Putting the argument back into argument structure constructions

1.1 Alternating argument structure constructions: lexical and constructional semantics

The issue of alternating argument structures has been dealt with in different ways in the literature: verb-centred approaches focus on the verb and the various argument structures it can be used with whereas construction-based approaches choose to focus more on constructions themselves, and the semantics associated with these constructions. In this subsection I will show how construction-based approaches can also gain from including a finer-grained analysis of the lexical items that occur in argument structure constructions.

Put simply, lexical-projectionist approaches (e.g., Levin 1993; Levin and Rappaport-Hovav 2005) assume that a verb’s semantics is sufficient to predict the various argument structure constructions it can be used in. Valency theory (Faulhaber 2011; Herbst 2011) also takes a somewhat verb-centric approach and argues that a verb’s ability to occur in an argument structure construction is based on both item-specificity and broader generalisations. The general assumption behind lexical-projectionist approaches is that it is not necessary to posit more schematic structures at a higher level of abstraction. However, by not doing so, these approaches fail to identify the characteristics associated with each argument structure construction and cannot account for the productivity of argument structure constructions. There are, however, more hybrid approaches to the issue, including Hampe and Schönefeld (2006) and Herbst (2018) which both make room for, or even insist on, the importance of collocations between verb and arguments slots.

Construction-based approaches tend to be more concerned with the semantics of each argument structure construction rather than verbs’ ability to alternate. A somewhat extreme example of this stance is Goldberg’s (2002) surface generalisation hypothesis which states that there is more to be gained by looking at various instances of the same construction with different verbs than by looking at various uses of the same verb in different constructions. While constructionist approaches give pride of place to constructional semantics, they also emphasise the role played by the verb in the elaboration of constructional meaning. As Goldberg (1995: 24) explains: “it is clearly not the case that the grammar works entirely top-down, with constructions simply imposing their meaning on unsuspecting verbs.” Rather, the general assumption is that the central schematic meaning of these constructions is closely related to the meaning of the verb that prototypically occurs in said construction and less central meanings are extensions from this prototype. According to Goldberg (1995), argument structure constructions contribute meaning to the verb they occur with, and verbs profile a certain aspect of the basic event denoted by the construction. Stefanowitsch and Gries (2003), for example, have shown that the verb give is strongly attracted to the ditransitive construction, whose most central meaning is that of transfer and thus close to the meaning of give. In a similar vein, in a study on alternating argument structure constructions and the productivity of these constructions, Yi et al. (2019: 603) posit the Verb Anchor Hypothesis, which states that “semantic similarity to the anchor verb modulates speakers’ choice of syntactic frame for these verbs.” In this hypothesis, the anchor is the verb that occurs the most frequently with the frame such that it is considered representative of that frame.

Other researchers have argued, however, that it is not necessarily the case that constructional meaning is centred around a prototype. Croft (2003) suggests that constructional polysemy is better captured through family resemblances at a verb-class or even verb-specific level, rather than extensions from a single prototype. Even though they do identify give as prototypical of the ditransitive construction, Stefanowitsch and Gries (2003) also comment on the various classes of verbs that occur in the ditransitive and to-dative constructions. Gries and Stefanowitsch (2004) also identify differences in the semantics of each construction based on the classes of verbs that are most attracted to each member of an alternation. Perek (2014) makes a similar observation in his analysis of the conative construction (e.g., She sipped at her drink) based on a modulation of Stefanowitsch and Gries’s (2003) collexeme analysis which includes verb-class specific constructional meaning. In another vein, Gilquin (2013) focuses on verb senses rather than simply verb lemmas in her analysis of periphrastic causative constructions and finds that different verb senses are attracted to different constructions. Similarly, Bernolet and Colleman (2016) propose a collostructional analysis of the dative alternation in Dutch that takes into account the different senses of each verb studied and they show that these different senses show “quite different degrees of association to the constructions involved in a grammatical alternation” (Bernolet and Colleman 2016: 177). Overall, these lower levels of generalisation, at verb-class or verb-sense level, are assumed to provide a more accurate representation of speakers’ knowledge. The argument presented in Croft (2003) reconciles the verb-centred and the construction-centred approaches, as it is quite plausible that speakers are aware both of a verb’s possible distribution across various argument structure constructions and of generalisations at (a) more abstract level(s).

