2.1 Grammatical weight and word order

The study of the role of grammatical weight (or heaviness) on the linear order of the constituents in a sentence is an important research topic in language sciences, including theoretical linguistics, psycholinguistics, as well as linguistic typology. It is an old topic, yet continues to motivate ongoing theoretical debates. Interest in the topic within general syntax goes far back. The first generalization on the matter, known as the “end-weight principle”, was originally formulated by Behaghel (1909) as the “law of increasing constituents” (Gesetz der wachsenden Glieder). According to Wasow (1997: 82), who presents one of the key contributions to the topic, the term end-weight was first used by Quirk et al. (1972) in their description of the English grammar where they find a “tendency to reserve the final position for the more complex parts of a clause or sentence” (Quirk et al. 1972: 943).

Early accounts of end-weight are formulated in terms of language processing and comprehension (Bever 1970; Kimball 1973; Frazier & Fodor 1978), assuming that postponing heavy elements facilitates processing. Meanwhile, some studies do not find parsing-oriented (or hearer-oriented) accounts entirely convincing and argue that postponing heavy elements serves to facilitate the speaker’s utterance planning and production (Wasow 1997; Stallings et al. 1998; Arnold et al. 2000; Wasow 2002; Chang 2009; Stallings & MacDonald 2011). Important early contributions to this view are two experimental works by Stallings et al. (1998) and Wasow (2002), who studied the heavy NP shift in English in a series of sentence production experiments, so as to credit the hypothesis that end-weight cannot only be explained by hearer-oriented accounts.

Indeed, a straightforward account of the end-weight effect can be framed in terms of availability-based incremental models of sentence production (e.g. Garrett (1980); Kempen & Hoenkamp (1987); De Smedt & Kempen (1987); Levelt (1989); Bock & Levelt (1994); De Smedt (1994); Kempen & Harbusch (2003)). In this view, the linear order of constituents reflects the order in which they become available for production, as long as grammar rules do not intervene. Constituents that become available at an earlier point in time can occupy an earlier linear position than constituents emerging later. Phrasal length is one factor modulating availability: everything else being equal, short constituents require less processing time and thus become available for production sooner than longer ones. As a matter of fact, the short-before-long or simply “short-first” preference (more commonly used in the psycholinguistic literature) presents a strong empirical argument for availability-based models of sentence production.

However, this account implies that the end-weight preference should be universal, as it has indeed been explicitly or implicitly assumed in many studies (see Hawkins (2007) for an overview). In other words, since the architecture of the language production system is assumed to be universal and the availability-based explanation is grounded in general principles of cognition, the short-before-long preference is expected to hold in all languages. Yet, as has been underscored by Hawkins’s typological survey (Hawkins 1994 and subsequent work), OV languages like Japanese show the mirror image of the preference observed in VO languages such as English, that is, the long-before-short preference.

2.2 Long-before-short preference in OV languages

Two types of hypotheses are available to account for the long-before-short preference in OV languages: 1) parsing-oriented and 2) production-oriented.

1. Dependency-based distance-minimizing accounts motivated by efficient processing, such as the Early Immediate Constituent (EIC) or the Minimize Domains (MiD) principles (Hawkins 1994; 2004; 2014), the Dependency Locality Theory, hereafter DLT, (Gibson 1998; 2000) or the Dependency Length Minimization tendency, hereafter DLM, (Temperley 2007; Gildea & Temperley 2010; Temperley & Gildea 2018), measure the complexity of parsing on the basis of the (linear) distance between the head verb and its complements (in number of words) and predict mirror-image preferences in head-initial vs. head-final languages.⁶ These models, however, differ in the measure they use to rank different alternative sentences.

   Hawkins (1990 and subsequent work) proposes a theory of the human parser based on efficient processing of complexity in grammar. In a nutshell, between two equally grammatical constructions, the human parser prefers the one that can be processed with optimal efficiency. To measure complexity, Hawkins takes into account the number of words that are needed to be parsed in order to recognize the syntactic structure of a sentence, or, in other words, all its immediate constituents (IC). The Early Immediate Constituent (EIC) principle,⁷ or its more recent version Minimize Domains (MiD),⁸ provides a metric to predict word order preferences between alternative/equally grammatical sentences (e.g. V-NP-PP vs. V-PP-NP). This metric depends on the Constituent Recognition Domain (CRD),⁹ which measures the size of the processing domain, roughly, for a VP in a VO language, the distance between the verb and the head of its last/rightmost complement (see Section 3.2 for illustrative examples). MiD predicts that, with an equal number of words, a sentence with a smaller CRD should be preferred.

   DLT and DLM, on the other hand, calculate the cumulative distance between the verb and the heads of all its dependents, referred to as the dependency length, and predict a preference for the sentence with a shorter dependency length. The two models differ substantially. While Gibson’s DLT only takes into account constituents with “new discourse referents” (Gibson 1998: 12), the DLM model includes all words when calculating the dependency length (see Temperley 2007: 303–304).¹⁰

2. Recall that accessibility-based incremental models of sentence production predict a universal end-weight preference. To account for both long-before-short and short-before-long preferences in a unified manner, that is, via the same model of sentence production, Yamashita & Chang (2001) propose a language-specific version of the accessibility-based model, assuming that: a) longer constituents are conceptually more salient, and therefore more accessible, than shorter ones, and b) in flexible languages and in the preverbal domain, conceptual factors have a stronger effect on sentence production than formal factors. Departing from the observation that long-before-short invalidates availability-based models of sentence production, the authors suggest that both of these preferences can receive an accessibility-based account within the framework of incremental production.

