Extending structural priming to test constructional relations: Some comments and suggestions

2.1 Why use structural priming?

The main reason why priming effects are regarded as an important source of evidence for constructional relations is that they shed light on how patterns of activation and information exchange unfold within speakers’ mental networks during individual usage events. Under a spreading activation account of language processing (Collins and Loftus 1975), priming effects can be explained in the following way: on encountering a linguistic stimulus, speakers not only activate their mental representation for the specific instance, but activation also spreads to neighbouring network units – which can either form part of the same abstract construction as the original stimulus, or instantiate other related constructions. As a result of this residual activation, speakers’ subsequent processing of the same or similar constructions is facilitated (or, depending on the process, potentially inhibited, see Section 3.3).

In Bock’s (1986) original structural priming study, for example, participants were more likely to produce double-object sentences after previous exposure to a double-object prime like (1a) than after a to-dative prime like (1b). Similarly, participants produced more actives after active primes like (2a) than after passive primes such as (2b). This suggests that speakers represent different instances of the same construction as interrelated in their mental networks; these network links thus encode generalisations that correspond to the notion of an abstract constructional schema.

The use of structural priming to test constructional relations is also supported by evidence that priming effects are sensitive to both formal-syntactic and functional-semantic similarities between grammatical patterns. While early studies indicated that structural priming effects are primarily syntactic, reflecting the shared constituent structure of the stimuli (Bock and Loebell 1990; but see Ziegler et al. [2019] for potential counter-evidence), more recent investigations have suggested that priming effects can also arise from overlap in thematic roles (Hare and Goldberg 1999), event structure (Bunger et al. 2013), information structure (Bernolet et al. 2009) and animacy (Vasilyeva and Gámez 2015; see Ungerer [2021] for further examples and discussion). Moreover, Ziegler et al. (2018) have argued that priming effects may be additive, with stronger priming effects emerging between sentences that share several formal and/or functional properties than between sentences that overlap only on a single dimension. This aligns naturally with a cognitive-linguistic view of constructional relations as multi-dimensional and gradient in strength (Goldberg 2019; Trijp 2020).^[2]

2.2 Advantages over alternative methods

Structural priming has several advantages that set it apart from other experimental methods that have been used to test constructional relations (see also Ungerer [2022] for a detailed discussion). First, priming experiments provide implicit measures of processes that are assumed to be automatic, unconscious and resource-free (Branigan and Pickering 2017: 6). The paradigm thus avoids the challenges faced by explicit metalinguistic methods such as sorting tasks, in which participants are asked to group similar sentences into the same category (Bencini and Goldberg 2000; Gries and Wulff 2005; Perek 2012). As Perek (2012: 629) points out, these experiments leave open the question of whether speakers’ sorting behaviour reflects their implicit grammatical knowledge or the ad-hoc categories that they may consciously form in response to the specific task.

A second advantage is that structural priming can, in principle, be applied to any kind of construction and tested in any type of speaker population. Priming effects have been observed in children as well as adults, in L1 speakers as well as non-L1 and bilingual speakers, and in neurotypical speakers as well as participants with aphasia, autism spectrum disorder and specific language impairment (see Branigan and Pickering [2017] for discussion). Other methods, in contrast, are often restricted to particular phenomena or participant groups. For instance, given the metalinguistic nature of the above sorting tasks, these techniques may be difficult to use with younger children or speakers that have limited working memory resources. Other methods, such as acquisition studies that investigate in what order the members of constructional families are acquired (Diessel and Tomasello 2005), are naturally restricted to children (or possibly L2 learners). Similarly, artificial learning experiments (Perek and Goldberg 2015, 2017) have been used to test how speakers generalise across novel clause-level constructions that contain nonce verbs (Tomasello and Brooks 1998); but clearly, these methods cannot be used to investigate already existing constructions in natural languages.

Finally, priming reflects how the activation patterns in speakers’ mental networks are affected by individual linguistic stimuli in a way that seems difficult to demonstrate with other methods. In acquisition experiments (both natural and artificial), for instance, the learning effects accrue over longer periods of time and researchers can only observe the overall outcome of those cumulative processes. In priming experiments, on the other hand, a single exposure to a prime can already have an observable effect on speakers’ behaviour. Moreover, by studying priming both in production and comprehension, researchers can not only examine the “outcomes” of individual activation processes (via speakers’ production rates), but they can also track their exact time course by obtaining “online” measures of speakers’ real-time processing (such as reading times in comprehension; see Section 3.3).

