2.1. A link-based network view of constructions

Usage-based Construction Grammar (CxG) assumes that linguistic units are organized in a network called the ‘constructicon’ (Langacker Reference Langacker1987; Goldberg Reference Goldberg1995). The constructicon consists of constructions, i.e. learned form–meaning pairings, where aspects of their form and/or meaning are not predictable from their component parts (Goldberg Reference Goldberg1995: 4). Constructions are linguistic units of varying degrees of schematicity and complexity, ranging from abstract grammatical patterns (e.g. the ditransitive, [NP V NP NP]) to concrete morphemes (house, de-, -ment). Following from the network idea, to know a language means to not only know the network’s nodes (constructions), but also their relationships (links). Links in CxG are understood in two different but related ways: the first is the traditional use, in which links refer to relationships between constructions, such as horizontally between the ditransitive (John gives Mary the book) or the to-dative (John gives the book to Mary), or vertically between the into-causative (They pushed them into fighting the cause) and the caused-motion construction (He pushed them out of the way), from which it inherits a number of semantic and syntactic properties. In the second use, links refer to usage aspects such as statistical relationships, combinatorial possibilities and constraints, idiomaticity, or situational appropriateness. Knowledge of nodes and links in either reading is assumed to arise from, and be modified by, language use and experience (Hopper Reference Hopper1987; Tomasello Reference Tomasello2003; Bybee Reference Bybee2006, Reference Bybee2010; Beckner et al. Reference Beckner, Blythe, Bybee, Christiansen, Croft, Ellis, Holland, Ke, Larsen-Freeman and Schoenemann2009; Diessel Reference Diessel2019; Schmid Reference Schmid2020).

Historically, CxG has placed a greater emphasis on taxonomic relationships between nodes, as evidenced for example by the vast amount of research on alternations (Goldberg Reference Goldberg1995; Croft Reference Croft2001; Croft & Cruse Reference Croft and Cruse2004; Perek Reference Perek2015; for a discussion about the state of the art on horizontal/vertical links, see Ungerer Reference Ungerer2024). Empirical studies of language use have highlighted the probabilistic nature of grammar. The growing evidence that speakers are sensitive to statistical information has raised the question as to where this information is stored. Assuming that usage information is chiefly a property of constructions leads to a ‘fat node’ problem (Hilpert & Flach Reference Hilpert, Flach, Depraetere, Hilpert, Cappelle, Denis, Dehouck, Flach and Grabar2023). For example, a modal auxiliary is followed by a verb or an adverb, all of which are separate nodes. But some combinations, such as the idioms may well or would rather are considered one node, that is, one form–meaning pair. Thus, it is far from clear at which point the decreasing idiomaticity from may well to could possibly to could likely stops being a construction and becomes a collocation of two constructions. In other words, it is unclear how much and what of the statistical bond should be considered a property of the individual nodes (Hilpert Reference Hilpert2016; Flach Reference Flach2021a).

A potential solution to this problem is to assume that probabilistic information is a property of the links between the nodes, which can vary in number, degree and strength (Diessel Reference Diessel2019; Schmid Reference Schmid2020).Footnote ¹ Elements are more or less closely related and thus more or less strongly integrated in the network: may and well in the idiom may well, undoubtedly a construction under any CxG view, share a stronger bond than the collocation of could and possibly. The link-focus models statistical information as link strength between co-occurring items, which are often far more gradual than assumed (Wulff Reference Wulff2008; Flach Reference Flach2021a). More generally, different communicative pressures will activate different links or subparts of the network and make the combination of elements more or less likely. It also accounts for individual differences, as speakers’ constructicons will vary along finer details.

Under this view, creativity and productivity are facilitated and constrained by network relationships and the complex interplay of link existence, link strength and the potential to establish and modulate connections. Independent of their status as constructions in the CxG sense, links range from ‘unbreakable’ connections between units in cases of ‘obligatoriness’ (e.g. present third-person singular -s) or fossilized expressions (e.g. kith and kin), to ‘creative’ language use that encompasses the vast majority of language use, to rule-breaking errors beyond the conventional (or ‘system’).

