Not all English Resultative Constructions (ERCs) are equal: The acquisition of ERC by Spanish speakers

ABSTRACT The English Resultative Construction (ERC) is a satellite-framed structure with no identical equivalent in Spanish. In a series of studies, we analyzed and compared recognition (acceptability judgment task) and comprehension (sentence comprehension task) of three ERC subtypes with the English Depictive Construction (EDC) (which has a Spanish counterpart) by Spanish speaker learners of English as a Foreign Language (EFL). Results showed that: 1) EDCs were better recognized than ERCs by L2 learners, but highly proficient participants were closer to English native speakers’ performance, 2) Less proficient EFLs comprehended EDCs better than those ERC subtypes that were further from Spanish (ERC Property and Fake Reflexive). We interpret our findings in terms of an interlinguistic distance gradient, where those constructions present (EDC) or closer (ERC-Path) to L1 are more readily acquired. This effect seems more prominent at lower EFL proficiencies, and fades as proficiency increases, evolving towards a more native-like pattern.


Introduction
Speakers of different languages recourse to different grammatical structures to describe the same events they witness in the world. Spanish and English sit on opposite sides in a verb-versus satellite-framed continuum (Talmy 1985(Talmy , 2000. Regarding the expression of motion events, satellite-framed languages such as English usually express Manner of motion in the main verb of a sentence ("walked," in (1)), while verb-framed languages encode bounded paths in the main verb (entró 'entered', in (2)).
(1) The boy walked into his bedroom.
The boy entered to his bedroom. 'The boy entered his bedroom.' RRG is a grammatical theory that describes linguistic phenomena in terms of psychological and typological adequacy. It focuses on characterizing the syntax-semantics interface, in particular, the systematic realization of the semantic arguments of verbs into syntactic functions in a sentence, a correlation that is also known as 'linking'. In RRG, linking is a projective operation that relies essentially on the interplay between Vendlerian's lexical aspect structures (i.e., Aktionarten [Vendler 1967;inter alia]) and the Actor-Undergoer Macroroles. Verb meanings are described in terms of structures containing an Aktionarten category and argument slots. The relative positions of these slots reflect different levels of activity, which in turn determine their projection into Actor or Undergoer, namely that the most active argument is projected to Actor while the least active as Undergoer. In consequence, the Actor is linked to the Subject while the Undergoer to the Direct Object of the sentence. In a complex predicate structure, two predicates share the same linking domain and they behave as one single but complex verb. Consequently, all the arguments are linked to the same syntactic realization as Subject and Direct Object (París 2012).
ERCs can be classified in three different groups: Path Resultatives (ERC-Path) as in (1) or (4) (Goldberg & Jackendoff 2004;Napoli 1992), Property Resultatives (ERC-Property) as in (3)-repeated here as (5)and Fake Reflexive Resultatives (ERC-Fake) as in (6) (Rappaport Hovav & Levin 2001). 1 We shall also contrast these subtypes of ERC to the English Depictive Construction (EDC), exemplified in (7). We argue that the grammatical properties of these ERC subtypes allow us to order them in a gradual complexity gradient. Complexity can be grounded on grammatical properties, on processing cost, and on acquisition (Spada & Tomita 2010); these different types of complexity are typically correlated but they might not need to be. We follow the insight that guides the literature on the topic (Bulté & Housen 2012;Hulstijn & de Graaff 1994;Housen & Simoens 2016;Palotti 2015) and assume that complexity is governed by the number of restrictions that are needed in order to arrive at a correct form (Hulstijn & de Graaff 1994). We only address here grammatical complexity and propose that it can be determined in terms of the sheer number of grammatical properties of a given structure. 2 In order to operationalize this theoretical construct, we make explicit the grammatical restrictions that are needed to characterize each of the target constructions; this means that we filtered out those properties that are shared by (almost) every simple sentence (e.g., having a subject or having a finite verb). These restrictions can be either formal, semantic, or part of the syntax-semantics interface. 3 ERC-Path-exemplified in (1) and (4)-is the least complex of all three ERC types. It consists of a verbal phrase (VP) modified by an (optional) adjunct, typically a Prepositional Phrase (PP). This is a simple and ubiquitous structure. The meaning of the verb "push" in (4) describes an Activity containing the motion of an entity along a Path that is bound by a PP ("off the stage"), boundary 1 Goldberg & Jackendoff (2004) propose another subtype of ERC (that is verbal, not constructional), namely the Verbal Resultatives as in They made Peter angry. In addition, within the constructional subtypes (Path, Property and Fake Reflexive ERC), Goldberg & Jackendoff (2004) propose other special-or marginal-cases. They propose Spit case like Bill spit out of the window; Follow case Path like Bill followed the road into the forest; Intransitive Sound Emission like The trolley rumbled through the tunnel; and Dancing Mazurka as Mark danced mazurkas across the room. In our experiments, we included transitive ERC-Path, ERC-Property and ERC-Fake Reflexive subtypes as they are broader categories of resultatives that hold significant differences among themselves. 2 The goal in some of the cited works on complexity (e.g., Palotti 2015) tends to be more ambitious than the one we set out for us here. While Palotti talks about the complexity of a linguistic system (i.e., English) or, for instance, the morphology of a system, we just aim at contrasting the complexity of subtypes of the same construction (ERC) and a construction that shares central grammatical properties (i.e., EDC), as both correspond to secondary predicates. 3 Furthermore, these properties are part of the inventory of any grammatical theory and they are not exclusive of ERC or EDC nor even of English or Spanish. It is the case that the description of a particular property can be theory dependent; yet, we should be able to agree on the underlying phenomenon, whether described in terms of one theory or another. that the Theme argument ("the piano") has crossed. 4 The meaning of the sentence is telic, but this telicity is typically obtained by a lexical aspect type shift (de Swart 2011) since an Activity verb turns into an Accomplishment. This shift takes place in (1) by the addition of a telic PP and in (4) by the combination with a delimited NP complement followed by a telic PP. This is also possible in cases where the direct object is not lexically licensed by the main verb (as in Jim danced Carol off the stage). 5 Example (4) contains a causal relation; in contrast, (1) shows that not every ERC-Path involves a causal relation. 6 ERC-Property-exemplified in (5)-is more complex than ERC-Path. It contains a complex sentence-specifically, a secondary predication structure (Rapoport 2019)-with two predicates ("dance" and "tired"). The second predicate does not project into a full-fledged clause since it does not have an explicit subject expressing its semantic argument ('tired(x)'). Rather, the reference of this argument is fulfilled by an antecedent NP ("John") that is the direct object of a different predicate, that is, the matrix verb. This direct object is said to be the 'controller' and it is coindexed-that is, coreferential-with the argument of the second predicate. There are further restrictions regarding this 'controller'; first, it cannot be the Actor of the matrix verb, that is, the matrix subject ("*Mary i danced John tired i "); second, it cannot be the object of a PP ("*John shot at the mouse j dead j "). In addition, the temporal interpretation of the second event (i.e., becoming tired) is realized by extending the tense operator in the matrix verb to the second predicate 7 ( van Valin 2005). Furthermore, the ERC-Property in (5) displays a non-lexically licensed direct object ("John") of an unergative intransitive verb ("dance"). As we see in (8), "dance" is an intransitive verb that cannot take by itself a direct object as it needs the final satellite to form a grammatical sentence. Therefore, the direct object in (5) is not a semantic argument of the matrix verb, which in RRG amounts to a significant piece of evidence of the 'constructional' (non-compositional) nature of this structure, a claim that is shared with Construction Grammar (Goldberg & Jackendoff 2004). Semantically, the ERC-Property in (5) comprises a lexical aspect type shift ("dance" is a pure Activity that turned into an Accomplishment in "Mary danced John tired"). Crucially, it also includes a causal relation that is not contributed by the non-causative predicate "dance" nor by "tired"; it is constructionally inserted. Another restriction concerns the aspect of the AP: the adjective cannot denote an individual state property (e.g.: "intelligent," "wise") but it has to be a stage-state one. Finally, ERC-Property encompasses a change in the lexical type of the predicate: the AP denotes a state (be-tired(x)) but ERC turns it into a change of state (Become be-tired(x)). Finally, we argue that ERC-Fake is the most complex ERC subtype. As the prototypical example in (6) shows, ERC-Fake has all the salient formal properties of ERC-Property: it is a secondary predicate structure with a non-lexically licensed direct object of an intransitive verb and a non-lexically contributed causal relation. Yet, ERC-Fake further adds the fact that the direct object is a reflexive 4 A quick search in the COCA corpus (Davies 2015) of the form "walked + preposition" resulted in 56.350 tokens from which 39.558 (70.2%) were undoubtedly Goal/Source introducing prepositions (e.g., "into," "to," "onto," "from," "away"). The other 30% is distributed among prepositions with unbounded Path meanings (e.g., "towards") and any other meaning. If unbounded Path prepositions with Goal/Source are included (e.g., "towards"), the percentage rises to 85.68. In sum, the vast majority of the walking event descriptions involve a specification of the Goal/Source. 5 Non-lexically licensed DOs are also known as non-argument DOs (since they are not arguments in the semantic of the verb) or as ECM (Exceptional Case Marked) (Wechsler 2005). ECM assumes the claim that they are 'deep subjects' of the small clause associated with the AP in ERC-Property or Fake, but receive Case from the matrix verb. 6 If we use the causative paraphrase as a test for the presence of causality ( van Valin 2005, among others), (1) does not allow it (#The boy caused himself to be in the bedroom) whereas (3) allows it since "Mary caused John to be tired by dancing" captures the basic meaning of (3) even if it does it in a less specific way since the cause does not need to be direct (Shibatani & Pardeshi 2002). 7 In terms of RRG, this scope extension is another evidence of a tightest syntactic relation between two predicates: a co-subordinated nuclear juncture ( van Valin 2005). pronominal form that does not have its usual reflexive interpretation (i.e., it is 'fake'), as it does not hold any semantic role assigned by the verb. The necessary presence of the reflexive pronoun can be explained by a strong tendency of ERC called 'Direct Object Restriction', or DOR (Rappaport Hovav & Levin 2001, which states that ERC requires a direct object controlling the predicative phrase. The English Depictive Construction (EDC) illustrated in (7)-and repeated below-is a secondary predicate structure (Rapoport 2019) that holds all the syntactic properties attributed to ERC-Property with the exception of a non-lexically licensed DO. As regard the control relation, EDC is not restricted since either the subject or the direct object of the matrix clause can control the reference of the missed subject of the AP (e.g., in contrast to (7), the controller is the matrix Subject in "John i clean the house naked i "). The controller cannot be though a Beneficiary of a three-argument verb ("*John gave Mary k the letter tired k ") and EDC does not contain a causal relation; the AP "damaged" in (7) is not the result of the main event denoted by "return." 8 In addition, the AP does not include a lexical aspect type shift operation, namely that it does not change the lexical aspect of the sentence as inherited from the verb phrase. In terms of its semantic complexity, EDC -just like ERC-Property and Fake-requires the extension of the tense operator of the matrix over this AP. Finally, the aspect of the AP is also constrained; it cannot be an individual state predicate ("John i cooked the chicken *wise i /naked i ").
(7) John returned the book damaged. Table 1 sums up the grammatical restrictions that characterize each ERC subtype and EDC. There are eleven (11) grammatical properties that are relevant for one or the other construction under consideration.
We distinguish between restrictions that are required to every grammatical version of the subtype from those restrictions that are possible but not obligatory to make a grammatical sentence. We assign two (2) points to the former ones and one (1) point to the latter. Those restrictions that are absent in a given construction are assigned zero (0) points. We end up with the complexity scale from low to high Table 1. Syntactic-semantic properties required (2), possible but not obligatory (1), and absent (0) in English Resultatives and English Depictive constructions.