2.3 Two limitations of previous priming research

In the light of these advantages, it is not surprising that structural priming is considered an important tool for investigating constructional relations. Nevertheless, as mentioned above, previous priming studies also display two important limitations that restrict the extent to which these findings can, to date, inform cognitive-linguistic models of constructional networks. These limitations will be outlined in more detail in the following; together, they motivate the suggested extensions of the priming paradigm discussed in Section 3.

The first limitation is that previous structural priming studies have focused on a relatively small set of constructions, while other types of patterns have been largely disregarded. In particular, most studies have so far tested constructional alternations, i.e., pairs of constructions that vary in their syntactic form but express near-equivalent meanings. In the domain of argument structure constructions, for instance, the by far most commonly studied phenomenon is the dative alternation, consisting of the double-object construction and the to-dative (see Bock’s [1986] classic study described in Section 2.1). In Mahowald et al.’s (2016) meta-analysis of 69 production priming studies, 63% out of all reviewed experimental conditions instantiate the dative constructions (with actives/passives making up another 22% of the sample). Other argument structure constructions that have been (though less frequently) investigated are the members of the benefactive alternation (Bock 1989), the locative (or spray/load) alternation (Chang et al. 2003) and the “fulfilling” alternation (Ziegler and Snedeker 2018); see below for examples.

Relying on a relatively small set of target constructions is arguably unproblematic if the goal of the investigation is to illustrate general processing mechanisms that should (in theory) apply equally to all constructions. This has indeed been the purpose of many structural priming studies, which have, for instance, compared production processes with comprehension processes (Bock et al. 2007) and spoken with written production (Cleland and Pickering 2006), or assessed individual processing differences due to age (Hardy et al. 2020) and language disorders (Cho-Reyes et al. 2016). In these cases, reusing the same constructions across different studies is not only convenient but also makes the experiments more comparable. The situation is different, however, when the research goal is to investigate speakers’ representations of specific constructions, and the relations between those constructions. Especially within the context of cognitive-linguistic network models, it is strongly desirable to obtain priming data for a large and varied set of constructions, beyond the relatively narrow scope of alternations. Not only are constructional relations expected to differ depending on the phenomena involved, but the eventual goal of the research programme must also be to map out speakers’ mental networks in their entirety, rather than restricting the analysis to smaller sections of these networks.

To understand why previous priming studies have favoured alternating constructions, it is necessary to take a closer look at their methodologies. The majority of structural priming studies to date have tested priming in production, with the two most popular methods being the picture description task and the sentence completion task (Mahowald et al. 2016). As suggested by Branigan and Pickering (2017: 7), these production methods usually “rely on the existence of structural alternatives”. In a picture description task, for instance, participants need to have two (roughly) equally felicitous constructions at their disposal with which they can describe the target picture (e.g., the double-object and the to-dative construction in the case of ‘transfer of possession’ events). This technique therefore lends itself naturally to the investigation of (near-synonymous) alternating constructions. At the same time, these methodological considerations already hint at the fact that researchers may be able to extend priming to a wider range of constructions by using alternative techniques, especially by testing priming in comprehension (see Section 3.3). Moreover, while production methods may generally be biased towards studying constructional alternations, it remains to be seen whether some production techniques can nevertheless be applied to specific cases of non-alternating constructions (see Section 3.2).

The second major limitation of previous structural priming studies – in the context of cognitive-linguistic network models – is that most previously observed priming effects reflect similarities between instances of the same construction (“within-construction priming”) rather than between instances of different constructions (“cross-constructional priming”). Consider again the results of Bock’s (1986) study described in Section 2.1: the fact that participants produced more double-object sentences after double-object primes than after to-dative primes suggests that they encode similarity relations between instances of the double-object construction. As such, however, the results do not reveal anything about the relationship between the double-object construction and the to-dative. This is a major limitation when it comes to testing cognitive-linguistic accounts of constructional networks, since the most interesting relationships in these models are arguably those that hold between distinct (but partially similar) constructions.