A widely discussed classification carves up this creativity continuum into F-creativity (‘fixed’) and E-creativity (‘expanding’) (Sampson Reference Sampson and Hinton2016). While F-creative language use involves ‘established material’ and encompasses all cases that create novel and/or unheard but grammatical sentences in the Chomskyan sense (Chomsky Reference Chomsky1965), E-creativity encompasses the ‘ungrammatical’ or ‘rule-breaking’ use that is ‘outside the system’, but which potentially expands the inventory of constructions (Bergs Reference Bergs2018; Herbst Reference Herbst2018; Hoffmann Reference Hoffmann2018; Uhrig Reference Uhrig2018; Bergs & Kompa Reference Bergs and Kompa2020).

Most of the discussion focuses on the border between F-creativity and E-creativity: which constructional combinations fall on which side of that border under which conditions? Is the semantically and syntactically incompatible or incoherent (Goldberg Reference Goldberg1995: 50) use of intransitive sneeze in He sneezed the foam off the cappuccino F-creative because it combines ‘established material’ that is simply ‘unmined potential’ (De Smet Reference De Smet2018: 330)? Or is it E-creative because this or similar combinations are ‘not found in the constructicon’ of most speakers (Uhrig Reference Uhrig2018: 297)? Even though they use ‘established material’, constructions such as Because reasons were E-creative at some point, but are now F-creative as they have become part of many speakers’ constructicons as schematic slot-filler constructions (e.g. because X; see Bergs & Pentrel Reference Bergs and Pentrel2025).

Hence, how would we decide which violations of selectional restrictions qualify as which type of creativity, assuming we can identify selectional restrictions reliably in the first place? Even if sneeze-cappuccino is F-creative, its use also expands the constructicon for people who have never used or heard it before. A common solution is to postulate additional categories – along the lines of ‘marginally E-creative’ – to increase descriptive accuracy in order to account for cases that cannot clearly be put on either side of the fence. While analytically helpful, postulating additional categories creates additional borders with additional sides.

The problem may well be that the distinction between F- and E-creativity, as analytically helpful a crutch as it may be, dichotomizes a phenomenon that is fundamentally gradient and probabilistic (an analogous argument can be made for terms like ‘rules-compliant’ and ‘rule-breaking’). That said – and unsurprising from a usage-based perspective – the gradient nature of the creativity scale is acknowledged by all authors in one form or another (Bergs Reference Bergs2018; De Smet Reference De Smet2018; Uhrig Reference Uhrig2018; see also Laws Reference Laws2025 and references therein). The perceived need for categorical distinctions most likely stems from the fact that neither the creativity scale, nor the amount, nor the type of data that falls onto various sections of the scale is linear and/or evenly distributed. After all, there is a noticeable qualitative difference between what is prototypically F-creative and what is prototypically E-creative.

From a link-based perspective, the distinction may be overemphasized. Both F- and E-creativity are facilitated and constrained by network relations and/or the absence thereof. Many utterances with ‘rule-breaking’ violations of selectional restrictions involve ‘established material’, just in novel or unconventional ways, for a multitude of socio-pragmatic reasons (Uhrig Reference Uhrig2018; Bergs & Pentrel Reference Bergs and Pentrel2025), and the judgement regarding their acceptability is subjective and context-dependent. But any novel combination, whether F- or E-creative, creates new links or reinforces (weak) existing links. For example, most analyses will accept He sneezed the foam off the cappuccino as F-creative, because sneeze has a semantic compatibility with – or lacks strong statistical pre-emption from – the caused-motion construction. By contrast, explain is so strongly associated with the to-dative which adversely affects its potential to establish and maintain a relationship with the ditransitive (Goldberg Reference Goldberg2019: 3). But whether He sneezed the foam off the cappuccino or She explained him the problem are judged F- or E-creative does not change the fact that both verbs have the potential, albeit with vastly different probabilities, to create new links, however weak, to other constructional hosts. It is likely that every utterance ever so slightly alters the structure of individual or communal constructicon(s), and nodes generally do not spring into existence. What is worth noting is that in both cases, a ‘creativity effect’ results because the uses go against speaker expectations of established norms and conventions (Uhrig Reference Uhrig2018: 298). In link-terminology, the ‘creativity effect’ arises because no or only weak links have previously been established between the construction and its (potential) slot-filler(s). In other words, it is crucially not just a new node that expands the constructicon, but also newly created links or links with altered weights.

