Glossa Psycholinguistics

2.1 The role of constituent length

Different factors have been linked to the occurrence of extraposition (for review, see Francis, 2021, on English; Strunk, 2014, on German). Here we focus on those that have been linked to processing constraints. One such factor is constituent length: it has been proposed that relative clauses are extraposed in order to reduce the distance between the host noun and the verb on which it depends. For example, in (1), the relative clause is fifteen words, while the main verb phrase (underlined) is three words long. Thus, extraposition in (1b) increases the dependency distance between the host noun evidence and its modifying relative clause from zero to three words. But, crucially, it reduces the distance between the noun and the verb from seventeen to two words. Thus, extraposition is advantageous in (1b), because it reduces dependency distances overall, as compared to the adjacent version in (1a).

The role of constituent length has been attributed to a processing bias towards local dependencies, according to which longer dependencies are harder to process than shorter dependencies (Futrell et al., 2015, 2020; Gibson, 2000; Hawkins, 2004; Temperley, 2007). Under such accounts, sentences like (1b) represent a more computationally efficient ordering that benefits both producers and comprehenders (Hawkins, 1994, 2004). Similarly, for language production, Arnold et al. (2000) and Wasow (1997a, 1997b) have suggested that the tendency to place shorter constituents before longer ones – known as the “short-before-long” or “end-weight” preference – reflects a processing mechanism that facilitates sentence planning by postponing the production of longer, more complex constituents (see Jaeger & Norcliffe, 2009, and references cited there). Together, accounts in both production and comprehension predict that extraposition should be more likely when the relative clause is long and/or the extraposition distance is short.¹

The prediction above is empirically well supported by corpus and experimental comprehension and production studies. Corpus studies have shown that relative clause length and extraposition distance influence the likelihood of extraposition in English (Francis, 2010; Francis & Michaelis, 2014), German (Strunk, 2014; Uszkoreit et al., 1998), and Persian (Rasekh-Mahand et al., 2016). In comprehension, acceptability studies in English and German have shown that even though adjacent sentences are usually preferred, compared to extraposed ones, the acceptability of extraposed sentences improves with increasing relative clause length and decreasing extraposition distance (Francis, 2010; Francis & Michaelis, 2017; Konieczny, 2000; Uszkoreit et al., 1998). Similar results are found in elicited production experiments in English and German (Bader, 2014; Francis & Michaelis, 2017). For example, Francis & Michaelis (2017) conducted a production experiment in which English participants had to build sentences by combining a noun phrase, a relative clause modifier and a verb. The results showed that extraposed relative clauses occurred most often when the relative clause was long, and the extraposition distance was short. In German, Bader (2014) conducted two production-from-memory experiments in which participants could place a relative clause in whatever position seemed most natural. The results showed that sentences with an extraposed relative clause occurred most often at short extraposition distances.

2.2 The role of verb type

With regard to verb type, previous work has focused on whether extraposition depends on the syntactic and semantic properties of the main clause verb. Concerning syntactic properties, verbs can be copular, transitive, or intransitive. Copular verbs link a subject to a predicate, while transitive verbs have an obligatory object argument in addition to a subject argument. Intransitive verbs have only a subject argument, and they can be further subdivided into unaccusative and unergative verbs, depending on the thematic role of their sole argument. Unaccusatives have a theme/patient subject, while unergatives have an agent subject (Perlmutter, 1978). Differences between unaccusative and unergative verbs might play a role in extraposition. For example, it has been claimed that in neutral contexts, extraposition with an unaccusative verb (2a) is more acceptable than extraposition with an unergative verb (2b) (Culicover & Rochemont, 1990; Guéron, 1980).

These informal reports are supported by a corpus study in English (Francis, 2010), which found that extraposition rates were overall low (15%), but increased with unaccusatives and passives (37%) compared to other verb types (6%). It was argued that this finding was in line with Culicover & Rochemont (1990), because unaccusatives and passives can be classified together, based on their syntactic and semantic properties: both have a single theme/patient argument. Further, a corpus study with 757 relative clauses in Persian found that for subject-modifying relatives, extraposed clauses occurred more often with copular than with transitive verbs – in fact, relative clauses with copular verbs were more likely to appear extraposed than in an adjacent position (Rasekh-Mahand et al., 2016).

