2.1 Sign languages within linguistic research

Research into sign languages is a young field when compared to spoken language studies. With our study, we also aim at contributing to the understanding of the nature of sign languages with respect to spoken languages.

Over the past decades, an ever-growing body of research has shown that sign languages resort to mechanisms that are sufficiently similar to those observed in spoken languages as to allow cross-modal generalizations in all linguistic domains. At the same time, the visual modality does have an impact on language organization and articulation, and its unique features afforded by the modality should not be downplayed in linguistic descriptions. This is particularly clear when it comes to the grammatical use of the signing space, as reflected by the glossing conventions that have been elaborated specifically to report on sign languages. We here follow the conventions adopted in the SIGN-HUB grammars (https://sign-hub.eu/grammar), that we briefly introduce now to facilitate the reader of this paper.

First of all, signs are glossed using non-inflected English words in small capital letters. As for spatial information, it is encoded through the use of indices. When a nominal sign is articulated in a specific location in the signing space, this is signaled with an index letter. When a nominal sign is body anchored, it might still carry an index to indicate possible agreement or coreferential relations the sign might entertain. Pronominal pointing pronouns are referred through the gloss “IX”, followed by an indexical number indicating person marking and/or a letter to indicate the specific location in the neutral signing space where they are articulated. Shared indices signal co-referring elements. The same conventions apply to directional verbs (i.e. verbs that display spatial agreement between a starting and an ending location, see Padden 1988) or any sign realized in a specific location in space. Finally, non-manual markers are linguistic forms that do not imply the use of the hand; they can include movement of the body or the eyebrows, eye-gaze direction, and any other facial expression. Non-manual markers and their spreading over manual signs are signaled through a line extending above the signs they co-occur with. These glossing conventions are exemplified in the following French Sign Language (LSF) sentence, in which “bt” refers to body turn, “re” to raised eyebrows and “rg” to right gaze.

(5)	French Sign Language (Hauser 2019: 170)
	‘Marie says to me that her sister is nice.’

2.2 The structure of this paper

The paper is organized as follows: Section 2 presents an overview of the different questions we are addressing, starting with what is known about biases in processing externally headed RCs, followed by an outline of the puzzle of internally headed RCs (§2.2), and a brief overview of the relevance of age of first language exposure in syntactic competences (§2.3). Section 3 offers a description of how relativization works in the three sign languages under investigation, LSF (§3.1), LIS (§3.2) and LSC (§3.3). Section 4 presents our study on the comprehension of RCs in these three languages and Section 5 details its results, by presenting the analysis of accuracy (§5.1) and that of errors (§5.2). Section 6 provides a general discussion and Section 7 briefly concludes the paper.

3.1 The puzzle of internally headed relative clauses

Internally headed RCs are clauses involving no gap, which are however nominal in nature (Cole 1987; Kayne 1994; Keenan & Comrie 1977; or Lehmann 1986). They can even show nominal inflection, as in Japanese, where a case marker (accusative) is attached to the internally headed relative clause in (12).

Internally headed RCs are analyzed by Kayne (1994), Cole (1987) and Culy (1991), among many others, as the mirror image of externally headed RCs: just like them, they modify a head noun (keeki, ‘cake in 12) that is shared semantically and syntactically between the main and the relative clause. The only difference between the two constructions from a structural perspective is the overt position of the head, which is either external or internal. For proponents of raising theories, like Cecchetto & Donati (2015), De Vries (2001) or Bianchi (2000), internally headed RCs involve the covert movement of the relative head while externally headed RCs involve the overt movement of the relative head. For proponents of head-matching theories, like Sauerland (2003) or, more recently, Cinque (2019), internally headed RCs and externally headed RCs are both base-generated with two head nouns, one internal and one external, with one of them deleted at PF.

In both families of analyses, there is no gap within internally headed RCs at least in overt syntax. If the Subject Advantage attested for externally headed RCs is due to the structural distance between the head noun and the gap (possibly because the identification of a null category is a complicated business), we would predict internally headed RCs to display no difference in comprehension between subject RCs and object RCs. However, both families of analyses posit that internally headed RCs involve an abstract dependency between two positions. Is comprehension affected by the abstract dependency involved in internally headed RCs in the same way it is affected by the concrete dependency involved in externally headed RCs? If it is, we would predict a Subject Advantage in internally headed RCs as in externally headed RCs.

To this date, very little research has been conducted with these questions in mind, and all come from the domain of L1 and L2 acquisition studies. Jeon Seon & Hae-Young Kim (2007) analyzed the elicited production of 40 L2 learners of Korean, a language displaying both internally headed RCs and externally headed RCs. The aim was to verify whether there is an asymmetry in the acquisition of Subject RCs and Object RCs in head-internal and head-external constructions. What they found was a clear Subject Advantage for externally headed RCs, with L2 learners producing Subject RCs more frequently and more accurately (91.1% of correct answer) than Object RCs (83.7%), but a less clear pattern for internally headed RCs. The authors do not report the accuracy scores, but note that Object RCs are slightly more frequent (6%) than Subject RCs (4. 8%). Similarly, Hu Shenai, Cecchetto & Guasti (2018) made an elicited production study of RCs in Wenzhounese, a language that allows both internally headed RCs and externally headed RCs, and report a Subject Advantage in externally headed RCs, but an object preference in internally headed RCs.

The order of acquisition of internally headed RCs and externally headed RCs itself seems to be different, as shown by a longitudinal study on three Korean children’s L1 acquisition conducted by Kim Youn-Joo (1987), who found that Korean children start to produce headless relative clauses around 2 y.o., before producing internally headed RCs, and that externally headed RCs come last. This pattern is also found by Lee Awen Ock’s (1991) investigation of the spontaneous production of 36 Korean children, with a mean age of acquisition of internally headed RCs around 27 months, and around 36 months for externally headed RCs. The same pattern is confirmed by Jeon Seon & Hae-Young Kim (2007) for L2 learners of Korean.