ERC-Property
In the generative tradition, EDC involves just like ERC a small clause and raising from subject of SC to DO of the main clause (Hoekstra 1988;Williams 1997). However, EDC does not need a hidden causative morpheme. In RRG, ERC is a complex predicate whereas EDC is not: AP is a peripheral modifier of the Core of the main Clause ( van Valin 2005). Even though our analysis stands from an RRG perspective, both Generative Grammar and RRG reflect differently the contrast in complexity of these structures. in (9) that reflects the number assigned to each structure: ERC-Path: 6, EDC: 12, ERC-Property: 19, and ERC-Fake: 22.
(9) ERC-Path < EDC < ERC-Property < ERC-Fake Once the contrastive complexity among the structures is established, we can go a step further and correlate grammatical complexity and proficiency: the more complex the structure, the more proficiency is required to acquire it. On the contrary, low proficient subjects can acquire less complex structures. Syntax studies provide plenty of evidence of the effects of complexity on L2 acquisition. It has been shown that simpler structures are acquired at earlier learning stages, as in the case of active vs passive voice (Wang 2016). In addition, syntactic complexity in production is strongly related to L2 proficiency (De Clercq & Housen 2017;Kuiken & Vedder 2019;Ortega 2003), expanding from coordination to subordination to phrasal elaboration (Wolfe-Quintero et al. 1998). Processability theories (Lenzing 2013;Pienemann 1998) explain these findings proposing that L2 acquisition follows complexity gradients that are constrained by the learners' processing limitations.

ERC and the typological differences between English and Spanish
The typical architecture of a Spanish event expresses the Result in the main verb whereas the Activity leading to it, if needed, is encoded in a peripheral form (typically adverbial). This pattern is observed in (2) and (10)-(11) below, which are the unmarked versions of the ERCs (5) and (6) in any kind of context (either formal or informal), respectively.
Mary let to John tired of so-much to-dance 'Mary danced John tired.' (11) Susy quedó tonta de tanto reír.
Susy came-about silly of so-much to-laugh 'Susy laughed herself silly.' The lack of a Spanish structure that mirrors ERC is not an accident. It is rather part of the fact that English is overwhelmingly a satellite-framed language whereas Spanish is devotedly a verb-framed language (Slobin 2000;Talmy 2000). The difference is about 'construal' in the sense that each language conveys different representations about the same situation in the world. Moreover, the processing of the sentences in each language is also different. The incremental process of understanding an ERC sentence requires speakers to think first of its participant as an Actor of an Activity (i.e., unergative verb "walk" in (1)). In contrast, the same event requires Spanish listeners to think first of that same participant as the Patient of a change of location (e.g., unaccusative verb entrar "enter"). There is also a hierarchical contrast: either an Actor or a Patient are expressed as subjects. Furthermore, the Activity in Spanish can be expressed only if it is the central communicative point of the assertion, that is, the Focus. This is not the case in (1), because the Manner of motion of a person moving into a bedroom is inferable. If the information is not inferable, as in (12) and (13), Manner is highlighted as potential Focus. Contrary to ERC, the verb in (12) does not express Manner but Result, while Manner is codified in an Adjunct de un salto, which attracts pragmatic Focus.
María entered to her bedroom from a jump (13) Mary jumped into her bedroom.
We claim that some ERC subtypes are more 'distant' from Spanish than others. An unmarked form F 1 in language A has an equivalent form F 2 (a form that captures the central meaning of F 1 ) in language B.
It is the case that we can evaluate to which extent F 1 and F 2 depart from one another (Alexiadou 2020;Chiswick & Miller 2005;Putnam et al. 2017). This evaluation requires, first, describing all the grammatical properties of F 1 and F 2 ; second, identifying those properties that are shared and the ones that are not shared (contrastive properties); and, third, analyzing qualitatively the impact of those contrastive properties in the codification of the Spanish rendering of each construction (Alexiadou 2020;Chiswick & Miller 2005;Putnam et al. 2017).
The aim is to establish a qualitative ranking in terms of distance. At the top, we set grammatical properties that are at the core of the sentence like any contrast in the syntactic functions, type of main verbs, syntax-semantic linking of arguments to syntactic functions, aspectual meaning of the main verb, and structural meaning of the main verb. These are core contrastive properties. Contrasts in the periphery of the sentence (e.g., prepositions, adverbs) are considered to be lower in the ranking, and contrasts in pragmatic factors like information structure (marking of Topic and Focus), register (colloquial vs. formal) and markedness (acceptable in any context or restricted to some specific contexts) are even lower. Furthermore, we assume that if F 1 and F 2 share the same grammatical underlying structures, meaning and register, information structure and unmarked status, then F 2 mirrors F 1 .
We use the theoretical construct 'interlinguistic distance' for contrasting the ERC subtypes and EDC with their respective Spanish rendering. ERC-Path examples like (14) can be mirrored in Spanish as shown by the word-by-word rendering. Yet, there are some provisos. First, the preposition hasta ("up-to") instead of a ("to")) has to be used in order to reflect the lexical aspect type shift that turns the Activity conveyed by the meaning of "walk" into an Accomplishment (Fábregas 2010;Ursini 2013). Second, the use of a ("to") sounds literary to Argentinian Spanish whereas Mexican speakers might find it more colloquial.
Javier walked to the school. 'Javier walked to school.' There is a relatively small subset of ERC-Path that involves boundary crossing like example (1) and (4) that cannot be mirrored in Spanish (Aske 1989). The Manner main verb needs to be replaced in Spanish by one that contains boundary crossing like entrar ('enter') or sacar ('take out') and, thus, the underlying syntax-semantics interface is the opposite than the one in the English expression. We refer to this contrast 'reverse linking' since Manner and Result are projected into different syntactic constituents. Notice that example (4) can have a word-byword Spanish counterpart if the boundary crossing PP is replaced by one that portrays a static relation (no boundary crossing) like in (15). This means that, first, the reverse linking for the Spanish version of ERC-Path is restricted to the presence of boundary crossing and, second, the Result (change of location) is codified even if there is no boundary crossing. Therefore, boundary crossing is a subsidiary semantic property in ERC-Path.
(15) Harry empujó el piano contra la pared. Harry pushed the piano against the wall 'Harry pushed the piano against the wall.' In this sense, reverse linking in the Spanish version of ERC-Path is 'downgraded' in the sense that it is very much restricted to a specific semantic property-that is, boundary crossing-that is not central nor required for the notion of Result. Given its restricted semantic nature and its subsidiary meaning regarding Result, this downgraded reverse linking (as in (1) and (4)) sets ERC-Path closer to Spanish than straight reverse linking (as in ERC-Property described in the following section). In addition, ERC-Path sentences that do not involve boundary crossing share all their grammatical properties with their Spanish counterparts except for some minor adjustments needed in the periphery (i.e., using a telic preposition)-like in (15)-and certain pragmatic considerations (i.e., it has a literary use in some Spanish variations).
The Spanish rendering of ERC-Property differs in substantial properties from its source. Verbs in ERC-Property do not entail the Result and, hence, they build 'strong' resultatives (Washio 1997) as shown in (5) or (6). The Spanish renderings of these ERC-Property systematically entail a reverse linking, namely that the Activity/Manner event is expressed by an adverbial adjunct whereas the Result is expressed by the matrix verb. This contrast in the expression also carries an information structure distinction since Manner becomes Focus in Spanish. Yet, some ERC-Property can be mirrored by Spanish. The Resul-the house painted-is entailed by the verb "paint" in (16) while the AP only specifies this entailed final state. This kind of ERC-Property is referred to as "weak" (Washio 1997) or "fake resultative" (Rapoport 2019).
My neighbor painted his house green. 'My neighbor painted his house green.' The weak subtypes of ERC-Property are very much restricted since most verbs that entail Result cannot be part of this construction (i.e., throw, arrive, fix, explode, etc.). In addition, few instances of weak ERC-Property have a Spanish equivalent like the one in (16). Sentence (18) is a weak example of ERC-Property since "break" entails the Result and "open" just specifies it; yet this sentence cannot be replicated in Spanish. In consequence, weak ERC-Property is not productive but very restricted in English. Furthermore, its replication in Spanish is erratic and there is no hypothesis in the literature suggesting a possible pattern.
(18) Peter broke the bottle open. "*Peter rompió la botella abierta" In short, the restricted nature of the downgraded reverse linking in ERC-Path makes it closer to Spanish than ERC-Property since its productive version (strong ERC-Property) comprises reverse linking. The account of the meaning expressed in ERC-Fake by a Spanish form amounts to 'reverse linking'-just as described for ERC-Property-but there is one further difference: Spanish does not make use of a special form (a reflexive pronoun) to express that meaning. In addition, the possible presence of a lexically unlicensed direct objects cuts across all ERC subtypes, but it is reduced to a small set of verbs (i.e., motion verbs) in ERC-Path whereas this set is much larger in ERC-Property, and it is the case for every ERC-Fake.
In sum, ERC-Path with the non-boundary crossing meaning shares fundamental properties with its Spanish counterpart while ERC-Path with boundary crossing requires reverse linking. Yet, this reverse linking is triggered by a single semantic property (crossing a boundary) that is entirely additional to the notion of Result. This is why we refer to it as 'downgraded reverse linking' and contrast it to the reversed linking shown by the examples of strong ERC-Property. The weak subtype of ERC-Property can be mirrored by Spanish, but it is not a productive form in English but very much restricted to some of the verbs that entail a Result. Furthermore, many of the weak ERC-Property instances simply cannot be mirrored in Spanish. This justifies taking only strong ERC-Property into consideration for the distance hypothesis. 9 A summary of the distance analysis is displayed in Table 2. 9 In fact, most of the ERC-Property examples (twelve (12) out of sixteen (16) items) in our data set correspond to strong instances of ERC-Property, and the four ones that are weak are instances that cannot be mirrored in Spanish as Result is not entailed by the verb. In the case of ERC-Path, most of the items selected codify boundary crossing. In this sense, no experimental ERC item in our data set can be replicated by a mirror image sentence in Spanish. Table 2 relies on a qualitative approach, it is not about the number of differences but about the nature of the grammatical properties that are not shared by the English and Spanish versions. We determine interlinguistic distance according to the following cline from the furthest to the closest to Spanish: i) reversed linking with no special use of a form in Spanish (as in ERC-Fake); ii) straight reversed linking (as in ERC-Property); iii) downgraded reversed linking (ERC-Path with boundary crossing) or no reversed linking with peripheral changes (ERC-Path non-boundary crossing); and iv) mirror image structure in Spanish (all properties shared with no reverse linking) like EDC as described in the following section.