A finer-grained analysis that includes verb classes or verb senses is particularly relevant to the causative alternation as it is very frequent, it occurs with many different types of verbs and may thus have several related meanings. This is why I chose to work with verbs from different semantic groups, as described in Section 2. Another interesting feature of this alternation is that its two argument structure constructions, namely the intransitive non-causative construction (henceforth INCCx) and the transitive causative construction (henceforth TCCx) do not have the same number of arguments. In the next subsection, I will show how the choice of argument structure construction is thus closely related to the types (and number) of participants one includes in their description of an event.

1.2 The causative alternation: event conceptualisation and argument realisation

In this subsection I will show the importance of taking event conceptualisation and argument realisation into account when trying to account for speakers’ choice of an argument structure construction. After all, as Langacker (1991: 286) points out, although participants can occur outside of events, “an event is conceptually dependant vis-à-vis its participants.” Before moving on to theoretical considerations, I will introduce in more detail the two constructions that make up the causative alternation. The intransitive non-causative construction (INCCx) can be illustrated by the examples in (1) and the transitive causative construction (TCCx) by the examples in (2):

As is illustrated in these examples, the INCCx requires only one argument which is realised in subject position. The TCCx takes a minimum of two arguments: an agent, realised in subject position, and the entity undergoing the event, realised in object position. In both constructions, I will refer to the entity undergoing the event as the theme. In the examples presented in (1) and (2), the reader will notice that the themes bottle(s), cheeks and cable can be found in both constructions but I will show in the next section that this is not always possible. I choose to call the participant that is shared by the two constructions the ‘theme’ following Langacker’s (1991: 288) basic thematic roles presented in Figure 1. Langacker proposes that in an event where an entity undergoes a change of state, this entity (the theme) forms together with the predicate a minimal unit which he calls the thematic relationship. In a thematic relationship, the theme can be any of the first four thematic roles represented in Figure 1 where the different types of arrows denote the different types of events these entities can undergo: change of position (mover), internal change (patient), and mental process (experiencer).

Figure 1:

Basic thematic roles (Langacker 1991: 288).

The thematic relationship corresponds to the INCCx to the extent that the entity in subject position can undergo the event denoted by the verb without mention of an external cause, as illustrated in (1). It could be argued that the theme is simply a patient realised in subject position but patients are generally assumed to be prototypically passive (Dowty 1991) which is not entirely true of the theme in this construction, which could be construed as co-instigator (see for example Lemmens 1998) or facilitator. It is this specific property of the theme that is explored in this paper and that motivates the somewhat more neutral label ‘theme’. The term theme also makes it easier to describe the entity in both the INCCx and the TCCx.

Depending on whether or not the speaker wishes to include an external cause in their description of the event, they can choose to use either the INCCx or the TCCx. These two constructions are illustrated in Figure 2, with the INCCx in (a) and the TCCx in (b). This representation shows that the glass breaking event exists pre-linguistically and, depending on how speakers conceptualise this event, they may choose one of several linguistic structures. If the speaker wants to include the cause of the event, they will use the transitive causative construction. If the speaker’s conceptualisation of the event only includes the change of state, without causation, then they will use the intransitive non-causative construction. Therefore, depending on the speaker’s conceptualisation of the event, only certain arguments will be expressed. As shown in Figure 2, the INCCx (a) only profiles the theme (in bold) whereas the TCCx (b) profiles both the theme and the agent.

Figure 2:

The intransitive non-causative construction and the transitive causative construction (adapted from Langacker 1991 and Lemmens 1998).