These scattered results, although interesting, are not sufficient to draw any conclusions regarding the Accessibility Hierarchy and its predictive power when it comes to internally headed RCs. It is however interesting to note that there seems to be a difference across RC types, and some studies suggest that no Subject advantage or even an opposite Object advantage might hold for internally headed RCs. In the narrow light of these studies, the predictions regarding the Accessibility Hierarchy and the Subject Advantage (if any) when it comes to internally headed RCs are rather blurred. If difficulty arises from structural distance or intervention, there are two possibilities: if what matters is the structural distance between the head noun and the edge of the relative clause where a silent copy is interpreted, we expect to observe a Subject Advantage in internally headed RCs similar to what is observed in externally headed RCs. If on the other hand what matters is that in internally headed RCs there is no gap to fill, we should expect no asymmetry between object RCs and subject RCs.

Additionally, if internally headed RCs are really easier to understand than externally headed RCs, as the studies reported above appear to suggest, this might make the detection of a Subject Advantage more difficult with the former, with participants being too accurate to observe any difference. Yet, on the other hand, the head noun is not easy to identify in internally headed RCs, since it is not singled out by its position. This might involve an extra burden that is not required in the processing of externally headed RCs.

Our study is likely to be able to provide some answers to these questions since both LSC and LIS exploit internally headed RCs. Besides, as we shall see below, Deaf signers include very heterogeneous populations, where differences in AoE might make visible even small processing biases.

3.2 Age of first exposure to a sign language

For the vast majority of hearing infants, language exposure starts already within the womb (see DeCasper & Spence 1986, or Gervain et al. 2008 on that matter) and language deprivation is restricted to exceptional cases of severe child abuse. But for the vast majority of deaf infants born in a hearing family, a delay in language exposure occurs, which is due to an incompatibility between the perceptual abilities of the child and the surrounding language. Even with early intervention through hearing aids or cochlear implants, language access is delayed if not provided from birth in a fully accessible way, which is the visual modality (see Hall et al. 2019 or Humphries et al. 2016 a.o.).

As a result, Deaf⁴ signing communities are very heterogeneous: few individuals have signing parents/relatives and are exposed to a sign language from birth (i.e. they are native signers in the narrow sense), whereas the vast majority, having non-signing parents, are exposed only to a spoken language in their early life and discover sign language later. This group is composed of people with quite different linguistic profiles, since outcomes of exposure to spoken language through hearing aids only are very variable across individuals, and late exposure to a sign language does not ensure full language acquisition (see Humphries et al. 2016).

The long-lasting effects of this delayed acquisition in adulthood have been discussed in a number of studies focusing mainly on American Sign Language (ASL) and British Sign Language (BSL). In 1993, Mayberry reported that native signers are better at recalling complex ASL sentences than late learners and that, more broadly, the age of ASL exposure is correlated to task performance. This result was confirmed in Boudreault & Mayberry’s (2006) experiment, and in its BSL version (Cormier et al. 2012), where participants had a limited time to determine the grammaticality of a set of sentences they were presented with. Stimuli ranged from simple declarative sentences to more complex ones like interrogatives or relative clauses, and participants’ performance decreased as their age of exposure to a sign language increased. At the morpho-syntactic level, Emmorey et al. (1995) highlight the difference in perceiving agreement errors in ASL sentences when comparing native and non-native signers. These results also extend to the morpho-phonological level, with native signers being quicker and more accurate in distinguishing between words and pseudo-words (Emmorey & Corina 1990). It is thus at all levels of the linguistic signal that effects of a delayed sign language exposure can be traced.

These effects have been mainly found in ASL and BSL signers even though their home countries are considered as very advanced in the recognition and diffusion of sign languages: the United States even host the only fully-ASL accessible university, and the United Kingdom is a pioneer in the development and use of sign language in linguistic assessment (see Haug 2005 for a review). This narrow focus makes it all the more important to investigate adult linguistic performance in European sign languages, where deeply rooted public policies favoring oralist education⁵ have largely impeded the recognition of sign languages and their importance in individuals’ cognitive and linguistic development. It is thus essential to investigate the consequences of a delayed exposure to sign language to better inform policy makers and support Deaf people’s efforts in this direction.

4.1 Relative clauses in French Sign Language

LSF allows both SVO and SOV word orders with preference for one or the other varying across individuals (Hauser 2019). It mostly presents externally headed RCs which can be marked in three different ways: through non-manual markers only, with the relative pronoun PI or with the classifier PERSON-CL when the relative clause head is a human entity (Hauser & Geraci 2017; Hauser 2019). All three strategies come with a specific set of non-manual markers which can optionally spread over the head noun as well, glossed as “rel”: raised eyebrows, body and head turned towards the locus of the relative pronoun or the head noun with which it agrees in space (see Figure 1).

Figure 1

Relativization markers in LSF, from Hauser (2019).

In the present study, we focused on the PI-strategy as it is the only strategy that receives an unambiguous restrictive interpretation with any type of head, whether human or not. The sign PI is glossed after its mouthing component since signers co-articulate the syllable [pi] when signing it. It starts with a hand configuration similar to a rounded pinch (F-shape but with non-selected fingers bent) and ends as a pointing sign (similar to the French gesture called “pichenette”) agreeing in space with the head noun. An example of subject and object relative clauses in LSF is given in (16) and (17).

(16)	Subject relative clause, LSF (Hauser 2019: 58)
	‘I prefer the man who pets the dog.’

(17)	Object relative clause, LSF (Hauser 2019: 59)
	‘I prefer the dog which the man pets.’

To avoid variation, all the stimuli we built for our study follow the structure described in the literature, with an SVO order and the relative clause in its base position (Hauser 2019). An example from our dataset is provided for subject relative clauses (see 18 and Figure 2) and for object relative clauses (19 and Figure 3).

Figure 2

Screen captures of the sentence in (18), a subject relative in LSF.

Figure 3

Screen captures of the sentence in (19), an object relative in LSF.

(18)	Subject relative clause, LSF
	‘(You) have to choose the giraffe that licks the cow.’

(19)	Object relative clause, LSF
	‘(You) have to choose the giraffe that the cow licks.’

