Creativity, paradigms and morphological constructions: evidence from Dutch pseudoparticiples

3.1 A constructionist approach to creativity

In this paper we take a usage-based, constructionist approach to the analysis of Dutch pseudoparticiples. Usage-based linguistics (e.g., Kemmer and Barlow 2000) sees the relationship between language structure and language use as bidirectional. Existing generalizations allow speakers to produce tokens of use, while at the same time, our capacity to make generalizations across usage events can lead to the creation of new pieces of linguistic structure. We will show that this distinction is helpful in understanding the patterns in use of Dutch pseudoparticiples as a whole, as well as differences between the two sets of pseudoparticiples.

Usage-based approaches to language change place great emphasis on creativity as a means to create new words and patterns “beyond the resources in memory” (Goldberg 2019: 61). Speakers may do so by means of “intentional manipulation of linguistic structure” (Bergs 2018: 290), which may be acknowledged by hearers in order to distinguish between creativity and errors (Uhrig 2020: 8).^[4] As argued in Hoffmann (2020: 2–3) however, there may be creativity without deliberateness, e.g., in the grammaticalization of English to be going to as a future auxiliary. In generative linguistics, creativity is used in yet another sense, i.e., to refer to the exclusively human ability to create an infinite number of expressions from a finite number of means (Chomsky 1975: 7–8). In order to reconcile these different views, Sampson (2016) proposes a model which encompasses both creativity in the Chomskyan sense and creativity as the source of new forms or schemas. The first is F-creativity (fixed creativity), defined as “activities which characteristically produce examples drawn from a fixed and known (even if infinitely large) range” (Sampson 2016: 17). The second is E-creativity (enlarging or extending creativity), or “activities which characteristically produce examples that enlarge our understanding of the range of possible products of the activity” (Sampson 2016: 17). In CxM terms, F-creativity involves the formation of new words that are fully sanctioned by existing, conventionalized schemas, while E-creativity implies the establishment of a new schema as a generalization over such new words.^[5] In this paper, we will propose an adapted version of Sampson’s model which, besides E-creativity, includes two types of F-creativity: apart from full sanction by an established schema (F1-creativity), speakers may form words that are only partially sanctioned by that schema (F2-creativity). In Section 6 we propose a hypothetical scenario for changes in Dutch pseudoparticiples involving F1-, F2-, and E-creativity at different stages of development.

Recent theorizing about creativity has stressed the overlap between creativity and extravagance (e.g., Eitelmann and Haumann 2022: 4), and Kempf and Hartmann (2022: 20) explicitly refer to pseudoparticiples as (potentially) extravagant expressions with extravagant meaning. In our view, however, it is important to keep the two concepts apart. Extravagance is a communicative strategy to attract attention, as in Haspelmath’s well-known ‘maxim of extravagance’: “talk in such a way that you are noticed” (Haspelmath 1999: 1,055). We argue that F2-creativity can be triggered by extravagance, but extension of an existing schema is not, by itself, extravagant. Another way in which creativity and extravagance intersect is that increasing frequency of a product of F2-creativity may lead to entrenchment, with the result that it is no longer perceived of as extravagant, as in the example of German benachbart ‘neighboring’ (literally: ‘be-neighbor-ed’) (Kempf and Hartmann 2022: 22). A similar case can be made for Dutch high-frequency types such as bejaard ‘aged’ (see Section 5.5).

3.2 Sanctions, schemas, and paradigms

In this section, we consider some key concepts in cognitive linguistics of relevance to the structure of pseudoparticiples. We begin (Section 3.2.1) with a discussion of full and partial sanction (Langacker 1987). We then consider the place of schema unification and paradigms in Construction Morphology (Section 3.2.2).

4.1 Introduction

As stated above, the purpose of this paper is to explore the relationship between creativity, entrenchment, and paradigm formation, including how schema unification and back-formation work in this specific construction. To this end, we identified be- and ont-pseudoparticiples in a web-based corpus of informal Dutch and compared these to regular be- and ont-participles.