   Their rationale is that: 1) since both conceptual and form-related factors are shown to influence word order preferences (see Bock 1982), a production system’s sensitivity to these factors can be viewed as being language-specific; 2) longer constituents, while being more complex, are lexically richer, hence conceptually more accessible than shorter constituents. Accordingly, they claim that, being more sensitive to conceptual factors, Japanese speakers prefer long-before-short ordering, while English is more sensitive to formal factors and English speakers prefer the opposite ordering. This difference, the authors argue, resides in the fact that: 1) English has a more rigid word order than Japanese; 2) In English weight-based shifts in ordering take place after the verb, and they assume, following Stallings et al. (1998), that “verbs exert strong influence over the phrases that follow them, including their order” (Yamashita & Chang 2001: B54)). More precisely, the authors suggest that “[s]ince accessibility of meaning and form can have different influences in production (Bock 1982), it could be that Japanese speakers are focusing more on conveying meaning (putting enriched material earlier), while English speakers are focusing on sequencing forms (putting easily accessed word earlier).” (Yamashita & Chang 2006: 6).

Each of these two hypotheses (parsing-oriented or production-oriented) tackles the problem from a different angle. Hence, they are not contradictory and can be assumed to hold simultaneously, as long as they are not falsified by the data. Furthermore, in previous studies, the predictions of these hypotheses converge for the majority of data discussed – coming from strictly head-final languages such as Japanese (Hawkins 1994; Yamashita & Chang 2001), Korean (Choi 2007) or Basque (Ros et al. 2015) – and makes it difficult to tease them apart.

Parsing-oriented (distance-minimizing) accounts nevertheless seem to have gained more approval among researchers because these accounts maintain the same mechanism for VO and OV languages. Moreover, a couple of recent large-scale cross-linguistic corpus studies conducted on available dependency tree-banks support the universality of dependency-distance minimizing hypothesis (e.g. Futrell et al. (2015); Haitao Liu et al. (2017)). The latter have also included Persian data as providing support for this hypothesis. However, it is important to note that there is a strong tendency for the postposition of subordinate clauses in Persian. For object clauses, this is a grammatical constraint. But even relative clauses that modify a preverbal NP can be extraposed to the postverbal domain.¹¹ This can easily tip the overall balance in favor of dependency-distance minimization in such large-scale corpus-based analyses. Note that there are also corpus studies that report contradictory results for mixed type languages (see Yao (2018) for Mandarin Chinese and Zoey Liu (2019) for a large-scale cross-linguistic study of PP ordering across 31 languages including a number of languages with mixed type patterns, e.g. head-initial PPs appearing before the verb).

Yamashita & Chang’s (2001) language-specific production-oriented account, on the other hand, stipulates a difference in the production system of OV and VO languages, to which not all studies subscribe (e.g. Tanaka et al. (2011); Ros et al. (2015); Tachihara & Goldberg (2020)). Moreover, this model can also be called into question as a post hoc accommodation of existing models, namely, it does not provide independent underpinning for the claim that longer constituents are conceptually more accessible (but see e.g. Karimi & Ferreira (2016)). The claim that salience plays different roles in typologically different languages needs to be empirically justified.

However, there are also studies that pay credit to this language-specific availability-based model and adhere to the conceptual accessibility hypothesis. For instance, Kempen & Harbusch (2003), in their account of word order scrambling as a consequence of incremental sentence production, maintain the availability-based model but with a more inclusive definition considering conceptual factors in line with Yamashita & Chang (2001) to determine constituents’ availability along with formal complexity and heaviness. Moreover, Stallings & MacDonald (2011) reiterate that: 1) it is the relative accessibility of constituents that influences word order during sentence production and 2) several factors other than length and complexity, including lexical-semantic properties, can possibly modulate accessibility.¹² They argue that (relative) length is a strong modulating factor of accessibility in English, while it has a weaker effect in Japanese. They also suggest that how (relative) length affects order in other languages is an empirical question.

It is important to bear in mind that salience is a complex and multi layered notion, and it is crucial to capture its various dimensions and the ways in which they interact. Conceptual accessibility of constituents, for instance, remains a vague and sometimes even confusing notion in the literature, especially with regard to its relation with discourse-related accessibility and the information structure of the sentence (e.g. topics are accessible entities and are generally realized by short/simple constituents such as pronouns).¹³

There however seems to be an agreement on the definition of accessibility in psycholinguistic literature, as “the ease with which the concept associated with a noun phrase (NP) can be retrieved from memory”, as well as a consensus that the latter is “one of the most influential factors” in the processing and resolution of ambiguous pronouns (Karimi & Ferreira 2016: 507).¹⁴ Importantly, in this line of studies, the informativity and/or specificity of a referring expression, such as an NP, is considered as a potential factor to enhance its accessibility. The length of an NP directly reflects the amount of information attached to it, and experimental studies (on pronoun ambiguity resolution) suggest that longer (referring) NPs are more accessible by virtue of being more informative – roughly because the additional information can “make it easier to retrieve the associated representation from memory” (Karimi & Ferreira 2016: 520). In this paper, we view conceptual accessibility along the same vein and assume that the informativity and/or specificity of a referring expression enhances its conceptual accessibility.