4.2 Relative clauses in Italian Sign Language

LIS is an SOV language. Its most common relativization strategy is to use a fronted internally headed RC⁶ marked on its right edge by the relative determiner PE (Figure 4). The relative marker PE agrees in location with the head noun. The whole relative clause is marked by a set of non-manual markers glossed as “rel” described as “raised eyebrows, a specific tension of the eyes and upper cheeks, and a forward head nod co-occurring with the sign PE” (Branchini 2014: 189).

Figure 4

The sign PE in LIS, from Hauser (2019).

An example of a subject (20) and an object (21) RC in LIS is provided below.

(20)	Subject relative clause, LIS (Branchini 2014: 191)
	‘I took the dog that eats a lot to the vet.’

(21)	Object relative clause, LIS (Branchini 2014: 192)
	‘Paolo washed the dog that I found yesterday.’

For the present study, in order to avoid variation across stimuli, we generalized the use of a classifier handshape (CL)⁷ that directly follows the object within the RC (see Benedicto & Brentari 2004, for a description of a similar phenomenon in ASL), as this was strongly preferred by most of our LIS informants. Example (22) and Figure 5 illustrate a subject RC and sentence (24) along with Figure 6 present an object RC using this classifier, both examples are taken from our data set.

Figure 5

Screen captures of the sentence in (22), a subject relative clause in LIS.

Figure 6

Screen captures of the sentence (23) in LIS, an object relative clause.

(22)	Subject relative clause, LIS
	‘Click on the giraffe that licks the cow.’

(23)	Object relative clause, LIS
	‘Click on the giraffe that the cow licks.’

4.3 Relative clauses in Catalan Sign Language

Catalan Sign Language (LSC) is also an SOV language having fronted internally headed RCs as the most common strategy for relativization. Here, relatives are signaled either through the use of non-manual markers only, or through the use of a sign glossed as MATEIX⁸ in the literature, meaning ‘same’. This sign is realized on the shoulder with an arc movement (Figure 7). Despite this sign being body-anchored and hence not displaying any agreement, Mosella Sanz (2012) notes that the head noun is signaled through body leaning (i.e. the signer leans in the direction of the head locus). The whole relative clause is marked through raised eyebrows, squinted eyes and slightly raised shoulders (Mosella Sanz 2012).

Figure 7

The sign MATEIX/SAME in LSC, from Barberá (2020).

According to the author, SAME can appear either at the left (see 25, in an object relative clause) periphery of the relative clause or at the right periphery of the relative clause (as in 24, a subject relative clause) without affecting the meaning.

(24)	Subject relative clause, LSC (adapted from Mosella Sanz 2012: 186)
	‘Yesterday we gave onions to the girl who now lives in the city.’

(25)	Object relative clauses, LSC (adapted from Mosella Sanz 2012: 186)
	‘I will put the mushrooms I found in the basket.’

The strategy with the relative marker on the left was preferred by our informants, therefore we adopted it consistently across stimuli. We also minimized variation by always fronting the relative clause before the main clause. An example from our dataset is provided for subject relative clauses (see 26 and Figure 8) and for object relative clauses (27 and Figure 9).

Figure 8

Screen captures of the sentence in (26), a subject relative clause in LSC.

Figure 9

Screen captures of the sentence in (27), an object relative clause in LSC.

(26)	Subject relative clause, LSC
	‘Click on the giraffe that is licking the cow.’

(27)	Object relative clause, LSC
	‘Click on the giraffe that the cow licks.’

5.1 Design and procedure

For each sign language, we developed a sentence-to-picture matching task based on Friedmann & Novogrodsky (2004). In each picture, three characters are displayed: two of them are identical and are either performing an action or undergoing that action with respect to a third different character standing between them (Figure 10).

Figure 10

Three characters, one giraffe licks the cow, and the other one is licked by the cow.

Participants are asked to choose one of the characters based on a sentence containing a relative clause. The head of the relative clause invariably corresponds to one of the matching characters (the giraffes in Figure 10), and the RC is either a subject RC (i.e. ‘Choose the giraffe that licks the cow’) or an object RC (i.e. ‘Choose the giraffe that the cow licks’). Some of the pictures were used as fillers, hence associated with a simpler request not including any relative clause (i.e. ‘Choose the black and white animal’). The function of the fillers was three-fold. 1) As fillers were simple, participants did not need to keep the same level of concentration throughout the whole experiment and could rest. 2) Performance in fillers was used as a baseline measure of attention and overall comprehension. Participants who were lower than 75% correct in the fillers were excluded. 3) Fillers provided some trials targeting the middle character. Since the experimental design tends to bias participants to always click on the left or right characters, half of the fillers were targeting the middle character while the other half were targeting side characters.¹⁰

The protocol had a within-participant design, such that the experiment contained two blocks, administered to each participant in two separate sessions or even on separate days, which were the mirror image of each other: if a picture was presented with a subject RC stimulus in the first block, it was then presented with an object RC stimulus in the second block, so that the same picture was never displayed within the same block. Within a single block, half of the items were subject RCs and the other half were object RCs. The same fillers were used across the two blocks. Items and fillers were fully randomized. All the sentences used in the experiments have been elicited with native informants of the investigated sign language. Non-manual markers typical of relativization strategies in each language (see Section 3) are not indicated in the gloss.

The test always started by a video presenting the instructions in the relevant sign language, followed by a short training phase. At the end of the training phase, participants could ask clarification questions. There were minor variations in the general design and procedures across sign languages that we detail now.

5.1.1 LSF

For the LSF experiment, each block contained 42 trials (14 subject RCs, 14 object RCs, 14 fillers) preceded by 3 training trials – one subject RC, one object RC and one filler. The full list of items is given in the Appendix. The duration of each testing session was around 20-25 minutes and was embedded within a larger testing session (~ 1h30) during which the participants performed other lexical and syntactic tests.

Participants were left alone in the experiment room to perform the test on their own. They were sitting approximately 45 cm away from the screen (4: 3 display, 22” screen) and answered using a mouse by clicking on the selected character.