It is no surprise that examples with a creativity effect are at the centre of the creativity literature. An interesting question then is which properties make link creation and reinforcement more or less likely across the full ‘creativity range’, and not only at the border between F- and E-creativity. To investigate this question, we will attempt to measure link relations in the network and connect the creativity potential of slot-fillers with the productivity potential of the into-causative. Specifically, we ask whether a verb’s link potential in general predicts its likelihood of occurring in a construction with which it is not primarily associated. To this end, we make a distinction arguing that productivity pertains primarily to the number of (potential) links between nodes, and that creativity pertains primarily to processes of creating, maintaining and reinforcing these links.

2.2. The into-causative

The vast majority of uses of argument structure constructions that lie on the ‘ordinary’ F-creativity section of the cline are rarely contentious. Although no exception, the into-causative is rather special among the argument structure constructions as many matrix verbs have other preferred semantic and syntactic environments. Following from the discussion above, we assume that the construction is ‘creative’ in the sense that, with very few exceptions, it is not a typical construction for any of the verbs that occur in it, and hence establishes and maintains a number of links with those items when they do occur in it. In this sense, the into-causative is creative in the ‘creativity without extravagance’ sense (Trousdale & Norde Reference Trousdale and Norde2025).

The into-causative encodes events where an animate causee is compelled by a cause(r) to perform the action encoded in the oblique, while the matrix verb specifies the manner of causation. In traditional CxG notation, the form [SUBJ V OBJ OBL _{into V-ing} ] is paired with the meaning ‘X causes Y to do Z by means of V’, illustrated by examples from the Corpus of Contemporary American English (COCA):

As an argument structure construction that has been around for centuries (Rudanko Reference Rudanko2011; Flach Reference Flach, Sommerer and Smirnova2020, 2021b), the into-causative is at the less creative end in terms of ‘extravagance’ or ‘aberrancy’, and is thus not the first pattern one would think of in the context of creativity. Yet ‘creativity’ is a recurrent topic, with some authors asking what, if anything, constrains speaker creativity and innovation (Hunston & Francis Reference Hunston and Francis2000; Rudanko Reference Rudanko2005; Kim & Davies Reference Kim and Davies2016; Rickman & Kaunisto Reference Rickman, Kaunisto, Kaunisto, Höglund and Rickman2018).

Three things are important to keep in mind. First, it is not ‘anything goes’ – rather, the productivity is limited to a few verb classes (Gries & Stefanowitsch Reference Gries, Stefanowitsch, Achard and Kemmer2004a; Stefanowitsch Reference Stefanowitsch, Herbst, Schmid and Faulhaber2014; Flach Reference Flach2021b). The matrix verbs are mostly verbs of communication (e.g. talk, coax, charm, flatter), trickery (e.g. cheat, con, deceive, fool), force (e.g. pressure, torture, arm-twist) and fear (e.g. shame, badger, guilt, scare). The miscellaneous class contains manner-specific verbs that do not fit into the other classes (e.g. drive, desensitize, galvanize, gulf, huckster, catapult, legislate).

Second, the senses of the matrix verbs do not include ‘causation’, and the primary verbs of causation, make and cause, are notably absent (Stefanowitsch Reference Stefanowitsch, Herbst, Schmid and Faulhaber2014). This implies that the meaning of causation is contributed by the construction, which is in contrast to other argument structure constructions. For example, the majority of verbs in the ditransitive encode the constructional semantics of transfer, whether successful, intended or denied (e.g. give, offer, deny). That said, verbs from the force or fear classes imply an effect on the causee to act, while others, especially communication and miscellaneous verbs, either do not or only in individual cases (e.g. persuade more so than chit-chat).

Third, many matrix verbs are syntactically less compatible with the construction. This is either because they prefer other environments or other word classes. The clearest example of the former is talk, whose primary use is complex intransitive (They talked to the manager) and which does not take animate objects outside the into-causative (*we talk him). The latter refers to the fact that many matrix verbs are not primarily used as verbs (e.g. gulf, school, back, careful, mousetrap), which is the reason the matrix verbs are referred to here as V1. In other words, the into-causative facilitates conversion-prone V1s.

The frequent reference to the construction’s creativity in the literature stems from two sources. On the one hand, the number of hapaxes especially in the miscellaneous category do characterize the construction as productive in traditional corpus linguistic terms. On the other hand, the impression of creativity is primarily due to the creative effect that arises when semantically and/or syntactically incoherent verbs are used in a construction with which they are not, or not predominantly, associated with. In other words, the into-causative goes against speaker expectation (Uhrig Reference Uhrig2018) with a substantial proportion of its V1s. So regardless of whether the into-causative is quantitatively more productive than other argument structure constructions, it certainly appears to be qualitatively more creative.