3.1 The interest of Persian data

Persian is an SOV and a three-way pro-drop (subjects, direct objects and obligatory PP arguments can be omitted) language with flexible word order and mixed head direction. While NP objects precede the verb, clausal objects follow it. Furthermore, NPs and CPs are head-initial¹⁵ and Persian has prepositions instead of postpositions. As such, Persian shares several properties with both VO and OV languages studied so far and makes the comparison promising for our understanding of length-based effects on word order variations.

In addition, Persian exhibits differential object marking (DOM),¹⁶ realized by the enclitic = rā and triggered (roughly) by definiteness.¹⁷ Note that Persian does not overtly mark definiteness. Definite NPs can be formed either by different definite determiners, like demonstratives, e.g. in medād ‘this pencil’, or by no overt determination, medād ‘the pencil’. In Persian, a definite DO is always rā-marked (1a).¹⁸ A DO lacking = rā, and carrying no determination or quantification, like medād or medād=e qermez in (1b), will necessarily receive a “bare noun” reading, that is, a nonspecific (existential or a kind-level/generic) reading. Note also that Persian bare nouns are underspecified for number and can consequently yield a mass reading (even if countable). Contrary to bare DOs, indefinite DOs are always specified for number and have an existential reading (1c). Indefiniteness is overtly marked, by the enclitic = i, e.g. medād=i, the cardinal yek e.g. yek medād, or the combination of both, e.g. yek medād=i ‘a pencil’. Naturally, indefinite NPs are also formed by different indefinite quantifiers, e.g. čand medād ‘a few pencils’.¹⁹ Indefinite DOs can also be rā-marked to receive a specific reading, however, here we only use the label “indefinite” to refer to non-rā-marked indefinite DOs.

DOM has important bearings on the canonical order between the NP and PP complements. While, in previously studied OV languages, the canonical order is given as S-IO-DO-V with no nuances or controversies, positing a canonical order in Persian ditransitive constructions is neither straightforward nor uncontroversial. In this paper, we rely on the following generalizations supported by quantitative studies (e.g. Faghiri et al. (2018)): 1) Bare DOs have a strong bias towards PP-NP-V, 2) Rā-marked DOs have a strong bias towards the NP-PP-V order, and 3) Indefinite DOs are more versatile with a fair inclination towards the PP-NP-V order.²¹ Faghiri et al. (2018) argue that these ordering preferences can be captured by a cline on the basis of the degree of determination/definiteness of the DO that maps into a given-first as well as a salient-first preference (Faghiri et al. 2018: 182–183).

Interestingly, as we will show in the next section, different distance-minimizing accounts discussed in Section 2.2 do not predict the same length-based ordering preferences for Persian, contrary to previously investigated OV languages such as Japanese, Korean or Basque. Hence, the Persian data provide us with cases that make it possible to tease apart between, on one hand, Hawkins’s MiD and, on the other, DLT and DLM. We limit our discussion to the predictions of DLM as a less restrictive version of Gibson’s DLT and will not discuss DLT separately.²²

As noted by Temperley (2007), one way to compare MiD and DLT/DLM is to examine mixed-branching constructions, containing both head-final and head-initial constituents, for which the predictions the two models differ (p. 323).²³ Temperley uses a small set of data (87 occurrences) from Turkish provided by Hawkins (1994) to test the predictions of these models but his results are not conclusive (Temperley 2007: 323–324).²⁴ In the next section, we present the predictions of these models for Persian data following Temperley (2007: 303–304).

3.2 Predictions of distance-minimizing models

In this section, we present the predictions of MiD and DLM for Persian data discussed in this paper. In Section 2.2, we saw that these models both account for length-based ordering preferences in OV and VO languages from a processing point of view by positing that some orders are less complex to process. However, they build on different measures of complexity. While both measures depend on the relative length between the constituents involved, they differ in the way the dependency distance is operationalized.

Recall that MiD depends on the size of the Constituent Recognition Domain (CRD) and takes into account the number of words needed to be parsed in order to obtain all immediate constituents (IC) of the sentence. Between two alternative word orders, MiD votes for the sentence that yields a greater IC-to-Word ratio, that is, for an equal number of ICs, a sentence with a smaller CRD is preferred. This is illustrated by the pair of examples in (2) from Hawkins (2014: 104–105). The CRD and IC-to-Word ratio are calculated for the VP. The latter contains three ICs: Verb, PP1 and PP2. Note that the preposition (or postposition in head-final PPs) is straightforwardly considered to be the constructing category of a PP constituent.²⁵ In other words, a (left-to-right) linear word-by-word parser recognizes the PP constituent when it parses the preposition.²⁶

In (2a), Verb, PP1 and PP2 can be recognized on the basis of five words (italicized in the example). The CRD contains 5 words and hence the sentence has an IC-to-Word ratio of 3/5 (60%). (2b), on the other hand, has an CRD of 9 words and an IC-to-Word ratio of 3/9 (33%). Consequently, (2a), which reflects a short-before-long ordering, is considered to be less complex and easier to process than (2b), and thus should be preferred.