The question that arises here is whether there are constraints at the level of participant/argument that may limit a speaker’s choice of construction. These constraints can be argued to be based on the semantics associated with the role at a schematic level or at a very local level, i.e., that of the interaction with the verb. The discussion around the various semantic roles that are necessary to capture generalisations in linguistic structures has been ongoing for quite a while in the literature (Boas 2003; Croft 1991, 1998, 2012; Cruse 1973; Davidse 1991, 1992; Delancey 1984; Dowty 1991; Fillmore 1977; Halliday 1967a, 1967b; Langacker 1991; Lemmens 1998, 2006; Levin and Rappaport-Hovav 2005; Talmy 2000 to name a few). Langacker argues that there are, to some extent, two extremes in the number of semantic roles that are relevant: at one end of the spectrum it can be posited that “each verb defines a distinct set of participant roles, that reflect its own unique semantic properties”; at the other end, “a role conception is arrived at by abstracting away from the peculiarities of individual examples” (Langacker 1991: 284). In a similar vein, Croft (2012: 206–207) posits that a finer-grained analysis is preferable as it “places role designation in the semantic representation of verbs (or more precisely, verbs in particular argument structure constructions).” In other words, in order to better understand what is at stake with thematic roles, it is crucial to pay attention to their interaction with verbs in context. I will argue here that by investigating the semantic (dis)similarities between the themes that occur in the same position in one construction and the themes that occur in different positions in the two constructions, it is possible to abstract away from local generalisations to identify a more schematic meaning for each position in each construction. This view is compatible with a constructional approach in that the participant roles of a verb (lower generalisation) can be matched with the argument roles of a construction (schematic representation) (Goldberg 1995: 50).

The identification of more abstract generalisations as to the constraints on the realisation of the theme in each construction should be helpful to better understand how much information is actually shared by the two constructions and to what extent they differ.

2.1 Data collection, annotation and distribution of verbs across constructions

The constructions that make up the causative alternation, INCCx and TCCx are very frequent and they occur with a large number of verbs (Levin [1993] lists 355), which makes it virtually impossible to analyse all of them. Therefore, it was necessary to restrict the dataset to certain verbs. A total of 29 verbs were selected from Levin’s list. These 29 alternating verbs are further divided into five groups, four of which were identified by Levin while the fifth, change of temperature was created for the purposes of this study. I chose to select groups of verbs as they can be considered as verb classes in the sense of Croft (2003) and might help identify constructional meaning at a finer-grained level of generalisation. I extracted the data from the Corpus of Contemporary American English (COCA). The INCCx and the TCCx are virtually impossible to automatically extract with mere Part of Speech tags. Therefore, I extracted a number of sentences that contained the verbs under study and then manually annotated them for construction and theme. Table 1 presents the number of instances retained for each verb and their distribution across the two constructions. During the annotation process I had to discard a number of constructions that are similar to the INCCx and the TCCx but are not instances of these. These constructions include the middle construction illustrated in (3) which requires an adverb to be grammatically acceptable (Goldberg 1995; Lemmens 1998, 2005; Yoshimura and Taylor 2004), objectless constructions as in (4), where the entity undergoing the event is omitted (Goldberg 2001; Lemmens 2006) and the setting construction as in (5), where the participant in subject position is not actually the agent (Davidse 1992; Guerrero-Medina 2014), which all had to be manually excluded.

Table 1 shows that, at least with our 29 verbs, the TCCx is slightly more frequent than the INCCx, and that most verbs actually do alternate, although some verbs such as slide and crush barely do.

2.2 Theme overlap (Jaccard index)

While we see from the annotated dataset that the distribution of instances across constructions varies from one verb to the next, it is not yet clear how much overlap in meaning there is between the two constructions. I thus follow the method put forward by Lemmens (2021: 106–112) to get a clearer idea of this overlap and look at the themes that are shared by the constructions, as they are, together with the verb, the one element that is present in both constructions. To measure this overlap, we use the Jaccard index, that is, a count of the number of themes that are exclusively used with INCCx, with TCCx or that are shared by the two constructions.