Finally, there is evidence that the semantic properties of the verb also play a role in extraposition. Specifically, extraposition has been proposed to be more acceptable with so-called verbs of appearance – a subset of unaccusative verbs – such as appear and arrive (Culicover & Rochemont, 1990; for a similar claim for extraposition of prepositional phrases, see Guéron, 1980). There are also reasons to assume the existence of such a category from a lexico-semantic point of view. For example, Levin & Rappaport Hovav (1995) suggest that appearance verbs are special cases of intransitive (unaccusative) verbs that lack a causative argument in their lexical semantic representation.

The distinction between appearance and non-appearance verbs is supported by an acceptability study in English (Walker, 2013) that compared acceptability judgments of sentences with an extraposed relative clause that contained appearance vs. non-appearance verbs: appear, enter, arrive vs. faint, stumble, and smile, respectively. The results showed that extraposed sentences were more acceptable with appearance than non-appearance verbs, supporting a role for information about verb semantics. Unfortunately, this study did not elicit judgments to adjacent relative clauses, and it also collapsed across different syntactic classes within the non-appearance category (e.g., faint and stumble are unaccusative, but smile is unergative). The current study addresses this issue by jointly considering the semantic and syntactic properties of verbs.

2.3 The role of the grammatical function of the host noun phrase

A final relevant factor for our study concerns the grammatical function of the noun phrase hosting the relative clause. Previous research largely focused on extraposition from grammatical subjects, probably because historically, these studies started by investigating English extraposition. Although extraposition from non-subjects is possible in English (e.g., I saw a man in the bank who claimed to have a gun), extraposition across a main clause verb is not possible, because non-subjects in English are post-verbal. In this regard, languages that allow preverbal subjects and objects – e.g., German, Dutch and Persian – are better suited for an investigation of the role of the grammatical function of the host noun phrase.

In the literature on German, an asymmetry in extraposition from object vs. subject nouns was noted, but it was linked to focus-related factors. Based on a corpus of 1,200 German relative clauses, Shannon (1992) noted that the antecedent of extraposed relative clauses often occurred sentence-medially in the typical focus position in German, directly in front of the clause-final verb, but not in sentence-initial position. Specifically, extraposition seemed very frequent from predicate nominals (3) and, to a lesser extent, from direct objects (4) – both with antecedents in sentence-medial positions. By contrast, grammatical subjects in sentence-initial position seemed resistant to extraposition: (5) is one of the few cases found by Shannon (1992). Thus, although Shannon (1992) did not report rates of extraposition from different grammatical functions, his observations suggest that in German, extraposition from subjects is less common than from other grammatical functions.

Further, Bader (2015) observed that extraposition from objects was preferred when only verbal material intervened (compatible with acceptability data in Konieczny, 2000). He contrasted this with extraposition from subjects of transitive verbs, in which there was at least an intervening object noun phrase. He provided length-, discourse- and prosody-based explanations for this pattern. However, he did not quantify extraposition from subjects of intransitives or passives, in which no non-verbal material intervened. This comparison is relevant, because it could reveal whether grammatical subjects and objects – depending on verb type – behave differently with respect to the effect of length in extraposition. Our study bridges this gap.

Our study proposes that the role of the grammatical function of the host noun phrase should be considered together with verb type, thereby suggesting a role for argument structure in extraposition. Before we introduce our proposal, we describe previous accounts of extraposition which have attributed the effects of verb type and grammatical function to discourse structure and/or prosodic constraints.

2.4 The role of discourse and prosody

Some accounts have tried to explain the occurrence of extraposition in terms of discourse and pragmatic factors (Culicover & Rochemont, 1990; Guéron, 1980; Lee, 2019; Poschmann & Wagner, 2016; Takami, 1999). For example, it has been proposed that extraposition functions as a device to introduce a new subject into the discourse, with the intervening verb phrase conveying old or backgrounded information. This claim is supported by the observation that extraposition is more frequent from indefinite noun phrases, which typically introduce new referents (for related syntactic proposals, see Baltin, 2006; Guéron & May, 1984; Kayne, 1994), or more acceptable from focused noun phrases (Bolinger, 1992; Huck & Na, 1990). Discourse-based accounts can also explain the increased likelihood of extraposition with verbs of appearance, because such verbs often introduce a new referent.