(2)	a.
	b.

Hawkins has also proposed a more fine-grained metric that allows the ranking of two sequences with the same CRD score.²⁷ An aggregate IC-to-Word ratio is calculated for the CRD instead of a simple IC-to-Word ratio: IC-to-Word ratios are calculated at each word in the CRD and the mean of all the IC-to-Word ratios is taken into account (see Hawkins (1990: 233–234)). In (3), IC-to-Word ratios are calculated at each word for the pair of examples in (2) above. Aggregate IC-to-Word ratios are 75.3% and 58.2%, respectively for (3a) and (3b).

(3)	a.
	b.

DLM, on the other hand, represents complexity of a sentence as the total sum of all the lengths of its dependencies. The length of a dependency is defined as the number of words spanned; a dependency connecting adjacent words is considered to have a length of 1 (Temperley 2007: 303). Between two alternative orders, the one with the shortest dependency-length is preferred. For two minimal-pair sentences that only differ in the order of their constituents such as those in (2), DLM only needs to compare the sum of the lengths of the two dependencies – between the verb and the head of each PP – that differ between the two options. This is schematically illustrated by the pair in (4), adopted from Temperley (2017: 317), where A and B can represent PP1 and PP2 above. In (4a) the dependency lengths are of 1 and 4 words and in (4b) of 1 and 8 words. Consequently, (4a), which reflects a short-before-long ordering, has a shorter total dependency length (5 vs. 9).

(4)	a.
	b.

In a consistently head-final language such as Japanese, as illustrated by schematic examples in (5) from Temperley (2007: 323), both measures predict a preference for the long-before-short ordering. (5) presents two alternative orders for a verb final sentence with two preverbal dependents (of 2 and 5 words long each). (5a) has a smaller CRD compared to (5b) as well as a shorter total dependency length. Hence, according to both measures, (5a) is less complex and should be preferred.

(5)	a.
	b.

Finally, it is worth highlighting that besides the direction of length-based preferences, both models predict that the strength of preference depends on the size of reduction in complexity (or the gain in efficiency), which amounts to the difference between the complexity/efficiency measures of the two alternative orderings, that is, the difference between total dependency lengths in DLM, and the difference between CRDs or IC-to-Word ratios in MiD. This difference is in turn directly related to the relative length of the constituents involved. In other words, the rate of length-based shifts is expected to increase with the relative length of the constituents.

In what follows, we present the predictions of these two models for a number of cases of constituent ordering variation in Persian where the constituents differ in length. Distinct predictions are provided for rā-marked and non-rā-marked DOs, given that due to their formal difference the latter cannot be treated in the same way. Note that while a non-rā-marked DO can only be viewed as a head-initial NP, a rā-marked DO can either be viewed as head-initial or head-final depending on whether rā is considered as the head (constructing category) of the NP or not.²⁸ Here, we consider both possibilities. For non-rā-marked DOs, we discuss the relative order between the DO and the PP argument. For rā-marked DOs, in addition, we discuss the relative order between the DO and the subject.

For illustration, we take a length difference of 4 words between the two constituents and consider the two possibilities where one constituent has a phrasal length of 6 and the other 2 and vice versa. We use Hawkins’s more fine-grained metric when relevant. Note that the latter is not taken into account by Temperley (2007).

We calculate the measures in minimal-pair sentences using schematic examples similar to (4) and (5) for simplification. Each sentence contains 3 ICs including the V(erb). We use bold to mark the head (constructing category) of the other constituents. In PPs, this would be the P (reposition), and in NPs, the first/leftmost element of the NP, which can be a determiner or a noun (represented by X that stands for any word). For rā-marked NPs, we are also considering the analysis in which rā is the constructing category. In short rā-marked NPs, the latter would appear at the right-edge. But in the case of long rā-marked NPs, which here we assume to contain relative clauses of 4 words length, rā appears just before the relative clause.

In (6) and (7), we consider the relative order between an NP and a PP complement in the preverbal domain. In (6), we consider the case where the PP is longer than the NP and in (7) the reverse.

(6)	a.
	b.

(7)	a.
	b.

Predictions:

In both (6) and (7), both pairs have the same CRD but (b) has a greater aggregate IC-to-Word ratio. The gain in efficiency is the same (10.6%) in each pair; (a) has a shorter total dependency length. The reduction is the same (4 words) in each pair.

• MiD predicts a preference for (b) that reflects a short-before-long ordering.

• DLM predicts a preference for (a) that reflects a long-before-short ordering.

• The strength of preference depends on the relative length between the two constituents – regardless of the direction of the difference (NP>PP or NP<PP).