By traditional corpus linguistic measures, the into-causative is not even particularly productive (in addition to the mathematical problems of productivity measures on constructions of different sizes from different sources; see Barðdal Reference Barðdal2008; Perek Reference Perek2016; Hilpert Reference Hilpert, Engelberg, Lobin, Steyer and Wolfer2018; for alternative approaches, see Säily et al. Reference Säily, Perek and Suomela2025). The type–token ratio is low (ttr = 0.06) and the hapax–type ratio below the average of 0.5 (htr = 0.44). It may, however, be more productive than related constructions from the perspective of the slot-fillers, which we will discuss in section 3.4. It is our main assumption that it is not (primarily) the into-causative that is productive, but that the creativity potential of the V1s contribute considerably to its productivity potential.

In sum, regardless of how high or low the construction scores on traditional measures of productivity, the into-causative has a number of properties that make it suitable for an attempt to relate a V1’s creativity potential with its propensity to occur in the construction: the into-causative attracts a varied number of V1s that give rise to a consistent ‘creative effect’. This effect is due to the fact that the into-causative is not the primary syntactic or semantic environment for most V1s. The expectation is that a V1’s statistical association with the into-causative is contingent on (a) its relative semantic and/or syntactic compatibility, i.e. being high in syntactic transitivity and able to express two-participant scenes, but crucially also on (b) its general greater flexibility with regard to its position in the constructicon, which makes it ‘easier’ to use in an unusual constructional environment. In link-view terminology, a V1 should have a greater association with the into-causative if it has general properties that allow it to establish and maintain links with other network elements. The next section discusses the operationalization of these properties.

3.1. Data sources

The association measures of V1s are based on 342 V1s in 6,215 observations of the into-causatives in the 2015 offline version of COCA (Davies Reference Davies2008–). The V1s’ creativity potential was determined from samples of their uses outside the into-causative. A few V1s were excluded, either due to their syntactic status as well as their high corpus frequency, but low construction frequency (e.g. get, will), or because they are too infrequent to occur in enough full sentences in COCA to reliably determine the metrics discussed below (sign-talk and sphroxify). Additional metrics were taken from the British Lexicon Project (Brysbaert & Biemiller Reference Brysbaert and Biemiller2017; Keuleers et al. Reference Keuleers, Lacey, Rastle and Brysbaert2012). To compare the slot-filler creativity with that of other argument structure constructions, we used data from Gries & Stefanowitsch (Reference Gries and Stefanowitsch2004b) on the ditransitive and the to-dative.

3.2. Measuring V1 association with the construction

A V1’s association with the into-causative is expressed in the log odds ratio (OR) from a Collostructional Analysis (Stefanowitsch & Gries Reference Stefanowitsch and Gries2003; Flach Reference Flach2021c). As recently discussed (Schmid & Küchenhoff Reference Schmid and Küchenhoff2013; Gries Reference Gries2019; Reference Gries2022), OR is preferable over traditional measures like log likelihood (G ²). G ² conflates frequency and association, as shown in figure 1, which plots the relationship between construction frequency and G ² (top panel) and OR (bottom panel). G ² is near-perfectly predicted by construction frequency (R ²_gam = 0.91), but only weakly by OR (R ²_gam = 0.14).

Figure 1. Association and frequency in the into-causative

However, as the size and colour of the dots show, OR is strongly negatively correlated with corpus frequency (R ²_gam = 0.70). This is mathematically due to the fact that OR overestimates rare events. It also partially reflects constructional semantics: the V1 encodes manner of causation, and more specific words also tend to be less frequent (e.g. the type of talking in jive-talk or chit-chat). Thus, in order to avoid identifying factors that predict association merely as a side-effect of corpus frequency, corpus frequency is included as a control variable.

