completely independent of each other: verbs that already show a lot of weakening are more likely to show weak forms). Though the random effect for verb stem and verb lemma catches part of this noise, the model does not know that there is a chronology in the observations. We controlled for this by adding a value to each preterite representing the amount of weakening of the preterite of that verb in the previous century and to each past participle the percentage of weakening of the participle of that verb in the previous century^[14]. This way the model can take into account what has been happening to the verb when it encounters a new form (cfr. Baayen and Milin 2010: 19; De Smet and Van de Velde forthcoming). This variable allows, in the words of Baayen and Milin (2010: 21): “a more precise estimation of the contributions of the other, theoretically more interesting, predictors.”

Finally, of course, our variables of interest, metaphor and aspect were added to the model. In order to answer our research questions, we also had them participate in an interaction with principal part. To answer our research question about the conservativeness or innovativeness of metaphorical versus literal meanings or punctual versus durative meanings, we will compare the model with originally strong verbs with the model with originally weak verbs.

6 Results and discussion

6.2 Discussion

6.2.1 Metaphor

Figure 1 shows the main effect of metaphor in both models. We see that for the originally strong inflection, verbs with a metaphorical meaning tend to show more strong forms, while for the originally weak inflection, verbs with a metaphorical meaning tend to show more weak forms^[17]. This means that in both cases the verbs with a metaphorical meaning tend to show the more conservative forms, while the verbs with a literal meaning tend to show more innovative forms. This is a nice example of Kuryłowicz’s fourth law of analogy (Kuryłowicz, 1947) which states: “When as a consequence of a morphological [= analogical] change, a form undergoes differentiation, the new form takes over its primary (‘basic’) function, the old form remains only in secondary (‘derived’) function.” (Hock 1986: 223). A typical case of this law is the often-cited brothers – brethren example where brethren survives only in a secondary meaning, while the newer form brothers takes over the original, basic meaning. An example in the domain of strong and weak verb morphology are cases like molten – melted, where molten is still used in the secondary function as an adjective, while the new form melted continues in the original function, namely as the past participle of the verb. Here the literal meaning is the primary or basic function, which lives on in the newer form, while the metaphorical meaning is the secondary or derived function, which takes over the original form. This same effect was already observed by Conradie (1985: 76) in Afrikaans, metaphorical meanings are more often expressed by strong past participles and literal meanings by weak past participles, compare for example n gebroke hart ‘a broken heart’ with n gebreekte bord ‘a broken plate’.

Figure 1:

Effect of metaphor for originally strong verbs (left) and originally weak verbs (right)^[18].

This effect of metaphor is at first sight slightly counterintuitive. We hypothesized that, because the meaning of metaphorical verbs lies further away from the original, basic meaning, their form might also move away faster from the original form. However, Kuryłowicz’s law shows that in most cases, the secondary, less basic meanings are stuck in the original form, while the more basic meanings are more prone to change. As Conradie (1985: 76) rightly points out, metaphorical meanings are more often used in idiomatic or (semi-)fixed expressions which are more conservative. These verbs might also fossilize in their metaphorical or idiomatic expression and become disconnected from the original lexical item in the mind of the language user. As a consequence, the fossilized verb would not undergo the same changes as the original lexical item.

Figure 2 shows the interactions in both models with principal part. We indeed see that preterites and past participles behave differently. In the model with only originally strong verbs, the difference in metaphor is hardly noticeable for the past participles (it even slightly tends to the other direction). This is unexpected, especially in light of Conradie’s (1985: 76) research on Afrikaans where the past participles are the ones that are affected by metaphor. However, we also know from previous research (Dammel et al. 2010) that past participles are more conservative. This may explain why literal past participles show less weakening than literal preterites. What remains unexplained for now, is why the metaphorical past participles show so much more weakening than the metaphorical preterites. In the model with originally weak verbs, on the other hand, the effect of metaphor is most clear for the past participles. We do not yet have a straightforward explanation for why this would be the case either.

Figure 2:

Interaction principal part and metaphor for originally strong verbs (left) and originally weak verbs (right).

6.2.2 Aspect

For aspect, the main effects are not significant in either model. The interactions with principal part are significant though. In Figure 3, we can see that both for the originally strong and the originally weak verbs, punctual preterites tend to be strong more often and durative preterites tend to be weak more often, whereas, for the originally strong verbs, we can also see that punctual past participles tend to be weak more often and durative past participles tend to be strong more often. This is an example of diagrammatic (or relative) iconicity (Haiman 1980, 1983) of quantity where a greater quantity of form expresses a greater quantity of meaning (e.g., singular book – plural books).

Figure 3:

Interaction principal part and aspect for originally strong verbs (left) and originally weak verbs (right).