When discourse-based accounts are evaluated with empirical data, the results are mixed. For example, using a corpus of 345 English sentences containing subject-modifying relative clauses, Francis (2010) and Francis & Michaelis (2014) showed that extraposition was more likely when the subject was new and the verb phrase was contextually accessible – when a semantically related verb phrase had been mentioned in the previous context. However, Rasekh-Mahand et al. (2016) did not find that the likelihood of extraposition in Persian was affected by the givenness or accessibility of the subject or the verb phrase. Further, extraposition is not always limited to indefinite subjects: a corpus study in German with 2,555 relative clauses found that approximately 35% of extraposed relatives had a definite antecedent (Strunk, 2014). Another concern with accounts based on givenness/newness is that they cannot explain extraposition from noun phrases that serve other grammatical functions, such as object, which are part of the verb phrase. Therefore, it is unclear how they generalize to languages like Dutch, German, and Persian, in which extraposition from non-subject nouns is more frequent.

Discourse-based accounts of extraposition have also been proposed for Dutch and German, but based on different observations. Perlmutter & Zaenen (1984) suggested that extraposition in these languages was possible only if the antecedent was a non-subject and occurred in a non-initial sentence position. Corpus data from Shannon (1992) confirmed that sentence-initial arguments do not usually show extraposition in German, but it suggested that it was not subjecthood – but rather initial position – that was resistant to extraposition. Specifically, although extraposition from sentence-initial positions was rare in German, it was not limited to subject nouns: it was attested with other elements, such as direct and indirect objects, prepositional phrases, and adverbs. Based on this observation and observations about definiteness, Shannon proposed a pragmatic account according to which extraposition is motivated by sentence focus. The claim was that extraposition was only possible with focused antecedents, which are new and/or more important information and thus tend to occur in a non-initial sentence position.

While the focus-based theory of Shannon (1992) can account for extraposition from non-subject positions, its empirical support is mixed. When Shannon asked a German native speaker to read aloud sentences from his corpus, the noun modified by the extraposed relative clause was not always focused and did not always carry the sentence stress, contrary to his expectations. An effect of focus was reported in an acceptability experiment conducted by Poschmann & Wagner (2016). In this experiment, participants listened to a question, read aloud a predetermined answer and then provided an acceptability judgment. The results showed a higher acceptability of extraposition in response to subject focus questions, as compared with object focus and wide focus questions. However, there was no evidence that accent on the antecedent of the relative clause affected extraposition. In the domain of production, Bader (2024) reported a small effect of context on the extraposition rate from object nouns. In a constrained production experiment in German, participants silently read a context scenario followed by a set of phrases in a jumbled order. Their task was to arrange the phrases into a sentence and then utter it. In the focus condition, the main verb was mentioned in the context, while the noun phrase and relative clause in the target sentence were new. In the topic condition, the main verb was new, while the noun phrase and relative clause were previously mentioned in the context. Participants extraposed the relative clause more often in the focus condition than in the topic condition, although the effect was small (56% vs. 47%).

Finally, other accounts have tried to explain extraposition based on prosodic considerations (Féry, 2015; Göbbel, 2013; Hartmann, 2013, 2017; Truckenbrodt, 1995). For example, Hartmann (2013, 2017) suggested that extraposition was used as a re-ordering strategy to avoid ill-formed prosodic structures. In this account, information structure is only indirectly related to extraposition, because the alleged prominence on the host noun, which is sometimes realized as focus, results from the prosodic restructuring that happens due to extraposition.

The prosody of the intervening material has also been argued to affect extraposition, such that prominent interveners block extraposition more than unaccented ones. For example, extraposition is more likely over a locational or directional adverbial compared to a noun phrase (Kathol & Pollard, 1995), or over verbs of appearance compared to other verbs (Guéron & May, 1984). In both cases, the former element is typically assumed to be unaccented (Bolinger, 1992; Poschmann & Wagner, 2016). Others have argued against this idea (see Féry, 2015, for a review). Empirical data, has not provided much support for these hypotheses. For example, Bader (2014) showed in a production-from-memory experiment that even an unaccented indefinite pronoun (etwas, ‘something’) decreased the likelihood of extraposition from subjects when the pronoun was the only intervening material before the verb.

