Semantic and contextual constraints on the causative alternation in English: a multifactorial analysis

2.1 Theoretical approaches to constraints on the causative alternation

Previous approaches to the causative alternation fall into two major classes. The first class assumes that the main constraint on the alternation derives from the lexical characterization of the cause argument. This class of approaches has been most fully developed in Levin and Rappaport Hovav (1995) and Reinhart (2002, 2016). The second class assumes that the constraints on the alternation follow from the lexical classification of the verb together with nonlexical factors, stressing the interplay of lexical and nonlexical factors in explaining the limited productivity of the alternation. This class of approaches is represented by Alexiadou (2010), Alexiadou and Doron (2012), Alexiadou et al. (2006), Rappaport Hovav (2014, 2020), and Rappaport Hovav and Levin (2012), despite significant differences among their analyses.

A key assumption shared by many accounts of the causative alternation that represent these two classes of approaches is that the alternation is enabled by the underlying lexical structure of the alternating verbs (e.g., Alexiadou et al. 2006; Levin and Rappaport Hovav 1995; Reinhart 2002, 2016; Samardžić and Merlo 2018, among others). In their seminal work, Levin and Rappaport Hovav (1995) first proposed that the crucial lexical-semantic distinction for characterizing the class of alternating verbs is between verbs denoting internally caused eventualities and those denoting externally caused eventualities. They assume that verbs denoting internally caused eventualities, such as bloom and blossom, are lexically monadic, selecting only the argument denoting the entity undergoing the change specified by the verb. Conversely, verbs denoting externally caused eventualities, such as destroy and open, are lexically dyadic, additionally selecting a cause argument. On their account, a subset of verbs denoting externally caused events, such as break and open, undergo a process of lexical binding of the external argument, which prevents the external argument from being expressed syntactically, thus resulting in the noncausative variant of the alternation. Reinhart (2002, 2016) provides a similar characterization of the class of alternating verbs. Like Levin and Rappaport Hovav’s (1995) analysis, her analysis of the causative alternation takes the causative variant to be basic, deriving the noncausative variant from it via a lexical operation of decausativization. This operation applies to verbs that simply specify that their subject is a cause, subsuming agents, natural forces, and instruments (Reinhart 2016: 25).

Thus, both Levin and Rappaport Hovav’s (1995) and Reinhart’s (2002, 2016) accounts characterize the class of alternating verbs over the causative variant of the alternation as verbs with no lexical specification for the causing event. This characterization predicts that alternating verbs can appear with a wide range of semantic types of NPs, as shown in (2a) and (2b). In contrast, the verb murder specifies something about the causing event (it must involve intention). Hence, this verb cannot undergo lexical binding or decausativization, and does not participate in the alternation, as shown in (3).

A challenge encountered by such lexically-oriented accounts is that some verbs which fit the characterization of the class of alternating verbs do not alternate, as illustrated for destroy and kill in examples (4) and (5). As noted out by Rappaport Hovav (2014: 15), English verbs of destruction and killing are often classified as externally caused and do not lexically specify any particular cause of the state change, as indicated by the fact that these verbs allow a range of NPs as cause as in (4a) and (5a). Nevertheless, these verbs do not alternate, as shown in (4b) and (5b).

Another challenge for Levin and Rappaport Hovav’s (1995) and Reinhart’s (2002, 2016) lexically-oriented accounts arises because the causative alternation is not entirely constrained by the lexical specification of the verb; contextual factors also play a critical role. If the lexical specification of the verb were solely responsible for constraining the alternation, then all externally caused verbs that are compatible with NPs, bearing a range of semantic roles in subject position, would be expected to appear in both variants, regardless of the properties of the NP chosen as the theme. However, this is not the case. Change of state verbs that may select agents, natural forces, and instruments as subjects may still not exhibit noncausative uses for some choices of causative variant themes. As Levin and Rappaport Hovav (1995) note, causative sentences of change-of-state verbs − the prototypical causative alternation verbs − lack a noncausative counterpart, as demonstrated in examples (6) and (7).

Rappaport Hovav and Levin (2012) demonstrate that similar patterns are found with other change of state verbs such as empty, lengthen and shorten. These verbs conform to the lexical specification of the class of alternating verbs, and do indeed alternate, as exemplified by clear in (8). However, in certain cases, a noncausative variant is ill-formed in isolation, as in (7b) above.

Levin and Rappaport Hovav (1995: 102) propose that the rule of lexical binding is limited to cases where the verb denotes an “eventuality [that] can come about spontaneously without the intervention of an agent” (cf. Haspelmath 1993; Smith 1978). This accounts for the contrast between (8b), which describes an event that can take place without an agent, and (6b) and (7b) which cannot.

However, the spontaneity condition on application of lexical binding is problematic on both theoretical and empirical grounds. First, it is not the verb’s meaning that determines spontaneity or agent involvement; instead, world knowledge and properties of events tell us that certain events, such as clearing of the sky and lengthening of the days, do not need the intervention of an agent, while others, like clearing of the counter/the table and lengthening of the skirt, necessitate such intervention. Thus, the constraint on the availability of variants appears to be non-lexical in nature (Rappaport Hovav and Levin 2012; Schäfer 2008). Rather, as Rappaport Hovav (2014) accurately observes, it seems to be a constraint on the suitability of the noncausative variant in describing events of a certain type. The non-lexical nature of this constraint on the alternation thereby challenges the basic assumptions of Levin and Rappaport Hovav’s (1995) and Reinhart’s (2002, 2016) lexically-oriented accounts, which assume that all alternating verbs are inherently dyadic and the noncausative form is derived through a lexical rule that removes the cause argument.

A further empirical problem of the spontaneity condition on the application of lexical binding is the prevalence of noncausative uses of alternating verbs that describe events involving an unspecified external agent. Consider the following examples of noncausative sentences of open from the BNC in (9), discussed by Lee (2023: 360).^[2] These examples describe events of opening that require the participation of an agent: opening shops and opening a village; these contrast with opening a door which could be done by an agent, an instrument, or even a natural force (see (2b)).