The motivation behind this measure is that, even though most verbs are shared by the two constructions they might actually be used with very different themes in each construction. For example, we know that the verb break occurs in both constructions but some uses are restricted to one construction or the other, as shown in examples (6) and (7).

As is clear from (6) and (7), the theme law is restricted to the TCCx and the theme day is restricted to the INCCx when used with the verb break. An exaggerated assumption could be that this is true of most instances, which would subsequently reduce the actual alternation strength of verbs. If verbs only occur in each construction with specific themes, then the amount of shared meaning between the two constructions would be drastically different from what one might expect based solely on the numbers in Table 1. A limitation, however, is that Theme Overlap only takes themes at face value. This limitation is all the more relevant that my dataset only contains 11,554 instances and it might be the case that some of the themes that appear not to be shared are actually semantically similar to themes restricted to the other construction. It is therefore necessary to identify semantic classes of themes. It will then be easier to see what kinds of themes occur in each construction, and which are shared. To achieve this, it is important to find an objective measure of semantic similarity, which is explored in the next subsection.

2.3 Vector space models

As Goldberg (2001: 516) observes: “Bubbles, TVs, breadsticks and hearts break in very different ways and with very different consequences.” While it is true that these things break in very different ways, they are also part of groups of entities that break in similar ways. For example, TVs break in a similar way to computers while windows and mirrors break in a similar way to each other, but very differently from TVs and computers. The fact that different things break differently leads to positing different verb senses for a verb such as break. The assumption here is that if we can identify clusters of semantically similar nouns, it should be possible to see whether these clusters are actually shared by the two constructions and thus whether a verb’s various senses are also used in both constructions. The collocation between theme(s) and verb(s), combined with their use in one or two specific argument structure constructions is taken as a solid ground for the elicitation of verb senses. While I could have used Behavioural Profiles (Gries and Divjak 2009) for the identification of verb senses, I chose not to for two reasons: (i) due to the size of my dataset, this would have taken an unreasonable amount of time to annotate manually and (ii) the dataset was already reduced to two specific argument structure constructions. I thus opted for a distributional semantics approach through vector space models.

Vector space models are an implementation of the principles of distributional semantics and were developed by Salton et al. (1975). The concept underlying vector space models is the representation of words (vectors) in a semantic space (for various application of vector space models see, among others, Gries and Stefanowitsch 2010; Hilpert and Perek 2015; Levshina and Heylen 2014; Padó and Lapata 2003; Peirsman et al. 2010; Perek 2016a, 2016b; Sahlgren 2008; Schütze 1992; Turney and Pantel 2010). The idea behind distributional semantics can be roughly summed up by Firth’s (1957: 11) famous observation that “you shall know a word by the company it keeps.” As proposed in Lenci’s (2008: 3) Distributional Hypothesis: “The degree of semantic similarity between two linguistic expressions A and B is a function of the similarity of the linguistic contexts in which A and B can appear.” The expectation is thus that words that occur in similar contexts will be semantically similar. This notion of similarity is to be taken broadly and not be restricted to (near-)synonymy as such a method will consider antonyms to be semantically similar as they tend to occur in the same contexts (e.g., things that open usually also close). As such, the type of semantic similarity one obtains with vector space models is based on the words’ frequently shared collocates in a given dataset, as shown in the toy co-occurrence matrix in Table 2, where the rows represent the themes and the columns their collocates. We see here that the semantically close nouns dress and shirt share certain collocates such as denim, white and the verb wear. Miller and Charles (1991) have shown the psychological/cognitive validity of this notion of similarity through a series of experiments.