In sum, the available empirical evidence in English and German suggests that neither prosody nor discourse and pragmatic factors can fully explain the occurrence of extraposition across different verb types and grammatical functions. Thus, it is useful to explore additional explanations. The next section outlines a processing-based account of extraposition in terms of argument structure.

4.1 Data preparation and coding

In the first step, 4,020 sentences containing a relative clause were extracted from the two corpora. The sentences were automatically coded for relative clause position (adjacent/extraposed), relative clause length (number of words in the clause), (potential) extraposition distance, grammatical function of the host noun phrase (subject/object/nominal copular predicate), verb voice (active/passive), and whether the object noun had direct object marking (dom). Note that (potential) extraposition distance refers to the number of words between the antecedent of the relative clause and the relativizer ke in the extraposed version as shown in (9b) (from PerUDT corpus). In the adjacent version (9a) (constructed from the original extraposed sentence), the potential distance for extraposition was the number of words in the post-relative clause region of the sentence.

In the second step, a list of unique main clause verbs was created from the extracted data. These verbs were manually classified as transitive active, transitive passive, unergative, unaccusative appearance, and unaccusative non-appearance (see the next section for definitions of verb classes). Copulas and verbs that did not fall into any of these categories were classified as: (i) be-have stative, which comprised all adjectival copular constructions, such as ŝād budan ‘be happy’, and light verb constructions with the light verb dāŝtan ‘have’, such as qarār dāŝtan ‘be located’, (ii) existential, which comprised budan and vojud dāŝtan ‘exist’, and (iii) copular with a predicate nominals (henceforth, copula).

During the second step, all sentences with a copula were manually checked to ensure that the verb type was correctly categorized. Additionally, sentences with a main clause verb that could potentially take a complement clause, such as tasmim gereftan ‘decision make’, were manually checked. Sentences were discarded if they did not include a relative clause, but rather a wrongly coded complement or appositive clause that modified a proper noun. Other discarded cases included incomplete or incomprehensible sentences, sentences with an ambiguous relative clause attachment site, and sentences in which the noun phrase hosting the relative clause was in a non-canonical order (e.g., topicalized). Automatic annotations were also manually checked if they included unusually long relative clauses and/or extraposition distances, or incompatibilities such as a noun phrase being coded as an object when the main clause verb was not transitive. This resulted in the manual checking of 1,986 sentences (out of 4,020) and the exclusion of 990 sentences. From the remaining 2,034 sentences, a random subset of 102 tokens (5% of the data) was manually checked to evaluate the accuracy of the automatic annotations. The rate of exclusion based on the criteria described above was 15% (15 out of 102 tokens). Additionally, of the 87 included tokens, only four relative clause lengths and eight extraposition distances had to be manually corrected. This means that the accuracy of automatic annotation for relative clause length and extraposition distance was 95% and 91%, respectively.

Finally, relative clauses were coded for the argument type associated with their host noun phrase: internal vs. external. Existentials, nominal copular constructions, and be-have statives were coded as lacking argument structure (Dechaine, 1993; Hazout, 2004). Note that adjectival copular predicates are sometimes considered to have an argument structure, because it is argued that the adjective behaves like a verb and can take complements (e.g., Chomsky, 1981; Haegeman, 1999; Stowell, 1983). Because this proposal is controversial in the theoretical literature (Pesetsky, 1982; Rothstein, 1999), we chose not to code adjectival copular predicates as having an argument structure.

4.2 Analysis

The statistical analysis was performed in R (R Core Team, 2023). Logistic regression models were used (Jaeger, 2008), because the dependent variable, RC Position, was binomial: whether a relative clause was adjacent (0) or extraposed (1).

The analysis was conducted as follows. First, we addressed our research question about the role of argument structure by evaluating whether a model with the predictor Argument Type provided a better fit to the data as compared to a model that included only the factors previously studied in the extraposition literature: RC Length, [extraposition] Distance, Host NP Function (subject/object), and Verb Type (transitive/unergative/unaccusative). Thus, two models were compared: a model with these four factors vs. a model that included these four factors plus Argument Type (external/internal). Model comparisons were performed using Chi-Square tests and the Akaike information criterion, or AIC (Akaike, 1998). AIC quantifies the change in the goodness of a model fit as a result of changing the number of predictors: a lower AIC indicates a better model fit (Navarro, 2019).

