Russian Children and Their Relatives: What Can a Free Word Order Language Reveal About the Subject/Object Asymmetry?

ABSTRACT

The present study revisits the well-attested phenomenon of Subject/Object Asymmetry in relative clause comprehension by monolingual Russian pre-school children. Additionally, it examines the NP-V and the V-NP word order types when the embedded NP precedes or follows the verb. Both the word orders are grammatical for Russian subject and object relative clauses alike. Object relative clause comprehension difficulty was evidenced across all the age groups. Word order alternation (V-NP vs NP-V) did not significantly affect the accuracy rates. Besides the efficacy of the agent-first strategy, growing morphological knowledge compels children to update their ongoing structural analysis, which results in embedded NP errors. Yet, even at 6 years of age, children do not display word order preferences to go with a particular relative clause type that are found in Russian-speaking adults. Restricted exposure to written texts at pre-school age as well as low validity of word order as a cue to thematic role assignment are discussed as possible reasons for children’s insensitivity to word order manipulation in Russian relative clauses.

Introduction

One common property of a language that preschool children are believed to exploit for structure building is the agent-first information sequencing. Existing studies demonstrate that children up to a particular age display great reliance on word order as a cue to thematic role assignment (Hirsch & Wexler, 2006; Slobin & Bever, 1982). Canonical structures are easier to comprehend than non-canonical ones even in languages where word order is relatively flexible and cannot be exploited as a reliable cue to structure building (Cristante, 2016; Dittmar et al., 2008; Edeleva, 2023; Gamper, 2016; Grünloh et al., 2011; Roesch & Chondrogianni, 2015; Sauermann & Höhle, 2016; Schipke et al., 2012; Schaner-Wolles, 1989 for younger children; Thelen, 2019). In other words, in languages with the neutral subject-verb-object (SVO) word order, the noun that is encountered first will be treated by children as the agent to the verb by default. The agent first strategy may work well with structures where the first encountered head NP (the goose) is indeed the agentive subject to the verb (is treating), such as subject relative clauses (SRC, 1).

(1) … the goose that is feeding the fox …

However, when the agent first strategy is applied to structures in which the order of the constituents departs from the canonical SVO pattern, the developing parser is sent down the wrong path. A typical example of a structure type where the head NP does not necessarily map onto the role of the agentive subject is the object relative clause (ORC, 2). As a result, object relative clauses are commonly misinterpreted by children in terms of thematic role assignment. Though the head NP (the goose) performs the role of the object in the relative clause structure, it is nevertheless treated by children as the relative clause subject and the agent to the verb “is feeding” while the embedded NP (the fox) receives the patient interpretation.

(2) … the goose that the fox is feeding …

Considerable comprehension disadvantage to ORCs as compared to SRCs in thematically reversible sentences is also known as “Subject/Object Asymmetry.” Significantly, it shows up both in languages where word order signals a particular structure type (e.g., English, Brown, 1971; Diessel & Tomasello, 2005; Sheldon, 1974; de Villers et al., 1979) and in languages where cues other than word order are more reliable for correct thematic role assignment. Those languages, including Russian, which will be the focus of the current investigation, may allow for word order variation. For example, in Romance languages, the verb follows the embedded NP in subject relative clauses (3) but can precede (4b) or follow (4a) the embedded NP in object relative clauses as shown in examples below (from Adani, 2011).

Table

(3) [SRC] … the horse that is chasing the lions …
il cavallo	che	sta inseguendo	i leoni
the horse.SG	that	is chasing.SG	the lions.PL
(4) a. [ORC; NP-V] … the horse that the lions are chasing …
il cavallo	che	i leoni	stanno inseguendo
the horse.SG	that	the lions.PL	are chasing.PL
b. [ORC; V-NP] … the horse that the lions are chasing …
il cavallo	che	stanno inseguendo	i leoni
the horse.SG	that	are chasing.PL	the lions.PL

Yet, though these languages display no steady syntax-word order correlations, recent studies observed that the two ORC word orders are not comprehended equally well either by children or by adults and the V-NP word order causes greater difficulty (Adani, 2011; Gavarrò et al., 2012; Guasti et al., 2008, 2012, 2018; Pozniak et al., 2016a, 2016b). Those experimental studies typically test relative clauses where both the head and the embedded NP are full NPs and are reversible. That is, the head and the embedded NP can be swapped around still yielding a plausible sentence interpretation. Object relative clauses with two reversible head NPs are relatively infrequent in natural language use giving precedence to those with inanimate heads or pronominal embedded NPs (Adani et al., 2017; Brandt et al., 2008 in child-directed speech; MacDonald & Christiansen, 2002; Reali & Christiansen, 2007; Roland et al., 2007). The V-NP word order is also the only constituent order that is possible for subject relative clauses. At the same time, subject relative clauses are more frequent in natural language use than their object-extracted counterparts. Consequently, the comprehender’s expectation of the V-NP word order to go with the all the more frequent subject relative clause remains relatively strong. A number of explanations have been proposed to account for the V-NP word order difficulty in ORCs. A detailed review of them will be provided in the subsequent section. Yet, children might simply be very likely to associate the V-NP word order with the subject relative clause type and not be fine-tuned to other features (e.g., verb morphology in Italian or the acoustically distinct relative pronoun que.that.ACC in French) that signal an ORC resulting in erroneous judgments.

The current study addresses this confound and presents a study of Russian relative clauses that allow for word order variation with either RC type (subject or object). Therefore, the V-NP word order is not statistically biased toward the all the more frequent subject relative clause type as is the case in Romance languages. Since there is no strong association between the SVO word order and the subject relative clause syntax, the relative strength of word order canonicity as a cue to thematic role assignment in Russian is reduced (E. Bates, 1999). Before we turn to specific accounts that have been proposed to explain the difficulty of the V-NP word order in Romance languages, we will briefly describe some of the properties of Russian relative clauses that will be relevant for the current study.

Relative clauses in Russian

Russian relative clauses represent instances of head-external structures (i.e., the head NP is located outside the relative clause) and represent instances of long-distance dependencies. The relative pronoun (kotor-) is an obligatory element of the relative clause and agrees with the head NP in gender and number. Depending on its syntactic function in the relative clause structure, it inflects for case. A full paradigm of inflectional forms for the relative pronoun kotor- is presented in Table 1.

Table 1. Declension paradigm of the relative pronoun kotor- in three genders.

		SG
		MASC	NEUT	FEM	PL
NOM		kotor-yi	kotor-oye	kotor-aya	kotor-ye
GEN		kotor-ogo	kotor-ogo	kotor-oi	kotor-yh
ACC	Animate	= GEN	= NOM	kotor-uyu	= GEN
	Inanimate	= NOM			=NOM
DAT		kotor-omu		kotor-oi	kotor-ym
INSTR		kotor-ym		kotor-oi	kotor-ymi
PREP		kotor-om		kotor-oi	kotor-yh

In subject relative clauses with a transitive verb (5), the relative pronoun receives the nominative inflection, and the embedded NP is marked for the accusative. In object relative clauses (6), the order of the case-marked elements is reversed, such that ACC appears first on the relative pronoun and the embedded NP is NOM.