Compared with within-construction priming effects, investigations of cross-constructional priming, i.e., priming between instances of distinct constructions, have so far been relatively sparse (see Ungerer [2022] for an in-depth discussion of the relevant results). The most detailed study of such effects has been conducted by Ziegler and Snedeker (2018), who present several experiments that test priming effects between members of different argument structure alternations. The results of their Experiment 6, for example, indicate that speakers are more likely to describe a ‘transfer of possession’ event with the double-object construction after hearing a benefactive double-object prime like (3a) than after a for-dative prime like (3b). While the dative alternation and the benefactive alternation are often regarded as distinct phenomena (Levin 1993), Ziegler and Snedeker’s findings suggest that speakers encode strong similarity relations between the two double-object members and/or between the two prepositional members of each alternation (see Bock [1989]; Chang et al. [2003] for similar results).

Targeting another pair of alternations, Ziegler and Snedeker’s (2018) Experiment 7 provides evidence that “theme-second fulfilling” primes such as (4a) prime speakers’ double-object productions, compared with “theme-first fulfilling” primes like (4b). This cross-constructional effect was weaker, however, than within-construction priming between dative primes and dative targets, suggesting that speakers represent the fulfilling alternation and the dative alternation as partially similar but still distinct sets of constructions (see Hare and Goldberg [1999]; Cho-Reyes et al. [2016] for related results). Finally, Ziegler and Snedeker’s Experiment 4 illustrates that the absence of cross-constructional priming, too, can be instructive: the fact that the two members of the locative (or spray/load) alternation in (5) did not prime participants’ dative productions suggests that the locative and the dative alternation are not (or at best weakly) related in speakers’ constructional networks.

Together, these findings support the view that structural priming can provide relevant evidence about constructional relations. At the same time, Ziegler and Snedeker’s (2018) study of cross-constructional priming is a relatively rare case within a literature that has primarily focused on within-construction priming between instances of the same construction. This is evidenced, for example, by the fact that cross-constructional effects are excluded from Mahowald et al.’s (2016) meta-analysis of production priming studies, and that the distinction between within-construction and cross-constructional priming is not explicitly addressed in structural priming reviews (Branigan and Pickering 2017; Pickering and Ferreira 2008). In addition, while Ziegler and Snedeker (2018) investigate cross-constructional effects, they mostly restrict their analysis to priming between the members of constructional alternations, thus mirroring the first limitation discussed above (but see Section 3.1 for one exception among their experiments).

In summary, the preceding sections have addressed both the advantages of using structural priming to test constructional relations, and the limitations that constrain the ability of previous priming studies to inform large-scale models of constructional networks. In particular, the use of priming evidence in the context of cognitive-linguistic network models is limited by the fact that previous studies have focused primarily on alternating constructions and that they often target priming between instances of the same construction rather than priming between different but related constructions. To overcome these limitations, the crucial question is therefore how cross-constructional priming can be extended to a wider range of – particularly non-alternating – constructions. In the next section, three possible strategies that may enable such extensions will be discussed and illustrated.

3.1 Strategy 1: pairing non-alternating primes with alternating targets

As argued in Section 2.3, production priming methods such as the picture description task typically require that participants choose between two ways of encoding the same event, thus lending themselves to the investigation of alternating constructions. Strictly speaking, however, this constraint only applies to the target trials of the experiments, where participants are asked to describe the pictures. The prime sentences, on the other hand, are simply presented to participants (in either written or auditory form). As a result, the primes do not necessarily have to instantiate alternating constructions.

Based on this logic, a first and relatively simple strategy for partially extending priming to non-alternating constructions is to pair non-alternating primes with alternating targets. Such experiments involve at least three constructions, represented schematically in Figure 1. As the diagram shows, participants are presented with two alternating constructions A and B on prime trials and they choose between the same patterns A and B on target trials. In addition, however, they also encounter a third prime construction C, which does not participate in the alternation but which is nevertheless hypothesised to be related to one of the alternating targets (which it should therefore prime). The advantage of this approach is that researchers can continue to use well-established production methods like the picture description task; the difference from most previous applications of the method is that one of the prime constructions is a non-alternating pattern which is itself not available as a possible target.

Figure 1:

Schematic representation of Strategy 1, where a non-alternating prime construction is paired with alternating target constructions in a production setting.