The model comparison revealed that the model with the predictor Argument Type provided a better fit to the data. Thus, its output is presented as the main model in 4.3.3. In this model, all categorical predictors were sum-coded. Numeric predictors were centered and coded as continuous. Note that due to the inclusion of the factor Argument Type, the main model could only be run on the corpus sentences that had a main clause verb with a clear argument structure (i.e., sentences with be-have statives, existentials or nominal copular verbs as main clause were excluded). Random effects were not used, because there was no relevant information – such as text style – in the corpora.

Since the first model included only the subset of the data whose argument structure could be unambiguously determined, a second or exploratory model was built to include the data without a clear argument structure – i.e., predicate nominals, be-have statives, and existentials – as well as to diagnose the potential contrast between appearance and non-appearance verbs. All predictors of the main model were kept, except for Argument Type. In the exploratory model, the factor Verb Type had the levels ‘transitive/unergative/unaccusative/existential/stative’. The levels ‘copula’ for Verb Type and ‘nominal predicate’ for Host NP Function had to be removed, due to convergence issues. This was due to a “complete separation” or “quasi-complete separation” issue in the data, which happens when an outcome is absent or very rare in the data (Albert & Anderson, 1984). In our case, for sentences with a nominal copular predicate, the subject noun phrase never appeared with an extraposed relative clause (0 out of 118 tokens), while the predicate nominal only appeared in 0.5% of cases with an adjacent relative clause (2 out of 428 tokens).

Additionally, the factors Verb Voice (active/passive) and Semantic Type (non-appearance/appearance) were entered into the exploratory model. Note that verb voice and semantic type are meaningful only for transitive and unaccusative verbs, respectively. Therefore, they were entered into the model as nested variables for the relevant verb types. The data and analysis code are publicly available at https://osf.io/6fmtz.

4.3 Results

Out of 3,015 relative clauses, 1,942 (64%) were adjacent, and 1,073 (36%) were extraposed. The results of the main and exploratory models are described below. For ease of interpretation, the figures show descriptive summaries of the data in the percentage scale.

4.3.1 Main model

To evaluate the usefulness of Argument Type as a predictor of extraposition likelihood, we compared a full model with all the predictors described in 4.2 with a model that differed only in that it lacked the factor Argument Type. The model with the factor Argument Type substantially lowered AIC, showing that it provided a better fit to the data: 1984.1 vs. 1991.8. An analysis of variance using a Chi-Square test confirmed that the model with argument type was significantly better (Deviance = 9.652, Df = 1, p = 0.002). Thus, this model was used as the main model. It examined the likelihood of extraposition as a function of argument structure in addition to grammatical function, verb type, and length-related factors (Table 1). The results showed a main effect of RC Length and a negative main effect of Distance: increasing RC Length and decreasing Distance increased the likelihood of extraposition. These effects match well with the descriptive results: extraposed relative clauses were, on average, 2.7 words longer than adjacent relative clauses. The (potential extraposition) distance was on average 3.7 words shorter for extraposed relative clauses than for adjacent relative clauses (Figure 1).

Table 1: Estimates of the main model in log odds. The intercept represents the likelihood of extraposition at the average relative clause length and extraposition distance for verbs that have an argument structure – this corresponds approximately to a probability of 12.56%. Positive estimates reflect an increase in the likelihood of extraposition. Model formula: RC Position ∼ RC Length + Distance + Host NP Function + Verb Type + Argument Type.

Coefficient	Estimate	Std. Error	z value	Pr(>|z|)
(Intercept)	–1.940	0.191	–10.144	<0.001*
RC Length	0.117	0.012	9.881	<0.001*
Distance	–0.113	0.019	–6.006	<0.001*
Host NP Function object–subject	0.688	0.439	1.565	0.118
Verb Type unaccusative–unergative	–0.784	0.976	–0.804	0.422
Verb Type transitive–unergative	0.860	0.581	1.480	0.139
Argument Type internal–external	1.986	0.174	11.429	<0.001*

Figure 1

Descriptive summary of the distribution of extraposition distances and relative clause lengths in adjacent vs. extraposed clauses. Diamonds display mean values. Each dot represents a corpus token.