For our quantitative analyses, which are strictly corpus-driven,^[8] we use an operational definition of pseudoparticiple: a pseudoparticiple is a participial form, attested in a static corpus, which is not paradigmatically linked to verbal forms (infinitives or finite forms) in the same corpus.^[9] This approach has two important implications. First, we do not consider data from outside of the corpus. We are aware that verbal forms corresponding to the pseudoparticiples in our data sets may be found in Google searches, and also that there are pseudoparticiples that are not attested in the corpus. Second, our definition is based on formal properties and (the absence of) paradigmatic links. We make no a priori assumptions about the categorial status of pseudoparticiples – pseudoparticiples may, as we wrote previously, be treated as adjectives in reference grammars, but our interest is in their syntactic behavior (which may include verbal constructions).

The corpus we used to compile our sets of pseudoparticiples as defined above is the Dutch 2014 TenTen Web Corpus (nlTenTen14), available on the Sketch Engine platform^[10] (Kilgarriff et al. 2014). In spite of its being limited to “only” 2,253,777,579 words, a corpus the size of nlTenTen14 can be seen as representative of informal written Dutch, and hence the be- set and the ont- set can be seen as representative of (pseudo)participles in informal written Dutch generally.^[11] The corpus consists of randomly crawled sentences that have been lemmatized and POS-tagged. With ca. 75 % of the corpus extracted from the .nl domain, the corpus primarily represents Netherlandic Dutch.^[12] Extensive and largely informal, nlTenTen is very well suited for our purposes because it is possible to find pseudoparticiples with low token frequency, including so-called ‘hapax legomena’ (henceforth ‘hapaxes’), i.e., types that occur only once in a corpus, which are possibly new formations (Kempf and Hartmann 2022: 23).

4.2 Corpus queries

Pseudoparticiples, as defined in Section 4.1, are very rare, and hard to find in a corpus because of two practical challenges: (i) pseudoparticiples are not tagged as such, so they needed to be queried as character strings and (ii) Sketch Engine only allows limited downloads of maximally 10,000 concordance lines. With two different prefixes (be- and ont-), two different dental suffixes (-d and -t), and two inflected forms (-ø and -e), pseudoparticiples come in eight different forms, as the basic regular expressions in (18–21) illustrate.

These regular expressions yielded millions of results, and because of said download limitations, strings were queried in batches; for example, the regular expression in (22) gives all participles with the prefix be- and the suffix -t or -te, with one of the characters [a,b,c,d] as the first character following the prefix.

From these results, all tokens that were not pseudoparticiples were removed from the results by adding the ‘exclude word’ command (word !=) to the regex. The irrelevant results include regular participles such as betaald ‘paid’, or forms that were not participles at all, such as nouns ending in the derivational suffix -heid (e.g., meerderheid ‘majority’). Excluding those from the queries typically resulted in very long regular expressions of more than 5,000 characters.^[13]

Due to the sheer frequency of these strings, several rounds of expanding the regular expression with tokens to be excluded were necessary to arrive at a set of less than 10,000 results, which were then downloaded and filtered further in Excel. Eventually, we had lists of potential pseudoparticiples, i.e., participles with a nominal base, but we still checked the corpus for infinitives and finite forms of these participles, and even if we only found a single one, the corresponding participle was removed from the set.

To give just one example, our initial set of potential pseudoparticiples included ontstengeld ‘with the stem removed, hulled’ (<stengel ‘stem’), which is also mentioned in Van Haeringen (1949: 189) as a ‘preverbal’ (i.e., pseudoparticipial) form. However, since we also found a verbal (imperative) form in ontstengel de aardbeien en snij ze in twee ‘hull the strawberries and cut them in half’, we no longer consider ontstengeld a pseudoparticiple and removed it from the dataset. Especially for ont-pseudoparticiples which, as we will see later on, are particularly prone to paradigmaticization, this meant a substantial reduction of the dataset.

The resulting lists of pseudoparticipial types were then inserted in new regular expressions, to obtain concordances that only contained pseudoparticiples. To these concordances we added pseudoparticiples with very high token frequency (such as bejaard ‘aged’), that had been queried separately (again, because of download limitations in Sketch Engine).

All pseudoparticiples were annotated on the type level for semantic categories (see Section 5.2). But since we are also interested in morphological properties of pseudoparticiples (see Section 5.3), we ran further queries in nlTenTen14, in order to find out whether the pseudoparticiples in our data sets feature comparative and/or superlative forms.