The approach is illustrated by a few precursors in the literature, among them Bock and Loebell’s (1990) study. Using a picture description task, the authors tested to what extent speakers’ double-object and to-dative productions were primed by double-object sentences such as (6a), to-dative primes like (6b) and caused-motion sentences like (6c). Crucially, the latter construction does not alternate with either of the two dative constructions, since it expresses a distinct meaning of ‘change of location’ rather than ‘transfer of possession’. At the same time, the caused-motion construction may be related to the to-dative given the syntactic similarity of the two constructions and a potential metaphorical connection between their meanings (Goldberg 1995). Bock and Loebell’s caused-motion primes thus fulfil the role of construction C in the schematic scenario above, while the double-object and the to-dative construction correspond to constructions A and B. The crucial question in the study was whether caused-motion primes led to an increase in speakers’ to-dative productions, thus suggesting that the two constructions are related in speakers’ networks.

The emphasis here is primarily on the logic of Bock and Loebell’s (1990) design rather than on their specific results. With respect to the latter, the authors found that caused-motion primes led to an increase in participants’ to-dative productions, in fact to a similar degree as to-dative primes did, suggesting that speakers represent a strong similarity relation between the two constructions. These results, however, stand in direct contrast to Ziegler and Snedeker’s (2018) Experiment 11, in which they replicated Bock and Loebell’s design with some changes among the stimulus materials. In Ziegler and Snedeker’s experiment, no evidence emerged that caused-motion primes facilitated speakers’ subsequent to-dative productions, thus casting doubt on the strength of the similarity relation between the constructions (see Ziegler and Snedeker [2018: 231–232] for a discussion of possible confounds in Bock and Loebell’s materials, which may explain the discrepancies).

A suggestion for how future studies could extend this first strategy to other phenomena concerns the three constructions illustrated in (7), consisting of a double-object example in (7a), a to-dative in (7b) and what could be called an instance of the “predicative complement” construction (Himmelmann and Schultze-Berndt 2005) in (7c).

The predicative complement pattern corresponds to construction C in the schematic Figure 1: it does not alternate with the datives and in fact expresses a very distinct meaning (something along the lines of ‘property ascription’ rather than ‘transfer of possession’). Nevertheless, as highlighted by examples (7a) and (7c), the predicative complement construction contains the same syntactic constituents as the double-object construction (NP–V–NP–NP). The two patterns therefore constitute what has been referred to as “homonymous constructions” (Ellis 2008) or “homostructions” (Percillier 2020): similar to lexical homonyms, they display identical surface forms, while being semantically distinct and etymologically unrelated (see Percillier [2020: 220–221], who shows that the two constructions had differential case marking in Old English).

Testing cross-constructional priming between the two constructions could provide evidence of whether the purely form-based relations between homostructions have some cognitive reality within speakers’ constructional networks. This question is particularly relevant given that the overlap between the constructions is arguably quite ‘superficial’ and that the constructions differ in other of their form-related properties. Among the latter are the fact that the direct object of double-object sentences can be passivised while the predicative complement in the other construction cannot, and that the predicative complement slot can alternatively be filled with an adjective while this is not possible in the double-object construction.

In order to test the strength of the potential homostructional link, experimenters would need to compare the effects of the three prime constructions in (7) on speakers’ subsequent dative productions.^[3] If participants are sensitive to the formal similarities between homostructions, their double-object productions (relative to their to-dative productions) should increase after predicative complement primes, potentially at a similar rate as after double-object primes. On the other hand, the absence of such priming by predicative complement sentences would suggest that homostructional relations do not form part of speakers’ grammatical knowledge. On a more general level, this type of experiment may help address the recurring question in cognitive-linguistic research of whether speakers can form generalisations based on form only, or whether such “meaningless” schemas are cognitively implausible (Goldberg 2006: 166–182; Hilpert 2014: 50–57; Sommerer and Baumann 2021: 125–126).

3.2 Strategy 2: extending sentence completion to non-alternating constructions

A second strategy for extending structural priming to non-alternating constructions can be derived by taking another critical look at the view that production priming methods usually rely on the existence of structural alternatives (see Section 2.3). This assertion seems to be true in particular for picture description tasks, which specify a target event that can usually only be described with two alternating constructions. Other production methods, however, especially sentence completion tasks, may offer some more flexibility: in these experiments, participants are presented with target fragments that only partially specify an event. As a result, participants may sometimes complete these fragments with constructions that are not necessarily alternating variants of each other.