For each trial, participants saw first the stimulus video, automatically followed by the picture on which to click to provide their answer (see Figure 11). The stimulus video always started by a small context introducing the characters present in the video (see 28), followed by a request embedding a subject RC (28a), or an object RC (28b) or, if it was a filler, a simpler sentence (28c). Notice that the context did not give any indication regarding the spatial disposition of the various characters, while the request embedding the RC displayed spatial location associated to the arguments of the clause. This spatial information did not always match the spatial arrangement of the characters in the picture (meaning that the left character on the picture was not always signed on the left side of the signer) since we did not want participants to be biased in their answers by this parameter. To avoid participants interpreting spatial information topographically nonetheless, the picture was only made visible after the end of the recorded sentence.

Figure 11

Illustration of an item in LSF.

(28)	Context, LSF
	HERE THREE CHARACTER. TWO GIRAFFE ONE COW…
	‘There are three characters, two giraffes and a cow.’
	a.	Subject relative clause
		… HAVE-TO CHOOSE GIRAFFEi [rcPIi___iLICKk COWk].
		‘(You) have to choose the giraffe that licks the cow.’
	b.	Object relative clause
		… HAVE-TO CHOOSE GIRAFFEi [rc PIiCOWk kLICKi___].
		‘(You) have to choose the giraffe that the cow licks.’
	c.	Filler
		… HAVE-TO CHOOSE ANIMAL WHITE BLACK.
		‘(You) have to choose the black and white animal.’11

Experiments took place at the Université de Paris and at the Université de Nantes.¹¹

5.1.2 LIS

For the LIS experiment, each block contained 32 trials (10 subject RCs, 10 object RCs, 12 fillers) preceded by 3 training items (one example of each type of trial). The full list of items is given in the Appendix. The duration of each block was around 15 minutes and it was embedded in a larger testing session (1h) in which participants were undertaking other lexical and syntactic tests.

Participants were sitting in a bright room in front of a laptop with a 15’’ screen and answered using the touch screen. For each trial, participants first saw the stimulus video (no context provided) accompanied by the corresponding small-size picture in the bottom left corner of the screen, and then the picture alone in the middle of the screen on which they could touch the correct character to provide their answer (see Figure 12). In half of the items the arrangement in space of the various arguments in the request sentence matched that of the various characters in the picture; in the other half, there was a clear mismatch. This pattern was adopted to avoid that the simultaneous presentation of a miniature of the picture and the stimulus video induce a topographical interpretation of the use of the signing space in the latter. The stimulus video consisted of a request embedding a subject RC (29a), an object RC (29b) or, if the item was a filler, a simpler sentence (29c) targeting the middle character.

Figure 12

Illustration of an item in LIS.

(29)	LIS
	a.	Subject relative clause
		[RC CHILDk MANjCLj kPUSHj PEk] CLICK
		‘Click on the child that pushes the man.’
	b.	Object relative clause
		[rc CHILDk CLk MANj jPUSHk PEk] CLICK
		‘Click on the child that the man pushes.’
	c.	Filler
		[CL TALLEST] CLICK
		‘Click on the tallest (animal).’

5.1.3 LSC

For the LSC experiment, each block contained 25 trials (10 subject RCs, 10 object RCs, 5 fillers) preceded by 2 trainings (one example of each type of relative clause). The full list of items is given in the Appendix. The duration of each block was around 15 minutes and was embedded in a larger testing session (3h30) in which participants undertook other lexical and syntactic tests.

Participants were left alone in the experiment room to perform the test on their own. They were sitting approximately 45 cm away from the screen (4: 3 display, 22’’ screen) and answered using a mouse by clicking on the selected character.

For each trial, participants first saw the stimulus video (without context), then clicked on an arrow to transition to the answer picture where they could click on the correct character (see Figure 13).

Figure 13

Illustration of a filler in LSC.

To avoid that participants might interpret any spatial information included in the items topographically, i.e. as referring to the actual position of the characters in the picture rather than a grammatical representation arguments’ reference and agreement relation, the picture was only made visible after the end of the recorded sentence. The stimulus video consisted in a request embedding a subject RC (30a) or an object RC (30b). Fillers were either subject RCs and object RCs targeting the middle character or simpler sentences without any embedding like (30c).

(30)	LSC
	a.	Subject relative clause
		[RC SAME ELEPHANTi iSPLASHk BOYk] IX2 CLICK
		‘Click on the elephant that splashes the boy.’
	b.	Object relative clause
		[RC SAME ELEPHANTiBOYk kSPLASHi] IX2 CLICK
		‘Click on the elephant that the boy splashes.’
	c.	Filler
		ANIMAL BLUE IX2 CLICK
		‘Click on the blue animal.’

5.2 Participants

Across all three sign languages, participants were mainly selected based on their hearing status (deafness had to be attested before 3 years of age) and the age at which they were first exposed to a sign language (no later than 15 y.o.). They were recruited either through social media (in France), through various associations for the Deaf in various cities (in Italy) or within the local Deaf community (Catalonia). They received a compensation of 50 euros (France, Catalonia) or 150 euros (Italy) for their participation in the complete battery of tests.¹² We collected metadata regarding their signing background through a questionnaire in sign language (AoE, existence of deaf relatives, modality of communication with parents, siblings, other relatives, type of education whether oralist, bimodal or mostly SL, etc…). The metadata were anonymized.

Additionally, participants had to perform a short, timed, Odd One Out Cognitive Task¹³ to ensure that they were not presenting potential cognitive issues that would affect their performance. This test was made of 2 training items followed by 28 items in which four simultaneous tokens were presented, one of which did not fit the general pattern obeyed by the three others (i.e. an intruder, see Figure 14).

Figure 14

Example of an item within the Odd-one out test.

Among the participants matching our inclusion criteria (i.e. hearing status, AoE and absence of detected cognitive impairment), we created three “language” groups, based on the age of first sign language exposure, what we call their language group:

1. participants with deaf parent(s), exposed to a SL from birth (natives);

2. participants exposed before primary school (early learners, with an AoE ranging between 1 and 5 y.o.) and

3. participants exposed later, between 6 to 15 y.o. (late learners).

   The distribution of signers across the different age groups slightly varies across the three sign languages. They are detailed below.