For the syntactic analysis, in which we compare pseudoparticiples to participles with the prefix be- or ont- generally, we queried the corpus using the regular expressions in (23–24). These find both uninflected and inflected forms that are tagged as either adjective (“adj”) or past participle (“verbpapa”).^[14]

We randomized the results in Sketch Engine, then downloaded two samples of 1,500 tokens each, of which 500 were annotated in Excel (see further Section 5.4).

5.1 Overview

The method described in Section 4 yielded two data sets, which are very different in size. Due to a number of lexicalized types with very high frequency, such as befaamd ‘famous’ and behaard ‘hairy’, the set of be-pseudoparticiples consists of 46,448 tokens across 212 types (144 of which are hapaxes – i.e., types that occur only once in the corpus), whereas ont- has only 143 tokens, across 84 types (69 of them hapaxes). Further, we find substantial differences across semantic categories, to the extent that, rather than lumping all results together, it makes more sense to consider separately the productivity of the be- and ont- sets per category, as shown in Section 5.2.

5.2 Semantic categories and productivity

As mentioned above, pseudoparticiples are very unevenly distributed across the two prefixes and various semantic categories of the nominal bases. The latter can be divided into seven categories: animal, body, clothing, concept, object, person, and substance. We used these categories because we are interested in the semantic niches in which be- and ont-pseudoparticiples are found to cluster. We were aware, looking through the data, that clothing appeared to be such a niche, and so distinguished this from other objects. The same holds for distinguishing person from animal. One example for each category is given in (25–31):

animal:

body:

clothing:

concept:

object:

person:

substance:

The quantitative results of our corpus queries are given in Tables 2 and 3, which list the numbers of types, tokens, and hapaxes for each semantic category as well as three measures of productivity. Apart from the basic computation of the Type Token Ratio (TTR), i.e., the number of types divided by the number of tokens, we give the Potential Productivity (PP), i.e., the number of hapaxes divided by the number of tokens, which can be seen as a proxy for vocabulary growth (Baayen 2009).^[15] Since both TTR and PP can be skewed by types with (very) high frequency (Gyselinck 2018: 177–178), we furthermore give the Hapax Type Ratio (HTR). We note that for be-, body, concept, and substance have the highest token frequencies, which are largely due to a handful of very frequent types. For body, these are the lemmas behaard (4,384) ‘hairy’, bezweet (2,170) ‘sweaty’, bebloed (1,066) ‘covered in blood’, and bebaard ‘bearded’ (760); for concept these are befaamd ‘famous’ (15,239), begaafd ‘gifted’ (6,077), bejaard ‘aged, elderly’ (5,922),^[16] bezorgd ‘concerned’ (2,866),^[17] and belust ‘keen’ (1,209). The most frequent lemmas for substance are besneeuwd ‘covered in snow’ (6,418), berijpt ‘frosted’ (263), and bedauwd ‘bedewed’ (171). Some of these are (strongly) lexicalized – belust ‘keen’, for instance, has the noun lust ‘lust’ as its base but lacks the connotation of sex. The high token frequencies in these two categories clearly have an impact on productivity, as can be seen in Table 2. Where types are concerned, the categories of object and (again) body and clothing stand out. Since object and clothing lack high-frequency items (for object the highest frequency is bebladerd ‘leafy’ [162], for clothing it is bebrild ‘bespectacled’ [209]), they have higher TTR and PP scores. For smaller categories, particularly animal, all frequencies are very low, so that their productivity scores are not very informative (they are merely included for the sake of completeness).

For ont-, we do not find distributional differences on the scale we observed for be-. Instead, nearly half of the ont-pseudoparticiples are hapaxes and the highest frequency in this set is for ontlichaamd ‘disembodied’ (38). The reason for this discrepancy is that ont-pseudoparticiples are much more prone to back-formation, with the result that they become regular participles in a verbal paradigm. We return to this issue in Section 6.

All in all, there are substantial differences in productivity, ranging from an extremely small PP (0.0000948) for be-pseudoparticiples in the category of concept, to the maximum score of 1 for the TTR and PP for ont-pseudoparticiples in the categories of animal and clothing, which in this case are not reflecting exceptionally high productivity but very low numbers of both types and tokens. In such cases, the HTR is likewise high, but again that does not signal high productivity.