We can quantify the major assumption that the into-causative is an unusual syntactic environment for (almost) all V1s. If we set a somewhat arbitrary threshold at 10 per cent of a V1’s uses in the into-causative for the construction to count as ‘typical’ for that V1, then the into-causative is a typical construction for only 24 of the 342 V1s (7 per cent). By comparison, based on the data from Gries & Stefanowitsch (Reference Gries and Stefanowitsch2004b), 28 per cent of types in the ditransitive and 51 per cent of types in the to-dative have these two constructions as a typical host. Figure 2 illustrates this: it plots ranked lists of the slot-fillers’ ΔP: a higher ΔP indicates that a greater share of that verb occurs in the construction. For the into-causative, the curve drops sharply, as only a handful of (very rare) V1s happen to occur in the into-causative when used as a verb (e.g. careful, dropkick, sign-talk, slick-talk, sphroxify). The following flat line for most V1s indicates that the into-causative is not typical for most of its slot-fillers. This can be compared to the shape of the curves for the to-dative and the ditransitive, where both the drop and the long tail are substantially less pronounced. While traditional measures of productivity do not identify the into-causative as particularly productive, the picture changes when we shift the perspective towards the slot-fillers: the into-causative is productive in the sense that it maintains a relatively greater number of connections with potential slot-fillers, because it hosts elements from a relatively larger pool of candidates.Footnote ² The next section outlines how we capture the properties of a V1’s network integration and how these properties are expected to affect a V1’s association with an unusual context that gives rise to the creativity effect.

Figure 2. Comparison of ΔP (word to construction) of three argument structure constructions

3.3. Measuring V1 network integration

To measure the network integration of V1s, the strings of all inflectional forms of 338 V1s were queried in COCA with no restriction on part-of-speech. From these concordances, we created 25 batches, each containing a maximum of 100 complete, randomly selected sentences per V1. For 317 V1s (94 per cent), 100 complete sentences could always be extracted, resulting in about 33,000 sentences per batch. The ~33,000 sentences of each batch were then dependency-parsed using spaCy’s transformer model for English (en_core_web_trf). A dependency parser attaches morpho-syntactic features and syntactic relations to each token. The dependency relation labels and their areas are based on established universal dependencies used in most dependency parsers.Footnote ³

To illustrate, spaCy returns the output in table 1 for the sentences We talked about the project, They were ambushed by the police and We thought about buying them a present. The head column indicates the tok_id of the word that a given token is a dependent to. In these examples, talk and ambush occur as the root (‘anchor’) of their sentence, and both have prepositional objects as dependents (about the project, by the police); ambush has a passive auxiliary as a dependent, from which we can determine voice. Non-root buy occurs in a prepositional complement to thought and heads two NP objects (them, a present).

Table 1. Sample output of the spaCy dependency parser

Transitivity is determined from the dependents of the V1. While a dependency parser depends on definitions about objects, complements, or adjuncts that not all syntactic theories agree on, the parsing information nevertheless allows us to approximate transitivity profiles based on general definitions that are used across syntactic descriptions (Goldberg Reference Goldberg1995; Huddleston & Pullum et al. Reference Huddleston and Pullum2002: 52–4). A custom script classified a V1 as intransitive if the V1 has neither NP nor PP dependents (They talked.), as transitive if it has one NP and no PP dependent (They told them.), as complex intransitive if it only has a PP dependent (They talked to each other ), as transitive complex if it has both NP and PP dependents (They told them about the project ), and as ditransitive if it has two NP dependents (They forwarded their kids the parcel.). This classification produces some inaccuracies, since parsers poorly distinguish between complements and adjuncts (which is also far from clear-cut for human annotators). We accept the inaccuracies as noise, as it is not feasible to hand-code 800,000+ concordances. The assumption is that the sizes of the data sets and the sampling procedures return robust patterns for the purpose of the current analysis.

We can calculate three types of metrics from cross-tabulating the parser output, illustrated for ten verbs and their word class uses (table 2). The first type of metric is simple ratios: for example, in batch 1, abuse occurs as a noun in 81 of 100 concordances (Noun_Ratio: 0.81). Of its 19 verb uses, 5 occur in the passive (Passive_Ratio: 0.23). The second type are Relative Entropies (Hrel), which measure the distribution across a given category, where higher values indicate a more even distribution across the values of that category. For example, alarm (0.38) is more evenly distributed across three word classes than argue (0.0). Relative Entropies are non-directional: they return quantitative information about a distribution, but cannot provide qualitative information about which end of the scale a distribution is skewed towards. Compare agitate and armtwist: they have comparable WordClass_Hrel values, but while agitate gravitates towards the adjectival end, armtwist is considerably more ‘noun-y’.Footnote ⁴

Table 2. Example summaries and metrics for ten V1s from batch 1

Thus, the third type of metric that captures the qualitative magnitude of a distribution, comes from Correspondence Analysis (Greenacre Reference Greenacre2017; Nenadić & Greenacre Reference Nenadić and Greenacre2007). CA is a dimension-reduction technique for categorical data in large contingency tables. While CA is mathematically complicated, the idea is straightforward: it quantifies (and visualizes) the relationship between two variables. We illustrate the general idea using the distribution across transitivity (table 3). What CA adds over Relative Entropies (column Transitivity_Hrel) is that it calculates a value for each V1 based on its relative (dis)similarity to all other V1s (column Dim1).