6.1 Accuracy

The analyses were realized on each sign language independently, but following the same scripts.¹⁴ We had a binary dataset in which correct answers were coded as “1” and incorrect answers as “0”. We started by conducting an item analysis to detect problematic items on which native signers performed below 50% on average. If the score obtained was due to a technical error, we removed the item in both blocks (the subject RC and object RC version of it). If no technical errors could explain the result, the items were kept. In LSF, this led to the removal of four items, hence leaving 48 items (24 subject RC and 24 object RC) to analyze. In LIS, three items were removed, hence leaving 34 items (17 subject RC and 17 object RC) within the analysis. In LSC, only one item had a technical issue such that we performed the analysis on the remaining 38 items (19 subject RC and 19 object RC).

Accuracy and error results were analyzed with the R software (R Core Team 2020) using generalized linear mixed models from the binomial family via the glmer function (package lme4, Bates et al. 2015). The independent variables were language group (native, early & late), condition (subject RC & object RC) and chronological age (as a continuous, mean centered corrected variable). Random variables were intercepts for participants and items.¹⁵

Using ANOVA, we started by comparing minimally different models (df = 1) in which one of the independent variables was removed to control for the overall impact of a given parameter in the data distribution. Keeping only relevant independent variables (more details within each language subsection), we then conducted pairwise comparisons using the package emmeans (Lenth 2020), to obtain finer grained information regarding differences between natives, early, and late learners in their understanding of subject RCs on one side, and of object RCs on the other side.

The results concerning accuracy are presented and analyzed separately for each language here after, error results are presented in Section 5.2.

6.1.1 LSF

Accuracy in LSF is illustrated in Figure 15.

Figure 15

Accuracy in object (blue) and subject relative clause (orange) comprehension in LSF, for native (left, n = 13), early (middle, n = 14) and late (right, n =15) learners. Black crosses represent the mean, and black horizontal lines represent the median of each group. The whiskers extend to the nearest data point that is no more than 1.5 times the Inter Quartile Range from the hinges.

In LSF, models including chronological age as a variable did not provide a better fit of the data, so we removed this parameter from the independent variables of the model (𝜒2(1) = 3.13, p = 0.07699). Language group (𝜒2(2) = 8.06, p = 0.01775), condition (𝜒2(1) = 346.16, p < 0.0001) and the interaction between these two factors (𝜒2(2) = 8.96, p = 0.01134) were significant. All pairwise comparisons using the resulting generalized mixed model across language group (natives, early, late) and condition (SRC, ORC) are presented in Table 1.

Table 1

Pairwise comparison of accuracy across language group (native, early, late) and conditions (SRC, ORC), using generalized mixed model and the binomial family in LSF.

Comparison
By condition	Est.	SE	z	p
Native SRC vs. Native ORC	–3.374	0.544	–6.204	<.0001
Early SRC vs. Early ORC	–2.410	0.286	–8.419	<.0001
Late SRC vs. Late ORC	–3.659	0.337	–10.861	<.0001
Population Natives vs. Early	Est.	SE	z	p
Native SRC vs. Early SRC	1.891	0.766	2.467	0.0136
Native ORC vs. Early ORC	0.927	0.557	1.664	0.0961
Population Natives vs. Late	Est.	SE	z	p
Native SRC vs. Late SRC	1.433	0.790	1.814	0.0697
Native ORC vs. Late ORC	1.719	0.563	3.053	0.0023
Population Early vs. Late	Est.	SE	z	p
Early SRC vs. Late SRC	–0.458	0.631	–0.725	0.4682
Early ORC vs. Late ORC	0.792	0.532	1.489	0.1365

Each line in Table 1 presents the estimate, standard error, z-ratio and p-value associated with the use of the model to compare the two datasets referred to in the corresponding row. For example, the first row displays the value of the parameters obtained when comparing the score realized by native signers in subject relative clauses accuracy to their performance in object relative clauses.

To help the reader better understand how to interpret all these values, significant differences are in bold font.

The table shows that for all three groups the difference between subject RCs and object RCs was significant, such that subject RCs were better understood. In the comprehension of object RCs, natives performed significantly better than late learners. For completeness sake: native signers outperformed early signers in understanding subject RCs.

6.1.2 LIS

Accuracy in LIS is shown in Figure 16.

Figure 16

Accuracy in object (blue) and subject relative clause (orange) comprehension in LIS (n= 39), for native (left, n = 15), early (middle, n = 13) and late (right, n = 11) learners. Black crosses represent the mean, and black horizontal lines represent the median of each group. The whiskers extend to the nearest data point that is no more than 1.5 times the Inter Quartile Range from the hinges.

In LIS, the presence/absence of the factor of chronological age did not impact the general fit of the data (𝜒2(1) = 1.35, p = 0.245). On the other hand, language group (𝜒2(2) = 7.18, p < 0.001) and its interaction (𝜒2(2) = 14.6, p < 0.0001) with condition, significant in itself (𝜒2(1) = 3.30, p < 0.01), were both significant. LIS presents more widespread distributions, which do not allow for the identification of a general Subject Advantage in this language. All pairwise comparisons are presented in Table 2. Significant differences are in bold font.

Table 2

Pairwise comparison of accuracy across language group (native, early, late) and conditions (SRC, ORC), using generalized mixed model and the binomial family in LIS.

Comparison
By condition	Est.	SE	z	p
Native SRC vs. Native ORC	0.321	0.226	1.422	0.1551
Early SRC vs. Early ORC	–0.106	0.205	–0.516	0.6061
Late SRC vs. Late ORC	–0.864	0.220	–3.920	0.0001
Population Natives vs. Early	Est.	SE	z	p
Native SRC vs. Early SRC	0.995	0.430	2.313	0.0207
Native ORC vs. Early ORC	1.422	0.434	3.276	0.0011
Population Natives vs. Late	Est.	SE	z	p
Native SRC vs. Late SRC	0.420	0.441	0.951	0.3414
Native ORC vs. Late ORC	1.605	0.444	3.615	0.0003
Population Early vs. Late	Est.	SE	z	p
Early SRC vs. Late SRC	–0.575	0.447	–1.288	0.1979
Early ORC vs. Late ORC	0.183	0.445	0.411	0.6808

In LIS, we observe that natives significantly outperformed early learners in subject RCs, and that they outperformed both early and late learners in object RCs. The difference between early and late learners was not significant. The only population in which we observed a Subject Advantage is late learners, where we see a significant difference between subject RCs and object RCs.