Crucially, the main model showed a main effect of Argument Type, with a higher likelihood of extraposition from internal arguments than external arguments. This demonstrates that argument structure affected extraposition rates. Figure 2 shows that for verbs with an argument structure, rates of extraposition increased for internal arguments (subjects of passive and unaccusative verbs, as well as objects of transitive verbs) compared to external arguments (subjects of active transitive and unergative verbs). No significant effect of Verb Type or Host NP Function was found. This means that once Argument Type was entered into the model, there was no evidence that the likelihood of extraposition was further modulated by either of these factors.

Figure 2

Descriptive summary of extraposition rates from different argument types. The numbers below the bars indicate the number of corpus tokens. Error bars show binomial confidence intervals. Extraposition rates from subjects of copulas are not shown, because they were absent from the corpus.

4.3.2 Exploratory model

The exploratory model was run to additionally include corpus sentences that could not be assigned a clear argument structure – i.e., predicate nominals, be-have statives, and existentials – and to test for potential differences between appearance and non-appearance verbs. The results of the exploratory model showed that increasing RC Length and decreasing Distance increased the likelihood of extraposition, similar to the results of the main model (Table 2).

Table 2: Estimates of the exploratory model in log odds. The intercept represents the average likelihood of extraposition at the mean relative clause length and extraposition distance for subject nouns of active transitive verbs – this corresponds approximately to a probability of 6.04%. Positive estimates reflect an increase in the likelihood of extraposition. Model formula: RC Position ∼ RC Length + Distance + Host NP function + Verb Type + Verb Type: Verb Voice + Verb Type: Semantic Type.

Coefficient	Estimate	Std. Error	Z Value	Pr(>|Z|)
(Intercept)	–2.770	0.182	–15.226	<0.001*
RC Length	0.120	0.012	10.372	<0.001*
Distance	–0.123	0.019	–6.550	<0.001*
Verb Type unergative	–0.311	0.365	–0.852	0.394
Verb Type unaccusative	0.762	0.244	3.119	0.002*
Verb Type stative	0.284	0.390	0.730	0.465
Verb Type existential	4.161	0.342	12.182	<0.001*
Host NP Function object	2.269	0.197	11.502	<0.001*
Verb Type transitive: Verb Voice passive	1.593	0.469	3.398	0.001*
Verb Type unaccusative: Semantic Type appearance	1.204	0.309	3.898	<0.001*

With regard to the Host NP Function, extraposition rates were the highest from nominal predicates of copular constructions, followed by objects and subjects (Figure 3). The exploratory model showed that the likelihood of extraposition increased when the noun hosting the relative clause was a grammatical object compared to a grammatical subject.

Figure 3

Descriptive summary of rates of extraposition as a function of the grammatical function of the host noun phrase and the verb type. The numbers below the bars indicate the number of corpus tokens. Error bars show binomial confidence intervals. Extraposition rates from subjects of copulas are not shown because they were absent from the corpus.

Concerning Verb Type, the descriptive results suggested that extraposition from subjects was modulated by the type of predicate: the highest rate of extraposition was found with existentials, followed by passives, unaccusatives, be-have statives, active transitives, and, finally, unergatives (Figure 3). The model showed that the likelihood of extraposition from subjects increased when the verb was unaccusative non-appearance, unaccusative appearance, existential, or passive, as compared to when it was active transitive.

Finally, extraposition was more likely with appearance than non-appearance verbs: 32% vs. 13%, respectively. Nested comparisons between these two semantic classes supported an increase in the likelihood of extraposition for appearance verbs.

5.1 The role of dependency length: A processing preference

We reported results from two statistical models. Both models found that the rate of extraposition increased with increasing relative clause length and decreasing extraposition distance. Following previous proposals, we assume that the role of length-related factors can be explained by an end-weight processing preference (Arnold et al., 2000; Wasow, 1997a, 1997b). Specifically, longer sentence constituents might be more difficult to process, and thus speakers might opt to postpone them when their grammar allows it. This processing preference has been consistently found with other constructions, such as verb particle movement (Lohse et al., 2004; Wasow, 1997a) and heavy NP shift (Arnold et al., 2000; Wasow, 1997a).