(5)  a. [SRC NP-V] … gus,’ kotor-yi  lis-u   ugoščajet …

        goose that.NOM fox.ACC is feeding

        … the goose that is feeding the fox …

  b. [SRC V-NP] … gus,’ kotor-yi ugoščajet lis-u …

        goose that.NOM is feeding fox.ACC

        … the goose that is feeding the fox …

(6)  a. [ORC NP-V] … gus,’ kotor-ogo  lis-a ugoščajet …

        goose that.ACC fox.NOM is feeding

        … the goose that the fox is feeding …

  b. [ORC V-NP] … gus,’ kotor-ogo ugoščajet lis-a …

        goose that.ACC is feeding fox.NOM

        … the goose that the fox is feeding …

In most cases, morphological marking is distinct and, therefore, highly reliable for thematic role assignment. The exceptions are instances of case syncretism when the nominative and the accusative forms coincide (e.g., for neuter nouns). Word order in Russian relative clauses can alternate between the V-NP and the NP-V configuration both for subject and for object relative clauses. Generally, Russian uses transitive sentences with two NPs that are ordered linearly. Yet, since word order is quite flexible, cases where the initial NP takes on the syntactic function other than the pre-verbal subject are not uncommon. Thus, word order canonicity is available as a cue, but it is not very reliable.

Under the “Competition Model” (E. Bates & MacWhinney, 1989; MacWhinney, 2005), the information value of different cues such as word order or case marking that can be used for structure building is captured in the concept of validity (MacWhinney et al., 1984). The overall validity reflects instances where the cue is available and where it is reliable, i.e., it correctly signifies the thematic role. At the same time, a given cue may be more or less reliable compared to some other cue. In non-canonical structures, word order and case marking go against each other in terms of the thematic role that they signal. In this case, a more reliable cue receives higher conflict validity. An earlier study into German (Dittmar et al., 2008) as a language with relative word order flexibility showed that older children and adults rely more on case marking as a cue with a higher conflict validity. By contrast, younger children are more likely to follow the word order cue with a high overall validity to interpret non-canonical sentences. In Russian, the ordering of the constituents is even more flexible than in German. Therefore, the overall validity of word order is not so high even for canonical structures. It is also overridden by case marking in terms of conflict validity. Thus, case marking generally appears to be more valid than word order.

Regarding word order alternation between the V-NP and the NP-V type, the right edge of the embedded clause has been claimed to mark the focus (Adamec, 1966; Kovtunova, 1976). In subject relative clauses, the V-O word order receives a wide focus reading and is, thus, communicatively neutral. In the O-V word order, the verb bears the contrastive reading, i.e., the goose is FEEDING the fox (rather than e.g., taking the cake away from it). In a similar vein, the S-V word order signifies the contrastive reading in object relative clauses with full embedded NPs, i.e., the goose that the fox is FEEDING (rather than taking the cake away from it). Yet, the focus can be shifted to the embedded NP, i.e., the goose that the FOX is feeding (rather than the cat). In this case, embedded NP is accentuated. The embedded NP can also receive contrastive reading in the V-S word order if it is accentuated. Existing corpus studies of Russian relative clauses in adult language evidence a frequency-based preference of the V-NP word order type to go with the subject and the NP-V word order – with the object relative clause (Levy et al., 2013; Price & Witzel, 2017). The preference is stronger for the subject than for the object relative clauses. Respective frequency counts are provided in Table 2. Additionally, Levy et al. (2013) report that the distribution pattern found in written texts is mirrored in adult native speakers’ online performance on subject- and object relative clauses if word order is manipulated. In a self-paced reading paradigm, the participants preferred the default V-NP and NP-V word orders for SRCs and ORCs, respectively, while contextually unsolicited word order configurations (Kaiser & Trueswell, 2004; Slioussar, 2011) resulted in longer reading times for context-free sentences (Levy et al., 2013). Since children are sensitive both to input frequency (see Diessel, 2007 for the role of sensitivity to statistical distributions in language acquisition; Adani et al., 2017 for the input frequency approach in the acquisition of transitive relative clauses) and to contextual constraints (Córrea, 1995; Hamburger & Crain, 1982; Kidd, 2003), it is reasonable to expect that respective adult native speakers’ tendencies will also be reflected in Russian children’s comprehension of relative clauses. As a result, children might display an asymmetrical word order preference for subject and object relative clause types and reject contextually constrained alternatives in context-free settings. On the other hand, children’s performance may be influenced by some structure-related factors that will be discussed in detail in the subsequent section.

Table 2. Frequency of relative clauses in Russian by 1) syntax; 2) word order for relative clauses with embedded NPs as full NPs (pronominal NPs excluded).

	SRC_VNP	SRC_NPV	ORC_VNP	ORC_NPV
(Levy et al., 2013)	147	4	41	29
(Price & Witzel, 2017)	581	33	58	91

The counts in Levy et al. (2013) are based on Russian Dependency Treebank collection of late 20th century texts (35,000 sentences; 1 million words) from a mixture of genres including fiction, news, and a small amount of scientific literature. The counts in Price and Witzel (2017) are based on the syntactic component of the Russian National Corpus (928 sentences). Number of words and text genre are not reported.

Word order variability in relative clauses from the developmental perspective and explanations of the V-NP ORC difficulty

Word order alternation for relative clauses has been exploited in selected languages (Greek, Italian, Catalan and French) to provide still more empirical evidence toward children’s sensitivity to grammatical properties of selected relative clause constituents. As has been illustrated in the previous section (examples (3) and (4a–b) for Italian), in Greek, Italian, Catalan and French the verb can precede or follow the embedded NP in ORCs, which results in two distinct ORC types that are both grammatical. In the absence of other cues, an Italian sentence “Indica il cavallo che sta inseguendo il leone” (“Point at the horse that.SUBJ is chasing the lion/ … that.OBJ the lion is chasing”; from Adani, 2011) is globally ambiguous between the subject and the object RC reading. It might prompt the comprehender to point at the horse that is chasing the lion, or the horse being chased by the lion. However, if the embedded NP is plural (i leoni “lions” instead of il leone “a lion”), the verb receives overt number agreement marking and thematic roles are disambiguated. In SRCs, the verb agrees with the singular head NP in number. In ORCs, the verb agrees with the plural embedded NP which follows it. Thus, word order flexibility for ORCs brings about three distinct structure types (examples (3) and (4a-b)) that have been contrasted in experimental studies.