The actual co-occurrence matrix used for the present study was taken from Perek’s model (Perek 2021; see Hilpert and Perek 2015 for an illustration of this model) which contains data extracted from the Corpus of Contemporary English (COCA). It contains all the nouns that occur in COCA and the collocates were limited to the 10,000 most frequent nouns, verbs, adjectives and adverbs found in COCA within a two-word window of the target nouns; function words are deleted as they do not bear enough relevant meaning and might create too much noise. Pronouns referring to people were not retained either for this part of the analysis but they were kept in the dataset. Two additional transformations were applied: (i) the co-occurrence counts were weighed by means of Positive Pointwise Mutual Information to give more weight to collocates that co-occur more often with one theme compared to the other themes and, (ii) the number of columns was limited to 300, thereby saving only the most relevant collocates (Hilpert and Perek 2015). These transformations make the co-occurrence matrix more salient. The similarity between the word vectors is measured by calculating the distance between the different rows. This distance is measured via the ‘cosine’ function of the R package lsa (Wild 2007); the shorter the distance between two word vectors, the more similar they are. Similar themes are then clustered together via the hclust function in R (R-Core-Team 2013) and represented on a semantic map plotted in R. As a result, semantically similar themes will be placed close together on the map and at a distance from semantically dissimilar themes. This analysis was run for each verb in each construction, which gave two maps per verb (granted there were enough themes to compare in each use). This method presents the advantage of visualising the entire space and helps identify semantic subspaces that are more or less populated. That is, clusters of lexemes will be more or less dense depending on the number of lexemes that populate this cluster in the semantic space.

In the next section, I will present semantic maps for a few verbs and discuss how they help visualise the distribution of each verb in the alternating constructions and what this entails for constructional meaning.

3.1 Theme overlap

The Theme Overlap analysis yielded results that indicate low overlap between constructions. Figure 3 shows the distribution of themes across constructions for each verb. The yellow portion represents the number of themes (types) that are actually found with both constructions in our dataset. Clearly, for many verbs, only a handful of themes are shared by the two constructions. For more than half of these verbs, the proportion of shared themes is below 10% and only three verbs have a theme overlap equal or above 20%.

Figure 3:

Theme overlap between the two constructions.

This measure shows more clearly the actual distribution of verbs and their themes across the two constructions, and highlights the limited amount of overlap between the two constructions. The next sections will zoom in on the kinds of themes that are restricted or shared.

3.2 Vector space models

The five verbs presented here are meant to be representative of a larger tendency among verbs that alternate between the two constructions: they are found in both constructions but some of their senses are restricted to one construction or the other. Even though I cannot present all my results here, I will offer a brief overview of the 29 verbs analysed. The manual annotation of the 11,554 instances of these 29 verbs led to the identification of 4,781 different themes. Based on the clusters of themes automatically generated by the vector space model for each verb in each construction, I identified 126 different verb senses, among which only 55 are shared by the two constructions. Within these 55 shared verb senses, 12 are shared but with different themes. Finally, 31 senses are restricted to the INCCx and 40 are restricted to the TCCx. The definitions used for these verb senses come from three different sources: the Oxford English Dictionary Online (“OED Online” 2018), WordNet (Princeton University 2010) and my own interpretation when a verb sense was identified that did not match any of the definitions in these resources.

Each map presents the themes that were found with each verb in each construction. The colours were generated automatically, which is why they do not always match the clusters perfectly. As explained in Subsection 2.3, the similarity of the themes was not measured based on their use with the target verb, but in all sorts of contexts, which may lead to themes expected to undergo the event denoted by the verb in a similar way to be placed somewhat far apart on the map. The method is not perfect, but it offers an overview of the types of themes used with each verb at a glance, which is crucial in our study.