In example (9), the addressee may not be able to determine the specific identity of the agent, but can infer the type of agent, such as the owner of a shop and the local government. In such cases, the unexpressed agent is type-inferable, though its exact identity remains unidentifiable, irrelevant and unimportant. This kind of inferable agent is generally disfavored in causative uses, if not entirely excluded. However, contrary to this observation, Levin and Rappaport’s (1995) account predicts that agentive uses of verbs lack a noncausative variant, as on their account a process of lexical binding is restricted to verbs denoting an event that can come about spontaneously.

Rappaport Hovav and Levin (2012) and Rappaport Hovav (2014) propose a non-derivational approach to the alternation which diverges from their own earlier analysis and those of Reinhart (2002, 2016), as well as non-derivational analyses embodied in Alexiadou et al. (2006) and others. In the realm of change-of-state verbs, Rappaport Hovav (2014) makes a three-way classification, summarized in Table 1.

According to Rappaport Hovav (2014), verbs such as murder and assassinate, which lexicalize both a volitional action and a change of state, are lexically associated with an external cause argument, and this argument cannot be omitted. Verbs of destruction and killing share this pattern, as their external cause argument cannot be omitted. Consequently, the verbs of the first two classes are indistinguishable by lexical adicity in English. In contrast, verbs that allow alternation in English are lexically associated solely with their patient. Rappaport Hovav and Levin (2012) and Rappaport Hovav (2014) provide systematic evidence supporting this view. The addition of the cause is non-lexical, driven by the well-known constraint that the cause must be construable as a direct cause – a cause with immediate control over the eventuality. Building on Rappaport Hovav and Levin (2012), Rappaport Hovav (2014) suggests that a causer can be realized with a change-of-state verb just when the event being described involves direct causation in the sense defined by Wolff (2003: 5). This constraint contributes to the determination of the semantic type of NPs allowed as the external argument in given contexts. It does not, however, explain why alternating verbs differ with respect to how often they are used as a causative and as a noncausative and which variant is appropriate in a given discourse context. For this, we need a theory which accounts for the (non-)appearance of the cause argument.

Rappaport Hovav (2014) provides a pragmatic account of the (non-)appearance of the cause argument. Drawing on Grice’s (1975) maxims of conversations, she argues that if the expression of the cause is deemed relevant, it is typically preferred since the causative variant is more informative. The degree of informativeness of the causative variant is influenced by factors such as contextual identifiability and predictability of the cause. For instance, if the cause is recoverable in some way from context, then the sentence with the cause expressed is no longer more informative than the corresponding sentence that expresses just the change of state. In that case, mentioning the cause seems superfluous and the noncausative may be preferred from economy considerations like Grice’s (1975) Maxim of Manner, which dictates avoiding prolixity.

Sometimes the speaker will leave the agent unmentioned though the causative variant would be more informative. This happens, according to Rappaport Hovav (2014), when the speaker does not know the cause; consequently, the noncausative is required because the speaker cannot truthfully specify the cause. Rappaport Hovav (2014, 2020) shows that this account covers constraints on the alternation which have been taken to be lexical and also those which are more clearly non-lexical in earlier accounts.

The inclusion of contextual constraints in the analysis of the causative alternation has recently gained ground in an empirical work by Lee (2023). In the following subsection, we discuss this work and other quantitative studies on the causative alternation.

2.2 Quantitative studies on the causative alternation

Quantitative studies on the causative alternation have focused on the morphosyntactic form of alternating verbs, which has emerged as a major locus of cross-linguistic variation. In his typological work, Haspelmath (1993), distinguishes five types of encoding of the causative alternation: (i) the marked causative type, where the causative variant is formally marked as opposed to the noncausative variant (illustrated by Georgian in Table 2); (ii) the marked anticausative type, where the noncausative variant is formally marked as opposed to the causative variant (as in the Polish example); (iii) the labile type, with no formal change in the verb occurs (as seen in English); (iv) the equipollent type, where both the causative and the noncausative variant bear special morphology that is attached to a shared stem (as in the Japanese example); and (v) the suppletive type, where the two variants are expressed by verbs which are formally not related (as in the Russian example). The different marking strategies illustrated in Table 2 represent only the most prevalent markings. Languages may adopt different strategies for different verbs. For instance, in Korean, the intransitive version of the verb meaning ‘open’ is morphologically marked, whereas the transitive version is not marked (i.e., yel-li- ‘open (intransitive)’ versus yel- ‘open (transitive)’). On the other hand, the transitive version of the verb meaning ‘freeze’ is morphologically marked, while the intransitive version remains unmarked (i.e., el- ‘freeze (intransitive)’ versus el-li- ‘freeze (transitive)’).

Numerous empirical studies have shown that the morphological marking of the causative alternation correlates with spontaneity (Croft 1990; Haspelmath 1993, 2008, 2016; Haspelmath et al. 2014; Nedjalkov 1969). Haspelmath (1993: 103) contends that the more spontaneous an event is, the more likely it is to be expressed with a marked causative:

[A] factor favoring the anticausative expression type [=marked anticausative] is the probability of an outside force bringing about the event. Conversely, the causative expression type [=marked causative] is favored if the event is quite likely to happen even if no outside force is present. (Haspelmath 1993: 103; modified)

This study has been extended, and its results have been reinterpreted with a novel explanation in Haspelmath (2008, 2016) and Haspelmath et al. (2014). In Haspelmath’s view, the recurrent pattern in the encoding of the alternation reflects a principle of efficient communication: unmarked forms remain unmarked due to their high frequency of use (as measured by corpus frequency) irrespective of the lexical properties of verbs.