6.1.3 LSC

In Catalan Sign Language yet another pattern emerged, as can be seen in Figure 17.

Figure 17

Accuracy in object (blue) and subject relative clause (orange) comprehension in LSC, for native (left, n = 12), early (middle, n = 11) and late (right, n = 11) learners. Black crosses represent the mean, and black horizontal lines represent the median of each group. The whiskers extend to the nearest data point that is no more than 1.5 times the Inter Quartile Range from the hinges.

In LSC, the factors of chronological age (𝜒2(1) = 9.23, p < 0.001) and its interaction (𝜒2(1) = 27.9, p < 0.0001) with condition were both significant. Similarly, condition (𝜒2(1) = 388.61, p < 0.0001) and its interaction (𝜒2(2) = 44.46, p < 0.0001) with language group (which, taken alone, is not predictive of the data distribution, 𝜒2(2) = 0.75, p = 0.6861), were significant, as can be seen very clearly in the graph in Figure 17. This language seems to present a strong Subject Advantage that increases with age of exposure to LSC. All pairwise comparisons highlighting the language group by condition interaction are presented in Table 3. Significant differences are in bold font.

Table 3

Pairwise comparison of accuracy across language group (native, early, late) and conditions (SRC, ORC), using generalized mixed model and the binomial family in LSC.

Comparison
By condition	Est.	SE	z	p
Native SRC vs. Native ORC	–1.605	0.250	–6.433	<.0001
Early SRC vs. Early ORC	–2.736	0.267	–10.236	<.0001
Late SRC vs. Late ORC	–4.316	0.361	–11.954	<.0001
Population Natives vs. Early	Est.	SE	z	p
Native SRC vs. Early SRC	–0.368	0.521	–0.707	0.4795
Native ORC vs. Early ORC	0.762	0.502	1.520	0.1285
Population Natives vs. Late	Est.	SE	z	p
Native SRC vs. Late SRC	–1.003	0.554	–1.810	0.0703
Native ORC vs. Late ORC	1.708	0.553	3.088	0.0020
Population Early vs. Late	Est.	SE	z	p
Early SRC vs. Late SRC	–0.634	0.520	–1.219	0.2229
Early ORC vs. Late ORC	0.946	0.520	1.818	0.0691

In LSC, the Subject Advantage was significant across all three groups. When it comes to object RCs, late learners had a significantly lower performance than native learners. The difference between late and early learners only approaches significance, while there is no significant difference between natives and early learners.

Regarding chronological age, it does not interact with language group but it does impact the performance on conditions, as can be seen in the effect plot in Figure 18.¹⁶ The plot shows how the score is affected by the Age variable, which has been scaled and centered, in the object relative clause condition (left panel) and in the subject relative clause condition (right panel). We see that the impact of Age is greater in object RCs than in subject RCs, since the performance is sharply decreasing as age increases.

Figure 18

Plot of the effect of the interaction between Age and Condition on the accuracy within ORC (left) and SRC (right) in LSC. On the x axis, Age has been scaled and centered.

6.2 Error analysis

Remember that internally headed RCs involve a further complexity that is independent from the subject RCs vs object RCs contrast, that of determining the head which is not explicitly marked through a fixed position. The design of our experiment allows us to tease apart these two difficulties. Participants had the choice between three characters. Hence there were two types of possible errors: (1) participants could click on the wrong side character, which meant they correctly identified the head but not its function within the RC (subject RCs vs object RCs); or (2) participants could click on the middle character, which meant they misunderstood which was the head of the RC (e.g interpreting the stimulus ‘the giraffe that the cow is licking’ as ‘the cow that is licking the giraffe’).

Since LSF presents externally headed RCs, the head is always identified by its position on the left of the RC. We thus expect participants to make few errors concerning head identification (few clicks on the middle character), and more errors concerning the gap position, i.e. many clicks on the wrong side characters. This latter type of errors should be more frequent in object RCs due to the Subject Advantage.

In LIS and LSC, on the other hand, since the head is internal, we expect participants to have troubles identifying it and hence they should make more errors clicking the middle character, independently from the subject RCs or object RCs condition. If there is also a Subject Advantage, we expect the distribution of errors concerning only the side character to depend on the subject RCs/object RCs condition, with more errors of this kind in object RCs.

All the analyses have been realized on each sign language independently, following the same procedure. We have created a subset of the data only containing the wrong answers in the object RCs condition. We have a binary dataset in which answers targeting the middle character were coded as “1” and answers targeting the left/right character were coded as “0”. This was our dependent variable.

The independent variables were language group (native, early & late), condition (SRC & ORC), and chronological age. We again controlled for relevant parameters and their interactions by conducting ANOVA analysis across minimally different models. Random variables were intercepts for participants and items, except for LSF whose model only had “items” as a random effect (instead of both “items” and “participants”). The great disparity across participants’ error figures, with some presenting only one datapoint in SRC, did not leave enough data against which to test for an effect of the participant variable. The results concerning error patterns in the two conditions are discussed separately for each language.

6.2.1 LSF

In LSF, models including chronological age as a variable did not provide a better fit of the data, so we removed this parameter from the independent variables of the model (𝜒2(1) = 2.24, p = 0.1341). The effect of condition approached significance (𝜒2(1) = 2.97, p = 0.0846), language group had an impact on the data distribution (𝜒2(2) = 25.05, p < 0.0001) and the interaction between the two parameters was not significant (𝜒2(2) = 3.362, p = 0.1862). We see in Figure 19 that the type of errors made in subject RCs are equally distributed between the head noun and the position of the gap for all three populations. In object RCs however, a different pattern seems to emerge for late signers, who present a larger proportion of mistakes targeting the side character than those targeting the middle character. This difference in pattern is significantly different from native signers (Est. = –1.37, SE = 0.34, p < 0.0001) and from early signers (Est. = –1.08, SE = 0.27, p < 0.0001), while the latter two do not differ between them (Est. = 0.29, SE = 0.33, p = 0.3779).