In the case of relative clauses, extraposition might be advantageous because it allows speakers to keep the verb and its arguments close while postponing the planning of another clause to a later point in time. Thus, delaying the relative clause enables speakers to more quickly integrate the verb and its arguments, which frees up memory resources for the later planning and production of the relative clause. A processing-based explanation of the end-weight preference predicts that the longer the relative clause – and, thus the higher memory burden – and/or the less pre-verbal material to be integrated with the verb, the higher the processing advantage of extraposing the relative clause.

5.2 The role of verb type and grammatical function

In the corpus, relative clauses were extraposed more frequently from grammatical objects than subjects (on average, 42% vs. 15%), as shown by an effect of the factor Host NP Function object in the exploratory model. This result is compatible with findings from German (Shannon, 1992). We failed to find an effect of verb type on the likelihood of extraposition from subjects of unergatives and active transitives, but we found a higher likelihood of extraposition with unaccusatives and passives compared to transitive actives. This is also compatible with findings from English (Culicover & Rochemont, 1990; Francis, 2010).

Past research has attributed these patterns to discourse factors, such as givenness/newness and sentence focus. We propose that the emerging pattern for the role of verb type and grammatical function might instead reflect a role of verb argument structure: extraposition is more likely from internal arguments (subjects of passives and unaccusatives and objects) than external arguments (subjects of active transitives and unergatives).

5.3 Argument structure plays a role in extraposition

The evidence supporting our hypothesis that argument structure affects relative clause extraposition comes from the main model, which showed an effect of argument type: internal arguments were more likely than external arguments to host a noun modified by an extraposed relative clause. Furthermore, the model including argument structure as a predictor provided a better fit to the data compared to the model without it.

We propose a production-based processing explanation for this result. This explanation is based on previous studies, which have shown that internal and external arguments are planned differently during language production (Momma & Ferreira, 2021; Momma et al., 2014, 2016, 2018; Schriefers et al., 1998). These studies suggest that internal arguments are planned together with verbs, due to their closer semantic and syntactic relation. Meanwhile, external arguments (as well as other material whose encoding does not depend on the verb) might be planned independently of the verb, at a different point in time. This assumption aligns well with the theoretical literature, in which an external argument is given a special status and sometimes not considered a true argument of the verb (Grimshaw, 1990; Kratzer, 1996; Williams, 1981). By contrast, internal arguments are considered true arguments. This implies that internal arguments have a closer relationship with the verb than external arguments.

We extend the processing explanations above to verb-final structures containing a relative clause. Our account of relative clause extraposition is as follows. First, we assume that verbs and their internal arguments are planned together in speech, thus forming a syntactic dependency. In verb-final structures – where verbs are uttered at the end of a clause – this dependency has to be retained in memory while intervening material is planned and produced – e.g., the relative clause, which can be syntactically complex, because its verb and argument structure can differ from those of the main clause.

Crucially, producing the relative clause in a noun-adjacent position should be more disruptive when the noun phrase is an internal argument, as both the verb and noun were planned jointly and thus are already activated in speakers’ memory, as in example (13). In such situations, speakers might opt to first utter the already planned verb, freeing up memory resources for relative clause planning. By contrast, this process is less likely to happen when the relative clause attaches to an external argument, as in example (14), because external arguments do not need to be planned together with the verb. Therefore, extraposition should occur preferentially from internal as opposed to external arguments.

A testable prediction of this hypothesis is that the effect of argument structure should be greater in relative clause extraposition than in prepositional phrase extraposition. This is because, as argued by Momma & Ferreira (2021), planning a relative clause, which contains a verb, should be more costly than planning a prepositional modifier, which lacks verbal material. Future studies could test this prediction by directly comparing the role of argument structure in the extraposition of relative clauses vs. prepositional modifiers.

5.4 Extraposition in structures with an unclear argument structure

We discuss two categories in this subsection: copular constructions with a predicate nominal and be-have statives. As for copular constructions with a predicate nominal, the predicate nominal in our data always appeared with an extraposed relative clause, similarly to what Shannon (1992) found in German. Copular constructions, which are often assumed to lack a thematic structure, due to the reduced semantic content of the copula (Löbel, 2000), show a different pattern of extraposition that seems to be independent of relative clause length, but might be related to the short extraposition distance. Furthermore, the subject of copulas in our data always appeared with an adjacent relative clause. Since these constructions are composed of two noun phrases followed by the linking copula (e.g., That guy a professor is), a post-verbal relative clause can seemingly attach only to the closer noun phrase, which is the predicative one, and not the subject. We conjecture that this serves the purpose of ambiguity avoidance. To sum up, copular verbs with a predicate nominal appear to lack argument structure, so argument structure does not seem to play a role in the extraposition of these constructions.