The relation between the lexical properties of verbs and formal encodings of the causative alternation has been examined in a recent corpus study by Samardžić and Merlo (2018). They revisit the concept of external causation, defining it as the likelihood of an external agent’s involvement in an event described by a verb. They refer to this property as Sp for spontaneity (following Haspelmath [1993]), and consider the degree of Sp as a scalar component integrated into the lexical representation of alternating verbs, assigned as a common property across the causative and noncausative realizations (Samardžić and Merlo 2018: 905). Statistical analyses of data extracted from the English component of the parallel corpus Europarl reveal that the corpus-based measure of Sp correlates with the rank of verbs based on the C/A ratio calculated from the typology of the morphological marking of the verbs across languages.^[3] Samardžić and Merlo (2018) interpret this correlation as evidence that the probability of external causation is a grammatically relevant lexical property, influencing both the corpus frequency distribution of structural realizations of alternating verbs and the typological distribution of their morphological markings: both are seen as effects of the distribution of Sp-value across verb types. This interpretation posits that the ratio between causative and noncausative uses of a verb serves as a quantifiable indicator of a verb’s lexical property.

Another notable result of this work is that the Sp-value exhibits a normal distribution across a large set of verbs listed by Levin (1993). Given that a normal distribution is symmetric and concentrated around the mean, this finding suggests that most alternating verbs can be expected to describe events with approximately a 50 percent likelihood of involving an external causer. Samardžić and Merlo’s (2018: 914) analysis further shows that while most verbs are neutral regarding external causality, only some tend to be assigned extreme Sp-values. If the majority of alternating verbs indeed describe events whose external causer can be present or absent with similar probability, this raises several questions unaddressed in Samardžić and Merlo (2018): what determines the appearance of the cause argument in a given context? Why one variant is chosen rather than the other when, in principle, both might be possible? Is the choice a matter of lexical, semantic, discourse or contextual considerations?

These questions are taken up in Lee’s (2023) work, which tests several predictions on the relationship between the frequencies of causative and noncausative uses of alternating verbs and the ease with which the ultimate cause of an event can be identified in specific discourse contexts. Figure 1 illustrates the classification of cause types based on ease of identification (Lee 2023: 362).

Figure 1:

Classification of cause types according to the ease of identification.

Agents are regarded as the prototypical causes (Croft 1991; Lakoff 1990; Talmy 1976, 2000) and are typically isolatable as the ultimate cause because they often have intentions to bring about specific changes. In contrast, when an event is nonagentive, determining its ultimate cause can prove more challenging. Rappaport Hovav (2014) contends that sentences such as The window opened and The vase broke are easier to accept than sentences such as The table cleared and The skirt lengthened because it is easy to imagine a scenario in which the speaker sees the change of state but not the act which brings it about: i) the change of state is brought about by some nonagentive cause lacking intentions (e.g., by the wind), and then the cause may not be easily identifiable; (ii) the change of state is brought about by the action of an agent from a distance. When there are a number of causes normally present simultaneously, it becomes even more difficult to identify the ultimate cause of an event. This happens in various circumstances when a change is recoverable by default or unidentifiable from context. Many natural events of change have default causes. As noted by Rappaport Hovav (2014: 24), there are a number of causes normally present simultaneously for changes which come about in the normal course of events (e.g., hair growing longer, trees growing, skies clearing) or those that occur gradually over the course of time. Thus, it is generally difficult to identify a single, isolatable cause for such changes.

One of the main results of Lee’s (2023) corpus study of 12 change-of-state verbs is that there is a strong positive correlation between the ease of identifying the cause and the frequency of verb uses: verbs predominantly used as a causative tend to have a higher percentage of agent causers, a cause type characterized by the greatest ease of identifiability, whereas verbs that are more frequently used as a noncausative cause are associated with cause arguments that are less easily identifiable. A further notable result is the increased proportion of nonagentive causers in the causative uses of verbs that are used more frequently as a noncausative. Nonagentive causers are less frequent in verbs that are predominantly used as a causative. These findings are consistent with the results of Heidinger and Huyghe’s (2024) recent corpus study on French change-of-state verbs. Furthermore, they offer corroborative evidence of a relationship between the noncausative use of a verb and the semantic role of its transitive subject: the more frequently the transitive subject of a verb is a nonagentive causer (as opposed to an agent or instrument), the more likely the verb is used as a noncausative (as opposed to a transitive causative).

In addition to presenting these novel empirical findings, Lee (2023) proposes an account for the observed correlation between cause types and verb uses based on communicative efficiency. Developing the line of argumentation in Levshina (2022), she argues that the correlation between agents and the causative variant, as well as between nonagentive causers and the noncausative variant, stems from the principle of efficient language use, which posits a positive correlation between benefits and costs in communication. According to this principle, high costs correlate with high benefits, and low costs correlate with low benefits. Therefore, language users are incentivized to invest more effort and time in conveying information that yields greater benefits, and to expend less effort and time on information of less utility (Levshina 2022: 22).^[4] From the efficiency perspective, the causative variant, which specifies the cause of a change of state, involves greater costly but is simultaneously more informative than the noncausative variant, which expresses just the change of state. As Rappaport Hovav (2014) has noted, the choice of the causative variant is preferrable unless the extra information, namely, the cause of the change of state, is redundant, unimportant, or unidentifiable. Consequently, the high costs associated with producing a more complex structure are justified by the substantial relevance and significance of the identifiable agent in describing a change of state and the enhanced informativeness of the causative variant. These constitute the communicative benefits in a broadly encompassing sense and ultimately motivate the efficient division of labor between the two causative alternation variants.

This idea is compatible with Rappaport Hovav’s (2014, 2020) pragmatic account of the causative alternation based on Gricean maxims of conversation. However, previous corpus-based studies, including Lee’s (2023) recent work, have utilized monofactorial analyses to test the impacts of individual semantic or contextual factors. There remains a lack of research employing the multifactorial analysis necessary to determine whether these individual factors independently affect the corpus frequency distribution of the alternation variants when tested concurrently. The current study aims to fill this gap by utilizing three multifactorial methods to identify a range of factors influencing variant choice in the causative alternation and to assess the impact of each factor while controlling for others.

In summary, accounts of the causative alternation must explain its limited productivity. Accounts that incorporate both semantic and contextual constraints are most successful at meeting this challenge. Precisely what factors play a more important role and how they interact in constraining the alternation is a matter for continued exploration. The following sections discuss a corpus study that tested the effects of the semantic and contextual factors known from the literature – externally versus internally caused change of state, intentionality, and the contextual identifiability of the causer – against a large set of alternating verbs.