The gradient given by
In this sense, the different linking of syntax and semantics within the core of the sentence amounts to a radical contrast while the selection of one or another preposition exemplifies a minor contrast. If we assume the hypothesis that the ERC subtypes are acquired in a cline by Spanish native speakers modulated by interlinguistic 'distance', the same sequence as the one modulated by 'complexity' in (9) can be made.
(19) Distance: ERC-Path < ERC-Property < ERC-Fake This means that Complexity and Distance make the same predictions. Is there a way to determine which of these factors can make an accurate prediction for the acquisition of ERC? The key to untangle this puzzle is to incorporate the acquisition of the English Depictive Construction (EDC) into the equation. Spanish has a Spanish Depictive Construction (SDC) illustrated in (20), which translates the EDC in (7).
John returned the book damaged. 'John returned the book damaged.' According to the interlinguistic distance hypothesis, EDC should be easier to acquire than any subtype of ERCs since SDC mirrors all the grammatical properties of EDC. In particular, both EDC and SDC encompass constituents with the same categories (i.e., verb and adjective) and project the same meanings to the matrix verb and to the AP, respectively. Furthermore, the AP conveys a secondary predication and the semantics of this secondary predication is the same in both languages, namely, it describes a state of a matrix verb argument ("the book is damaged") while the main event unfolds ("John returned the book"). Furthermore, the APs in EDC and SDC are both restricted to nonindividual states. If EDC is taken into the equation, the prediction of the interlinguistic distance hypothesis is that the acquisition process should follow the sequence in (21). Reversed linking with no special use of a form Susy quedó tonta de tanto reírse.

ERC-Property
Mary danced John tired.
Reversed linking Mary dejó a John cansado de tanto bailar.

ERC-Path
Boundary crossing Harry pushed the piano off the stage.
No boundary crossing Javier walked to school.
Downgraded reversed linking Harry sacó el piano del escenario.
No reversed linking with peripheral changes Javier caminó hasta la escuela.

EDC
John returned the book damaged.
No reversed linking John devolvió el libro dañado.
(21) Distance Hypothesis: EDC < ERC-Path < ERC-Property < ERC-Fake The crucial difference in the prediction between (9)-repeated below as (22)-and (21) lies on the order of EDC and ERC-Path. Complexity predicts that EDC should be acquired later than ERC-Path in (22), whereas distance predicts exactly the opposite order, namely, first EDC and then ERC-Path.
Regarding the other ERC subtypes both hypotheses envision the same order.
(22) Complexity Hypothesis: ERC-Path < EDC < ERC-Property < ERC-Fake If it is the case that Distance hypothesis overrides Complexity, this result will be consistent with previous studies on the role of L1 in the acquisition of L2. The relevance of Interlinguistic Distance would indicate unambiguously that EFL learners use previous knowledge (L1) to acquire new knowledge (L2). This prediction is also consistent with the Competition Model of language processing (MacWhinney & Bates 1989) and the later developed Unified Model of Language Learning (UMLL) (MacWhinney 2005) which conceptualizes language learning as an emergent process both in the acquisition of first and foreign languages. Consistently with the RRG perspective on language description (Rispoli 1991), the Competition Model is a functionalist language processing model with highly interactive real time processing in which the different types of information (syntactic, semantic, phonological, etc.) are equally integrated to convey meaning. In this sense, languages vary in the different possibilities of form and function interaction (or in the relative force of form-function mapping) (Hernandez et al. 2007). This highly interactive integration of linguistic information is what is called 'interface' within RRG ( van Valin 2005). This form-function mapping is carried out by the linking algorithm which relates the language-specific syntactic components to the semantic ones (argument slots) and vice versa. In L2 learning, learners' proficiency allows them to mature their representations of the syntax-semantic interface of the L2 (i.e., to improve their recognition of different structural restrictions [ van Valin 1991van Valin , 2000], leading to a better way of thinking in a foreign language [Slobin 1996]). The Competition Model assumes that learning L2 is also an interactive process in which the representations of a language are formed as subjects' proficiency increases and learners are gradually able to integrate (to put into competence) the different functions of a language into a single form. Learners use their L1 grammatical system at the earliest stages of the L2 acquisition in a gradual process that ends up with a fully proficient speaker that entirely by-passes L1 while using L2 (MacWhinney & Bates 1989). Therefore, this model predicts that the functions that overlap between L1 and L2 should be easier to learn due to interlinguistic influence. As the L2 linguistic system is built upon the overall resources provided by L1, the overlap of functions helps to facilitate the creation of L2 representations (MacWhinney 2005). In our case, the model would predict that Low proficient Spanish speakers learners of English would recognize and comprehend the different conditions in a given sequence starting from the ones whose function-form mapping (syntax semantic interface) overlap in L1 and L2 (EDC > ERC-Path > ERC-Property > ERC-Fake). As proficiency increases, this sequence would show less and less differences among conditions to the point of a native-like performance (EDC = ERC-Path = ERC-Property = ERC-Fake).