Table 3. Distribution across transitivity for ten V1s (grey predictors in analysis)

Let us briefly discuss the general idea of CA, since we use it for multiple metrics. Table 3 shows the argument structure profiles for ten V1s with very different syntactic preferences. For example, starve is primarily used intransitively, while wish is primarily used in complex transitive environments. Figure 3 shows the corresponding CA biplot which is the reduction of multiple dimension (columns) to a 2D representation. For illustration purposes, the plot only shows the fifty V1s with the greatest contribution to the CA (‘distinctiveness’); all other V1s occupy the (empty) empty space near (0,0). In essence, CA biplots visualize similarities and show V1s in their ‘typical’ constructional ecology (but note that (absolute) distances between row and column labels are not meaningful). Important for the current purpose is the continuum from top to bottom along the y-axis, running from less to more complex environments. Thus, we take the Dim2 values of each V1, i.e. their position on that complexity cline, as a measure of their transitivity ecology. (While the Dim2 values correspond to the values on the y-axis, they are not identical, since the plot is a distorted 2D representation of a multidimensional data structure; the labels are also plotted such that they do not overlap.)

Figure 3. Biplot of a Correspondence Analysis of transitivity behaviour

To make the interpretation of Transitivity_CA easier, Dim1 is multiplied by -1, so that V1s which are ‘higher in transitivity’ have a positive value.Footnote ⁵ The grey columns in table 3 illustrate that Hrel and CA measure different things: a V1 can be low in transitivity, yet relatively evenly used in various argument structure constructions (e.g. starve), or it can be high in transitivity, but limited to relatively fewer argument structure constructions (e.g. persuade). Quantifying continua in this way is a useful property of CA that we use for a number of metrics (if a continuum exists in the data).

The following metrics were calculated based on cross-tabulations of the parser output, and are summarized in table 4 (we will discuss the expected effects in the next section). Passive_Ratio captures syntactic and semantic compatibility with the into-causative, which has a high passive ratio (~24 per cent): a higher propensity for passivization mirrors a V1’s ability to express multi-agent scenes. Noun_Ratio captures the construction’s attraction to manner-specific lexemes via conversion, and a V1 that can easily be used in several classes would have a higher compatibility with the construction. Root_Ratio measures how often a V1 is the root of the sentence in which it occurs: a higher ratio indicates that the V1 is more commonly used in ‘core’ parts of a sentence, i.e. arguably in a tighter syntactic integration.

Table 4. Summary of metrics and expected effects (italics for cautious expectations)

In a similar vein, we calculated two metrics to capture syntactic relationships. First, Deprel_Hrel measures a V1’s spread across types of dependency relations and Deprel_CA captures a V1’s position along a cline from tighter (‘root’) to looser integration (‘non_core_dep’). In addition, Dist_to_Head measures the (absolute) average distance of a V1 to its head.

Finally, a set of metrics captures general word properties. Genre_Hrel measures the spread across five genres in COCA (academic, news, fiction, magazine and spoken). Orality_CA is its directional equivalent, where a higher value indicates a higher association with spoken/oral discourse. From the British Lexicon Project (Brysbaert & Biemiller Reference Brysbaert and Biemiller2017), Age_of_Acq indicates the average age of acquisition of a V1. In addition, the corpus frequency of a V1 is used as a control for the frequency effect illustrated in figure 1, expressed as a log-scale Zipf value (Brysbaert, Mandera & Keuleers Reference Brysbaert, Mandera and Keuleers2018).

Before we discuss assumptions and hypotheses, it should be noted that the metrics for the analysis are only a subset of several possible ones, many of which had to be discarded due to collinearity. For example, metrics derived from morphological form, such as tense or finiteness, are correlated with metrics derived from dependency relations, because a root use of a verb will often be in a tensed form. Likely due to form syncretism, the respective CAs did not produce interpretable continua for finiteness/tense. A metric was discarded if it was positively or negatively correlated (r > |0.5|) with a mathematically simpler or conceptually more appropriate one.