In (8), we consider the relative order between the subject and the DO in the preverbal domain. Note that here we provide one pair of examples to illustrate both the case where the subject is longer than the DO and the case where the subject is shorter. In each example, NP1 and NP2 can represent respectively the subject and the DO (Subj>DO) and vice versa (Subj<DO).

(8)	a.
	b.

Predictions:

Both sentences have the same CRD but (b) has a greater aggregate IC-to-Word ratio. The gain in efficiency is the same in each case; (a) has a shorter total dependency length. The reduction is the same (4 words) for each case.

• MiD predicts a preference for (b) that reflects a short-before-long ordering.

• DLM predicts a preference for (a) that reflects a long-before-short ordering.

• The strength of preference depends on the relative length between the two constituents – regardless of the direction of length difference (Subj>DO or DO>Subj).

In (9) and (10), we consider the relative order between an NP and a PP complement in the preverbal domain. In (9), we consider the case where the PP is longer than the NP and in (10) the reverse. Recall that in the latter case the NP contains a relative clause of 4 words that appears after rā.

(9)	a.
	b.

(10)	a.
	b.

Predictions:

Sentence (9b) has a smaller CRD and hence a greater IC-to-Word ratio; Sentence (10a) has a smaller CRD and hence a greater IC-to-Word ratio. The gain in efficiency is the same in each case; Sentences (9a) and (10a) have shorter total dependency lengths. The reduction is the same (4 words) in each pair.

• MiD predicts a preference for NP-PP-V order regardless of the relative length.

• DLM predicts a preference for (a) that reflects a long-before-short ordering.

• The strength of the preference depends only on the relative length between the two constituents – regardless of the direction of length difference.

In (11) and (12), we consider the relative order between the subject and the DO in the preverbal domain. (11) illustrates the case where Subj>DO and (12) the case where Subj<DO.

(11)	a.
	b.

(12)	a.
	b.

Predictions:

Sentence (11b) has a smaller CRD and hence a greater IC-to-Word ratio; Sentence (12b) has a smaller CRD and hence a greater IC-to-Word ratio. The gain in efficiency is the same in each pair; Sentences (11a) and (12b) have shorter total dependency lengths. The reduction is the same (4 words) in each pair.

• MiD predicts a preference for the DO-first order regardless of the relative length.

• DLM predicts a preference for a long-before-short ordering.

• The strength of the preference depends on the relative length between the two constituents – regardless of the direction of length difference.

To sum up: in different pair-wise comparisons examined here, DLM consistently predicts a long-before-short preference that increases with the relative length between the two constituents, regardless of the choice of analysis for rā-marked NPs. MiD, on the other hand, predicts a short-before-long preference that increases with the relative length, provided all NPs are treated as head-initial constituents. Otherwise, MiD predicts a preference for putting the rā-marked NP first (in both cases of word order variation studied). Note that while the NP = rā-PP-V order corresponds to the canonical order in Persian (see Section 3.1), the DO = rā-S-V order is non-canonical, given that Persian is an SOV language.

3.3 Previous studies on Persian

A number of quantitative studies have already addressed the issue of grammatical weight in Persian. Rasekh-Mahand et al. (2016) investigate the role of weight in the extraposition of relative clauses in the postverbal domain²⁹ in a corpus-based study and argue that this extraposition supports Hawkins’s MiD and provides evidence that “Persian, a seemingly SOV language behaves typologically as a VO language, in which the heavy constituents shift rightward” (Rasekh-Mahand et al. 2016: 21).

In line with studies on the heavy NP shift in different SOV and SVO languages, Faghiri and colleagues (Faghiri & Samvelian 2014; Faghiri et al. 2014; 2018) study the effect of weight on the order between the NP and the PP complement (referred to as direct and indirect objects in a broad manner) and report a long-before-short preference in the preverbal domain. However, they also highlight that this effect is limited to specific cases. Importantly, these studies report zero length-based effects in constructions for which other SOV languages such as Basque (Ros et al. 2015) and Japanese (Yamashita & Chang 2001) are shown to have length-based word order variations. Below we present a summary of these studies and their main results.

Faghiri & Samvelian (2014) present a multifactorial study of word order preferences carried out on a sample of 905 NP-PP-V and PP-NP-V utterances extracted out of a journalistic corpus. They find a significant effect of heaviness, operationalized as the relative length between the NP and the PP in number of words,³⁰ corresponding to a long-before-short preference for bare and indefinite objects (Faghiri & Samvelian 2014: 226–227).³¹ Importantly, in their sample, rā-marked objects occur in the NP-PP-V order in more than 95% regardless of the relative length (Faghiri & Samvelian 2014: 226).

In the vein of constrained production experimental studies on the heavy NP shift, Faghiri et al. (2014; 2018) carry out two experiments to test length-based ordering preferences observed in Faghiri & Samvelian’s (2014) corpus study for indefinite and bare DOs respectively. They use a web-based version of the sentence production task used by Stallings et al. (1998) and Yamashita & Chang (2001) – that we are also using in our experimental studies (see Section 4.1). They study the relative order between the DO and the obligatory PP argument while manipulating the relative length between these two constituents by adding a relative clause to the PP and attributive (restrictive) modifiers to the DO.³² Both experiments find a significant main effect of the relative length. In the case of indefinite DOs (Est. 1.75, SE = 0.299, p < 0.001), the rate of the NP-PP-V order increases from 55.7% to 80.3% with longer DOs (Faghiri et al. 2014: 232). For bare DOs (Est. 1.479, SE = 0.308, p < 0.001), on the other hand, the rate of the PP-NP-V order increases from 85.5% to 94.2% with longer PPs (Faghiri et al. 2018: 175).