3.2.1 Break

After manually annotating 1,750 instances of the verb break, I identified 121 instances of the INCCx and 520 instances of the TCCx, for a total of 641 instances retained. Break was used with 239 different themes, among which 68 were found with the INCCx and 190 with the TCCx, for a theme overlap of 7.95%. Figure 4 shows the themes used with the INCCx and Figure 5 the themes used with the TCCx. I manually added a dashed line on each map to show the separation between concrete themes from abstract themes. As is obvious from a quick glance at the two maps, there are many more abstract themes found with the TCCx than with the INCCx. In the INCCx, we find the news family such as news, story and scandal and the people family with man and folk. These two families of themes are actually shared by the two constructions. The abstract themes found in the TCCx can be grouped under three different families: law (including curfew, law and rule), promise (as in promise, oath and pledge) and habit & situation (habit, routine or engagement). Interestingly, these themes are actually restricted to the TCCx and cannot appear in the INCCx.

Figure 4:

A semantic map of the themes that occur with break in the INCCx.

Figure 5:

A semantic map of the themes that occur with break in the TCCx.

Based on these semantic maps, I isolated eight different senses of break, which are summed up in Table 3. In this table, the senses that are specific to the INCCx are on the right, senses restricted to the TCCx on the left and shared senses are in the middle. Among these eight senses, half are shared by the two constructions. The first sense, “(cause to) become broken”, is found with a rather large variety of themes and specifically with bones, as illustrated in (8).

Table 3:

Shared and construction-specific verb senses of break.

break
Intransitive non-causative construction	Shared	Transitive causative construction
	1.1. bones 1.2. various themes
1.3. water
	2. motive-powered devices
	3. people & mind states, 3.1. heart
	4. news
5. day & atmospheric elements
		6. law
		7. promise
		8. habit & situation

(8)	You attacked me from behind and broke my arm this day […]

Another interesting example within this group is water which is restricted to the INCCx as illustrated in (9).

(9)	Juanita Massie can recall her baby’s kicks inside her belly, how her water broke […]

The second verb sense identified for break is found with themes that refer to motive-powered devices, and in which case the item becomes or is made inoperative, as illustrated in (10).

(10)

motive-powered devices:

a.	My tape recorder broke, but I’m going to get it fixed tomorrow.
b.	The intent here is for me to learn flight-testing and not break the airplane while I do […]

The next two senses are found with more abstract themes: people & mind states and news. These two senses are found with both constructions, as illustrated in (11) and (12).

(11)

people & mind-states, (cause to) lose emotional control

a.	Did police put their fists in women’s faces so the husbands’ spirits would break along with the women’s jaws?
b.	New York breaks a lot of people, but Spree seems to be thriving there.

(12)

news

a.	After the infidelity scandal broke, paparazzi showed James’ every move.
b.	[…] our White House correspondent that broke the news this morning about the departure of Jean Bertrand-Aristide.

The last four senses are restricted to one construction. The first of these four, “appear” is found with day & atmospheric elements and is restricted to the INCCx, as shown in (13). In the Oxford English Dictionary, the definition of break for this sense is To burst out of darkness, begin to shine; as of the day, morning, daylight. Lemmens (2006) argues that there is actually a now lost motion sense with this particular sense of break, and that the entity in subject position is not actually the entity that undergoes the event denoted by the verb. The assumption is that themes such as day, sun or dawn actually do the breaking, i.e., they break the darkness. Considering the opacity of this interpretation, I chose to keep this sense as part of the INCCx.

(13)	[…] as dawn breaks in Afghanistan […]

The last three groups are restricted to the TCCx and are all found with abstract themes. Sense 6 is found with themes such as law and rule and corresponds to infringe. Sense 7 is closely related and is found with themes such as promise and oath, for these the meaning of break is similar to betray. Finally the last sense is found with themes that refer to habits and certain situations, as illustrated in (14) and (15) respectively.

(14)	[…] find the joy in breaking your routine, Mellan suggests.

(15)	Steven’s then-fiance, Traci Greer, broke their engagement and married Walter Millbank.

To sum up, while it seems that the most prototypical senses of the verb break are found in both constructions, less prototypical uses are restricted to one construction or the other. Typically, we observe that elements that cannot be acted upon, such as dawn, are only found with the INCCx whereas things that do not seem to have the ability to break without an external cause, i.e., that have to be acted upon in order to break, are restricted to the TCCx.