3.1 Procedures

3.1.2 Data extraction

The data source for this study is the BNC, a 100 million word collection of samples of written and spoken language from a wide range of sources originally compiled by Oxford University Press in the 1980s − early 1990s. The BNC was selected as the data source for a first step toward multi-corpora analysis aimed at comparing the use of identical verbs additional English corpora representing more recent language use (e.g., the BNC2014, the ANC, and the COCA).

To conduct the multifactorial analysis required for this study, it was essential to extract the relevant corpus data from the raw BNC texts in a format amenable to further statistical analysis. Consequently, we established a pipeline for systematic data collection and processing; this ensures the reproducibility and replicability of our results. This section outlines the data processing pipeline depicted in Figure 2.

Figure 2:

Procedures for data collection and analysis.

To identify instances of the two patterns involved in the causative alternation, it was necessary to add syntactic information to the corpus data. For this purpose, the raw BNC text comprising unannotated sentences was parsed automatically using a dependency parser. In this study, dependency parsing was performed using Stanza, an open-source Python natural language processing toolkit that supports 66 human languages and was developed by the Stanford NLP Group (Qi et al. 2020). Dependency parsing builds a tree structure of words from the input sentence, representing the syntactic dependency relations between words. The resulting tree representations, which follow the Universal Dependencies formalism (https://universaldependencies.org), are useful in many downstream applications. More specifically, this initial parsing step produces a corpus in a format known as the CoNLL-U format. Such a metadata file consists of split sentences that specify various types of information about each word in the sentence, including the word’ form, its position in the sentence, its lemma, its part-of-speech (pos) tag, and dependency relations (deprel)). This information facilitates the identification of relevant corpus data of interest in this study, enabling the automatic retrieval of causative and noncausative uses of verbs. Figures 3 and 4 provide the graphical dependency representation for the example sentence in (8a) and (8b), respectively.^[5]

Figure 3:

Graphical dependency representation for the sentence: The wind cleared sky.

Figure 4:

Graphical dependency representation for the sentence: The sky cleared.

In Figures 3 and 4, all words in a sentence are connected through dependency relations, defined as binary relations that hold between a governor (also known as a head) and a dependent. For instance, the subject and object depend on the main verb; determiners depend on the nouns they modify; and so on. The labeled arc is directed from the governor to the dependent.^[6]

The second step involved extracting causative and noncausative instances of verbs classified as alternating in Levin (1993) from the parsed corpus. In this study, we will consider only instances of active transitive realizations and intransitive realizations. The passive instances were extracted but excluded from the analysis, as including passives complicates the interpretation of the regression model.^[7] Furthermore, we focused only on one-word uses of the verbs, excluding complex verbs composed of [verb + particle], known as verb particle constructions (VPCs), such as break down and melt away. The exclusion was based on the observation that not all multi-word uses of verbs under scrutiny exhibit the causative alternation. For example, the verb break is frequently used in VPCs, and shows some concentrated use as a sense that does not have a change-of-state interpretation, as in the war broke out. The VPC break out has a coming-into-existence sense and lacks the causative counterpart that can be paraphrased as ‘cause the war to happen’. As it is difficult to exclude only such VPCs, all VPCs were excluded from the analysis.

To retrieve only the relevant data, we wrote a Python script that identified the one-word causative and noncausative instances of 354 alternating verbs. Our script identified the causative instances by selecting sentences that contain dependency relations tagged as “nsubj” (nominal subject) and “dobj” (direct object) and governed by “root”. To extract the noncausative instances, the extraction script utilized an algorithm that filtered out sentences lacking a “dobj” dependency relation governed by “root”. Additionally, the script utilized an algorithm to exclude VPCs and passive sentences by identifying specific patterns that include the presence of “aux:pass” (passive auxiliary) and “compound:prt” (phrasal verb particle).

The data extracted during the initial two steps of the pipeline necessarily contain processing-related errors because they are collected automatically from an automatically parsed corpus. Consequently, the third step involved filtering the data to remove such errors. A manual evaluation of the dataset was conducted to assess the extent to which the analyses were accurate. The validation reveals a bias toward identifying the extracted instances as noncausatives when they are not, and this bias was found in instances of a majority of verbs included in the analysis. The noncausative bias is primarily attributed to the incorrect analysis of English passive constructions headed by the verb be such as The door was opened as noncausatives rather than passives. Other cases of incorrect analyses resulted from parsing errors that led to the assignment of an incorrect form. For instance, an intransitive form was assigned where a transitive form with an elided object should have been, or the actual forms found were not verbs but adjectives or nouns (e.g., open, close, clear, dry). As the current extraction method cannot automatically handle these cases yet, such errors were manually removed from the dataset.

Furthermore, we excluded the following constructions from the samples of relevant tokens eventually analyzed, because controlling for the direct causation interpretation and the change-of-state interpretation were paramount:

1. Transitive uses that are not causative: the subject is not the cause as in She broke her leg;

2. Uses that do not have a change-of-state interpretation as in the book opens (=begin) with an introduction to the theory of syllables;

3. Noncausatives modified by an adjunct phrase that does not express a direct cause as in The service in the bank improved as a result of her complaint: in this

4. example, her complaint is not construed as a direct cause of the change. Hence,

5. this sentence does not have the paraphrase Her complaint improved the service. A better paraphrase would be: The bank improved their service because of/as a result of her complaint.

Finally, we excluded idiomatic uses of the verbs to control for the influence of the non-compositional meaning on their ability to alternate and the interpretation of the sentences.

The filtering step produced a total of 15,763 extracted instances of alternating verbs, encompassing 4,566 causative instances and 8,761 noncausative instances.^[8] For practical reasons, we chose to analyze a subset of the data, specifically 3,864 causative and noncausative instances of break verbs, deadjectival verbs, and other alternating verbs of change of state. Some verbs within these classes do not occur in the causative or noncausative variants in the dataset and were thus excluded. The final data set includes 13 break verbs, 34 of deadjectival verbs, and 88 other alternating verbs of change of state. We analyzed corpus occurrences of causative and noncausative uses of these 135 verbs, following the methodology outlined in the subsequent subsection.