Structural explanations of ORC difficulty assume a filler-trace dependence between the head NP and the point in the relative clause structure where it should be integrated (cf. Active Filler Hypothesis, Frazier & Flores d’Arcais, 1989; the Dependency Locality Theory; Gibson, 1998, 2000). The distance between the filler and the gap has been argued as one of the factors that may potentially explain the comprehension difficulty of ORCs compared to SRCs. The distance can be conceived of in terms of a) the number of linearly intervening constituents (Gibson, 1998); 2) the number and type of intervening discourse referents (Gibson, 2000; Warren & Gibson, 2002) and 3) the depth of embedding in the hierarchical structure (Hu et al., 2016). In SRCs, the head NP trace is posited immediately after the relative pronoun. In ORCs, it is placed after the verb. Therefore, the gap is located lower down the structural hierarchy (which is also commonly referred to as depth of embedding) in ORCs than in SRCs. The representation for ORCs will be structurally more complex irrespective of word order. Another factor that has been claimed to contribute to ORC difficulty within structural complexity accounts is the intervening embedded NP which bears grammatical feature similarity to the head NP on the way of the filler to the trace (cf. Relativized Minimality Hypothesis, Friedmann et al., 2009; Rizzi, 1990). Indeed, a mismatch in grammatical features between the head and the embedded NP has been shown to improve children’s ORC comprehension. Adani et al. (2010) documented a facilitative effect of both gender and number mismatch in older children acquiring Italian, though the effect was greater when the NPs differed in number. Belletti et al. (2012) failed to confirm a facilitative effect of gender mismatch for Italian-speaking children but observed that gender mismatch between the head and the embedded NP significantly ameliorated Hebrew-speaking children’s performance on object relative clauses. Contradictory findings regarding Italian children’s object relative clause comprehension will be picked up in at a later point in the current paper.

In summary, the depth of hierarchical embedding implies greater structural complexity of object compared to subject relative clauses. The integration account (Gibson, 1998, 2000) discusses the role of cognitive resources, such as working memory, as a factor contributing to the processing difficulty of object relative clauses. In this case, the linear distance from the filler to the verb plays a greater role than the complexity of the hierarchical representation. The filler has to be maintained in memory up to the point where it can be integrated as the argument to the verb. If we consider word order alternation, the verb occurs earlier with fewer intervening constituents in the V-NP compared to the NP-V word order. Thus, the V-NP word order is less taxing in terms of working memory load. Secondly, the way of the filler to the trace is not occluded by any element with similar features (e.g., another NP) that might interfere when the parser attempts to recall the filler at the point of integration.

Thus, while the depth of hierarchical embedding does not predict a differential effect of word order alternation, the cognitive resources required to deal with surface intervention effect would render the V-NP constituent order in ORCs easier for comprehension than its respective NP-V counterpart due to ease of integration. On the surface, the filler-trace unit in the case of the NP-V word order (It. che i leoni stanno inseguendo__/that the lions are chasing) bears the intervention effect from the embedded NP. By contrast, in the case of the V-NP word order in ORCs the gap comes up just before the embedded NP (It. che stanno inseguendo__ i leoni/that the lions are chasing). Thus, the path to the trace is not occluded by any elements of a similar type (full NPs) which allows to avoid competition at the verb where the integration of NPs takes place.

Yet, against this prediction pre-school children were significantly better at interpreting NP-V structures compared to their V-NP alternatives. Table 3 gives an overview of studies where the NP-V advantage was observed in Italian-, Greek-, Catalan- and French-speaking children across the age span of three to seven years in a character selection paradigm (Guasti et al., 2008, 2012; Adani, 2010, 2011 for Italian; Guasti et al., 2008, 2012 for Greek; Gavarrò et al., 2012 for Catalan; Guasti et al., 2018 for French; Sevcenco & Avram, 2012 for Romanian).¹

Table 3. Percentage counts of children’s correct responses on the three relative clause types across the experimental studies that test word order alternations in ORCs.

Language	SRC	ORC_NPV	ORC_VNP
Italian (Guasti et al., 2008)	79,99	54,16	40,82
Italian (Guasti et al., 2012)	80,86	n.a.	40,74
Italian (Adani, 2009; 2010)	92,6	78,6	44
Greek (Guasti et al., 2012)	68,60	n.a.	48,86
Greek (Guasti et al., 2008)	65,31	52,25	45,97
Catalan (Gavarrò et al., 2012)	91	70	14,5
French(Guasti et al., 2018)	87	58	23
Romanian(Sevcenco & Avram, 2012)	91,51	78,4	65,9

Since the predictions regarding the linear distance and the intervention effect that largely stem from adult processing literature were not confirmed, alternative explanations such as the Diagnosis Model (Fodor & Inoue, 2000) were proposed. In accordance with the Minimal Chain Principle (De Vincenzi, 1991), children are assumed to assign the head NP the syntactic role of the subject within the relative clause by default. However, when they attempt to posit the subject trace, they encounter either the embedded NP (in case of the NP-V word order) or the verb which may violate morphological agreement with the head NP as the relative clause subject (in case of the V-NP word order). Thus, a reanalysis has to be initiated. However, children are notoriously less efficient in reanalysis due to their poorly developed inhibitory skills (Mazuka et al., 2009; Trueswell et al., 1999). Therefore, they persist with their first adopted SRC interpretation and fail to correctly interpret the ORC.

In the case of the NP-V word order, the head NP is followed by the embedded NP and the subject function can be immediately reassigned. In ORCs with the V-NP word order, the head NP is followed by the verb that is morphologically incongruent with the head NP as a putative subject and violates the subject-verb agreement. The head NP has to be neglected as the putative relative clause subject and the search for the actual subject has to be launched. While searching for a congruent subject to the verb, the parser establishes an empty [pro]-slot with respective morphological features. When it arrives at the embedded NP, it has to check with the previously established [pro]-slot whether the embedded NP is morphologically a good candidate as the subject to the RC verb. This additional step requires still more resources for the reanalysis to be accomplished.

On the other side of the coin, recent research into French relative clauses reveals that the two ORC word orders are not fully interchangeable. The inverted V-NP structure, though not necessarily more frequent than the NP-V alternative, is amenable to certain restrictions in terms of verb semantics and subject length (Pozniak et al., 2016a, 2016b). As has been mentioned earlier, children are sensitive to pragmatic felicity conditions for certain structure types (Hamburger & Crain, 1982) and would prefer the default NP-V sequencing in the absence of the lexical-semantic information that would support the alternative word order. French children might, thus, mishear or miss the acoustic difference between the subject relative pronouns qui and the object relative pronoun que and assign the SRC structure to the V-NP word order.

Though the structural explanation with an empty [pro]-slot to be established works well for the languages investigated so far, it has never been put to test independently of the statistical bias of the V-NP word order toward the SRC. In the current paper, we investigate a language where both SRCs and ORCs can alternate between the NP-V and the V-NP word order and yet display an asymmetrical word order preference in terms of canonicity and contextual constraints that should be met. We, therefore, attempt to tease apart language-specific word order canonicity effects from purely structural explanations and report children’s comprehension study of relative clauses in Russian.