Furthermore, Faghiri et al. (2018) present an additional experiment on the relative order between bare DOs and the PP argument. They focus on the behavior of simple bare DOs, ex. ‘syrup’, compared to modified bare DOs, ex. ‘icy mint syrup’, that are longer by two words, and report an important rise of shifted orders for the latter. The rate of the NP-PP-V order increases from 28.2% with bare single word DOs to 47.3% with bare modified DOs (Faghiri et al. 2018: 178).

Faghiri et al. (2018) also address the effect of weight on the relative order between the subject and the DO. They first note that the rate of the SOV order is overwhelmingly high, above 95%, in their corpus sample and that non-canonical (shifted) orders only occur with rā-marked DOs (see p. 171 for details). They carry out an experiment that allows them to test whether the rate of OSV increases with longer DOs, modified by a relative clause. Faghiri et al. (2018) use rā-marked human DOs to give the OSV order the highest chance (see p. 179). Their results show no weight-based effect rise in OSV orders (7.81% vs. 6.98%, respectively for long/complex vs. short/simple DOs).³³ It is important to note that their study has a high statistical power to detect a true small effect size (0.83 calculated with the power function in R for binary data with 877 observations and df = 1). Note also that the absence of a length-based effect is unlikely to be due to a ceiling effect, given that they do find a significant effect of animacy (Est. 0.4, SE = 0.172, p < 0.05): the rate of OSV rises from 4.2% with animate subjects to 10.64% with inanimate subjects (Faghiri et al. 2018: 180).

To resume, previous studies report length-based ordering variations corresponding to a long-before-short preference in the relative order between the DO and the PP argument, limited to non-rā-marked DOs. They show that 1) long PPs shift (leftward) more often than short PPs when the DO is bare or indefinite, 2) long bare DOs shift more often than short bare DOs. On the contrary, the relative length is reported to have zero effect on the linear position of rā-marked DOs (with respect to the PP argument) as well as on the relative order between the subject and the DO in transitive sentences.

In their attempt to provide an explanation for these observed length-based effects, Faghiri and colleagues favor a production-oriented account over parsing-oriented accounts. They argue that Yamashita & Chang’s (2001) conceptual accessibility hypothesis provides an adequate account of the Persian data, in particular because it can provide a compelling explanation for the cases where no length-based effects are observed (Faghiri et al. 2018: 183), while these cases are problematic for dependency-length-minimizing models (Faghiri 2016: 268–269). They also rightly note that Hawkins’s model falls short of accounting for the long-before-short preference in the preverbal domain in Persian (see e.g. Faghiri & Samvelian 2014: 228).

3.4 The focus of our study

Length-based ordering preferences depicted by available studies so far present some potential challenges for dependency-distance minimizing accounts.

1. The cases for which previous studies report zero length-based effects run counter to the predictions of DLM. According to the latter the long-before-short preference is expected to trigger ordering variations for all types of DOs.

2. Although MiD may account for the absence of a length-based effect in the case of rā-marked DOs (provided the latter are viewed as headed by rā), it makes wrong predictions with respect to the direction of length-based effects in the preverbal domain.

3. The relatively important rise in the rate of shifted orders observed for a two-word length difference, when the relative length is increased by adding two attributive modifiers to bare DOs, is intriguing.³⁴ Importantly, by comparing the size of length-based effects in the three experiments on the relative order between the NP and the PP argument, we observe that it is not possible to establish a correlation between the strength of the preference and the relative length.³⁵ This is potentially problematic for any dependency-distance minimizing account of length-based effects. Indeed, as we have seen in Section (3.2), the strength of a length-based preference is predicted to increase with the relative length between the two constituents regardless of the direction of the difference. On the other hand, the strong effect observed when adding (restrictive) modifications to bare nouns may be viewed as triggered by the added information and the increase in their degree of specification rather than by length per se. Therefore, the question rises as to whether all length-based effects can be viewed as triggered by dependency-distance minimization. An alternative account would be to view the effect of adding attributive modifiers as an effect of conceptual enrichment, in line with the salient-first preference in the preverbal domain.

We have conducted two more production experiments to follow up on these observations. These experiments are carried out in order to 1) pin down the effect of phrasal length in terms of conceptual enrichment vs. dependency-distance by comparing simple/short vs. modified/long non-rā-DOs and 2) replicate the absence of length-based effect in the case of rā-marked DOs in order to assess the relevance of dependency-distance minimizing accounts.

4.1 Procedure

Our experiments are implemented via web-based self-administrated questionnaires³⁷ conducted on the Ibex Farm (Drummond 2013). In these experiments, participants are asked to construct sentences with phrases that appear on the screen to complete a preamble.