In order to conduct a statistical analysis, the data destined for the multifactorial analysis were recorded in an Excel spreadsheet and manually annotated with five variables (verb, realization, intentionality of causation, contextual recoverability, and external causality of change of state) to compose the final dataset for this study. Figure 5 displays the first line of the file containing the final dataset for the verb abate:

Figure 5:

Screenshot of the file containing the final dataset (the first line for abate).

3.2 Coding predictor variables for multifactorial analysis

This section provides an overview of the three semantic and contextual factors included as predictor variables in the multifactorial analysis, as listed in Table 3. The reliability of the coding, as indicated by the high inter-rater agreement score (Light’s kappas) exceeding 0.80, establishes the credibility of the coding schema.

3.2.2 Contextual identifiability

This variable captures the degree to which the cause of a change of state is contextually identifiable, differentiating between recoverable cause (RC) and nonrecoverable cause (Non-RC). A recoverable cause is defined as one that can be contextually identified, whereas all other cases were considered a nonrecoverable cause and classified into unknown cause (UC) or identified cause (IC).

An identified cause is defined as a cause of a change of state explicitly singled out as the primary cause in the clause where the alternating verb is used. Identified causes and the change they bring about are specified in the same clause, enabling their clear identification as the ultimate cause of the change without having to rely on the surrounding context. The following two cases were considered identified causes: (i) causes realized as the subject or the external argument in the causative variant, and (ii) causes realized in other positions of the clause where the investigated verb is used. Research on lexical causatives has established that the subject of the causative is necessarily characterized as a cause that is conceptualized as both proximate and ultimate (Levshina 2022, Rappaport Hovav 2014, Wolff 2003).^[10] The data coding procedure adopted this widely accepted perspective by counting causes realized as the subject in the causative variant as instances of an identified cause. Further cases of an identified cause from the corpus are given in (12). In (12a) the cause (Bannister) fulfills two roles: it serves directly as the object of the matrix verb permit, and indirectly, through indexing, as the external argument in the complement clause containing the verb break (Radford 1988, 1997); in (12b) and (12c), the prepositional phrase containing the cause modifies the verb phrase headed by the verb (melt and broke).

(12)

a.	The record was broken by scientific training and pacing in which two first-class athletes sacrificed themselves to permit Bannister <to break the record>.  (BNC W:non-ac:soc_science, A6Y-24)
b.	… these crystals melt in hot water.  (BNC W:fict:prose, H8A-2435)
c.	The window broke from the pressure/from the explosion/from Will’s banging.  (Heidinger and Huyghe 2024: 191, (23a))

The restriction to causes realized in the same clause where the change is specified excludes causes specified or hinted in the surrounding context from cases of identified cause. Instead, these causes were coded as recoverable causes, that is, causes that can be contextually identified. The following types of cause were regarded as recoverable causes: (i) previously mentioned cause, (ii) hinted cause, (iii) default cause, and (iv) type-inferable cause. The first two types are contextually identifiable based on information available from the immediate discourse context, whereas the latter two types are considered recoverable because they are part of what discourse participants know about the way the world works.

Previously mentioned cause. A cause can be recoverable when it has been established in the surrounding context, that is, in the clauses or sentences preceding or following it. These are cases such as those illustrated in (10b) and (11a) above, wherein the cause of the change mentioned in the preceding clause licenses the noncausative variant of the verb. In (11a) the causing action was performed unintentionally, while in (10b) it was performed intentionally.

Hinted cause. A cause can also be recoverable when the cause might be subtly suggested in the surrounding context. An example is (13) uttered in the context of an event with many tables and food served by waiters. In such contexts, a noncausative variant may be used to describe the event.

(13)

As the night wore on, <the tables slowly cleared> and there was nothing left for the late comers to eat.  (Rappaport Hovav 2014: 26, (81))

Default cause. Another type of recoverability, as noted by Rappaport Hovav (2014), is the default cause, defined as causes that are part of discourse participants’ common knowledge about the way the world works: they are recoverable by default. Examples of a default cause of the event are causes of changes which occur in the normal course of events such as the lengthening of hair, the growth of a tree, and the clearing of the sky. According to Rappaport Hovav (2014: 24), one attribute that distinguishes default causes from identified causes is the presence of multiple simultaneous causes for changes that occur as part of the normal course of events. Consequently, it is typically difficult to pinpoint a single, isolatable cause for such changes. When the cause is recoverable by default, the use of the causative variant would be odd.

Type-inferable cause. Another type of cause that is recoverable from world knowledge about the properties of eventualities is termed a type-inferable cause. An example with a type-inferable cause is given in (14). The when-clause in this example describes the closure of a hospital effected through the involvement of an agent who can be inferred to be typically responsible for such an event, namely, the hospital’s owner (see also (9) above). The precise identity of the agent is unimportant and need not be specified. This seems to be the reason for the general avoidance of the type-inferable cause in causative uses, although it is found in agentless passives (refer to Thompson [1987: 499] for further discussion).

(14)

Careful plans have been made for these people so that when <the hospital eventually closes> they will not find themselves on the streets.

(BNC W:non-ac:polit_law_edu, HH3-9982)

The noncausative variant may also be appropriate when neither speaker nor the hearer is aware of the ultimate cause of the event. The following types of cause were considered unknown causes (UCs): (i) cause from a distance, (ii) cause for a gradual change, and (iii) unidentifiable cause.

Cause from a distance. As noted by McCawley (1978), the choice of the verbs come and go determines the orientation of the speaker: in (15) the speaker is probably positioned in the lunchroom. Consequently, this example allows the hearer to infer that the speaker does not know how the door was opened.