In Russian, the V-NP word order is not statistically biased toward the SRC interpretation and word order has a very low overall validity as a cue to structure building due to its high flexibility. If Russian children find the V-NP word order more difficult as has been shown for French and Italian, the effect can be interpreted as originating from its representational complexity and the difficulty to find a putative subject to the verb. However, if they do not show any preference toward a particular word order type, this would cast doubt on the structural explanation provided for other languages in favor of more general statistical constraints that are attributable to a particular word order type. In addition, we also consider the role of linear distance as well as the effect of the intervening NP as possible factors that might also account for children’s relative clause comprehension pattern in Russian. Finally, we also inquire whether Russian pre-school children display the comprehension advantage to subject relative clauses just as their peers acquiring other head-initial languages do (Adani, 2011; Diessel & Tomasello, 2005; Friedmann & Novogrodsky, 2004; Guasti et al., 2012; Kas & Lukácz, 2012; Theodorou & Grohmann, 2012, to name but a few) even though the data regarding the SRC benefit in Russian relative clause processing studies of adult populations are not entirely clear (Edeleva & Demareva, 2017; Levy et al., 2013).

Relative clause processing in monolingual Russian speaking children

Acquisition of Russian relative clauses has enjoyed only limited attention on the part of experimental linguistic research until now. Existing child language corpora (cf. CHILDES; MacWhinney, 2000) do not provide sufficient linguistic material to follow up on English (Diessel et al., 2000) and German (Brandt et al., 2008) and study the emergence of relative clauses in Russian children based on corpus data. To date, there is no investigation known to us of relative clause comprehension in Russian children based on spontaneous child language data. The available entries from two children contained in the CHILDES database do not suffice to make frequency-based predictions based on child-directed speech regarding Russian children’s performance in a controlled experimental setting.

The processing penalty toward ORCs in adult native speakers is dependent on the measurement type (Levy et al., 2013; Price & Witzel, 2017). Scarce experimental studies into children’s comprehension of Russian relative clauses (Polinsky, 2008, 2011; Rakhlin et al., 2016) provide conflicting evidence on the Subject/Object Asymmetry using a picture selection paradigm (see also Lau & Tanaka, 2021). These studies primarily address special populations, such as English-dominant heritage speakers (Polinsky, 2008, 2011) or DLD²-children (Rakhlin et al., 2016). In either case, typically developing Russian-speaking children were treated solely as a baseline group. Thus, to our knowledge, no studies have been conducted so far that would explicitly address relative clause acquisition by pre-school monolingual Russian children.

Polinsky (2008, 2011) investigated the performance of 6- and 7-year-old children on Russian relative clauses which is a much older age group than the age range examined in other studies reporting Subject/Object Asymmetry (cf. Lau & Tanaka, 2021 for a review). In a picture selection paradigm, children showed almost at-ceiling comprehension scores both on subject and on object relative clauses, with a slight downturn on the latter that was not statistically significant. Neither was there a difference between the two word orders or their interaction with the relative clause type (percentage counts for correct responses: 85.9% (SRC-V-NP), 87.8% (SRC-NP-V), 87.2% (ORC-V-NP), and 86.1% (ORC-NP-V)). Since children did not significantly deviate from adults in their performance, Polinsky concludes that “6- and 7-year-olds have adultlike control of relative clauses, with equal mastery of SRs and ORs” (Polinsky, 2011, p. 17).

Rakhlin et al. (2016) developed a more complex visual frame for their study of relative clause processing in Russian children. Their verbal experimental stimuli were also more complex and involved a propositionally filled matrix clause and a relative clause, which could either disrupt the matrix clause (center-embedding) or be attached at the matrix clause offset (right-branching) The visual displays were comprised of three individual images with three characters each.

One of the target images corresponded to the SRC interpretation and depicted the head NP character verb-ing the embedded NP character. In the second image, thematic roles were reversed, and the embedded NP character was verb-ing the head NP character, which is congruent with the ORC reading. The third image (a foil) was not designed to accommodate any relative clause reading. The head NP appeared twice in the image to pragmatically solicit the use of restrictive relative clauses (Hamburger & Crain, 1982). Typically developing children (avg. age = 10,69) were more accurate on SRCs than on ORCs, as was evidenced on their accuracy scores. They also experienced less difficulty understanding right-branched relative clauses as opposed to center-embedded ones. Neither effect showed up in the adult control group. To challenge Polinsky’s results (2008, 2011), Rakhlin et al. (2016) reason that “typically developing children with the mean age of ~ 10.6 still didn’t reach adult-level RC processing” and attribute children’s comprehension deficit with ORCs and center-embedding to inefficient deployment of grammatical knowledge in real time.

In sum, the studies deliver contradictory results on the age at which Russian children eventually comprehend SRCs and ORCs equally well. In addition, the results of Polinsky (2008, 2011) who didn’t find any word order effect in the comprehension of relative clauses would benefit a replication with a larger sample of monolingual children.

Current study

The current study systematically investigates relative clause processing in typically developing Russian-speaking children. Empirical evidence on Subject/Object asymmetry and word order variability in relative clauses will be considered against previous findings on relative clause acquisition for Russian and for other languages.

The study seeks to fill in the existing gap in the relative clause acquisition literature and trace typically developing children’s processing of relative clauses with alternating word orders in Russian across different age groups. The first question is related to the well-known Subject/Object Asymmetry. Polinsky (2008, 2011) examined a group of children with a mean age of 6,6 while Rakhlin et al. (2016) investigated still older children. While Polinsky (2008, 2011) and Rakhlin et al. (2016) deliver controversial results on this issue, it might be due to the differences in the experimental designs that they employed. Rakhlin et al. (2016) visual displays were more complex, and their stimulus material also included instances of center-embedding.

Earlier studies observed a difference in linguistic properties that children of different ages make use of to determine agent–patient relations in a sentence. For instance, Belletti et al. (2012) failed to confirm a facilitative effect of gender mismatch on ORC comprehension, which Adani et al. (2010) evidenced in Italian children. Yet, the age range of children examined by Adani et al. (2010) was between 6 and 10 years while the mean age in the study of Belletti et al. (2012) barely reached 3,7. Obviously, younger and older children differ in the extent to which they can put their existing morphological knowledge to work during structural analysis. Thus, rather than investigating still older children, we investigated younger preschoolers that were within their fourth or fifth year of life at the point of the experiment to trace whether and how they differ in their use of morphological knowledge for structure building.

Regarding word order, Russian allows for embedded NP placement before and after the verb in the relative clause structure. However, in contrast to Greek, Italian, Catalan, French and Romanian, word order alternation is possible both for SRCs and for ORCs. Assuming that reversible object relative clauses are comparatively infrequent and, therefore, cause greater interpretation difficulty with children, the frequency effect could also apply to word order alternation. Considering input frequency, one would expect to observe a processing penalty to the less frequent V-NP structure for ORCs and the NP-V one for SRCs. By contrast, the V-NP SRC and the NP-V ORC should be receiving a comprehension advantage. In addition, the frequency effect is concurrent with word order canonicity. Existing research on Russian word order (Levy et al., 2013; Slioussar, 2011) reports that though Russian sentences usually contain multiple instances of scrambling, resulting word orders are subject to context effects. Therefore, contextually constrained NP-V and V-NP structures for SRCs and ORCs, respectively, would cause comprehension difficulty outside a pragmatically appropriate context.