Figure 19

Proportion of clicks on the competing head noun (middle character), and on the competing gap (left/right characters) in errors for object RC (left panel) and subject RC (right) in LSF. Comparison between natives (left), early (middle) and late (right).

6.2.2 LIS

In LIS, the factor of chronological age did not provide a better fit when integrated in the model (𝜒2(1) = 0.66, p = 0.4164) and was thus removed. Condition (𝜒2(1) = 0.15, p = 0.6966) and language group (𝜒2(2) = 1.98, p = 0.3724), as well as their interaction (𝜒2(2) = 1.82, p = 0.4025) did not had any impact either, confirming what was already visible graphically in Figure 20. As expected, errors target mainly the middle character, the competing head noun (68.,5% of the errors) independently from the conditions or the language groups.

Figure 20

Proportion of clicks on the competing head noun (middle character), and on the competing condition (left/right) in errors for object RCs (left panel) and subject RCs (right) in LIS. Comparison between natives (left), early (middle) and late (left) learners.

6.2.3 LSC

In LSC, the factor of chronological age did not play a role in the data distribution of error types (𝜒2(1) = 1.38, p = 0.2393), similarly, the factor of language group (𝜒2(2) = 3.28, p = 0.2194) and its interaction with condition (𝜒2(2) = 2.15, p = 0.3407) did not affect the fit. We thus only kept the factor of condition as an independent variable, since it was the only parameter affecting the data distribution significantly (𝜒2(1) = 26.01, p < 0.0001). With the resulting model, we obtain a significant difference between subject RCs and object RCs (Est. = 1.46, SE = 0.3, p < 0.0001).

When we look at Figure 21, we indeed see that, in subject RCs, all participants (native, early and late) patterned like LIS participants: most errors were of head identification (the middle character represents 78% of errors in subject RCs).

Figure 21

Proportion of clicks on the competing head noun (middle character), and on the competing condition (left/right) in errors for object RCs (left panel) and subject RCs (right) in LSC. Comparison between natives (left), early (middle) and late (right) learners.

In object RCs, however, a difference between natives on the one side and early learners on the other side arose. Native learners present a pattern that is very similar to what we observed with subject RCs: around 62% or errors were of head identification (middle character). Early learners, on the other hand, present a pattern that is similar to what is observed in LSF, with most errors regarding the identification of the function of the head (60% of errors targeting the side characters). The difference between Native ORC error pattern and Early ORC error pattern is significant (Est. = –1.39, SE = 0.65, p = 0.0341).

7.1 Does the Subject Advantage in RCs comprehension extend across modalities?

The answer to this question is straightforward. We found that externally headed relative clauses present a Subject Advantage across modalities. LSF behaves exactly like other languages that display externally headed relatives, exhibiting a strong Subject Advantage, hence confirming the results reported by Hauser & Pozniak (2019). The Subject Advantage interacts with language group: in LSF, even if it is observable across groups, it is strongest in late learners. In LSC the Subject Advantage is also present in all three groups, and the strength of the effect increases with the age of exposure. In LIS, the Subject Advantage is significant only in the late group.

7.2 Does delayed exposure to (sign) language impact RC comprehension in adult signers?

The answer to this question is also straightforward: the Subject Advantage is modulated by age of exposure. As we summarized in 6.1, the later the exposure, the stronger the Subject Advantage. That difficult structures are more affected by age of exposure is in line with the literature (Emmorey & Corina 1990; Mayberry 1993; Emmorey et al. 1995; Boudreault & Mayberry 2006; Cormier et al. 2012) and with new further evidence by our research group (Authors, under review).

7.3 Does the Subject Advantage emerge across relativization strategies?

As for the third research question concerning internally headed RCs, the pattern of results that we found requires a detailed discussion. As expected, we found that it is more difficult to identify the head in internally headed RCs than in externally headed RCs. In LSC and LIS, more errors were made regarding the identification of the head noun (i.e. participants hesitated between the correct head noun and the middle character), while the opposite pattern was found in LSF, where signers made more mistakes of gap identification, hence targeting the wrong peripheral character.

The result that was most difficult to anticipate (and potentially the most interesting) is the existence of a Subject Advantage in the comprehension of internally headed RCs, which is observable both in LIS and in LSC although to a different degree. In LIS, this Subject Advantage is statistically significant only in late learners, while in LSC it is visible in all three acquisition groups. We think that this result can contribute indirect evidence both on the debate about the source of the Subject Advantage in RCs in general, and for the syntactic analysis to be given to internally headed RCs.

As for the debate about its source, the fact that the Subject Advantage is attested despite the absence of a gap in internally headed RCs is a powerful argument against all approaches that explain the Subject Advantage in languages like English as a reflex of shorter retention of the subject gap in the memory buffer given its vicinity to the relative clause head. Similarly, the presence of a Subject Advantage in internally headed RCs is also a challenge for any explanation that reduces it to a canonicity effect. Swinney, Prather & Love (1996) and Sekerina (2003) suggest that the language processor employs a canonical sentence schema to interpret sentences. Thus, structures in a language’s canonical word order (like subject relatives in English) are assumed to be easier to process than structures with a non-canonical order (like object relatives in English or in LSF). Crucially, both subject and object RCs follow the canonical word order in LIS and LSC and, this notwithstanding, a difference emerges between them. Therefore, the Subject Advantage is not likely to be a canonicity effect.

The type of explanation that seems to be best suited to explain the Subject Advantage in internally headed RCs is the account that assumes that (i) there is a processing cost when the subject intervenes hierarchically between the object and its gap (cf. Friedmann et al. 2009 for this explanation of the Subject Advantage) and (ii) the gap that is not visible in overt syntax is created by covert movement of the RC head (cf. Cole, 1987 a.o. for this approach to internally headed RCs).