Advances in the psycholinguistic field
Previous works have studied the interlinguistic effect in comprehension and production within the English-Spanish contrast, especially in the lexicalization of Path and Manner of motion (Alonso 2015;Larrañaga et al. 2012;Robinson et al. 2009). Studies show that learners of satellite-framed languages, even at high proficient levels of Spanish L2, seem to have difficulties in producing complete configurations of Path and Manner in target sentences in L2. In addition, Robinson et al. (2009) used an oral narrative task to identify the significant effect of different L1 in high proficient EFL learners. Speakers of Danish (a satellite-framed language) produced English L2 lexicalization patterns of motion better than speakers of Japanese (a verb-framed language), showing that L1-based lexicalization patterns influence the coding of motion events in the L2. The evidence displayed in the literature shows that the typological characteristics of a verb-framed L1, in terms of its syntactic-semantic properties, have effects on the comprehension and production of a satellite-framed L2.
In addition, some works have addressed the acquisition of ERC by L2 learners. Oliveira (2016) showed the effect of L1 on the acceptability of ERC-Property and EDC by Brazilian Portuguese (BP) subjects. Similar to Spanish, Portuguese licenses depictive constructions, but it has no mirror image of the ERC. This led to a higher acceptability rating for EDC than ERC-Property. Also, reaction times showed that unlicensed ERCs-Property were more difficult to process than unlicensed EDC, which confirms the effect of L1 on L2. However, this study only considered highly proficient subjects, disregarding a whole group of low and intermediate subjects whose judgment on ERC could enlighten our understanding of ERC processing. Similarly, Kim et al. (2019), applied an AJT to show that Korean L1 speakers of English as L2 carried their L1 knowledge on their recognition of ERC. As Korean codifies two types of resultatives, one licensed in English and one not allowed by the English grammar, high and low proficient subjects were not able to judge the unlicensed ERC as unacceptable, as they were part of their L1 grammar. However, the English target structure in this experiment is restricted to ERC-Property, which limits the interpretation of data to a selected set of EFL structures.
As for Spanish, Bautista et al.'s proceeding (2016) reports the results of an experiment on the acquisition of English change of state resultative constructions (what we call ERC-Property and ERC-Fake Reflexive with final AP) and depictive constructions (as control sentences). They applied a True Value Judgment Task and Acceptability Judgment Task to rate the grammaticality of ERC with or without Fake Reflexive. Results show that low and intermediate subjects performed better in the recognition of EDC rather than ERC, as proficiency increased the recognition of ERC in the high proficient groups. In addition, Spanish speakers acquire the resultative interpretations of secondary predication in English as they become more proficient, except for the constraints on the use of ERC-Fake Reflexives (Bautista et al. 2016). Interestingly, Bautista et al. (2016) deals with a small sample of immersed bilinguals rather than EFL students in order to draw conclusions on L2 acquisition. This makes data and predictions' applicability different from the claims regarding our paper. Additionally, authors do not include ERC-Path as a critical condition to assess L2 knowledge on ERC.
Proficiency has also been studied as a key variable in L2 comprehension in a number of studies. Oliveira & Penzin (2019), for instance, showed that Brazilian Portuguese (BP) subjects improved their discrimination of licensed and unlicensed target structures (English double object constructions) as their proficiency increased. Yet, their study does not focus on the ERC as a target structure. In addition, Yotsuya et al. (2014) showed that low proficient Japanese subjects were less accurate in the comprehension and acceptability of Strong ERC (Washio 1997) than their high proficient peers. Also, Kim et al. (2019) identified that both high and low-proficient Korean subjects failed to accept grammatical ERCs with AP. However, contrary to Spanish, both Japanese and Korean show some type of resultative construction as proper templates in their L1.
Regarding ERC recognition and comprehension by English native speakers, the only available evidence (Shi 2016) showed a considerable degree of performance variation across comprehension and production tasks with ERC and EDC constructions, both in native and non-native English speakers. However, Shi's study did not consider potential differences among ERC types, and the interpretation of the results is limited by the small sample size.
Taken together, no research has tackled the acquisition of the three big subtypes of ERC. Instead, they either take ERC as a homogeneous category (no subcategories distinction) or focus on ERC with final AP (what we call ERC-Property). A limited number of research includes ERC-Fake and we found no work dealing with ERC-Path. Regarding Spanish-English contrast, the only work performed on this subject (Bautista et al. 2016) lacks one critical condition, that of ERC-Path, at the same time that it does not contribute to the understanding of the foreign language acquisition as they test immersed bilinguals. Contrastingly, our work makes a systematic comparison of L2 proficiency levels of English as a Foreign Language.
In addition, acceptability judgements are quite usual in this field (Bautista et al. 2016;Kim et al. 2019;Oliveira 2016;Oliveira & Penzin 2019;Yotsuya et al. 2014), yet none of the work on ERC acquisition has attempted to tackle ERC sentence comprehension directly. Also, the scarce data on English native speakers' processing of ERC is striking and, consequently, no previous work has compared the recognition and comprehension of the three subtypes of ERC by English native speakers. It is impressive that the seminal works on ERC published by Boas in 2003, Goldberg &Jackendoff in 2004, andLevin in 2001 have led to no specific and significant psycholinguistic studies on the acquisition of this crucial structure for Spanish native speakers in particular. Our work then intends to fill this gap.
With the purpose of exploring how Spanish EFL learners acquire different subtypes of ERC at different proficiency levels, we conducted two experiments that could address different aspects of L2 linguistic knowledge in terms of ERC recognition and comprehension: An Acceptability Judgment Task and a Sentence Comprehension Task respectively.
On the one hand, we designed an AJT in order to shed light on the syntax-semantics interface. This task was implemented to explore how different ERC with equivalent syntactic structures were recognized as acceptable by Spanish native speakers and what differences of subjects' recognition of the selectional restrictions of each ERC and EDC can be found in terms of proficiency. In the following pages, subjects' judgements of acceptability will be referred to as the recognition of licensed and unlicensed sentences in L2. In order to match items as closely as possible to detect small differences in sentence recognition (Schütze 2019), we included only ERC-Property and ERC-Fake, and EDCs as control structure, leaving aside ERC-Path.
On the other hand, we designed a Sentence Comprehension Task to specifically address subjects' semantic interpretation of the different subtypes of ERCs and the EDC on a sample of EFL learners, and to examine its association with the participants' proficiency. We will refer to subjects' interpretation performance of each target structure as their comprehension of ERC and EDC. This task consisted of a multiple-choice task which included the three subtypes of ERC (ERC-Path, ERC-Property and ERC-Fake Reflexive) and EDC. Unlike experiment 1, ERC-Path was included in experiment 2 as this structure allowed us to enrich our interpretation further: ERC-Path is a critical ERC subtype that has the crucial property of being less complex than EDC and the other ERC subtypes (in terms of its syntax-semantics properties). At the same time, it is also the ERC closer to Spanish than ERC-Property and ERC-Fake. The prediction was that complexity and interlinguistic distance could have a role in the organization of the acceptability gradient. Yet, if complexity prevails over distance, ERC-Path should be better comprehended than EDC (at least at low English proficiency levels).
Even when both experiments address different grammatical aspects of a language and, therefore, show different designs, previous works have shown that evidence on L2 acquisition is more robust whenever we obtain consistent results among them (Kim et al. 2019). In the same line, Ionin & Zyzik (2014) have suggested that acceptability and interpretation tasks can be used in combination to yield a more complete and insightful picture of learners' knowledge, as we attempt to do in this research. In our case, as the acceptability task addresses the combinatory knowledge of the structure basis, the comprehension task lets us observe the semantic knowledge constraints needed for ERC and EDC interpretation.

Experiment 1: Acceptability judgment task
Our first objective was to explore the degree of recognition of syntactically similar ERC's (ERC-Property and ERC-Fake) and the EDC by Spanish EFL learners. In order to do so, we carried out an Acceptability Judgment Task (AJT) to examine potential differences in the recognition of licensed and unlicensed ERC's and EDC, and we compared acceptability ratings of high and low proficiency participants with the acceptability rating of an English native speakers' group. Since ERC Property and Fake Reflexive have no equivalent counterpart in Spanish, we expected to observe lower acceptability ratings for these structures when compared to EDC (which is indeed present in the participants' L1). In addition, we expect English native speakers to show no difference in recognizing the ERC and EDC, as these are all structures of the English language.

Participants
Participants were recruited through social networks (e.g.: Facebook). A total of 90 native speakers of Spanish from Argentina who learn English as a Foreign Language (EFL) were tested in their home country (61 female, mean age: 34.77 ± 10.46 years) took part in the study. Most of the participants had completed tertiary or university studies (67.77%), while the rest had not finished them yet (33.23%). No significant differences in age (t(64) = 1.039, p = 0.303), sex or education level (χ 2 (2)'s < 0.418, p's < 0.518) were observed between these groups. The use of English at home included 81% of the sample, whereas immersion in an English-speaking country corresponds only to 18%. Only a small number of EFL learners reported knowledge of other languages such as Portuguese (13.33%) and French (14.4%), both of them verb-framed languages. In addition, only 18% of the sample had formal instruction of EFL at Kindergarten, 62.2% at Primary school, 91.1% at Secondary school, and 58.9% also studied English at University.
In order to compare EFL learners' results with native speakers' performance, a sample of 30 English native speakers were also included (19 female, 1 non-binary, mean age: 38.78 ± 14 years). Sixteen subjects had learned their native language in the United States and were living in their home country at the time of the study, the rest were native speakers of British English and residents of the United Kingdom. None of the participants had graduate or undergraduate studies in linguistics. No significant differences on age (F(1,109) = 2.215, p = 0.140) or sex (χ 2 (2) = 6.563, p = 0.087) were observed among EFL learners and English native speakers. 10

Proficiency measure
Two English proficiency level measures were registered: 1) English level according to The Common European Framework of Reference for Languages (CEFR) exams, and 2) English Vocabulary performance, assessed by an online version of the LexTALE (Lexical Test for Advanced Learners of English; www.lextale.com) (Lemhöfer & Broersma 2012). This was administered to evaluate EFL participants' English lexical proficiency. The test consists of 60 items (40 words, 20 nonwords), and participants are required to decide whether each one is a real English word. It can be administered within 5 minutes and it has been validated as an indicator of English vocabulary knowledge and a strong predictor of general proficiency. In general, a total of 28 subjects indicated A1 to B2 CEFR levels, while 38 indicated C1 or C2 levels and 24 did not report their proficiency exam level. In addition, LexTALE scores had a mean of 46.65 ± 7.53 (range: 29 to 60). The English native speakers' group was not exposed to LexTALE (neither in this experiment or in experiment 2) since we do not expect performance differences in a task designed to evaluate EFL learners' proficiency.
In order to verify that LexTALE scores were an adequate index for proficiency for our sample, we checked them against CEFR scores for those participants who provided such information (76 EFL learners out of 90). In order to do so, we fitted a linear model with LexTALE score as the response variable and Group as the predictor, assessing the correlation between both measures. We defined "group" as either "low" (A1 to B2) or "high" (C1-C2) to keep the number of subjects per level balanced. LexTALE scores were significantly higher in the "high" group (F(1, 65) = 14.224, p < 0.0001, R2 = 0.194). We concluded that LexTALE was a representative and a reliable measure of our 10 We acknowledge that one cannot rule out potential differences on the basis of null results. However, the inclusion of subject random effects on mixed models allows control for between-subject sources of variation. In addition, item-level sources of variability are controlled by including item random effects in our models.
participants' CEFR level (as previously shown by Lemhöfer & Broersma 2012) given that both measures yielded similar results and LexTALE scores were available for the whole sample.