The linear embedded NP intervention effect would remain neutral to word order alternation in SRCs but impair the processing of NP-V ORCs compared to their V-NP counterparts in the course of integration. Yet, in structural terms, the depth of hierarchical embedding would generally render ORCs more challenging than SRCs. The V-NP ORC would be the most difficult to process since it requires the parser to posit an empty [pro]-slot for a putative subject to the verb (in a similar vein as suggested in Fodor & Inoue, 2000). Upon encountering the actual subject, the parser would need to check whether it is legitimated by verb morphology and can be filled into the [pro]-slot. Since this additional step is reported to be particularly costly (Adani, 2011; Guasti et al., 2012, 2018), we would expect Russian children to perform worse on V-NP ORCs compared the NP-V structure. By contrast, no such difference will be observed between the two word orders for SRCs where the head NP trace is established immediately after the relative pronoun.

Regarding the utility of cues (cf. the Competition Model, E. Bates & MacWhinney, 1989), we would predict a subject advantage due to cue redundancy as the initial agentive NP is additionally marked for the nominative (see Tal & Arnon, 2022 for a similar suggestion). However, as children become more reliant on cues with a higher conflict validity (e.g., case marking), their performance on ORCs should enhance.

Table 4 gives a concise overview of the predictions regarding different accounts and how they might affect word order processing in Russian relative clauses.

Table 4. Summary of predictions with regard to different factors affecting word order processing in relative clauses.

Factor	Syntax	Word order
Agent-First Strategy	SRC	ORC generally more challenging than SRC irrespective of word order
	ORC
Intervention of Embedded NP (for Linear Filler-Gap Distance)	SRC	NP-V = V-NP
	ORC	NP-V>V-NP
Depth of Hierarchical Embedding	SRC	ORC generally more challenging than SRC irrespective of word order
	ORC
Diagnosis Model (Fodor & Inoue, 2000)	SRC	NP-V = V-NP
	ORC	V-NP > NP-V
Word Order Canonicity and Frequency Distribution	SRC	NP-V > V-NP
	ORC	V-NP > NP-V
Competition Model (E. Bates & MacWhinney, 1989; MacWhinney, 2005)	SRC are facilitated due to cue redundancy; the performance on ORCs improves as children become more reliant on cues with higher validity	Word order plays a role in as far as it enhances the agent-first interpretation in coalition with the nominative case marking

“>” stands for “is more costly than;” “=” stands for “requires equal processing resources.”

More generally, the study will consider relative clause processing pattern observed for Russian pre-school children against a general pathway of relative clause development from early years into adulthood and, thus, relate the findings to the processing tendencies observed for adults, on the one hand, and for children acquiring other languages, on the other. Arguably, relative clause acquisition in Russian speaking children follows a general cross-linguistic pattern. Although adult native speakers of Russian do not always evidence additional processing cost to object relative clauses as compared to their reversible subject-extracted counterparts (Levy et al., 2013; Edeleva & Demareva, 2017, though cf.; Price & Witzel, 2017 for the subject advantage on a comprehension check), a monolingual L1 Russian learner is characterized by the phenomenon of Subject/Object Asymmetry found for the majority of rigid word order languages with initial head NP placement. With increasing age, children become more reliant on morphological information for structural analysis but remain indifferent to the word order that goes with a particular structure type. Thus, even at six years of age, children don’t display a full range of preferences that have been attested for adults.

Materials and methods

To probe children’s comprehension of Russian relative clauses, a touchscreen character identification³ task was developed contrasting the following two parameters: 1) relative clause syntax (subject vs object); 2) word order (V-NP vs NP-V), - resulting in 4 distinct experimental conditions summarized in Table 5.

Table 5. Overview of experimental conditions for the study of relative clauses in Russian children.

Syntax	Position	Main Clause		Relative Pronoun	Embedded NP/Verb	Embedded NP/Verb
ORC	NPV	Gde	gus,’	kotor-oroi	lis-a	ugoščajet___ti?
		Where	goose.NOM	that.ACC	the fox.NOM	treat. SG
		Where is the goose that the fox is treating?
	VNP	Gde	gus,’	kotor-oroi	ugoščajet___ti	lis-a?
		Where	goose.NOM	that.ACC	treat.SG	the fox.NOM
		Where is the goose that the fox is treating?
SRC	NPV	Gde	gus,’	kotor-yii___ ti	lis-u	ugoščajet?
		Where	goose.NOM	that.NOM	the fox.ACC	treat.SG
		Where is the goose that is treating the fox?
	VNP	Gde	gus,’	kotor-yii___ ti	ugoščajet	lis-u?
		Where	goose.NOM	that.NOM	treat.SG	the fox.ACC
		Where is the goose that is treating the fox?

The head NP referent was always a masculine singular noun to allow for phonologically distinct endings on the relative pronoun in the subject and the object condition (kotor-yi.that.NOM vs kotor-ogo.that.ACC) while the embedded NP could be of masculine or feminine grammatical gender. The gender of the embedded NP did not affect the verb ending due to paradigmatic syncretism. The main clause was the interrogative verb-omission structure ”Gde X?‘//’Where is X?” prompting the child to search for the head NP character as agent or patient to the relative clause verb. Thus, the two structure types contrasted in the study are the S-S and the S-O relative clause.

A set of 12 experimental six-object pictures as in Figure 1 were created using a commercially available clip-art database. The visual display contained two rows of animals Each row consisted of two characters on the sides corresponding to the head NP and contrasted visually for the agent or the patient thematic role (e.g., the goose with or without a cake), and one character in the middle, which designated the embedded NP. In one of the rows, the middle character corresponded to the embedded NP in the auditory stimulus. The other row was unrelated.

Figure 1. Visual display frame for the study of relative clauses in Russian children.

The designated embedded NP character and the unrelated one (the characters in the middle from the upper and the lower rows) were carefully selected in such a way that they did not significantly differ in their mean frequency across the stimuli (Welch Two Sample t-test: t = 0.58566; df = 21.98, p = .56; individual frequency counts for animals in each pair are contained in Appendix 1). The animals in the middle appeared twice across the stimulus material, each time with a different verb and in a different animal combination (e.g., “dog-up/snake/wash;” “dog-down/frog/tickle”). Their spatial location (whether they appeared in the upper or in the lower row) was counterbalanced across trials. The verbs were: feed, treat, heal, draw, kiss, watch, wash, tickle, dry, teach, dress, take a photo of. All the verbs are transitive in Russian and were used in the active voice.

Additionally, a set of four two-object images (as in Figure 2) was created where the animals were not interrelated to accommodate 8 filler items. The sentence frames for the filler items were subject relative clauses with prepositional argument verbs (Gde utka, kotoraja igrajet na gitare?//Where is the duck that is playing the guitar?); intransitive verbs (Gde ryba, kotoraja zvonit?//Where is the fish that is calling?); reflexive verbs (Gde belka, kotoraja smejetsja?//Where is the squirrel that is laughing?); transitive verbs with an inanimate argument (Gde petuh, kotoryi p’jot vodu?//Where is the cock that is drinking water?).