3.4. Assumptions and hypotheses

Given the exploratory nature of this analysis, there are few prior hypotheses. Nevertheless, we would expect the following patterns based on the assumption that the metrics at least approximate relevant dimensions of network integration such as compatibility, flexibility, or rigidity: a V1 that scores higher on flexibility and compatibility metrics should also have a higher association with the into-causative.

In general, the metrics fall into three (overlapping) categories: those that capture (i) core verb properties (Passive_Ratio, Root_Ratio, Transitivity_Hrel, Transitivity_CA), (ii) a V1’s relationships with other elements in the network (Deprel_Hrel, Deprel_CA, Dist_to_Head) and (iii) lemma-based general properties (WordClass_Hrel, Noun_Ratio, Genre_Hrel, Orality_CA, Age_of_Acq). Another way of looking at them is in terms of measuring semantic coherence (passivization or transitivity), syntactic flexibility (most Hrel metrics) and syntactic integration (metrics from dependency relations).

For some metrics it is easier to formulate expectations (see table 4), because they straightforwardly capture syntactic and/or semantic compatibility of a V1 with the into-causative. For example, a higher Passive_Ratio indicates that a V1 can readily encode two-participant scenes. A higher Transitivity_CA indicates greater semantic and syntactic compatibility with a complex transitive into-causative. Both their effects should be positive.

Some metrics are less clear. One type is less clear because of the non-directionality. While high values of Deprel_Hrel and Transitivity_Hrel indicate a greater flexibility to occur with different types of dependency relations or arguments, low values are difficult to interpret, as it is unclear which qualitative end of a range the V1 is skewed towards. Crucially, we argue that an even distribution in general indicates greater general flexibility and should hence predict greater constructional association.

The second type of metric that is less clear includes Deprel_CA and Root_Ratio, because we have no prior hypothesis regarding the effect of greater ‘root-y-ness’. That said, a higher propensity to systematically act as the verbal head – as opposed to being a dependent of another element – could reflect a more rigid, more central network position. In other words, greater syntactic rigidity could indicate that a V1 is less readily available to occur in syntactic environments with which it is not primarily associated with. Following this logic, we expect a negative relationship of Root_Ratio and Deprel_CA with constructional association.

Other metrics where it is more difficult to formulate expectations are Noun_Ratio and WordClass_Hrel. On the one hand, the into-causative is well suited for manner-specific lexemes and conversions, but whether that holds in general is an open question. In line with the previous arguments, the assumption is that a higher non-verb ratio and a greater flexibility to occur in various morpho-syntactic environments indicate a greater flexibility on the part of the V1 – the effect should thus be positive. Although there is no prior assumption about the effect of Distance_to_Head, but again in line with the arguments above, a greater average distance to one’s head might indicate greater flexibility and a looser integration with other network units, thus the effect should be positive as well.

Finally, for the metrics that aim to capture general word properties, we expect a positive effect on constructional association for V1s that occur in more informal contexts, which is likely a spontaneous locus of creativity (Orality_CA). We remain uncommitted about the effect of age of acquisition (Age_of_Acq) and genre distribution (Genre_Hrel). Finally, Corp_Zipf is a control variable, and its effect on constructional association is known and negative (see figure 1 above).

4.1. Choice of regression technique

A common problem for analyses like the present one is that traditional regression methods can be unreliable or unstable for datasets with a large number of predictors that have no (strong) theory-driven hypothesis. In addition, and certainly in this case, many predictors are sample-dependent and noisy approximations. The latter is due to the complexity of language that is reduced to a few low-dimensional numerical values. These problems can lead to overfitting, meaning a model will perform poorly on unseen data. The problem is succinctly discussed in Van de Velde & Pijpops (Reference Van De Velde and Pijpops2021) for common applications in corpus linguistics, and we will follow their suggestion to use LASSO regression.