Stimuli
Experimental items consisted of 48 sentence frames equally distributed across the following experimental conditions: 1) EDC object-oriented, 2) transitive ERC-Property, and 3) ERC-Fake transitivized by the construction. We selected these transitive constructions in order to keep syntactic differences at a minimum: EDC and ERC-Property are syntactically equivalent, whereas ERC-Fake has the same syntactic template but with a different realization of the Direct Object. In addition, all of these constructions codify their satellites by means of an AP. For each experimental item, two versions were created: licensed and unlicensed. The unlicensed sentences were created from their licensed counterpart by replacing the Adjectival Phrase for another one that violates the construction's selectional restrictions (Wechsler 2005) (see Table 3 for examples).
Sentences were selected and adapted from several previous works (such as Boas 2003; Broccias 2004; Goldberg & Jackendoff 2004). Examples of experimental sentences are provided in Table 3. Experimental sentences had a mean length of 8.35 ± 1.74 words and no significant length differences were observed among conditions (F (2,45) = 0.822, p = 0.446). In order to detect frequency biases, we compared the frequency of experimental items (verb + particle) estimated by the Corpus of Contemporary American English (COCA) (Davies 2015). A Kruskal Wallis test showed no frequency differences between experimental conditions (EDC, ERC-Property, ERC-Fake) in any of the estimates (χ 2 (3) = 4.25, p = 0.119).

AJT instrument and procedure
Data were collected through a Google Forms online survey and tasks were administered in the following order: sociodemographic questionnaire, Acceptability Judgment Task, LexTALE test, and English Background questionnaire based on Language History Test (Li et al. 2013) to describe their frequency of use of English language, immersion, and bilingual experience. Data on frequency of use of English were not included in the current study analysis, but are made available in the Online Supplementary Materials (see the Experiment 1: Participants section).
As for the AJT, participants were asked to rate each of the experimental and filler sentences individually, answering how acceptable these sentences would sound to a native speaker of English in a 1 (totally unacceptable) to 7 (totally acceptable) Likert scale. The task consisted of two lists made up of 16 experimental items for each construction type (48 items in total), alternating the licensed and the unlicensed version of each item in a Latin square design (no list contained both the licensed and unlicensed version of the same items). The presentation of the lists was counterbalanced between participants. In addition, 32 filler sentences were included, half of them well formed, 25% containing semantic violations and 25% containing syntactic violations. Two examples and five training trials were administered before the experiment. Target and filler items were presented in a pseudo-randomized order, so each condition was not repeated more than twice in a row. AJT was designed following Schütze & Sprouse (2014) recommendations.
The procedures of this study complied with the provisions of the Declaration of Helsinki regarding research on human participants. The study protocol was approved by the Scientific and Technical Center (CCT)-Mendoza's Ethical Committee from the National Scientific and Technical Research Council-Argentina (CONICET). Prior to the study, subjects were informed that their participation would be voluntary, anonymous and that they could withdraw from the study at any time without consequence. No economic compensation was offered.

Data analysis
Statistical analysis was carried out in the following way: first, we fitted a cumulative link mixed model comparing the acceptability rankings of EFL learners and native speaker performances. The model considers correctness, construction and group as fixed factors. Group is a categorical variable that indicates if the subject was a high or low proficiency EFL speaker, or an English native speaker. EFL speakers' proficiency was categorized as "high" or "low" by splitting LexTALE scores at their median (Med = 48, IQR = 12). Then, in order to address the role of proficiency on the acceptability ratings of EFL learners only, we fitted a second model replacing the factor "group" for "LexTALE." LexTALE score was introduced as a continuous predictor in this model. Acceptability ratings were the dependent variable, codified as ordinal (1-7) (Agresti 2002). All models described in this section and in Experiment 2 were built with the ordinal package (Christensen 2019) in R. IC (Akaike information criterion) was reported as an index of model fit. Main effects and interactions were analyzed with post-hoc comparisons ("emmeans" R package), and pvalues were adjusted using the Tukey method. Scripts for both analyses can be found in the Supplemental data.

Acceptability ratings by construction and correctness: Comparison of EFL and native speakers
We fitted an ordinal logistic regression with the acceptability score as the response variable. Factors were dummy coded and treatment contrast was applied. Correctness (correct, violation), construction type (EDC [reference level], ERC-Property, ERC-Fake), and group (EFL low [reference level], EFL high, native) were coded as factors, and all their interactions as fixed effects. We also added random intercepts for participants and items, and random slopes for all fixed effects and their interaction. The final model included random slopes for group × items and construction type × participants. Model was fit with the Laplace approximation (Log-Likelihood: -8447.53, AIC: 17017.05). Nested model comparisons showed that the triple interaction was significant (χ 2 (2) = 21.43, p = 0.0003). Full output for the model is provided in the Online Supplementary Material (Supplementary Table 3 and 4). The following main effects and interactions were found: Unlicensed (Estimate: -2.98, z = -8.515, p < 0.001), Unlicensed × Low (Estimate: 122.43, z = -4.221, p < 0.001), Unlicensed × ERC-Property (Estimate: -2.98, z = -8.515, p < 0.001), Unlicensed × ERC-Fake (Estimate: 1.185, z = 3.044, p = 0.0023), Low × ERC-Fake (Estimate: 122.43, z = -4.221, p < 0.001) (Reference categories: Licensed (Correctness); EFL High (High proficiency) (Group); EDC (Construction). Figure 1 shows the estimated mean acceptability scores by correctness, group, and construction type.
Post hoc analyses of correctness effects by construction type and group indicated that licensed items were scored as more acceptable for all constructions in the native speakers and EFL high group, while this was only true for EDC in the EFL low participants (p's < 0.001) (see Table 4). Acceptability ratings did not differ among groups for licensed EDC items, while ERC-Property was more accepted in EFL high and native speakers with respect to EFL low participants (p's < 0.03). Acceptability for ERC-Fake followed this pattern: low < high < native (p's < 0.039). Unlicensed items did not differ significantly between groups or constructions (see Table 5). Finally, within group comparisons indicated that EFL low participants scored licensed EDCs as more acceptable than both ERCs (p's < 0.001), while this was only true for ERC-Fake in EFL high (p = 0.033). No significant differences among construction types were observed in the native speakers' group. In addition, no significant effects were found for unlicensed items (see Table 6).

Acceptability ratings by construction type, correctness and LexTALE scores
We fitted a new model of acceptability ratings for EFL learners only, similar to the previous one but replacing "group" for "LexTALE" scores as a continuous predictor (English native speaker group was excluded for this analysis). The final model included random slopes for LexTALE × items and construction type × participants. Model was fitted with the Laplace approximation (Log-Likelihood: -6171.97, AIC: 12435.93). Nested model comparisons showed that the triple interaction was not significant (χ 2 (2) (2) = 0.45, p = 1), but the LexTALE × construction type (χ 2 (2) = 6.51, p = 0.03), the grammaticality × construction type (χ 2 (2) = 22.65, p < 0.0001) and the grammaticality × LexTALE (χ 2 (2) = 21.4, p < 0.0001) interactions were significant. Full output for the model is provided in the Online Supplementary Material (Supplementary Table 5 and 6).
The following main effects and interactions were found: ERC-Property (Estimate: -2.230, z = -2.457, p < 0.001), ERC-Fake (Estimate: -3.837, z = -4.029, p < 0.001), Unlicensed × LexTALE        Figure 2a shows the estimated probabilities of receiving each score considering all three independent variables. In order to have a clearer picture of the differences between conditions, Figure 2b shows the estimated probabilities and the 95% confidence intervals for scores 1 and 7. In order to compare acceptability of each construction, we ran post-hoc analyses of acceptability ratings for each Construction Type in three different scores across the range of the LexTALE distribution. These scores were 35, 45, and 55, and they were chosen because they are at a midpoint within the three subgroups that compose the 30-to-60 LexTALE range. Post hoc analyses indicated that licensed items received higher acceptability ratings for all constructions at LexTALE scores 45 and 55 (p's < 0.005), but this was only true for EDC at 35 (see Table 7). Acceptability ratings of licensed items did not differ among LexTALE scores for EDC or ERC-Property (although p-values were considerably closer to significance for these items). Conversely, licensed ERC-Fake items' acceptability increased for each LexTALE score with respect to the previous one (35 < 45 < 55, p's < 0.003) (see Table 8). Regarding unlicensed items, EDC violations were scored as less acceptable as participants' proficiency increased (35 > 45 > 55, p's < 0.05), but no differences were found for the rest of the items. Finally, within-LexTALE level comparisons of licensed items (see Table 9) indicated that EDCs were scored as more acceptable than ERC-Property and ERC-Fake (with no differences between the latter) at all LexTALE score levels (p's < 0.04) (but the significance of this effect was larger at LexTALE 35 and 45 scores (p's < 0.0003).