This type of iconicity has met some criticism by Haspelmath (2008a, b) who claims these phenomena are to be explained as an effect of frequency rather than as an effect of iconicity: the more predictable and thus the more frequent a word is, the shorter it will be (Haspelmath 2008: 5; Zipf 1935). In this case, one might be inclined to argue that the punctual meaning is more predictable and more frequent and therefore takes the shorter forms. There are, however, a number of reasons why this is not very plausible. First of all, the difference in frequency between punctual and durative meanings is negligible (at least in this dataset). Punctual meanings represent 56% of all instances, while durative meanings represent 44% of all instances. When we look at only the originally strong verbs, durative meanings even take the upper hand. Secondly, language users do not just pick one form to express the most frequent category, because it is shorter, but the form becomes shorter, because it is used more often (Pustet 2004). That way, it becomes more entrenched and the production becomes more routinized, which can cause reduction of this form (Croft 2008: 52; Bybee 1985). However, this is not how this variation came about. The shorter forms in this case are not a reduction of the longer form, but either came about by analogy (in case of weakened originally strong past participles or originally weak preterites that became strong) or are the original variants (in case of the originally strong preterites or in case of the originally weak past participles). So, the tendency to express durative meanings by longer verb forms cannot be explained by frequency.

But how does the iconicity emerge then? Of course, we do not claim that these verbs became variable in their inflection because there was a need for a more iconic system. There are verbs that show variation, but do not show a difference in aspect (like we said, barsten is almost always punctual) and there are also verbs that do not show variation, but where you could distinguish between different aspects. Even the verbs that do show an iconic differentiation between punctual and durative meanings, do not necessary hold on to this variation. We rather believe iconicity appears, once the variation exists, because it can. It makes use of a morphological differentiation which does not serve any other purpose and is thus an ‘exaptive’ change (Lass 1990; Norde and Van de Velde 2016). The possibility to use this differentiation to express aspect is available and it is exploited so in an iconic way, perhaps because it is aesthetic (Haiman 2008: 45), perhaps because it is natural (Dressler et al. 1987: 18) or perhaps because it is easier to process (Givon 1985: 189), though the jury is still out on the reality of these explanations (cfr. Haiman 2008: 45; Haspelmath 2008a, b: 25–26). There are, however, some experiments, mostly on sound symbolism, that show that iconicity indeed benefits language processing (see for an overview Dingemanse et al. 2015; Perniss et al. 2010: 6–8). Somewhat related to our specific case of iconicity is the research of Shintel et al. (2006) who find that speakers use a faster speaking rate when describing faster moving objects, while listeners used this information to guess the speed of the objects.

However, because there is not necessarily a need to formally mark the difference between durative and punctual meanings (some verbs never show the variation that can express this aspectual difference, other verbs do not even show aspectual differences), we also see that this refunctionalization is not pervasive enough to ensure both forms survive. When the two variants exist, they are used in the most aesthetic, natural or easy-to-process way, with the longer forms for durative situations and the shorter forms for punctual situations, or at least while the variation lasts. The analogical drift towards either one of the inflections seems to be stronger than this iconic exaptation. Only in a few cases, where other meaning differentiations have taken place (e.g., zinnen, plegen) both variants survive.

How does this iconic refunctionalization align with previous results of Baayen and Moscoso del Prado Martín (2005), Brodahl Nilsen (2012) and Karlsson and Sahlquist (1974), Tabak et al. (2005: 77–79), who all show that the strong inflection tends to correlate significantly with punctual verb semantics? From the results of our token-based study, we cannot conclude that strong variants are more associated with punctuality than weak variants. In the two models, the main effect of aspect does not reach significance. So, when looking at individual verbs with variable inflection, iconic exaptation seems to be a more accurate explanation for the distribution of the strong and the weak forms than a mere association between punctuality and the strong inflection. Of course, this does not disprove that the strong inflection as a whole is also correlated with punctuality, as is shown in previous research, it just does not show up in the very small and biased (only variable verbs) fraction of strong verbs we have looked at.

The reason for the coexistence of these two propensities is that they are both trends, rather than brute forces. The iconic exaptation might be a temporary tendency in periods of variation, while the association between punctuality and the strong inflection may be more stable diachronically, but does not amount to a strong constraint, and can easily be overridden as well.

7 Conclusion

Whether we say schuilde or school turns out to be (partly) determined by the semantic context of the verb. We see two discernable trends. Firstly, aspect plays a role. If the verb is used in a durative context, we tend to say schuilde, if the verb is used in a punctual context, we tend to say school. This way, we refunctionalize the strong-weak variation (which was not used to express any other function) in an iconic manner, even though the variation is usually not persistent. Often it has been claimed that in order for the variation to survive, the variants need to show some function or meaning differentiation (Bolinger 1968a: 127; Kroch 1994) (though lately, different voices have argued that variation an sich is functional too, see De Smet et al. 2018). It would be interesting, though, to see whether verbs where this differentiation occurs, tend to show variation longer than verbs where this differentiation does not occur. However, this would require more data for each verb than we now have, and will have to be postponed to future research. Other future research would consist of backing up this data with experimental research to see whether these patterns we find in our data, actually exist in the minds of language users.

Secondly, we can also observe that verbs with some meanings are more conservative than verbs with other meanings. More specifically, the results showed that verbs with a metaphorical meaning tend to be more conservative than verbs with a literal meaning. Though counterintuitive at first sight, this seems to be a more general trend in language change (see Kuryłowicz’s fourth law of analogy), though it has only been investigated on an anecdotal basis. It could be interesting to investigate this interplay between morphological and semantic change on a larger scale: for different morphological changes, for different languages, in different time periods.