We conclude this section by considering the ‘Global Productivity’ (Baayen and Lieber 1991), which can be visualized by plotting PP on the y-axis against the number of types on the x-axis. High Global Productivity is characterized by both high PP and high Type Frequency, in other words, by a position in the upper right corner in a Global Productivity plot. As we see in Figure 1, however, no construction meets these criteria – either PP is high but Type Frequency is low (e.g., the categories of animal and clothing for ont-) or Type Frequency is relatively high, but at a low PP (the category of object for ont-).

Figure 1:

Pseudoparticiples: Global Productivity.

5.3 Morphological constructions

In Section 2, we have seen that among the few unambiguously adjectival constructions are (pseudo-)participles with a negative prefix on- ‘un-’. To check whether the pseudoparticiples in our dataset are found in their negated forms in nlTenTen14, we ran queries of tokens of on- when prefixed to the pseudoparticiples in our data (including inflected forms in -e). This gave no results for ont-pseudoparticiples at all, and 29 tokens for be-pseudoparticiples; e.g., onbebaard ‘beardless’, onbezorgd ‘not worried’, or onbenaamd ‘unnamed’. Where comparatives and superlatives are concerned, once more we found no examples for ont- at all, although for be- they are extremely rare as well: only 12 comparatives (e.g., begaafder ‘more gifted’, bezweter ‘more sweaty’) and five superlatives (e.g., begaafdst ‘most gifted’, befaamdst ‘most famous’). It is also worth pointing out that these are strongly lexicalized pseudoparticiples. This lower degree of compositionality means that speakers are more likely to treat the pseudoparticiples as adjectives and less likely to derive a new verb as attested in related finite forms.

Another relevant set of examples concerns adjective-participle compounds as bruinbehaard ‘brown-haired’ or witbesneeuwd ‘covered in white snow’ schematized in (32). We queried such constructions for both prefixes and found no tokens for ont- and 2,803 tokens for be-. The relevance of this finding is, again, that be-pseudoparticiples are more adjectival than ont-.

5.4 Syntactic constructions

As mentioned in Section 2 above, participles have both adjectival properties (such as inflection) and verbal properties (such as usage as the main verb in perfect constructions). In this section, we review all syntactic constructions that past participles figure in, including constructions where participles behave like either adjectives or verbs, but also those that are ambiguous between a verbal and adjectival interpretation. For this comparison, we used 500-token randomized samples of regular be- and ont-participles. For the pseudoparticiples, we also analyzed a 500-token randomized sample for the be-pseudoparticiples,^[18] but for the ont-pseudoparticiples we only had 143 tokens in our set.

5.4.3 Ambiguous constructions

Types A and V1/V2 include pseudoparticiples that are fully sanctioned by the Adjective and Verb schemas respectively. In addition to those, we find pseudoparticiples that are partially sanctioned by those schemas. We label these AV, because the pseudoparticiple can be interpreted either as an adjective in a copular construction or as a main verb in a passive construction. The two AV types are as follows:

Type AV1: pseudoparticiples in predicative position in main clauses:

(36)

a.

Te midden van verbrokkelde muren staan vuile bedden. De vensters zijn betralied .

‘Standing among the crumbled walls are dirty beds. The windows have bars on them.’

[tralie ‘bar’]

b.

Het land moet beveiligd worden, ontvogeld , ontganst , ontkipt .

‘The country needs to be protected, stripped of birds, geese, chickens.’

[vogel ‘bird’, gans ‘goose’, kip ‘chicken’]

Type AV2: pseudoparticiples in predicative position in subordinate clauses with “green” word order (i.e., with the pseudoparticiple preceding the finite verb):

(37)

a.

Maar liefst 85 procent van de pisten (…) kan kunstmatig besneeuwd worden.

‘No less than 85 % of the ski slopes can be covered in artificial snow.’

[sneeuw ‘snow’]

b.

Wellicht is dit ook de manier waarop de Ommelanden van Groningen ontfriest zijn .

‘Perhaps this is also the way in which the area surrounding Groningen is/has been defrisianized.’

[fries ‘Frisian’ (adjective)]

A verbal interpretation becomes more likely in the presence of a prepositional agent, e.g., door de wind in example (38).