This is illustrated, for instance, by Gompel et al.’s (2012) study, who tested how intransitive primes like (8a) and monotransitive primes like (8b) affected participants’ responses to target fragments such as (8c). The authors found that participants were more likely to complete the targets with an intransitive structure after intransitive primes, and with a transitive structure after transitive primes.

Gompel et al.’s results as such are not central for the present context, given that they concern within-construction priming rather than cross-constructional priming. The main point here is that the study provides an example of a sentence completion experiment that tested participants’ processing of non-alternating constructions. This is because the intransitive and the monotransitive construction do not encode the same event; rather, the monotransitive construction extends the intransitive construction with an additional argument role (or, conversely, the intransitive construction can be seen as a “subpart” of the monotransitive; cf. Goldberg 1995). A schematic repre of the second proposed strategy could therefore look like Figure 2. Focusing on the boxes with the full black lines for the moment, and ignoring the greyed-out boxes, the diagram captures the fact that participants are presented with two non-alternating prime constructions A and B, the latter of which forms a larger extension of the former, and that they then choose between the same two constructions (i.e., non-extended pattern A vs. extended pattern B) when completing the target fragments.

Figure 2:

Schematic representation of Strategy 2, testing the effect of two or more primes on a target pair in which one construction forms a larger extension of the other.

In order to use this strategy to investigate cross-constructional relations, additional prime constructions could be added, marked by the grey boxes in Figure 2 (construction Bʹ and Bʺ). One potential area in which such a design may be feasible concerns adverbial complementation patterns. As noted in reference works (Herbst et al. 2004; Quirk et al. 1985) and also discussed in the cognitive-linguistic literature (Bergs 2021; Hoffmann 2007), adverbials vary along a cline of how closely associated with the verb, and therefore how “obligatory” they are in the sentence. One case of such variation is illustrated by the examples in (9), whose adverbials become increasingly more obligatory. In (9a), in the living room is an optional adverbial that can be dropped without rendering the sentence infelicitous; in (9b), in the cupboard can be regarded as semi-obligatory since the sentence would seem grammatical but semantically incomplete without it; while in (9c), in the drawer is a fully obligatory element without which the sentence would be ungrammatical.^[4] The obligatoriness of these adverbials plays an important role in the long-standing debate about their status as complements versus adjuncts (which, as a result of the gradual differences, has also been reconceptualised as a continuum).

Priming may inform this discussion by providing evidence of whether or not speakers are sensitive to differences in the obligatoriness of the adverbials. For example, following the schematic representation in Figure 2, participants could be presented with bare monotransitive primes that contain no adverbial (construction A) and primes like (9a) which contain an optional adverbial and thus extend the bare pattern (construction B). In addition, participants could be confronted with two further prime types that contain either semi-obligatory adverbials as in (9b) or fully obligatory adverbials as in (9c). These two prime conditions are represented as constructions Bʹ and Bʺ in Figure 2 since they also form extensions of construction A, but they differ from construction B in terms of the obligatoriness of their adverbials. Following these four prime types, speakers could then be asked to complete target fragments such as (10). In this example, the verb cut requires an object (e.g., the wood); beyond this, however, participants could decide freely whether they want to add an optional adverbial (e.g., in the forest) or not.

The question would be whether the likelihood that speakers use an optional adverbial in their target completions increases after an adverbial-containing prime, and whether the strength of such priming varies depending on the obligatoriness of the previously encountered adverbial. For example, it would seem feasible that optional adverbials are more strongly primed by other optional adverbials such as (9a) than by adverbials that are required by the argument structure of the verb, as in (9c) and perhaps to a lesser degree in (9b).

3.3 Strategy 3: testing priming in comprehension

The previous two strategies were based on extensions of production priming experiments beyond their traditional focus on alternating constructions. A third option is to move beyond production methods altogether and to test priming in comprehension instead, using methods such as self-paced reading and eye-tracking during reading. Not only do these methods have the advantage that researchers can pre-select the exact stimuli they want to test, rather than having to elicit them from participants with the help of pictures or sentence fragments (which can be difficult for complex or infrequent constructions). Even more importantly for the present context, the benefit of comprehension methods is that they are not restricted to alternating constructions but can, in principle, be applied to any construction. This is because, in a reading task, participants do not choose between two alternative ways of encoding the same event; rather, they process each target sentence on its own terms and independently from other constructions.