Figure 2. Sample visual display frame for filler items.

The sentences were audio-recorded by a professional speaker at a sound-recording studio. The stimulus material was Latin-square counter-balanced and pseudo-randomized. Each experimental list contained 20 experimental items (3 × each experimental condition; 1 × each of the eight filler sentences).

46 monolingual Russian children recruited at a Russian kindergarten in Nizhny Novgorod (Russia) participated in the study. They were stratified according to age (with age ranges corresponding to 3;6 to 4;11 years; 5;1 to 5;11 years and 6;1 to 6;11 years). Prior to the experiment, the parents of participating children were informed about the experiment procedure. They signed a consent form to document their consent for their children’s participation in the study. The study was approved by the ethics committee of the University of Bielefeld (Germany) and was administered in accordance with the Declaration of Helsinki (World Medical Association, 2013). All the children were normally developing, did not have any known history of language disorders and had not been exposed to any language other than Russian on a regular basis prior to the time of the study.

The children were tested individually by a native Russian experimenter. The task was administered to them via Open Sesame 2.9.6 Runtime for Android Software (Mathôt et al., 2012) on Telefunken Tablet Touchscreen PC TF-MID7805G; 7.85“; Android 4.2.2 Jellybean Operation System. The experimenter invited the children to play a computer game where they had to help the female speaker find different animals. They had to listen to the speaker carefully and choose the animal that they thought the lady was looking for via touch response. To ensure that the children could recognize all the animals, prior to the test phase they had to sequentially identify each of the 22 animals on the display in response to the speaker’s request of the type Gde gus‘, gus’ ? (”Where is the goose, the goose?”) The familiarization phase was followed by two practice items identical to those used in the test phase.

Results

Descriptive statistics for children’s accuracy on each of the four conditions are summarized in Table 6 and visualized in Figure 3. Non-target responses were put into three categories for further analysis: thematic reversal error, embedded NP error and other. Thematic reversal error corresponds to the reverse thematic role (i.e., the patient goose in SRCs and the agent goose in ORCs). Embedded NP error designates the choice of the character which corresponds to the embedded NP (i.e., the fox). The category “other” collected responses from the non-target row in the image.⁴

Figure 3. Children’s accuracy data aggregated by error type and summarized by condition. Error bars indicate standard error.

Table 6. Descriptive statistics for different response types (Target; Reversal; Embedded NP; Other) by condition (O_NPV; O_VNP; S_NPV; S_VNP).

Condition	Response type	Mean	SD
O_NPV	Target	0,55	0,5
	Reversal	0,28	0,44
	Embedded NP	0,12	0,37
	other	0,05	0,2
O_VNP	Target	0,52	0,5
	Reversal	0,23	0,42
	Embedded NP	0,19	0,39
	other	0,06	0,26
S_NPV	Target	0,8	0,4
	Reversal	0,08	0,28
	Embedded NP	0,07	0,22
	other	0,05	0,24
S_VNP	Target	0,81	0,39
	Reversal	0,06	0,24
	Embedded NP	0,09	0,29
	other	0,03	0,18

For statistical evaluation, a generalized linear mixed effect model (Bates et al., 2015) was fitted with accuracy scores as response variable and age, syntax, word order as well as the interaction of syntax and word order as predictor variables in the R programming environment (R Core Team, 2015). Categorical predictors were sum-coded and scaled. The maximal model that converged was fitted with random intercepts for participants and items. The summary of the model fit is provided in Table 7.

Table 7. Summary of LME output (fixed effects) for response accuracy on relative clauses in Russian children with age groups collapsed.

	Estimate	Std.Error	z value	p value
Intercept	−0.2343	0.2619	−0.895	0.3710
Syntax (Sub)	1.5646	0.2893	5.409	6.34e-08	***
Position (NPV)	0.1578	0.2510	0.629	0.5296
age5	0.2257	0.3008	0.750	0.4532
age6	0.7133	0.3093	2.306	0.0211	*
syntax*position	−0.2542	0.3976	−0.639	0.5225

Age groups collapsed, the model revealed a significant effect of relative clause syntax with a comprehension advantage to subject relative clauses. The effects of word order or the interaction of syntax and word order were not significant. The effect of age was significant indicating that children’s performance improved with age. The participants’ accuracy per age group is summarized in Table 8 and visualized in Figure 4.

Figure 4. Response accuracy per age group and condition. Error bars indicate standard error.

Table 8. Response accuracy per age group (4-, 5-, 6- year-olds) and condition (O_NPV; O_VNP; S_NPV; S_VNP).

Age	Condition	Mean	SD
4 y.o.	O_NPV	0,54	0,5
	O_VNP	0,39	0,49
	S_NPV	0,72	0,44
	S_VNP	0,79	0,41
5 y.o.	O_NPV	0,47	0,5
	O_VNP	0,46	0,5
	S_NPV	0,83	0,37
	S_VNP	0,83	0,37
6 y.o.	O_NPV	0,62	0,48
	O_VNP	0,68	0,46
	S_NPV	0,83	0,37
	S_VNP	0,81	0,39

In the next step, the same generalized linear mixed effects model as for the entire data sample was fitted for each of the three age groups separately. For each data sample, the maximal model that converged was computed which accounts for slight differences in the random effects structure between the models. Individual model fit summaries together with respective random effect structures are provided in Appendix 2. By-group analysis revealed that relative clause syntax was a significant predictor of children’s performance in all the three age groups. No other predictor produced a different statistical effect compared to the analysis of the entire data sample.⁵

In sum, we observe considerable processing penalty to ORCs compared to SRCs with no comprehension advantage to either word order. Neither did the children display any word order preference with regard to the subject or the object relative clause type.

Non-target response types

To determine whether a specific non-target response type was more common compared to other non-target response types, pairwise proportion comparisons were performed (Table 9). Thematic reversal error and embedded NP error were each more common than “other” error. Reversal errors were also significantly more common than embedded NP errors.

Table 9. Pairwise proportion comparisons for the three error types across the data sample.

	Df	Diff	Chisq.value	p value
Embedded NP error vs Reversal error	1	−0.045	4.906	0.027	*
Embedded NP error vs Other error	1	0.108	34.739	<.001	***
Reversal error vs Other error	1	0.062	14.454	<.001	***

To examine whether embedded NP Error was more common in a particular condition, an additional generalized linear mixed effects model was fitted with embedded NP error counts as response variable. Relative clause syntax, word order and their interaction as well as the participants’ age were introduced into the model as predictor variables. The maximal model that converged was fitted with random intercepts for participants and items. The model yielded a significant effect of relative clause syntax, but not of word order (cf. the model fit summary in Table 10).

Table 10. Summary of LME output (fixed effects) for embedded NP error counts in Russian children with age groups collapsed.