(38)

en	toen	ik	aankwam	bleek	door	de	wind	een	heel
and	when	I	arrived	turned.out	by	the	wind	a	whole
stuk	van	het	water	“ ontkroost ”	te	zijn!
part	of	the	water	“de-duckweeded”	to	be
‘and when I arrived it turned out the wind had removed a large swath of duckweed!’

The frequencies of all construction types are given in Table 4 (pseudoparticiples) and Table 5 (all participles) and illustrated in Figures 2 and 3.

Table 4:

Construction types: pseudoparticiples.

be-			ont-
Type	CXN	FREQ	Type	CXN	FREQ
A	Attributive adjective	377	A	Attributive adjective	54
V1	Main verb (with auxiliary hebben ‘to have’)	0	V1	Main verb (with auxiliary hebben ‘to have’)	14
V2	Main verb in subordinate clause with “red” word order	1	V2	Main verb in subordinate clause with “red” word order	4
AV1	Participle in predicative position	89	AV1	Participle in predicative position	53
AV2	Participles in subordinate clause with “green” word order	33	AV2	Participles in subordinate clause with “green” word order	18
Total		500			143

Table 5:

Construction types: all participles.

be-			ont-
Type	CXN	FREQ	Type	CXN	FREQ
A	Attributive adjective	141	A	Attributive adjective	77
V1	Main verb (with auxiliary hebben ‘to have’)	48	V1	Main verb (with auxiliary hebben ‘to have’)	150
V2	Main verb in subordinate clause with “red” word order	44	V2	Main verb in subordinate clause with “red” word order	54
AV1	Participle in predicative position	195	AV1	Participle in predicative position	176
AV2	Participles in subordinate clause with “green” word order	72	AV2	Participles in subordinate clause with “green” word order	43
Total		500			500

Figure 2:

Participles and pseudoparticiples: syntactic constructions (based on token frequencies).

Figure 3:

Participles and pseudoparticiples: syntactic constructions (based on lemma frequencies).

Figure 2 ^[19] shows a striking difference between be-participles and be-pseudoparticiples. Adjectival constructions make up only 28.2 % of the set of regular be-participles, but 75 % of the be-pseudoparticiples – for verbal constructions, this is 18.4 % against 1 %. When distributions are so dissimilar across syntactic constructions, we propose that be-pseudoparticiples are a new schema, which is suggestive of E-creativity. Recall also that we left out lexicalized constructions such as bejaard (cf. footnote 18); had we left them in the difference with regular be-participles would have been ever larger, since the lexicalized ones are predominantly used in adjectival constructions.

Although we find distributional differences in the two ont- sets as well, these differences are much smaller. Coupled with the observation that we find much more back-formation in ont-participles (i.e., they behave more like regular verbal participles), this makes the establishment of a new schema less likely.

As one reviewer pointed out, the comparison in Figure 2 has the disadvantage that frequency distributions are dissimilar for regular participles and pseudoparticiples because of the relatively high number of hapaxes in the latter sets. We therefore did the same for lemma frequencies (where each type is counted once), illustrated in Figure 3. Although the figures are different, we still see a majority of adjectival constructions in the pseudoparticiple sets, especially in be-pseudoparticiples. It is furthermore telling that the number of verbal constructions is still very small in the be-pseudoparticiples, which underscores our suggestion that be-pseudoparticiples are E-creative. We explore this issue further in the next subsections.

5.5 Sanctions, schemas, and paradigms revisited

In this subsection, we return to some of the issues raised in Section 3.2 above, and discuss them in light of the data collected from the Dutch corpus. In particular, we consider (a) full and partial sanction and the connection with schema unification (Section 5.5), and (b) category membership and paradigms (Section 5.5.2).

6.1 RQ1: what are the constraints on the formation of verbal paradigms from Dutch be- and ont-pseudoparticiples?

Our analysis in Section 5 has shown that be- and ont-pseudoparticiples differ significantly in terms of frequency (distributions) and morphosyntactic properties. The be- set is characterized by some high-frequency types (e.g., bejaard ‘aged’ or befaamd ‘famous’), but we also saw ‘pockets of productivity’ (Cappelle 2014), i.e., clusters of relatively high Type Frequency in some semantic categories (particularly body, clothing, and object). We also saw that be-pseudoparticiples are overwhelmingly adjectival, not only in comparison to ont-, but also in comparison to regular be-participles.