This also highlights another feature of comprehension priming methods, namely that they provide separate outcomes (e.g., reading times) for each target construction individually. This contrasts with production methods, which typically yield proportions of one target type relative to the other. As discussed in detail by Ungerer (2022), one advantage of independent outcomes is that they make it easier to distinguish between within-construction priming and cross-constructional priming effects. This is illustrated in Figure 3, which provides a schematic representation of the third strategy. As the arrows in the diagram show, comprehension priming methods allow researchers not only to test two non-alternating constructions A and B, but the within-construction priming effects (from A to A and from B to B) can also be assessed separately from the cross-constructional priming effects (from A to B and from B to A).

Figure 3:

Schematic representation of Strategy 3, testing within-construction and cross-constructional priming between non-alternating constructions in a comprehension setting.

Comprehension methods have been increasingly used in structural priming over the last two decades (Branigan and Pickering 2017). They are, however, most often applied to phenomena that arguably do not lie at the heart of cognitive-linguistic research, such as garden-path sentences (Traxler and Tooley 2008) and structural ambiguities, for example between “high-attached” and “low-attached” prepositional phrases (Branigan et al. 2005). In addition, these studies have investigated within-construction priming effects rather than priming of constructional relations.

Recently, however, two experiments by Ungerer (2021; 2022) have provided examples of how comprehension methods can be used to study cross-constructional priming between English argument structure constructions. Ungerer (2021) examined priming between the caused-motion construction in (11a) and the resultative construction in (11b), while Ungerer (2022: Experiment 1) tested priming between the resultative construction in (12a) and the object-oriented depictive construction in (12b). Both studies used a special version of self-paced reading called the “maze task” (Forster et al. 2009), in which participants read sentences word by word while selecting between a correct sentence continuation and an incorrect distractor at every step. Moreover, both experiments made use of the fact that the two constructions can be used in temporarily ambiguous sentences; as a result, any potential priming effects were expected to emerge at the later sentence regions where the constructions are disambiguated.

The results of the two experiments contrast in interesting ways. Ungerer’s (2021) study yielded symmetric inhibitory priming between the two constructions: after caused-motion primes, participants responded more slowly to the postverbal regions of resultative targets, and vice versa. One possible interpretation of this result is that speakers perceive the caused-motion and the resultative construction as rather distinct, thus not supporting previous claims that the constructions are related via an asymmetric metaphor (Goldberg 1995). Ungerer’s (2022) experiment, on the other hand, yielded asymmetric facilitatory priming: after depictive primes, participants responded faster to the final adjectives of resultative targets, but the same effect did not emerge in the other direction, i.e., from resultative primes to depictive targets. This suggests that speakers encode a strong similarity relation between the two “secondary predication” constructions, despite their semantic differences. Moreover, the asymmetry of the effect may be explained by the fact that less frequent constructions, such as the depictive, tend to give rise to stronger priming (the “inverse frequency effect”; Ferreira [2003]).

While the interpretation of comprehension priming effects can be intricate, the methods provide powerful tools for extending structural priming to a wider range of constructions. One phenomenon that could be investigated as a follow-up to the above-mentioned studies is the relationship between the caused-motion construction in (13a) and the into-causative in (13b). The into-causative is often regarded as a close relative, or even a subtype of, the caused-motion construction (Rudanko 2011), from which it most likely also emerged historically (Flach 2020). Nevertheless, present-day speakers may represent the into-causative and the caused-motion construction as distinct constructions, given their differences in constructional meaning (‘X causes Y to do Z’ vs. ‘X causes Y to go Z’), the syntactic complexity of their complements (nominal vs. sentential) and the fact that the caused-motion construction allows a range of different prepositions (13a), while the into-causative is restricted to into.

The strength of the similarity relation between the constructions could be tested with the help of comprehension priming, using either traditional self-paced reading or its extension via the maze task described above. To avoid lexical confounds, it would be feasible to select caused-motion instances that contain a preposition other than into. If the two constructions share a common semantic ground, they should show signs of mutual facilitation. Moreover, if the caused-motion construction acts as a metaphorical source for the into-causative (Kim and Davies 2016), priming may be asymmetric, with stronger effects from caused-motion primes to into-causatives than vice versa.