We acknowledge that with internally headed RCs, namely in LIS and LSC, the Subject Advantage was especially (or only) significant in late learners or non-native populations. Remember that our experiment is an off-line sentence-to-picture matching task and we measured only accuracy, therefore we cannot exclude that a Subject Advantage could emerge even with native LIS and LSC signers with a more sensitive on-line task. Still, it is true that the same off-line task revealed in LSF a Subject Advantage in all groups of signers, including natives. This crosslinguistic difference may suggest that the effect is stronger in externally headed RCs than in internally headed RCs. Notice that this is coherent with the acquisition pattern observed in Korean, where internally headed RCs emerge earlier than externally headed RCs and might perhaps explain the Object Advantage observed in Wenzhounese (Hu Shenai et al. 2018). The modulation of the Subject Advantage might also indicate that the two factors mentioned above (canonicity effects and presence versus absence of an overt gap) might play a secondary role, adding to the main effect due to hierarchical subject intervention in object dependencies.

As we mentioned, the Subject Advantage in internally headed RCs provides indirect evidence in favor of a certain kind of analysis of this structure, one that assumes a covert movement of the head. This has an important consequence: if we are right, the Subject Advantage, an effect observable in comprehension that is easily measurable with experimental methods, can be used as a reliable, if indirect, diagnostic for LF movement. As such, it could be used in addition to islands effects (traditionally employed to argue for a covert movement analysis of wh-in situ at least since Huang Charles 1982) in all those cases where a covert dependency is argued for and needs to be detected (cf. also Xiang Ming et al. 2014; 2015 for experiments aiming at other direct processing evidence for this covert dependency). This is an interesting case of cross-modal interaction in which a diagnostic detected in a sign language can be extended to the analysis of spoken languages having the same abstract configuration.

7.4 Comparing LIS and LSC

LSC and LIS did not yield identical results while both displaying internally headed RCs, and both exhibiting a Subject Advantage. In LIS only late learners display a significant Subject Advantage, while native and early learners appear to have a good comprehension of both subject and object RCs. In LSC, the Subject Advantage is visible across all signers, with a clear distinction between native and non-native comprehension of object relative clauses.

We might wonder here about the source of this difference. There are two possible ways to go. One is related to a methodological difference between the two experiments and the other concerns a grammatical difference between the two languages.

Comparing the design between the two experiments, we observe that, while in the LIS experiment, the sentence was shown together with a miniature of the response picture, this was never the case in LSC, where the sentence was shown before the picture. It could be the case that the simultaneous exposure to the picture and the sentence induce LIS participants to a direct (non-structurally mediated) mapping between position of the sign in the neutral space and position of the corresponding character in the picture. Notice however that this is not likely, because the LIS team controlled the possible mapping between spatial organization of the signs in the sentence and that of the characters in the picture: while in half of the items, the loci associated to the arguments in the sentence indeed matched the positions associated with the characters in the response picture, the other half of the items displayed a systematic mismatch between grammatical space and the disposition of the characters in the picture. Had the participants relied on such a topographical mapping, they would have made many more mistakes in general, and crucially more than in the LSC experiment, where any topographical mapping was discouraged by the non-simultaneous display of the sentence and the picture. This means that the explanation of the difference between LSC and LIS data has to be found elsewhere.

Turning to the linguistic side, the internal make up of internally headed RCs in the two languages is different, in a way that might contribute to explaining the differences we observe in comprehension data.

In LIS, the relative marker PE obligatorily agrees in space with the internal head in LIS, as shown in the object relative clause (23) visually represented in Figure 6, repeated below as Figure 22.

Figure 22

Screen captures of the sentence (23), an object relative clause in LIS.

In Figure 22 the relative PE points towards the locus in space associated with the object classifier co-referring with the giraffe, which is the (topicalized) head of the RC. Participants are likely to rely on this spatial agreement information to interpret the RC.

No such explicit spatial agreement is displayed in LSC relative clauses, where the relative marker MATEIX (‘SAME’) is articulated on the body. As can be observed in Figure 9, repeated here as Figure 23, corresponding to the object RC (27), the association between the relative marker SAME and the head (GIRAFFE in the example) is only signaled through a light head tilt and body turn.

Figure 23

screen capture of the sentence in (27), an object relative clause in LSC.

On the other hand, in this example the relative element SAME sits close to the (topicalized) head of the RC, and participants are likely to use the linear proximity to identify the head of the RC. This cue is however less reliable than spatial agreement in LIS since the position of SAME is not systematically adjacent to the RC head in LSC grammar (Mosella Sanz 2012), even if this was its position in our data set. We speculate that this difference in the grammar of the two languages, and in particular the absence of a grammatical marker of the head, like spatial agreement, explains why the Subject Advantage is stronger in LSC: in the absence of a univocal cue identifying the head of the RC, participants are more prone to follow their bias towards a subject interpretation.

If we look now closer to LSC, we find that the Subject Advantage is so strong as to take the shape of a categorical difference between the grammar of native signers and that of non-native signers who systematically misunderstand object relative clauses.

The accuracy and error data that we observed are compatible with the hypothesis that native signers have both subject RCs and object RCs in their grammar while non-native signers do not allow object RCs at all in LSC. This configuration offers yet another illustration of the Accessibility Hierarchy and its Primary Relativization Constraint corollary (Keenan & Comrie 1977), which states that:

1. A language must have a primary RC-forming strategy.

2. If a primary strategy in a given language can apply to a low position on the Accessibility Hierarchy, then it can apply to all higher positions.

3. A primary strategy may cut off at any point on the Accessibility Hierarchy.

The LSC situation appears to comply with the Primary Relativization Constraint: (1) is verified since all signers of LSC understand subject RCs; (2) is verified since native signers, who understand object RCs, also understand subject RCs and (3) points towards two different grammars between native and non-native signers: the native grammar cuts off the Accessibility Hierarchy after object RCs, while the non-native grammar cuts it off right after SRCs. The fact that different varieties of the same language realize different steps of the Accessibility Hierarchy should not come as a surprise given the exceptional circumstances of access to language experienced by a large part of the deaf population. In fact, this finding, which replicates language internally the conclusion based on the typological literature, appears as an extreme case of AoE effect.