Discussion
Analysis by vocabulary proficiency as a continuous variable and the comparison with native speakers as a group variable both indicated an advantage for EDC over ERCs among less proficient participants. Low EFL speakers were able to recognize licensed EDCs, but failed to differentiate licensed and unlicensed ERCs. High EFL speakers were closer to native speakers in their discrimination performance for all constructions. In addition, both ERCs were more accepted in the EFL high and native speakers' groups than in EFL low, and scores for ERC-Fake were also higher in native speakers when compared to EFL high. Furthermore, unlicensed EDC acceptability ratings were the only ones that decreased with proficiency (showing better recognition of violations). In short: a gradient of performance (EDC > ERC-Property > ERC-Fake) can be clearly identified in EFL low speakers, while this effect gradually fades for more proficient speakers and disappears completely for native speakers.
The pattern of results in experiment 1 can be interpreted as a consequence of L1-L2 interlinguistic differences since ERCs (the constructions with no equivalent Spanish counterpart) were not as readily recognized as acceptable as EDCs (structure with mirror image in English and Spanish) by less proficient EFL speakers. As vocabulary proficiency increases, the advantage for EDCs gradually fades and participants' ratings resemble more closely those of native speakers, who accept all licensed constructions without preference. This remains true even in the first analysis, where EDC was scored as more acceptable than both ERCs at the highest LexTALE score considered (55), because the significance level of this effect is much lower than the one observed at scores 45 and 35. It could be argued that these results are reflecting structural complexity differences between the EDC, the more complex ERC-Property (that adds a causal relation between the predicates), and the most complex ERC-Fake (that adds the exceptional use of a reflexive pronominal form) (see Table 1).
We decided to include an additional condition in Experiment 2 to inform the question further. It can be shown that the ERC-Path is closer to Spanish than the other ERC subtypes, while at the same time it is also the less complex of all the constructions in question, including the EDC (see Introduction section). Therefore, it was of interest to examine whether comprehension performance of ERC-Path by Spanish EFL learners falls between EDC and the other ERCs (following an interlinguistic distance gradient) or if this less complex construction is also better comprehended than EDC (following a complexity gradient).

Experiment 2: Sentence Comprehension Task
In order to explore and compare the comprehension (i.e., the semantic interpretation) of different ERC types and the EDC by Spanish EFL learners, we designed a Sentence Comprehension Task. To analyze potential proficiency effects and their interaction with construction types, we included participants with varying degrees of English level (low, medium, and high). Additionally, we decided to include ERC-Path to explore if performance patterns were better explained by the linguistic distance or complexity hypotheses. ERC-Path is a midway construction, that is, it is simpler than EDC (in terms of complexity hypothesis) at the same time that it is further from Spanish than EDC (distance hypothesis). Therefore, testing the comprehension of ERC-Path vs EDC would allow us to shed light on the subtle differences between our two hypotheses. In addition, the rationale for including a comprehension task drives us back to the general objectives of this paper. Our aim is to observe effects from the different types of grammatical knowledge that make up the syntax-semantic interface that non-native speakers of English need to handle in order to acquire the ERC: as the acceptability task addresses the combinatory knowledge of the structure basis, a comprehension task let us observe the semantic knowledge constraints needed for ERC and EDC.

Participants
Participant recruitment was the same as experiment 1 and a total of 287 EFL learners native speakers of Spanish from Argentina were tested in their home country (231 of them female, mean age: 34.94 ± 9.68 years) participated in the study. Most of the participants had completed tertiary or university studies (82.93%), while the rest had not finished them yet (17.07%). The use of English at home included 88.5% of the sample, whereas immersion in an English-speaking country corresponds only to 23.34%. In addition, a small number of subjects have studied other languages as L2 such as Portuguese (12.20%), French (17%), Italian (15.33%), and all verb-framed languages. Only 4.88% studied German, a satellite-framed language. Only 18,82% of the sample had formal instruction of EFL at Kindergarten,58.19% at Primary school,91.29% at Secondary school,and 58.19% also studied English at University. A complete description of participants' demographics and English background measures can be found in the Online Supplementary Materials (Experiment 2: Participants).
In order to compare EFL learners and native speakers' comprehension performances, a sample of 54 English native speakers were also included (48 female, 2 non-binary, mean age: 33.34 ± 13.92 years). Nineteen subjects were native speakers of American English and residents of the United States; twenty-two were native speakers of British English and residents in the United Kingdom, and the rest were distributed between native speakers and residents of Australia and Scotland. None of the participants had graduate or undergraduate studies in linguistics. No significant differences in age (F(1,338) = 1.055, p = 0.305) were observed among EFL learners and native participants.
Proficiency measure CEFR exams and LexTALE scores were considered as English proficiency measures. In order to avoid priming effects, participants were asked to refrain from taking part in this study if they had participated in Experiment 1.
Participants were classified according to their self-reported English level in three proficiency groups: Low (A1, A2; n = 34), Intermediate (B1, B2; n = 63) and High (C1, C2; n = 127). Sixtythree participants failed to inform their English level. There were no significant age (F(2,223) = 2.290, p = 0.104), sex or education level differences (χ 2 (2) 2.797, p's > 0.247) between groups. In addition, we considered Vocabulary proficiency as a continuous variable indexed by LexTALE scores (Mean: 46.87 ± 6.98). As in Experiment 1, we chose LexTALE scores as indicators of proficiency for the main analysis, but we ran a complementary model using CEFR group as predictor, to verify that both measures led to a similar pattern of results (see Online Supplementary Materials: Experiment 2; Results).
In order to show that LexTALE scores were an adequate index for proficiency for our sample, we fitted a linear model with LexTALE score as the response variable and Group as the predictor, assessing the correlation between both measures. Results showed an effect of Group in LexTALE scores (F(2, 221) = 36.46, p < 0.0001, R2 = 0.24). Scores were higher for speakers in the high proficiency group when compared to speakers in the intermediate (b = 6.04, p < 0.0001) and low (b = 8.06, p < 0.0001) proficiency group. There was no difference between intermediate and low proficient speakers (b = 2.01, p = 0.252).

Stimuli
Experimental items consisted of 24 sentences, equally distributed among the following experimental conditions: 1) EDC object-oriented, 2) transitive ERC-Path, 3) transitive ERC-Property, and 4) ERC-Fake transitivized by the construction. Sentences were selected and adapted from the same sources as in experiment 1. Experimental sentences had a mean length of 7 (±2.44 words), and neither significant length differences were observed among conditions (F(3,23) = 0.793, p = 0.512), nor frequency differences of experimental items (verb + particle contingency) were found (Kruskal Wallis χ 2 (2) = 9.47, p = 0.092) estimated by the Corpus of Contemporary American English (COCA) (Davies 2015). In addition, 12 garden-path sentences were included as fillers.

Sentence Comprehension Task: Instrument and procedure
Data collection was the same as in experiment 1 and the same series of questionnaires was administered, replacing the AJT with the Sentence Comprehension Task. In the Sentence Comprehension Task, participants were presented with each of the experimental and filler sentences individually, and asked to choose between three possible interpretations, indicating the one that best reflected the meaning of the target sentence. In addition, a does not know/does not answer option was added. Options were designed to provide a correct response, a plausible but incorrect interpretation and an implausible interpretation. Examples of multiple-choice alternatives are provided in Table 10. Sentences were presented in a pseudo-random order, and correct options were pseudo-randomized among items as well. Ethical considerations were the same as in Experiment 1. Lobsters were alive when people cooked them.
People needed to be very active when cooking lobsters.
Lobsters became alive because of the cooking.

ERC-Path
Fred tracked the leak to its source.
Fred found the source of the leak by tracking it.
The leak made Fred track it. Fred tracks the course of the leak from its source.

ERC-Property
Tom watered the plants flat.
The plants became flat because Tom watered them.
The plants were already flat when Tom watered them.
Tom was feeling flat when he watered the plants.

ERC-Fake Reflexive
The baby cried himself asleep.
The baby fell asleep due to so much crying.
The baby was asleep while crying.
The baby cried because he fell asleep. Filler Garden Path Until the police arrest the drug dealers control the street.
The drug dealers controlled the street until the police arrested them.
Until the police arrested the drug dealers, the police controlled the street.
The police's control on the street caused the drug dealers' arrest.

Data analysis
Data analysis procedures were similar to Experiment 1, but this time it was performed on the response accuracy, treated as a dichotomous variable (correct/incorrect). In this case, we added conditional and marginal coefficients of determination for Generalized mixed-effect models (Pseudo-R 2 , calculated using r.squaredGLMM function from Multi-Model Inference R package) (Nakagawa et al. 2017). We found no evidence of ceiling effects for participants' performances on the comprehension task. The frequency for the maximum possible number of correct responses (24) was 4.3%, skewness: -1.354, kurtosis: 1.387. In order to compare EFL learners and native speaker performances, we replaced the factor "LexTALE" for "group," a categorical variable that indicated if the subject's EFL proficiency was "high" or "low," or if he was an English native speaker. EFL speakers' proficiency was categorized as "high" or "low" by splitting LexTALE scores at their median (Med = 47, IQR = 15). In addition, we fitted a second model on the EFL learners' data, considering proficiency (LexTALE scores) as a continuous variable.