(15)

<The door of Henry’s lunchroom opened> and two men came in.
(McCawley 1978, cited in Rappaport Hovav 2014: 22, (59a))

Cause for a gradual change. Another type of unknown cause is the cause for a gradual change discussed by Rappaport Hovav (2014). Consider (16):

(16)

With the 1929 stock market crash, <skirts lengthened> but kept their narrow silhouette, with longer waistlines.  (Rappaport Hovav 2014: 27, (83))

This sentence does not describe a change in the length of an individual skirt, but rather a change in the kind of skirt. It reports a situation in which, over time, there are a variety of instantiations of the kind, and when comparing one to another over this period, the length is seen to increase. It is likely that multiple causes (and reasons) contribute to the gradual change in the kind, which are generally diffuse and not uniquely identifiable.

Unidentifiable cause. In numerous instances, an identifiable cause could not be discerned and the available contextual information did not indicate the cause of a state change. These instances, characterized by unclear and unidentified causes due to insufficient information, were categorized as cases of unknown cause. An example with an unidentifiable cause is given in (17).

(17)

People can enjoy reading about you even if they don’t like your music and if enough of them do, it improves your chances of more coverage the next time you need it. … Although the total press coverage of pop music increased dramatically over the last ten years, <the quality hasn’t improved> and the power of music papers to make and break an artist has greatly dissipated.

(BNC W:misc, A6A-449)

From the immediate discourse context, it is apparent that the influence of music papers augmented the probability of obtaining more press coverage for artists and their music. However, the reason for the persistently unimproved quality of pop music, despite a significant increase in total press coverage, remains unclear.

Determining whether the cause for a change is a recoverable cause or an unknown cause required a careful analysis of the context. We used the CQPweb interface of the BNC to access preceding and subsequent context. CQPweb (Hardie 2012) is an online corpus analysis tool that serves as an interface to the corpus workbench software (CWB) and its effective corpus query processor (CQP) search utility (https://cqpweb.lancs.ac.uk). The most convenient method to view context is by searching for query expressions and concordancing the results. As shown in Figure 6, by default, the CQPweb system displays only the immediate text unit containing the search element (in this case, But the age structure within this total has changed).

Figure 6:

Viewing a concordance with surrounding sentences in the CQPweb.

To examine more context, one can simply click on the search element (underlined and center-aligned in the concordance line). This action opens a new window displaying the entire text from which the search element originates (Figure 7). This configuration enables reading the context immediately before and after the sentences (highlighted in the window).

Figure 7:

Displaying more context in the CQPweb.

4.1 Descriptive statistics

We analyzed the distribution of the causative and noncausative variants in relation to the three factors described in Section 3.2.^[12] Figure 8 shows the proportional distribution of the two variants and the absolute frequencies according to the levels of the factors.

Figure 8:

Proportional distribution of alternation variants by intentionality, cause identifiability and spontaneity.

Intentionality. The proportional distributions in Figure 8 confirm our hypotheses regarding the influence of intentionality of causation as posited in (20). We can see that the causative variant is proportionally preferred over the noncausative variant when the causing action is intentional, aligning with (20a). Conversely, when the causing action is nonintentional, the noncausative is preferred, as outlined in (20b).

Recall that the category of nonintentional causes encompasses inanimate causes and nonintentional animate causers. While animates may act intentionally or unintentionally in precipitating a particular event, inanimates lack the capacity for intentional action. This difference between inanimate causes and animate causers suggests that inanimate causes are more closely linked to nonintentional causation. To investigate whether this is the case in our data, we have further analyzed nonintentional causes that are explicitly expressed in the clause of the COS verb or identifiable contextually by their ontological class. Figure 9 shows the frequency distribution of nonintentional causes among the two variants based on their animacy.

Figure 9:

Frequency distribution of animate and inanimate nonintentional causes.

We can see that there is a marked asymmetry between inanimate causes and animate causers: nonintentional causes are predominantly inanimates in both variants. There were merely 43 instances of a nonintentional animate causer realized as a causative subject, constituting only 6.56 % of the 655 causatives with a nonintentional causer subject in the dataset; the number of noncausative instances with a nonintentional animate causer totals 148, accounting for 7.86 % of all 1,884 noncausative uses with a nonintentional causer. All these cases are recoverable causes either overtly expressed in the local linguistic context or hinted at in the surrounding context, as exemplified in (11a) and (13) above, respectively. While inanimate causes can be expressed as adjuncts in noncausatives as illustrated in (12c) above, animate causers cannot (Alexiadou and Schäfer 2006: 41):

This difference between inanimate causes and animate causers is consistent with the proposal that inanimate causes are more closely associated with the noncausative variant than are animate causers (Alexiadou and Schäfer 2006; Heidinger and Huyghe 2024). Further evidence supporting this is provided by analyzing the realization patterns of prototypical examples of inanimate causes – natural forces and conditions. We analyzed a total of 215 instances of natural forces and conditions identified or identifiable as causes, examining the frequency of their realizations and the cause types. The results of this analysis of natural causes are summarized in Figure 10.

Figure 10:

Frequency distribution of subtypes of natural causes.

Figure 10 shows that natural causes predominantly appear in the noncausative variant. Concerning natural causes in noncausatives, the discrepancy between RCs and ICs (adjunct PPs in noncausatives) is striking: RCs significantly outnumber ICs. When assessing the relative frequency of subtypes of RCs, it was found that 102 instances were default causes (DCs), and 53 instances were RCs mentioned or suggested in the local linguistic context. It is noteworthy that noncausatives with these RCs take up 80.31 % of all instances of noncausatives with a natural cause. Examples of noncausatives with recoverable and identified natural causes from the BNC are given in (24) and (25), respectively. We revisit to these findings in Section 5 and propose an efficiency-based explanation for them.