By contrast, the ont- set is very small, whereby most types are hapaxes. Back-formation is extremely common, such that even a strongly entrenched and token-frequent form like ontheemd ‘displaced’ turned out not to be a pseudoparticiple on our corpus-driven definition. On the morphosyntactic level, differences with regular ont-participles were much smaller than what we found for be-. We argued that back-formation is constrained by semantic factors: a verbal form is implausible when the pseudoparticiple does not imply a human agent (as in bebaard ‘bearded’), or when a verbal construction is pre-empted by another construction (e.g., expressions for putting on items of clothing, blocking verbal forms such as ^? besjalen ‘to put on a scarf’). The ont- set is less likely to be affected by this, because the privative meaning makes it easier to back-form a transitive verb ontsluieren ‘to remove a veil’ out of a participle such as ontsluierd. On the basis of the differences we found, we suggested strong entrenchment for the be-pseudoparticipial schema, but weak entrenchment for the ont-pseudoparticipial schema.

6.2 RQ2: what do Dutch pseudoparticiples tell us about the relation between morphological creativity, sanction, and the development of new paradigms?

An appropriate answer to this question requires us to think about (a) general patterns across the set of (pseudo)participles and (b) differences between the be- and ont- set of pseudoparticiples. We address both these issues by returning to Sampson’s (2016) distinction between F- and E-creativity and our slight adaptation of it. Our corpus analysis revealed differences in the behavior of the be- and ont- sets, in particular variation in the extent to which new morphological paradigms have been created. We suggest this synchronic gradience is the result of gradual constructional change (Traugott and Trousdale 2010). This variation is also consistent with some ideas in usage-based constructional approaches to language as discussed in Section 3.1 above. Extrapolating from the synchronic patterns we see in our data, we now propose the following diachronic developments:

• Stage 0: speakers use the established participle schema to create regular instances, such as bepaald ‘determined’ (full sanction: F1-creative)

• Stage I: speakers create extensions of the established participle schema in which the lexical base is not a verb, such as bejaard ‘aged’ (partial sanction: F2-creative)

• Stage IIa: speakers perceive similarities across the set of extensions and form a generalization (i.e., a pseudoparticiple schema: E-creative)

• Stage IIb: speakers create new finite and infinitival verbal forms from these pseudoparticiples that align with established paradigms (full sanction: F1-creative)

Stages 0, I, and IIb are all cases where a piece of linguistic structure sanctions particular instances of use, while Stage IIa is a case where a generalization across usage events has led to the creation of a new schema. Interestingly, this aligns with the distinction between F- and E-creativity. F-creativity occurs when existing schemas fully or partially sanction new tokens (structure shapes use), while E-creativity occurs when speakers’ generalization over existing tokens creates new schemas (use shapes structure).

It appears that both the be- and ont- sets behave similarly with respect to Stages 0 and I, and the crucial differences are in Stage II. Specifically, we suggest that, following cases of F2-creativity in Stage I, there are two potential paths of development. One path involves E-creativity and the establishment of a new schema (Stage IIa); the second path involves F1-creativity, extending an existing schema (Stage IIb). In other words, the first path involves schema creation while the second path involves paradigm alignment. Given the distributional differences between the be- set and ont- set as laid out in Tables 2 and 3 above, we suggest that the be-pseudoparticiple schema is more entrenched than the ont- schema. This is E-creative use on the part of Dutch speakers, and suggests that Stage IIa is more characteristic of the be- set than the ont- set. By contrast, there is a high number of back-formations associated with the ont- set, but not with the be- set, and so a higher incidence of new finite and infinitival verbal forms beginning with ont-. This is F1-creative on the part of Dutch speakers, and suggests that Stage IIb is more characteristic of the ont- set than the be- set. We summarize this in Table 6.

Our adapted model of F- and E-creativity is novel, and it would be interesting to test it not only against diachronic data, but also against other instances of creativity, e.g., those that do not originate in F1-creative schemas (e.g., morpheme spawning from blends [Norde and Sippach 2019]) or instances of constructional contamination (Pijpops et al. 2018).