This production paradigm is inspired by the in lab cued sentence recall production. This is a common experimental task in psycholinguistics for the study of sentence production and is used in experimental studies on word order preferences, especially on heavy constituent shift (e.g. Stallings et al. (1998); Yamashita & Chang (2001)). In these experiments, participants are asked to make a sentence with given constituents that appear on a computer screen and to produce it orally after a lapse of time during which, for distraction, they are presented with a basic arithmetic operation. This design is meant to encourage participants to produce their sentences from meaning (Yamashita & Chang 2001: 48).

In the design used in our experiments, participants are asked to make a sentence to complete a preamble with phrases that appear on the screen. Participants see simultaneously a preamble, an incomplete sentence represented by blank boxes and a vertical list of phrases (that appear in blue). A screenshot of a (filler) item is given in Figure 1.

Figure 1

Screenshot of an experimental item (English equivalents are added).

Each experimental item contains a sentence in which three constituents are missing, represented by three blank boxes. To complete the sentence, participants are provided with four phrases (presented in a counterbalanced order). Participants are instructed to 1) read the preamble and the list of phrases, 2) complete the sentence with the most natural continuation that comes to their mind using three of the four given phrases, 3) fill in the blanks accordingly, and 4) click on “Continue” to go to the next sentence.

The list of options contains one element more than the number of blanks in order to prevent participants from guessing the purpose of the experiment and to push them to concentrate on the content of each sentence, in order to produce reasonably natural sentences. The relative order between these constituents (in the final sentence) is left to the participants and constitutes the response or the dependent variable.

4.2 Experiment 1

This experiment targets the effect of phrasal length by adding an attributive (restrictive) modifier to non-rā-marked (bare and indefinite) DOs. Faghiri et al. (2018) showed a particularly important effect of the phrasal length, when comparing the relative order of short bare DOs (one-word length ex. gol ‘flower’) with long bare DOs (three-word length, ex. gol-e orkide-ye sefid ‘white orchid flower’) with respect to inanimate PP arguments. The rate of the non-canonical NP-PP-V order increased from 28.2% to 47.3% for long bare DOs. The research question here is whether indefinite DOs show a comparable effect when lengthened by attributive (restrictive) modifiers.

Furthermore, in order to disentangle the effect of conceptual enrichment from the effect of increasing the dependency-distance, we manipulate the length by adding only a one-word attributive modifier. Recall that the strength of a dependency-length preference depends on the size of reduction in the length difference between two alternatives. A one-word length difference presents the least possible reduction in the dependency length. Note that the lowest rate of length-based shifts observed so far (in Persian data) is about 10% and was triggered by a three-word length difference (see Section 3.4). Consequently, a comparably strong effect in this configuration can hardly be viewed as a dependency-minimizing effect.

Finally, in Faghiri et al.’s study, PPs were inanimate in all sentences and the results showed an overall higher rate of NP-PP-V order with respect to another experiment with bare DOs but including only animate PPs (see Faghiri et al. 2018: 177). Here, we include the animacy of the PP as a between-items (control) variable in order to neutralize its effect and also to see to what extent animate and inanimate PPs behave differently.

4.2.2 Results and discussion

Table 1 presents the frequencies of different ordering choices in the data, and Figures 2 and 3, the rate of NP-PP-V order (NPV in short), respectively by DO length and by animacy of the PP.

Table 1

Distribution of word order for Experiment 1.

Bare Dos
	NPV	NVP	PNV	NA	Total
DO Length
Short	26	0	301	1	328
Long	55	1	268	4	328
PP Animcy
Anim	25	0	300	3	328
Inan	56	1	269	2	328
Total	81	1	569	5	656
Indef Dos
	NPV	NVP	PNV	NA	Total
DO Length
Short	154	1	148	1	304
Long	198	1	104	1	304
PP Animcy
Anim	164	0	139	1	304
Inan	188	2	113	1	304
Total	352	2	252	2	608

Figure 2

NP-PP-V rate by DO type and DO length in Experiment 1 (the error bars present 95% confidence intervals).

Figure 3

NP-PP-V rate by DO type and animacy of PP in Experiment 1.

The overall rates for NP-PP-V are 12.5% for bare DOs, and 58.3% for indefinite DOs. These rates are in line with the rates reported for these DO types in previous experimental studies. We observe that the rate of NP-PP-V significantly increases with longer/modified NPs for both bare and indefinite types, respectively, from 8.0% to 17.0% (χ² = 11.454, p < 0.001) and, from 51.0% to 65.6% (χ² = 12.59, p < 0.001). Likewise, the rate of NP-PP-V is significantly higher for inanimate PPs with both bare and indefinite types, 7.7% vs. 17.2% (χ² = 12.693, p < 0.001) and 54.1 vs. 62.5 (χ² = 3.976, p < 0.05).

We analyzed all the data (a total of 1254 data points excluding miscellaneous orders) using an GLMM model including items and participants as random effects and DO type, DO length and PP animacy (sum-coded) as fixed effects. The NP-PP-V order is coded as success. We find a significant effect for the DO type (Est. = –1.94, SE = 0.22, p < 0.001) as well as for length (Est. = 0.55, SE = 0.11, p < 0.001), but no significant effect for animacy nor any significant interactions between the variables (ps > 0.10).