Contextual identifiability. The proportional distributions in Figure 8 also confirm our hypotheses regarding the influence of contextual identifiability as discussed in (21). We observe that the causative variant predominates when the cause is clearly singled out and identified; specifically, when the cause and the change are specified in the same clause (NRC-IC). Conversely, in cases where the cause is either recoverable or unknown, the pattern shifts: the noncausative prevails if the cause is recoverable (RC), and it is exclusively selected when the cause is unknown (NRC-UC). There are 49 instances of causative uses with a recoverable cause, constituting merely 3.02 % of the total 1,620 causative instances in the dataset. Predominantly, such cases of recoverable cause subjects in the causative uses of the alternating verb in our dataset are omitted subjects of finite verbs, which refer back to the subject in the immediately preceding sentence or to a salient entity introduced earlier in the discourse, as illustrated in (26).^[13]

External causality. The proportional distribution of the realization of internally caused changes of state is in line with our hypothesis in (22b). Figure 8 shows that the noncausative is the categorical choice when describing an internally caused COS. There are no attested instances of causative uses describing an internally caused COS within our dataset. As already mentioned, we excluded instances of transitive uses of the analyzed verbs that are not causative. Examples of the transitive uses removed from the final dataset are given in (27):^[14]

However, the proportional distribution of the realization of externally caused changes of state is not consistent with our hypothesis regarding the influence of externally caused changes of state as outlined in (22a). As shown in the figure, the noncausative variant is marginally favored over the causative variant even when the sentence describes an externally caused COS.

In the subsequent subsection, we will use conditional inference trees and conditional random forests in order to identify the set of independent variables that significantly differentiate between the two variants.

4.2 Tree and forest analysis

Conditional inference trees (CITs) and conditional random forests (CRFs) were introduced to linguistic analysis in a paper by Tagliamonte and Baayen (2012) and have been fruitfully used in models of linguistic variation (Hundt et al. 2021; Levshina 2015, 2016, 2020, 2022; Szmrecsanyi et al. 2016). These methods are non-parametric permutation approaches that do not assume a specific distribution of the data, but rather build a model through resampling from the input.

CITs are methods for regression and classification based on recursive partitioning. They split the data recursively into smaller subsets based on predictor variables that co-vary most strongly with the outcome. For each binary split, the data is evaluated for the predictor that best preserves the homogeneity of each split (e.g., all causatives and all noncausatives) at a certain significance level. This splitting process continues until no further splits can significantly increase the subset’s homogeneity with respect to the outcome variable. Despite their utility, single trees have the disadvantage that they very much hinge on the dataset at hand and are hence subject to a high degree of variability. Conditional random forests (CRFs) mitigate this by resampling across a predefined number of trees through a conditional permutation scheme and thus particularly robust to, for instance, predictor collinearity (see Levshina [2020] for details).

CITs and CRFs can be particularly useful in situations where the use of parametric methods (in particular, regression models) may not be suitable, such as in situations characterized by ‘small n, large p’, (where n is the number of observations and p is the number of predictors), complex interactions, non-linearity, and correlated predictors (Tagliamonte and Baayen 2012). They are also noted for their robustness in the presence of outliers. Furthermore, CITs facilitate the intuitive interpretation of complex interactions involving more than two predictors.

Figure 11 shows an individual CIT fitted to the data using the function ctree ( ) in the party package in R. The ovals highlight the names of the variables selected for the best split, along with the corresponding p-values. The levels of the variables are specified on the ‘branches’. The bar plots at the bottom (‘leaves’) display the proportions of the two variants in each end node (‘bin’), which contain all observations with a given combination of features. The number of observations in each bin is noted in parentheses above the boxes.

Figure 11:

Conditional inference tree of the variant choice.

The plot is presented with all significant splits at the 0.05 level and should be interpreted from the top down.^[15] The interpretation proceeds as follows. First, the algorithm looks for the variable most strongly associated with the variant (causative or noncausative). This variable is contextual identifiability (see Node 1). It then divides the data along this variable into two subsets, separating observations with NRC_ICs (cases on the left side of the tree) from those with RCs and NRC_UCs (cases on the right side of the tree). We observe that causative situations involving NRC_ICs are expressed predominantly with a causative (Nodes 3 and 4), whereas those involving RCs and NRC_UCs are realized predominantly as a noncausative (Nodes 7, 8 and 9). Subsequently, the algorithm seeks another variable significantly associated with the variant. On both sides of the tree, the subsequent split occurs with the variable of intentionality (Nodes 2 and 5), indicating that in cases where the cause type is similarly identifiable, the intentionality of the causer becomes relevant. Descending the left-hand branch from Node 2, where the cause is an NRC_IC type, the proportion of causative outcomes is higher in situations of intentional causation (Node 3: 96.7 %) compared to nonintentional causation (Node 4: 80.9 %). Conversely, on the right-hand branch from Node 5, where the cause is recoverable or unknown, the proportions in the bar plots in the node leaves show that noncausative is uniformly preferred in instances of intentional causation with NRC_UCs (Node 8: 100 %) and that noncausative is more prevalent in instances of nonintentional causation (Node 9: 99.2 %), relative to intentional causation with RCs (Node 7: 92.2 %). Note that the variable external causality is not used in the data splitting because it does not significantly correlate with the outcomes (with p < 0.05) in the model. These results support the conclusion that contextual identifiability is the most predictive factor, followed by intentionality, while external causality appears to have a minimal impact in differentiating between the two variants.

The results of the CIT analysis were corroborated by the CRF analysis, where a CRF consists of an ensemble of multiple CITs. In addition to measures of classification accuracy, CRFs facilitate conditional variable importance scores, which show how important each variable is in predicting the outcome (in our study, the use of the causative and noncausative variants), taking into account all other variables and their interactions. To compute this measure for a predictor, the algorithm averages results across numerous trees and measures the reduction in predictive capability when the predictor is randomly permuted. As explained in Levshina (2016, 2020), a notable distinctive feature of CRFs is their ability to ascertain the conditional variable importance measures, offering certain benefits over non-conditional approaches (Strobl et al. 2007). Specifically, these measures are unbiased towards correlated predictors and with regard to the number of categories in a categorical variable. Furthermore, CRFs are superior to other popular random forest methods as they make predictions by retaining information about individual observations in each tree without pruning the trees. Consequently, terminal nodes with a high count of observations exert more influence because they provide more data points for calculating predicted values.