LASSO regression (Least Absolute Shrinkage and Selection Operator) is similar to the more widely used OLS regression (Ordinary Least Squares), but can protect against overfitting (James et al. Reference James, Witten, Hastie and Tibshirani2013; Van De Velde & Pijpops Reference Van De Velde and Pijpops2021). Unlike OLS regression which outputs a number of coefficients and p-values, LASSO performs k-fold cross-validation that iteratively (‘k-times’) splits the data into training and test data. It then adds a tuning parameter lambda, which shrinks (‘penalizes’) a set of coefficients towards zero. While this procedure makes the model fit worse, it also makes it more robust to predict unseen data (the test data). Predictors that contribute little or no improvement to the model fit will have zero or near-zero coefficients. The interpretation of the LASSO output is similar to OLS, but unlike (many implementations of) OLS, LASSO does not return p-values. Rather, it places emphasis on how much a predictor ‘contributes’ to the overall model fit and/or whether the predictor was retained in the final model (i.e. has a non-zero coefficient). This means that LASSO aims to identify variables that consistently add their share (or not) to explaining variance in the response variable. One disadvantage of LASSO is that, if predictors are correlated, it randomly selects one predictor over the others. This problem is minimal in the current data, as strongly correlated predictors were excluded beforehand.Footnote ⁶

We used the R package {glmnet} v4.1.8 (Friedman, Hastie & Tibshirani Reference Friedman, Hastie and Tibshirani2010) with the defaults for k-fold cross-validation and lambda determination. All predictors were standardized and scaled. We performed LASSOs on all twenty-five batches of 100 parsed sentences of V1. During each LASSO, we fetched each predictor’s coefficient or marked a predictor as ‘dropped’ if it was not retained.

4.2. Results and interpretation

The ridge and raincloud plot in figure 4 shows the output of twenty-five LASSO regressions. Each dot represents the coefficient from a batch in which the predictor was retained. Predictors are sorted from top to bottom in ascending order of the mean of their coefficients. The colour represents how often a predictor was dropped. The interpretation of coefficients is similar to OLS: a positive coefficient means that higher values of that predictor are correlated with higher values of the response variable, while a negative coefficient means that higher values of that predictor are correlated with lower values of the response variable. The predictor Corp_Zipf, which has by far the largest coefficients (mn = −0.76, sd = 0.02), is not shown to preserve the readability of the plot. The average R ² across twenty-five LASSOs was R ² = 0.66 (sd = 0.01).

Figure 4. Output of LASSO regression of twenty-five samples

It should be borne in mind that, because of random splits into training and test sets and k-times repetitions, every run of every of the twenty-five LASSOs will look slightly different, and the number of times each predictor is dropped will vary, but the overall pattern is stable. In general, we interpret the fact that almost all predictors were dropped at least once as ‘overall, there is a consistent signal between V1 properties and their association with the into-causative, although it might be weak’. Similarly, a higher number of retentions indicates that the predictor has a greater (relative) stability in making a genuine contribution. In this spirit, Dist_to_Head and WordClass_Hrel have a marginal effect at best. All other predictors hint at a more reliable, genuine pattern. For example, Deprel_CA and Root_Ratio have negative coefficients, whereas Transitivity_CA, Passive_Ratio, and Deprel_Hrel have positive coefficients.

The results can be interpreted as follows: higher uses as the root of a sentence (Root_Ratio) and greater ‘root-y-ness’ (Deprel_CA) correlate with a lower association of the V1 with the into-causative. Conversely, a higher passive use (Passive_Ratio) is strongly, and a higher level of noun uses (Noun_Ratio) is weakly correlated with constructional association. A higher transitivity of a V1 (Transitivity_CA) is also correlated with greater association. As discussed above, due to its non-directionality, the interpretation of the distribution across different levels of transitivity (Transitivity_Hrel) is less straightforward: while a greater flexibility across the range of types of transitivity is plausibly connected with a greater flexibility to occur in a syntactically quirky construction, the reverse does not hold for low values, which could indicate a strong skew towards any section of the intransitive–complex-transitive scale. Transitivity_Hrel might be a poor operationalization for the underlying concept. On the extralinguistic level, higher flexibility across genres (Genre_Hrel), a later age of acquisition (Age_of_Acq), and greater orality (Orality_CA) are positively correlated with constructional association.

Let us return to the expectations in table 4. With the exception of Transitivity_Hrel, the expectations were borne out for predictors for which we could formulate clear expectations. For predictors with less clear predictions, one turned out to not hold (WordClass_Hrel) and two turned out to be marginal (Dist_to_Head, Noun_Ratio). Overall, however, the results collectively show a clearer picture: flexibility is linked to greater, and rigidity is linked to lower constructional association.