With be-have statives, our data showed a low rate of extraposition from subjects. Further, there was no evidence that this rate differed from subjects of active transitives and unergatives. Since the argument structure of statives is debated in the literature (Rothmayr, 2009), they were not included in the main model that assessed the role of argument structure. However, we suggest that subjects of stative verbs could be seen as agent-like arguments, and that this is probably why they showed a very low rate of extraposition. This position is similar to the proposal by Levin & Rappaport Hovav (2005) and Rappaport Hovav & Levin (1998), who suggested analyzing stative verbs similarly to unergative verbs such as run and whistle, which they treat as activities. In this analysis, states and activities are events with a single structural argument: the ‘holder’ of the state. This argument is similar to the ‘actor’ of an activity and thus analogous to an agent argument.

5.5 Semantic factors affect extraposition in Persian: Existentials and appearance verbs

Previous studies have shown that extraposition from subjects is more acceptable and more common with unaccusative appearance and existential verbs (Culicover & Rochemont, 1990; Francis, 2010; Francis & Michaelis, 2017; Guéron, 1980). The exploratory model showed that these verbs do show the highest rate of extraposition from subjects. Specifically, existentials showed the highest rate of extraposition from subjects (87%), and appearance verbs showed the second highest rate (32%). This pattern has been associated with the discourse function of extraposition, that is, introducing the subject to the ‘world of discourse’ (Guéron, 1980, p. 654).

These results underscore the key role that discourse-related factors play in extraposition (Culicover & Rochemont, 1990; Francis, 2010; Francis & Michaelis, 2017; Guéron, 1980; Lee, 2019; Takami, 1999). Our proposal that argument structure affects extraposition should not be interpreted as claiming that syntactic and semantic factors can solely explain how and when speakers decide to extrapose a relative clause in speech. Instead, we believe that an appropriate cognitive model of such decisions should take into account different factors, including: (a) length-related and prosodic variables; (b) sentence-internal syntactic and semantic relationships; and (c) discourse-related factors, which include the discourse status of the sentence constituents, as well as the communicative intentions of a speaker. In our view, processing considerations related to argument structure are only one piece of the puzzle.

5.6 Limitations and future directions

Our study has limitations, and it also leaves some open questions for future research. First, one factor that we did not examine is definiteness, which features prominently in research on extraposition in other languages. Specifically, extraposition rates and acceptability have been found to correlate with the indefiniteness of the antecedent of the relative clause (Culicover & Rochemont, 1990; Guéron & May, 1984; Strunk, 2014; Walker, 2013). This factor could not be properly assessed in our study, because written Persian does not have definite articles, and also because the indefinite marker – the suffix -i – is homophonous with the restrictive relative clause marker. Therefore, the morphological markers in our data did not unambiguously indicate whether a noun was definite or indefinite. But as a preliminary assessment, we examined the distribution of -rā in our data, which is a definite direct object marker dom. We found only 30% of dom-marked nouns modified by an extraposed relative clause, while the pattern was reversed for dom-unmarked nouns, with a 70% rate of extraposition. This result is compatible with previous findings, but to rigorously assess the role of definiteness in Persian, future research should use experimental data, which allows us to construct stimuli controlled for definiteness.

Other limitations of our study relate to the use of written data. It was not possible to explore the role of focus or prosody, which have been reported to play a role in relative clause extraposition (Francis, 2010; Francis & Michaelis, 2017; Shannon, 1992; Takami, 1999). More critically, written corpus data cannot be used to assess the real-time choices of speakers during sentence processing. In addition to the potentially unlimited time available to speakers when writing, they may follow stylistic rules to form constructions that are more frequent in written than spoken language. For example, it has been reported that extraposition is sometimes used as a stylistic strategy to avoid center-embedded relative clauses with two consecutive verbs in sentence-final position (Najafi, 1992). Therefore, the conclusions drawn from our study need additional support from experimentally controlled studies.