3.2.2 Tear

The verb tear is less polysemous and less frequent than break in these constructions but presents some interesting specificities too. After manually annotating 1,600 instances of tear, I identified 22 instances of tear in the INCCx and 70 in the TCCx for a total of 92 instances retained. The majority of instances extracted from the COCA were, unsurprisingly, instances of the resultative construction (including also phrasal verbs such as tear apart). There were also many instances of the setting construction such as She tore her ACL playing golf. Within the instances of the INCCx and the TCCx I identified 66 different themes, 20 with the INCCx and 51 with the TCCx for a total overlap of 7.58%.

The maps in Figures 6 and 7 show the difference in the number of different themes found with each construction but also the three main clusters of themes found with tear, namely fabric, paper and flesh (such as skin, cartilage and tendon) which are actually the only ones found to be shared by the two constructions. As shown in Table 4, the main and only concrete meaning of tear, which is similar to rip is the only verb sense that is found with both constructions. However, it also appears that while this sense is shared by the two constructions, it is only found with themes that belong to the fabric, paper and flesh families. While the INCCx is limited to these three types of themes, it is clear that the TCCx is used with a wider variety of themes such as sandwich, as illustrated in (16). This shows that in order for a theme to be used with tear in the INCCx, it must be construed as likely to tear on its own or at least as having properties that may facilitate the tearing event. With the TCCx, the theme is not as restricted as to its ‘tearability’ but rather has to be something that can be acted upon. This further explains why the TCCx is also found with figurative senses.

Figure 6:

A semantic map of the themes that occur with tear in the INCCx.

Figure 7:

A semantic map of the themes that occur with tear in the TCCx.

Table 4:

Shared and construction-specific verb senses of tear.

tear
Intransitive non-causative construction	Shared	Transitive causative construction
	1. 1.1 fabric, 1.2 paper, & 1.3 flesh
		1.4. other themes
		2. “cause to break apart (figuratively)”
		3. “interrupt suddenly”

(16)	She stood on her chair, making small grunting noises as she tore the sandwich and squeezed the bread and peanut butter into paste.

As can be seen in Table 4, two other senses of tear were identified with the TCCX and both are figurative uses of tear, as illustrated in (17) and (18).

(17)	I’ve seen how it has torn my family when you don’t accept that people are going to fall in love with who they fall in love with.

(18)	[…] but a cough tore the stillness.

Crucially we see that only the most literal sense of tear is shared across constructions, and only with a limited number of themes.

3.2.3 Crease

Out of the 850 instances of crease extracted from COCA, I identified 112 instances of the INCCx and 247 instances of the TCCx, for a total of 359 instances retained. Interestingly, while my dataset contains substantially more instances of crease than tear, crease is clearly less varied as only 22 themes were found with the INCCx and 71 with the TCCx. This is explained by the fact that the meaning of crease is very specific and limited to one verb sense. This monosemy may also explain the unusually high amount of overlap of themes between the two constructions with 16.25% of themes shared.

As is visible from Figure 8, a majority of the themes that occur with crease in the INCCx relate to skin (cf. top right corner of the map). There are also themes that belong to the paper and fabric families. These three families are the only ones that are shared by the two constructions, as shown in Table 5. The themes foundation and formula both refer to creamy or liquid make-up products, as shown in (19) and (20).

Figure 8:

A semantic map of the themes that occur with crease in the INCCx.

Table 5:

Shared and construction-specific themes that occur with crease.

crease
Intransitive non-causative Cx	Shared	Transitive causative Cx
	fabric, paper, and skin
liquid products		surface

(19)	Beverly Hills Mineral Foundation, $65, is a mineral-based liquid formula that won’t crease or fade.

(20)	Blot your face to get rid of excess oil and blend in foundation that has creased using a clean makeup sponge in a circular motion.