We then fitted two models separately for bare and indefinite DOs (a total of, respectively, 650 and 604 data points), to see whether there is a numerical difference between the estimated coefficients and in what direction. The estimated coefficient of length is greater for bare DOs (Est. = 0.64, SE = 0.17, p < 0.001) than for indefinite DOs (Est. = 0.38, SE = 0.11, p < 0.01).

In sum, our data show a robust effect of phrasal length for both non-rā-marked DO types. We observe that adding modifications to the DO increases NP-PP orders for both bare and indefinite types, and while there is no significant interaction between the effects of length and DO type, we can say that it is likely to have a larger effect size for bare DOs than for indefinite ones. Indeed, if the effect of length is due to the increase in the degree of specification (of a non-specific object), it is safe to assume that with less specific objects (bare vs. indefinite) the contribution of length is likely to be more important.

With respect to the animacy of the PP, in line with the general animate-before-inanimate preference, we observe that overall in our data the rate of PP-NP-V orders increases with animate PPs. We are not going to comment further on the effect of animacy which is beyond the scope of this paper.

4.3 Experiment 2

In this experiment, we study the effect of the relative length (by adding a relative clause to the PP argument) for rā-marked vs. non-rā-marked DOs. In their experimental study, Faghiri et al. (2014) found an important effect of the relative length for indefinite (non-rā-marked) DOs, corresponding to a long-before-short preference (see Section 3.3). This preference is in contradiction with the prediction of MiD while it is in line with that of DLM (see Section 3.2). The latter predicts a long-before-short preference for rā-marked DOs as well. However, corpus studies found no length-based effects for these DOs suggesting that they may not be sensitive to length-based effects (Faghiri & Samvelian 2014: 230). Recall that MiD also predicts no length-based effects for rā-marked DOs in this configuration. The research question in this experiment is thus to test whether increasing the length of the PP argument favors more shifted orders for rā-marked NPs. Such result would provide further support for DLM while undermining Hawkins’s MiD.

4.3.2 Results and discussion

The overall rates of NP-PP-V order are 80.4% for rā-marked NPs, and 42.4% for indefinite NPs. The latter is lower than the rate reported previously for these NPs, but when we look at the distribution by the preposition involved (Figure 4), we observe a fairly important variation, suggesting that the default order is not the same depending on the preposition type. Importantly, for the preposition be ‘to’, the baseline rate tips for the NP-PP-V order in line with previous studies. Table 2 and Figure 5 present, respectively, the frequencies of different ordering choices in the data and the rate of NP-PP-V order by DO type and PP length. We observe that for both types the rate of NP-PP-V order is significantly higher with short PPs: 86.8% vs. 74.1% (χ² = 6.152, p < 0.05) and, 52.6% vs. 31.9% (χ² = 11.068, p < 0.001), respectively for rā-marked and indefinite NPs.

Figure 4

NP-PP-V rate by DO type and Preposition in Experiment 2.

Table 2

Distribution of word order for Experiment 2.

	NPV	NVP	PNV	NA	Total
Indefinite DOs
Short PP	71	0	64	1	136
Long PP	43	1	92	0	136
Total	114	1	156	1	272
Rā-marked DOs
Short PP	118	0	18	0	136
Long PP	100	0	35	1	136
Total	218	0	53	1	272

Figure 5

NP-PP-V rate by DO type and PP length in Experiment 2.

We analyzed all the data (a total of 541 data points excluding miscellaneous orders) using a GLMM model including preposition types, items and participants as random effects and DO type and PP length (both sum-coded) as fixed effects. The NP-PP-V order is coded as success. We find a significant effect for the DO type (Est. = 1.10, SE = 0.12, p < 0.001) as well as for PP length (Est. = –0.52, SE = 0.11, p < 0.01), but no significant interaction between the two variables (p = 0.85).

We also fitted the model separately for each DO type in order to check whether there is a numerical difference between the estimated coefficients of length between them. In addition, in order to have more homogeneous data for comparison, we also used a limited dataset including only items with the preposition be. In both cases, the estimated coefficient of PP length was slightly greater for indefinite DOs than for rā-marked DOs, respectively: –0.60 (SE= 0.16, p < 0.001) vs. –0.50, (SE = 0.18, p < 0.01), and –0.76 (SE = 0.23, p < 0.01) vs. –0.54 (SE = 0.276, p < 0.05).

In sum, our data show a robust length-based effect corresponding to the same long-before-short preference for both DO types. This effect is in accordance with the predictions of DLM, and importantly, it contradicts Faghiri & Samvelian’s (2014) findings regarding the absence of a length effect on the position of rā-marked DOs. Nevertheless, while we find no significant interaction between the effects of relative length and DO type, we observe that increasing the length of the PP is likely to have a larger effect in the case of indefinite DOs than rā-marked DOs. This is not surprising given that rā-marked DOs have a stronger bias toward the NP-PP-V order and are said to display less ordering variation than indefinite DOs (see e.g. Faghiri et al. 2018).