We created a random forest comprising 500 trees, and subsequently computed the variable importance scores.^[16] The variable importance scores are presented in Figure 12 in order of decreasing importance. The horizontal axis represents the conditional variable importance (CVI) for each predictor. The dashed line separates the important scores on the right from the unimportant ones on the left. If a variable is irrelevant, its importance values vary around zero.

Figure 12:

Conditional variable importance (CVI) scores of the variables based on CRFs.

Figure 12 shows that contextual identifiability is the most important predictor for the choice between the two alternation variants (CVI score: 0.337). This variable also predominantly facilitated the splits in the CIT. Consistent with the results of the CIT analysis, intentionality emerges the second most important variable (CVI score: 0.0036). Notably, external causality does not seem to have any discriminatory power in the CRF.

We compared the variable importance scores obtained from the CIF analysis with those derived from non-conditional random forests. The mean decrease accuracy plot in Figure 13 demonstrates that all three variables play a role in random forests:^[17] with contextual identifiability being the most important predictor, followed by intentionality and external causality.

Figure 13:

Gini scores of the variables based on RFs.

To summarize, the results of the CIT, CRF and RF analyses suggest that contextual identifiability is the most predictive factor, whereas the other two variables are less useful for distinguishing between the two variants. In the next subsection, we will use mixed-effects logistic regression modeling to gain a more robust insight into the size of the predictors’ effect, their interactions, and random effects (verb specific effects on variation).

4.3 Mixed-effects logistic regression model

Mixed-effects logistic regression enables us to model binary outcome variables with two types of factors that are mixed in the statistical analysis (Baayen 2008; Baayen et al. 2008; Jaeger 2008, 2011): fixed factors and random factors. Random factors are employed to systematically exclude variation that can be deemed as ‘random’ or unpredictable, thus only indirectly affecting the response variable. For example, factors that could arguably be considered random are the number of speakers in a given corpus or the number of verbs that are found in a given construction. The separation of the effects of random factors permits a more reliable assessment of the effects of the remaining fixed factors.

Our model included realization as a binomial response/outcome variable (i.e., causative vs. noncausative), intentionality (intentional vs. nonintentional), contextual identifiability (RC vs. Non-RC), and external causality (internally vs. externally caused COS) as independent variables, and verb as a random factor. All independent variables were coded using sum contrasts, in which the proportion of responses for each level is compared against the grand mean across all levels, as shown in Table 4.

We started out with a model that included all the fixed factors specified in Table 4 and their possible interactions, successively removing predictors that did not significantly contribute to the overall fit of the model. The final model includes two significant main effects and no significant interaction effect. A summary of the model is given in Table 5.^[18]

In the fixed effects section, we report the β coefficients and the odds ratios associated with each predictor variable along with the significance levels (p-values). Under the Estimate column, the β coefficient indicates the slope (effect size) of the effects of each predictor variable. A positive coefficient indicates that the predictor variable increases the probability of noncausatives (the outcome having value 1), whereas a negative coefficient means that the variable increases the probability of causatives (the outcome having value 0). The odds ratios show similar information, but have a lower bound at zero, and a value of one indicates an equal chance of either variant. If the odds ratio is greater than one, noncausatives are more likely than causatives; if it falls between zero and one, they are less likely. Odds ratios can be interpreted as factors with which the odds of a noncausative variant are increased or decreased given certain predictor changes. For instance, intentionality changes the odds of the noncausative realization by a factor of 7.76, thus making noncausatives 7.76 times more probable in situations of nonintentional causation (having value 1 of the predictor variable ‘intentionality’) compared to the overall average.

Table 5 shows that, consistent with the results of the CIT and CRF analysis, intentionality and contextual identifiability were significant predictors of the realization of the causative alternation, with contextual identifiability emerging as the most predictive factor. Specifically, a causative situation is 125.8 times more likely to be realized as a noncausative when the cause is contextually identified than when it is not. External causality, however, turns out not to be significant, indicating it does not play a significant role in distinguishing between the two variants as the noncausative variant is favored in situations of both internal and external causation (refer to Figure 8).

Figure 14 visualizes the fixed effects of the final model, with predictors situated to the right of the dotted line indicating a high probability of noncausatives. The horizontal bars extending from the points are 95 % confidence intervals. We can interpret the figure as follows. The odds ratios for both significant fixed factors are greater than 0, and their 95 % confidence intervals exclude 0, indicating that they significantly increases the probability of noncausatives. Contextual identifiability is particularly notable for its preference for noncausatives.

Figure 14:

Predictors for the final model, with 95 % confidence intervals for the odds ratios.

In the random effects section of Table 5, variance and Std. Dev. (standard deviation) indicate the variability from the predicted values due to the random effects added to the model. Consequently, these values reflect the fact that every observation has some unexpected factors affecting the realization of the causative alternation in addition to the fixed effects. The random effect variable, verb, has a standard deviation of 1.5, which translates into an average difference between verbs in their noncausative use of approximately 37.5 %. In other words, while genuine differences exist between verbs regarding their causative or noncausative realizations, the model takes this into account, such that the fixed effects in Table 5 represent the effects of our predictors beyond the verb specific effects. Figure 15 below shows the random effects for 9 sample verbs, with 95 % confidence intervals. Grow and explode are notable for their preference for noncausative realization, whereas open and fill are linked to a higher probability of causative realization. Increase and diminish exhibit no significant deviation from the overall average.

Figure 15:

Random effects for the final model, with 95 % confidence intervals.

To summarize discussion in this section, the results of the statistical analyses show that the contextual identifiability and intentionality of the causer influence the causative alternation as predicted by the hypotheses in (20) and (21) above. Multifactorial analyses further reveal that these two factors are significantly associated with the realizations in both the CIT/CRF and mixed-effects logistic regression models, with contextual identifiability emerging as the most important predictor: causative situations with a clear identifiable agent are predominantly realized as a causative, whereas those with a less clear, nonagentive cause are predominantly expressed with a noncausative. These findings underscore the pivotal role of contextual and semantic properties of the causer in distinguishing between the two variants over the involvement of an external causer.