Comprehension performance by construction type: Comparison of EFL learners and native speakers
Percentage of correct responses by construction and group is provided in Figure 3. A mixed-effects logistic regression model was built with response accuracy (correct/incorrect) as the response variable and Construction Type, Group (EFL low, n = 148; EFL high, n = 139; native, n = 54) and their interaction as predictors. Factors were dummy coded as stated in experiment 1 and treatment contrast was applied. Due to convergence errors, the final version of the model included Group × items and type × subject. Model was fit with the Laplace approximation (Log-Likelihood: -2882.8, AIC: 5793.7). Nested model comparisons showed that the two-way interaction was significant (χ2(6) = 124.96, p < 0.0001). Full model output is provided in the Online Supplementary Materials (Supplementary Tables 9 and 10). The following interactions were observed: ERC-Fake × Low (Estimate: -1.3100, z = -6.092, p < 0.001), ERC-Property × Low (Estimate: -2.0586, z = -8.626, p < 0.001), ERC-Path × Low (Estimate: -0.867, z = -3.932, p < 0.001), ERC-Fake × Native (Estimate: 0.867, z = 2.725, p = 0.006), ERC-Property × Native (Estimate: 0.811, z = 2.070, p = 0.038), ERC-Path × Native (Estimate: 1.470, z = 3.836, p < 0.001) (Reference categories: EDC; Group: High). Figure 3 shows response accuracy per Group and Construction Type. Post hoc analyses indicated that comprehension performance was better for EDCs than ERC-Property and ERC-Fake (p's < 0.04) only in the EFL-low group, while no significant differences were found among constructions for EFL-high or native participants. In addition, comprehension of EDCs was comparable among all groups, while it followed the pattern low < high < natives for all resultatives (p's < 0.035) (Tables 11 and 12).

Comprehension performance by construction type and LexTALE scores
A mixed-effects logistic regression model was built with response accuracy (correct/incorrect) as the response variable and Construction Type (dummy-coded as previously stated), LexTALE score and their interaction as predictors. As in experiment 1, this analysis was carried out for EFL learners only, and LexTALE was included as a continuous variable (n = 285, mean = 46.8, sd = 6.98, range: 18 to 60). Four participants were omitted from these analyses because their score was two standard-deviations  below the mean. 11 We also added random intercepts for participants and items, but we had to drop random slopes for fixed effects and interactions, due to convergence issues. Model was fit by Laplace approximation (AIC: 5056, Log-Likelihood: -2.518). In addition, we estimated goodness-of -fit by pseudo-R 2 for generalized linear mixed effects models (marginal  Table 13. Pairwise comparisons summed up on Table 14 showed that the effect of proficiency on accuracy was smaller in EDCs when compared to all ERCs. There were no statistical differences between ERC-Path and ERC-Fake regarding this same effect. Finally, the effect of proficiency on accuracy was greater for ERC-Property when compared to the other types of ERCs. Significant differences among EDC and ERCs can only be found at low proficiency levels (see Table 15) and performance differences among LexTALE scores are larger and more significant for ERCs (see Table 16). These trends are depicted in a linear scale in Figure 4a. Figure 4b shows the estimated probability of accuracy across LexTALE scores.
A complementary analysis, considering proficiency as indexed by CEFR exam can be found in the Online Supplementary Materials (Experiment 2: Results).  The same analysis described in this section was conducted later with the full dataset (including these four participants) and inferential results did not differ qualitatively.

Discussion
Results from Experiment 2 showed that proficiency modulated ERC comprehension: for more proficient speakers, accuracy did not vary significantly between constructions. Less proficient speakers performed better in EDC than in ERC-Property and ERC-Fake (Table 15), no differences were observed with ERC-Path at any proficiency level (Low: EDC ≅ ERC-Path; EDC > ERC-Property ≅ ERC-FR). In general, these results showed that EDC accuracy was higher than ERC-Property and ERC-Fake only when proficiency was low. As proficiency improves, there is an increase in correct responses for ERC-Path, ERC-Property, and ERC-Fake, but this increase is not as impressive for EDC (EDC ≅ ERC-Path ≅ ERC-Property ≅ ERC-FR) (Table 13). In fact, we see a significant increase in correct responses for EDC for higher LexTALE scores, but this effect was larger for the ERCs (Tables  14 and 16). These results indicate that: 1) EDCs were easier to understand than ERC-Property and ERC-Fake Reflexive (but not easier than ERC-Path) for less proficient English learners; 2) comprehension performances increased with proficiency, but these effects were larger for ERCs when compared to EDC; 3) among ERCs, the ERC-Property showed a greater effect of proficiency on accuracy.  This pattern of results seems to be better explained in terms of interlinguistic distance, since: 1) performance was better for those constructions closer to Spanish; 2) proficiency effects were more prominent in the ERCs when compared with the EDC, which is not the simplest construction but the only one that is present in Spanish language. In particular, proficiency effects were larger even for ERC-Path, a structure less complex than EDC, with no exact equivalent in Spanish (even though it is closer to the L1 than the rest of the ERCs). This interpretation is further supported by the comparison with native speakers, which showed that performance patterns of EFL-students resemble more closely those of native speakers (no significant differences among constructions) as proficiency increases. In the same line, only EFL-low participants exhibited an advantage for EDCs. Finally, between-group comparisons indicated that native speakers and EFL-high participants outperformed less proficient EFL learners for all ERCs, but not in the case of depictives. In addition, native speakers were better than EFL-high at comprehending ERCs. The findings from Experiments 1 and 2 are discussed jointly and in more detail in the following section.

General Discussion
The present studies attempt to make an initial contribution to the contrastive study of the acquisition of different subtypes of ERC and EDC by Spanish speaking EFL learners. In particular, this is the first attempt to examine differences in the acceptability and comprehension of different ERCs, and to contrast the acquisition of ERCs with another construction (EDC) that mirrors participants' L1 knowledge. In addition, it is the first study to analyze the role proficiency plays during the acquisition of these constructions, comparing acceptability (referred to as the recognition of licensed and unlicensed sentences) and comprehension measures of EFL learners with native speakers. The relevance of our findings is highlighted by the fact that empirical evidence of ERC L2 acquisition is scarce and, in the case of L1, almost nonexistent (however, see, Shi 2016).
The pattern of results can be summarized as follows (see Table 17): 1) Less proficient speakers recognized and comprehended the construction that is present in L1 (EDC) better than those that are further away from it (ERC-Property and ERC-Fake). As experiment 2 shows, comprehension performance for ERC-Path (a construction that is closer to Spanish than other ERCs) seems to be somewhere in the middle of a continuum that goes from EDC to the other types of ERCs, showing no significant differences with the rest of the conditions. 2) Proficiency had a greater impact on the recognition and comprehension of ERCs than EDCs. In particular, less proficient subjects perform worse at discriminating licensed and unlicensed ERCs, and their recognition and comprehension of licensed ERCs resembles more closely that of native speakers as proficiency increases. Nevertheless, ERC-Fake is still rated as more acceptable and better comprehended by native speakers than to high EFL participants; 3) Comparing the different ERC subtypes, proficiency had a larger impact on ERC-Fake recognition and ERC-Property comprehension. The implications of these results are discussed in detail in the following section.

Interlinguistic distance or complexity effects?
Our data shows that the acquisition of the different types of ERC for EFL learners is not homogeneous. If we consider the acceptability and comprehension rates of each structure as an indicator of their acquisition difficulty, our results can be interpreted to reflect different acquisition rates of ERC types by Spanish-speaking students. Two different explanations could be put forward to explain this. According to the complexity hierarchy hypothesis, the more intrinsically complex ERC should be more difficult to learn due to their syntactic and semantic properties. This interpretation predicts that subjects' performance should follow a complexity gradient based on the ERCs cumulative interface properties: ERC-Path > EDC > ERC-Property > ERC-Fake. Conversely, the interlinguistic distance hypothesis states that those constructions that are closer (or more similar) to native language (Spanish) should be acquired earlier and should be more easily recognized and comprehended. In this case, the expected pattern would be: EDC > ERC-Path > ERC-Property > ERC-Fake. Our results seem to suggest an interlinguistic distance effect. Both ERC-Property and ERC-Fake were more difficult to comprehend at low proficient levels, and were rated as less acceptable by all participants. Conversely, we did not observe clear comprehension differences between ERC-Path and EDC, nor a difference between ERC-Path and the other types of ERCs. It may be the case that ERC-Path belongs to a middle ground of sorts: not too different from Spanish to differentiate themselves clearly from EDCs (which are actually present in the L1), but not close enough to the L1 to be more easily comprehended than the other ERCs. Contrastingly, this same pattern cannot be explained purely in terms of complexity. If that was the case, we should have observed better performances in ERC-Path compared to EDC (a more complex structure). This hypothesis also predicts larger performance differences when comparing ERC-Path (6 properties) to both ERC-Property (19 properties) and ERC-Fake (22 properties) (see Table 1) that were not found at any proficiency level. Therefore, while our results do not perfectly follow a strict interlinguistic distance gradient, they certainly indicate better recognition (at all EFL levels) and comprehension (at low proficiency levels) of the English construction that shows a mirror image in Spanish (EDC), worse performance for constructions further away from Spanish (ERC-Property and ERC-Fake) and an intermediate situation for the construction that is more similar, but is not fully present in Spanish (ERC-Path). In addition, Experiment 1 and 2 showed that performance patterns resemble more closely those of native speakers as EFL proficiency increases, with the difference between EDC and ERCs gradually fading until it disappears completely. Therefore, we found evidence that possible interlinguistic effects are larger at earlier stages of English acquisition (as can be seen in the recognition and comprehension of ERCs), which seems reasonable in the light of the Competition Model of language processing and MacWhinney's Unified Model of Language Learning (UMLL) for bilingual processing (Hernández et al. 2007;MacWhinney 2005;MacWhinney & Bates 1989).
It is worth noticing that, along with complexity and interlinguistic distance, verb-construction contingency (the frequency of association of a verb within a given construction) is another factor that could have an effect on L2 language acquisition (Ellis & Collins 2009;Rescorla 1968;Sung & Kim 2022). It has been shown that the degree of association of a verb with a construction has an impact on subjects' comprehension of target structures. In particular, Sung & Kim (2022) showed that Korean L2 learners of English performed equally well in the recognition of strong and weak associated verbs with ditransitive constructions, whereas their performance in recognizing strong associated verb within ERC (such as "make" or "get") is significantly better than in weak associations (or less frequent combinations). Although we acknowledge the plausibility of this variable's effects, we do not consider it to be a problematic issue within our model. As verbs could only be used twice as part of the experimental items (in a licensed and an unlicensed version of a sentence), one would expect this frequency effect to constitute an item-level source of variation, which was considered in the models by item random effects. While a systematic study of these frequency contingency effects might shed light on their contribution to ERC acquisition, we found no strong evidence to support a frequency-based account of our current findings.