	Estimate	Std.Error	z value	p value
Intercept	−1.5518	0.2777	−5.588	2.30e-08	***
Syntax (Subj)	−1.3799	0.3178	−4.342	1.41e-05	***
Position(VNP)	0.1730	0.2789	0.620	0.535
age5	0.1511	0.3119	0.484	0.628
age6	−0.5376	0.3370	−1.595	0.111
syntax*position	0.3975	0.4897	0.812	0.417

In sum, embedded NP error was the second most common non-target response type after thematic reversal error. Importantly though, children opted for the embedded NP character significantly more often than for unrelated characters contained in “other” error counts. Thus, embedded NP error reflects a systematic response pattern in Russian preschool children that is not attributable to inattention or occasional slips in performance independent of experimental manipulations.

Discussion

The present study tested monolingual children’s interpretation preferences regarding subject and object relative clauses. Additionally, the study design included word order alternations (NP-V vs. V-NP). To extend the body of existing literature on the comprehension of syntactic dependencies by Russian children, we also studied a younger age group than those tested in previous studies. The so-called Subject/Object Asymmetry persisted across the three age groups, thus, indicative of the fact that even late preschoolers do not display full mastery of relative clause syntax. Nevertheless, the children’s performance in the current study generally improved with age. The emergent pattern is strikingly consistent with the one outlined by developmental studies into quite a number of languages (Adani, 2011; Diessel & Tomasello, 2005; Friedmann & Novogrodsky, 2004; Guasti et al., 2012; Kas & Lukácz, 2012; Theodorou & Grohmann, 2012, to name but a few, though this effect is subject to language-specific constraints, see; Kirjavainen et al., 2017 for Finnish; Carreitas et al., 2010; Gutierrez-Mangado, 2011 for Basque).

While the linear distance assumption is not compatible with our data on younger children, the representational complexity account is generally congruent with the pattern of results we observed. The general assumption is that object-extractions will always be deeper embedded than subject extractions and will contain an intervening NP in their hierarchical structure. Therefore, they will be more complex in terms of structural representation. Thus, ORCs will always be more taxing irrespective of the word order type, which is consistent with our data. Yet, it is not corroborated by children’s comprehension data for Romance languages that additionally evidence a well-pronounced word order preference for ORCs which we did not evidence in the current study. Probably, the [pro]-slot will only need to be established if the number marking on the verb overtly contradicts the number information of the putative subject (i.e., the head NP is in the singular and the verb in the plural). This assumption would, however, go against the facilitative effect of number mismatch reported earlier. Cross-linguistic inconsistency puts into question the universal nature of hierarchical complexity to account for children’s poor performance on ORCs. Regarding children’s better performance on SRCs compared to ORCs, a possible constraint in the current study might be that we used a propositionally empty main clause where the head NP occurred in its citation form (Gde gus,’ … //Where is the goose.NOM.SG, …). The use of the citation form increases the comparability across languages that do not use inflection to mark the object function. Yet, Rakhlin et al. (2016) note a possible anti-priming effect that emerges due to case-mismatch between the head NP and the relative pronoun. Though case-mismatch significantly degraded children’s performance on both the extraction types, it still cannot fully explain the SRC advantage. Unfortunately, Rakhlin et al. (2016) do not provide exact statistics regarding the amount of variance explained by the syntax type and its interaction with case match. Therefore, it remains for further research to investigate in more controlled way (e.g., by explicitly manipulating the syntactic function of the head NP in a propositionally empty main clause), which explanatory power can be ascribed to case-mismatch to account for the Subject/Object Asymmetry.

Though we expected different factors (structural, statistical and pragmatic) that have been argued to account for the Subject/Object Asymmetry to co-determine children’s word order preference, word order did not significantly affect children’s accuracy in the current study. A purely technical confound in the current study might be that accuracy scores are not sensitive enough to capture potential differences in word order preference. The investigation of word order preference might require more fine-grained methodologies such as, for example, the speed with which children direct their eye movements toward the designated character depending on the relative clause constituent order.

As such, children’s insensitivity to the word order manipulation is consistent with the assumption of relative cue strength proposed in E. Bates (1999). For Italian and English, they observed that children acquiring either language generally performed worse on non-canonical structures than on canonical ones. Yet, English-speaking children were quantitatively more prone to assigning the faulty SVO interpretation to a non-canonical structure than Italian-speaking ones. Since Italian relies on morphological information (e.g., number agreement on the verb for all the three tenses) for thematic role assignment on more occasions than English, the authors reason that users of languages with the same basic word order (SVO) differ in the extent to which they rely on word order information for structural analysis. Russian displays high constituent order flexibility for pragmatic purposes resulting in generally low correlations between thematic roles and a particular constituent order. Yet, case marking requires time to be sufficiently mastered in order to be utilized for structure building. Hence, it has been suggested that learners acquiring flexible word order languages would first adhere to one most salient type of information sequencing (e.g., agent-action-patient) and then relax it as more evidence from actual linguistic input comes in (Crain, 1993). The results of the current study are in keeping with this explanation, which suggests that cue redundancy on the initially occurring NP that additionally inflects for the nominative positively affects the comprehension of subject relative clauses in younger children. Further, the type of extraction (subject vs object) pertains to the core syntax while word order alternation is an element of information-structural organization of linguistic input. It can, thus, be hypothesized that children will first acquire the two extraction types and only later develop adult-like information-structural preferences which would be reflected in their sensitivity to word order alternation.

That children will be trying to acquire case morphology as the most reliable cue (Devescovi et al., 1999) as early as possible, is confirmed by at least one more finding that remained in shade in previous investigations of relative clause interpretation in Russian children. Though thematic reversal error was the most common error type attested in the study, non-target response pattern of Russian children also featured instances of embedded NP error. Embedded NP error barely reaches 20% with selected relative clause types. However, it is significantly more frequent than “other” error when children chose the characters from the irrelevant row that were not questioned by the experimental stimuli. Thus, it cannot be regarded as an artifact of the experimental design but requires an explanation as a systematic behavioral pattern in its own right. Neither can it be regarded a product of lexical activation for the character designated by the embedded NP. If we hypothesize that children point at the most recent referent, they would display a comparable response pattern for SRCs and ORCs alike. In both syntax types, embedded NP referent is mentioned last. However, the embedded NP error is significantly more common in ORCs compared to SRCs. In this respect, Russian extends the list of languages where embedded NP error has been attested on behavioral response data. As in Italian (Adani, 2011), Catalan (Gavarrò et al., 2012), Hebrew (Arnon, 2005), Chinese (Hu et al., 2016), Finnish (Kirjavainen et al., 2017) and French (Guasti et al., 2018), embedded NP error was more persistent in ORCs than in SRCs. Thus, a cross-linguistic account of ORC interpretation difficulty is needed that would also incorporate embedded NP errors and their role in relative clause acquisition (Arnon, 2005). However, rather than suggesting that children fail to grasp the modifying nature of the relative clause (Arnon, 2005), we reason that children’s commitment to the embedded NP character is indicative of their growing morphological knowledge and their ability to put it to work in real-time. Instances of embedded NP error are reflective of children’s attempts to use morphological information that becomes available later in the sentence in order to update their ongoing structural analysis. Since children’s cognitive resources are limited (Mazuka et al., 2009; Trueswell et al., 1999), they do not allow them to update it into an entirely new structure and generate an ORC instead of an SRC. However, those intermediate steps that become evident as embedded NP errors are mediated by morphological knowledge. It remains for further research to determine how embedded NP error probabilities are modulated by individual differences in morphological knowledge and morphological growth. Further, morphological knowledge might be taking effect as early as at the relative pronoun. Case marking on the relative pronoun will be used predictively to preempt either a subject or an object relative clause structure. Since a non-canonical structure is more complex, but also less frequent and, therefore, less expected, the prediction of this structure type might generally be weaker than of a symmetrical canonical structure. Rather than suggesting that children will be neglecting morphological cues in favor of word order, we suggest that children acquiring morphologically rich languages would be using case marking on the relative pronoun for structural prediction. The ORC structure will later be verified as soon as more information (e.g., case marking on the embedded NP) becomes available.