Aside from the obvious larger number of themes found with the TCCx in Figure 9, we also find a whole cluster of themes that refer to types of surface that various things can cause to crease, as in (21) and (22).

Figure 9:

A semantic map of the themes that occur with crease in the TCCx.

(21)	Six flashing oars creased the surface of the river in synchronized strokes […]

(22)	[…] the highway only barely creasing the snowed-over sagebrush flat land.

The distribution of crease across the two constructions seems to confirm our previous comments as to the constraints on the theme in each construction. Items such as fabric, paper and skin which seem to have certain properties that makes them very likely to crease but also likely to be creased by an external cause are shared by the two constructions. Interestingly, the group that is restricted to the INCCx is made of things that one does not want to crease, i.e., various makeup products.

3.2.4 Crinkle

Out of the 600 instances of crinkle manually annotated, I retained 330 instances: 204 in the INCCx and 126 in the TCCx. Crinkle was found with 73 different themes, 57 with the INCCx and 36 with the TCCx for a theme overlap of 27.4%. Again, a monosemous verb displays an unusually high theme overlap. Crinkle is semantically similar to crease, as both denote some type of intricate folding and just like crease, the only themes that are shared between the two constructions are of the fabric, paper and skin families, as shown in Table 6. Also, like crease, themes that refer to a type of surface are limited to the TCCx.

Table 6:

Shared and construction-specific themes that occur with crinkle.

crinkle
Intransitive non-causative Cx	Shared	Transitive causative Cx
	fabric, paper, and skin
		surface

As opposed to crease though, there are actually more themes that occur in the INCCx than the TCCx, as visible from Figures 10 and 11. Again, with the TCCx, we find themes that refer to a type of surface, as illustrated in (23).

Figure 10:

A semantic map of the themes that occur with crinkle in the INCCx.

Figure 11:

A semantic map of the themes that occur with crinkle in the TCCx.

(23)	[…] a light breeze crinkled the surface of an otherwise smooth and moonlit sea.

Again, themes that can easily be construed as being likely to crinkle because of their properties are more likely to be shared, or, for that matter, be found in the INCCx. On the contrary, the surface of the sea requires external intervention to be creased.

3.2.5 Wrinkle

A total of 1,100 instances of wrinkle were extracted from the COCA and manually annotated. From these, 701 instances were retained: 215 instances of the INCCx and 489 instances of the TCCx. Despite the large number of instances found in the dataset, the variety of themes is somewhat limited, with only 87 different themes identified, 55 with the INCCx and 50 with the TCCx for a theme overlap of 20.69%.

The two maps in Figures 12 and 13 show that three clusters of themes are shared by the two constructions: fabric, skin and surface. In this case, a small cluster of themes is restricted to the INCCx: people. As may appear evident from the examples in (24) and (25), people can wrinkle (metonymically at least) but it seems unlikely that something or someone can be construed as wrinkling people. Note, however, that certain things can wrinkle people’s skin.

Figure 12:

A semantic map of the themes that occur with wrinkle in the INCCx.

Figure 13:

A semantic map of the themes that occur with wrinkle in the TCCx.

(24)	Once people start to stoop and wrinkle they all look alike to me.

(25)	I didn’t know they could prove I was wrinkling.

Our analysis of wrinkle also leads us to conclude that constraints on the use of these two constructions go not only beyond the level of the verb, but also beyond individual verb senses (see Table 7).

Table 7:

Shared and construction-specific themes that occur with wrinkle.

wrinkle
Intransitive non-causative Cx	Shared	Transitive causative Cx
	fabric, skin, and surface
people

By means of conclusion to this section, it can be argued that a finer-grained analysis that includes each verb’s themes and thus various senses reveals a certain amount of discrepancy between the two constructions. The most prototypical verb senses tend to be shared while less central verb senses tend to be restricted to one construction. At an even lower level of generalisation, we see that even when a verb sense is shared, it is regularly the case that some of the themes that correspond to this verb sense are actually restricted to one construction.