Proficiency effects
Regarding the effects of proficiency, results from the acceptability and comprehension tasks showed larger improvements of ERCs performance at higher proficiency levels (when compared to EDCs). This suggests that EDCs are more readily mastered by Spanish learners, so their performance reaches its ceiling earlier. Furthermore, results suggest that the recognition and comprehension of ERCs evolves towards native speakers' performance as proficiency increases. Conversely, when specific ERCs are considered, different patterns emerge, which probably reflect differences in the underlying processes tackled by each task. Regarding the acceptability task, ERC-Fake ratings follow a pattern: EFL low < EFL high < natives, while ERC-Property ratings do not significantly improve from EFL high to native speakers.
In the case of comprehension, proficiency had larger effects on ERC-Property than ERC-Path and ERC-Fake. It might be the case that ERCs-Path are the easiest structures among ERCs (thus reaching an earlier ceiling), while ERC-Fake is the hardest structure to learn (thus benefiting less from English vocabulary or general knowledge), which could explain why there is more room for improvement for ERC-Property (see Figure 4 to compare ERCs learning curves). In fact, it could very well be the case that many instances of ERC-Fake are processed as idioms, namely, they are lexically stored. 12 We acknowledge that this is an ad hoc hypothesis and should be elucidated by further studies. Conversely, it should be noted that we did not find evidence of differences among ERCs or EDCs in our native speakers' group.
Taken together, Experiments 1 and 2 can be interpreted in the following way: 1) Recognition of licensed ERCs evolves towards native speakers' performance (who show no difference between EDC and ERCs) as proficiency increases. However, ERC-Property reaches its ceiling for high proficient EFL learners, while ERC-Fake is still better recognized by native speakers; 2) Comprehension of ERCs improves with proficiency, with this effect being larger for ERC-Property. These results are not necessarily at odds, since recognition and comprehension are not identical processes: being able to recognize the correct pattern of a construction is not the same as being able to access the full meaning of a sentence.
These claims are in line with the fact that the contrast between Spanish and English style depicts not only a formal difference between languages, but a semantic way of representing [[events]]. This implies, in turn, a different event conceptualization in each language as stated by the 'thinking for speaking' hypothesis (Slobin 1996). The progressive acquisition of ERCs can be considered to reflect the process of learning a new way of thinking for speaking (Cadierno 2012;Cadierno & Ruiz 2006) or even learning to rethink-for-speaking (Robinson et al. 2009) as it triggers a given cognitive representation which differs from the event representations in Spanish. In terms of RRG, the progressive acquisition of the different ERC is a reflection of the cognitive process of learning the particularities of each L2 structures' selectional restrictions ( van Valin 1991van Valin , 2000. The learning of these selectional restrictions seems to be achieved only at higher stages of English proficiency, where comprehension of ERCs is similar to a structure common to both languages, but their acceptability is still lower than that of EDCs.

Limitations of the study
We should acknowledge the following limitations in our study. Regarding the stimuli, the number of items per experimental condition might be considered low (particularly in Experiment 2). This was done due to time and study length constraints, considering the number of experimental conditions to be tested.
The sentence's verb transitivity might be a potentially confounding variable in our study, as verb transitivity differences within conditions might have an effect on the recognition and comprehension of the different items. However, intrinsic differences between ERC-Path, ERC-Property, and ERC-Fake prevented us from keeping the number of transitive and intransitive verbs balanced among conditions. While our study was not specifically designed to test the effects of transitivity, we ran additional mixed effects model analysis to check if transitivity was a confounding variable (see in the Online Supplementary Material section Experiment 2: Results for more details). Although we cannot draw strong conclusions from this analysis, the statistical models did not indicate a significant influence of this potential confound in our results.
Regarding our proficiency indicators, we relied on LexTALE scores which is a specific vocabulary measure but has been shown to be a robust predictor of placement test scores (Lemhöfer & Broersma 2012). In addition, we showed significant associations between LexTALE scores and self-reported CEFR level among our participants. While a Placement test might have been a better measure, we consider that the combination of indexes allowed us to describe the participants' English level adequately, and we did observe similar effects in both cases (see Online Supplementary analysis Experiment 2: Results). Conversely, we cannot discard the possibility of selection bias due to the voluntary nature of the task. 12 We have conducted an EEG experiment on ERC processing by EFL learners and its preliminary analysis strongly suggests that Spanish L1 subjects that process ERCs lexically perform better than those who process it syntactically. ERC-Path condition was not included in experiment 1 because we chose to keep syntactic differences between EDC and ERC target items at a minimum. We reasoned that, if AJT ratings mirrored comprehension performance, it would provide further empirical ground for our interpretation of these findings. Future studies could replicate experiment 1 including ERC-Path for greater empirical support of our claims. Even when we find it more parsimonious to assume that linguistic distance is the factor that better explains both patterns of findings, a replication of Experiment 1 with all ERCs would allow to completely discard the complexity hypothesis.
The inclusion of a native speakers' group allowed us to directly compare the recognition and comprehension rates of EDC and ERCs between EFL learners and English speakers, thus verifying that performance differences among constructions were only present in L2. While our results do not indicate processing difficulty differences between these constructions in native speakers, future studies with behavioral measures (such as response and reading times) or eye-tracking techniques might inform this question further.

Conclusions
We have examined the acquisition of three major ERC subtypes by Spanish native speakers, while comparing it to the acquisition of another construction -i.e., EDC-with a mirror image in the speaker's native language. In the first experiment, we observed better recognition of EDC and larger proficiency effects on ERCs. However, the difference between these constructions tends to disappear as proficiency increases, resembling more closely the pattern of native speakers. The second experiment detected comprehension differences only in the less proficient EFL speakers, who performed better in EDCs than in ERC-Property and ERC-Path. ERC-Path comprehension was intermediate in this group, showing no significant differences with the rest of the constructions. In addition, proficiency had a larger impact on ERCs recognition and comprehension than on EDC. We interpret these findings as indicators of an interlinguistic distance gradient for ERC acquisition, where those constructions closer to Spanish are more readily recognized and comprehended. Contrastingly, a complexity hypothesis would have predicted better comprehension performances and lower proficiency effects for the simpler construction (namely, ERC-Path), and more similar outcomes between EDC and the other ERCs. In a nutshell, our results show that the shorter the distance between an L2 construction to an L1 form, the less proficiency is required to acquire it. In addition, interlinguistic distance effects seem to be stronger at earlier stages of the acquisition, and give way to a more native-like pattern as proficiency increases. This is in line with Ringbom's (2006) claims regarding second language acquisition: the degree of congruence between L1 and L2 systems determines how much facilitation there will be in learning for comprehension. An important part of this work has been to make the notions of 'complexity' and 'interlinguistic distance' operational. The former has been defined in terms of the number of grammatical restrictions of a form, while the latter is centered on the set of (not)shared grammatical restrictions by two forms in different languages and the appraisal of those properties in relation to their status as either core or peripheral grammar.
Our results are in line with the thinking for speaking hypothesis (Slobin 1996). ERC is a prototypical instance of the satellite-framed language and, hence, it is expected to be particularly difficult to learn by speakers of verb-framed languages like Spanish. In contrast, the verb-framed/satellite-framed typology is irrelevant for EDC, which is consistent with, first, the presence of a depictive construction in Spanish and, second, the relative easiness in recognition and comprehension of the construction. Furthermore, the further the ERC subtype from Spanish, the more difficult it is to acquire it by Spanish speakers and, thus, it requires a stronger effort to learn the structures' selectional restrictions (van Valin 1991(van Valin , 2000. Once learners manage to reach high level proficiency in L2, they begin to actually "think in the second language" (Robinson et al. 2009: 551) and, as proposed by MacWhinney (2005) the L2 begins to assume a similar status that the native speakers attain for the L1.