Finally, relative clauses are an attribute of written rather than of spoken discourse (Roland et al., 2007). In written texts, they occur in their great variety and abundance including different types of oblique relativization. The age range examined in the current study lies between 4 and 6 years when children have only limited exposure to written texts. Recent research, however, confirms the role of reading proficiency as a significant predictor in the processing of complex syntactic structures. King and Just (1991) observed that language users with higher reading span showed shorter reading times on certain regions in sentences containing object relative clauses while no effect of the reading span was evidenced for sentences with subject relative clauses. Cilibrasi et al. (2019) report that children’s accuracy on ORCs may catch up as their reading proficiency increases. Further research should, therefore, reach out beyond pre-school age and examine whether other relative clause properties, for example, constituent order preference, require greater exposure to written texts to emerge (see also MacDonald & Christiansen, 2002 for the importance of direct experience with object relatives through exposure to sophisticated reading).

Conclusion

The present study tested monolingual Russian-speaking children’s ability to comprehend relative clauses with subject- and object extraction types. The study also manipulated constituent order (V-NP vs NP-V) in relative clauses and extended the body of existing literature on the comprehension of complex syntactic dependencies from the developmental perspective. An examination of children’s interpretation preferences has revealed the Subject/Object Asymmetry for all the three age groups. The emergent pattern is generally consistent with the one outlined by developmental studies into other head-initial languages, suggesting that the Subject/Object Asymmetry can be seen as a cross-linguistic developmental phenomenon that children display when they acquire their first language. Yet even six-year-olds still remain indifferent to adult-like preferences with regard to a particular word order to go with either the subject or the object relative clause. Thus, further research examining adolescent populations is needed to determine how increasing exposure to written texts and individual reading proficiency contribute to the development of adult-like preference in information-structural domain.

The error pattern preschool children displayed when they failed to correctly identify the target character revealed embedded NP error as the second most common error type after the thematic reversal error. Instances of embedded NP error are reflective of a possible developmental pathway toward ORC mastery when children first apply the canonical structure, but then discover an inconsistency at the morphological level and attempt to update their ongoing analysis.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data availability statement

The data that support these findings can be made available from the corresponding author Julia Edeleva, upon reasonable request.

Notes

¹ For Romanian, Sevcenco and Avram (2012) point out that the data are less clear-cut. The experimental results provided in the study attest to the NP-V advantage in ORCs. At the same time, the authors note that instances of ORCs with post-verbal subjects appear in child speech quite early. Generally, children show a comprehension-production asymmetry since the number of NP-V ORCs produced by children is quite low.

² Developmental Language Disorder.

³ A character identification task has some obvious advantages as compared to a picture selection task used in the original study by Friedmann and Novogrodsky (2004) since it is more sensitive to children’s performance when they point at the correct picture, but not at the correct referent (Arnon, 2005)

⁴ Occasionally, children also opted for one of the two geese from the non-target row, i.e., they pointed to the goose which was “fed by the owl” when they were asked to point to the one “fed by the fox.” Since the visual relation between the animals in the image is not sufficiently transparent, the choice of the goose in the intended syntactic function might be interpreted as correct. However, none of the children systematically adopted this strategy and the statistics on this response type barely reach 3%. Since it is not entirely clear what motivated those rather sporadic responses, they will also be treated in the “other” category.

⁵ One of the anonymous reviewers pointed at the numeric difference in four-year-old children’s comprehension accuracy on two word order types in ORCs. Though the NP-V word order indeed seems to display a comprehension advantage, applying a different contrast specification (sliding contrasts, cf. Schad et al., 2020) did not statistically validate this difference (Estimate = −0,58; Std.Error = 0,41; z = −1,42; p = 0,15).

Appendix 1.

Frequences of animal nouns for each animal pair (in wpm) based on the basic component of the Russian National Corpus

Table

Item	Animal 1	Frequency	Animal 2	Frequency
1	LISA.fox	9,1	SOVA.owl	11,81
2	LJAGUSHKA.frog	13,39	AKULA.shark	7,096
3	KOSHKA.cat	44,85	DELFIN.dolphin	5,8
4	KURITSA.hen	35,23	MYSHKA.mouse	19,82
5	ORJOL.eagle	41,88	SLON.elephant	21,37
6	ZMEJA.snake	36,3	SOBAKA.dog	148,73
7	SOBAKA.dog	148,73	LJAGUSHKA.frog	13,39
8	LISA.fox	9,1	KURITSA.hen	35,23
9	MYSHKA.mouse	19,82	ZMEJA.snake	36,3
10	ORJOL.eagle	41,88	AKULA.shark	7,096
11	KOSHKA.cat	44,85	SOVA.owl	11,81
12	DELFIN.dolphin	5,8	SLON.elephant	21,37

Appendix 2.

Summary of LME output (fixed effects) for response accuracy on relative clauses in Russian children per age group

Table

	Accuracy (4 y.o.)					Accuracy (5 y.o.)					Accuracy (6 y.o.)
Random effects structures	(1+syntax|subject) + (1+syntax|item)					(1|subject)					(1subject) + (1|item)
predictors	Odds Ratios	std. Error	CI	Statistic	p	Odds Ratios	std. Error	CI	Statistic	p	Odds Ratios	std. Error	CI	Statistic	p
(Intercept)	2.10	0.66	1.14–3.87	2.37	0.018	2.32	0.59	1.41–3.83	3.30	0.001	3.44	0.88	2.08–5.70	4.81	<0.001
syntax	6.21	3.90	1.81–21.28	2.91	0.004	7.13	2.74	3.35–15.16	5.10	<0.001	2.68	0.97	1.32–5.47	2.71	0.007
position	1.21	0.46	0.57–2.56	0.49	0.626	1.05	0.38	0.52–2.13	0.13	0.893	0.92	0.33	0.46–1.86	−0.22	0.824
syntax * position	0.24	0.19	0.05–1.16	−1.78	0.075	0.91	0.65	0.22–3.73	−0.13	0.893	1.60	1.15	0.39–6.52	0.66	0.512
Observations	192					192					192
Marginal R^2/Conditional R^2	0.179/0.396					0.196/0.335					0.062/0.212