J Psycholinguist Res (2011) 40:137–154 DOI 10.1007/s10936-010-9160-0
Disambiguating Information and Memory Resources in Children’s Processing of Italian Relative Clauses Fabrizio Arosio · Maria Teresa Guasti · Natale Stucchi
Published online: 9 November 2010 © Springer Science+Business Media, LLC 2010
Abstract We investigated the role of number agreement on verb and of animacy in the comprehension of subject and object relative clauses in 51 monolingual Italian-speaking children, mean age 9:33, tested through a self-paced listening experiment with a final comprehension question. A digit span test and a listening span test were also administered to examine the role of memory in comprehension. Subject relative clauses were easier to comprehend than object relative clauses; animacy of the relative clause head improved comprehension of object relative clauses; memory, as measured by the digit span test, modulates comprehension of object relative clauses, both with animate and inanimate heads, as shown in response accuracy. Although all children process number agreement morphology on the verb, only some perform a correct reanalysis, as shown by the accuracy measures. We argue that number agreement disambiguation is particularly taxing for children, as it provides a negative symptom in the sense of Fodor and Inoue (J Psycholinguist Res 29(1):25–36, 2000) and reanalysis requires them to hold two dependencies in memory. Keywords Relative clauses · Animacy · Agreement disambiguation · Digit span · Diagnosis and repair model · Self-paced listening Introduction Children’s comprehension of relative clauses has been extensively studied mainly using offline methods. By contrast, the ongoing comprehension process has been mostly investigated in adults, with only recent research changing this trend (Booth et al. 2000; Felser et al. 2003; Love 2007; McKee et al. 1993; Roberts et al. 2007; Sekerina et al. 2004; Traxler et al. 2002). Very little is known so far about how the comprehension process unfolds in children and which features they use in assigning structural representations to incoming material even though this information is crucial to deepen our knowledge of the acquisition process. In this
F. Arosio (B) · M. T. Guasti · N. Stucchi Department of Psychology, Università degli Studi di Milano-Bicocca, P.zza dell’Ateneo Nuovo 1, 20126 Milano, Italy e-mail:
[email protected]
123
138
J Psycholinguist Res (2011) 40:137–154
paper, we aim at investigating how children process subject and object relative clauses in Italian, focusing on the features that contribute to local disambiguation. Like adults (e.g., Frauenfelder et al. 1980 for French; King and Kutas 1995, for English; Schriefers et al. 1995 for Dutch), children find object relative clauses (RCs) more difficult to comprehend than subject RCs (Booth et al. 2000; Friedmann and Novogrodsky 2004; De Villiers et al. 1979). This asymmetry is likely to arise from the way sentence processing in children and adults deals with temporarily ambiguous sentences, such as RCs. Syntactic based processing theories explain the subject advantage in terms of economy principles of gap prediction that drive the analysis of filler-gap dependencies (De Vincenzi 1991; Frazier and Flores d’Arcais 1989). According to these theories, shorter dependencies are computationally less demanding than longer ones, since the filler has to be maintained active in memory for a shorter period of time. Therefore, when the parser sees a relative pronoun following an NP, he/she postulates an RC with the gap in the embedded subject position as in (1), in agreement with the Active Filler Hypothesis (Frazier and Flores d’Arcais 1989) or with the Minimal Chain Principle (De Vincenzi 1991). When the incoming input after the relative pronoun consists of a verb, like in (1), it will be easily integrated in the already parsed structure; on the contrary, when the incoming input consists of an NP, like in (2), a reanalysis must be started that results in a revision of the parsed structure into an object RC with the gap in the object position. The detection of the temporary incongruity or ungrammaticality with respect to the previously postulated structural analysis and the revision of the structure into an object RC is computationally costly; in addition, the filler gap distance in object RCs is structurally longer than in subject RCs and therefore computationally more demanding. These facts explain why object RCs are more difficult to process than subject RCs (see MacWhinney and Pléh 1988; Gibson 1998 for alternative proposals). (1) The woman [who/that twho admires the clown]RC SUBJECT (2) The woman [who/that the clown admires twho ]RC OBJECT Object RCs are not all equally difficult to process and this might depend on the nature of the disambiguating features. In German RCs, the finite embedded verb always occurs in the clause final position and the linear position of the embedded NP is compatible with both a subject and an object interpretation. To unambiguously obtain either of the two interpretations, German speakers might make use of the Case morphology on the embedded NP, as in (3) and (4), and/or of the agreement morphology on the embedded final verb, as in (5) and (6): German Subject–Object Disambiguation Through Case (3) Die Frau [die den Clown sieht]RC The woman [who theACC clown watches]RC ‘The woman who is watching the clown’
SUBJECT RC
(4) Die Frau [die der Clown sieht ]RC The woman [who theNOM clown watches]RC ‘The woman who the clown is watching’
OBJECT RC
German Subject–Object Disambiguation Through Agreement (5) Die Frau [die die Kinder sieht]RC The woman [who the children watch3SING ]RC ‘The woman who is watching the children’
123
SUBJECT RC
J Psycholinguist Res (2011) 40:137–154
(6) Die Frau [die die Kinder sehen]RC The woman [who the children watch3PL ]RC ‘The woman who the children are watching’
139
OBJECT RC
In (3) and (4), the head NP, the embedded NP and the embedded verb “sieht” share the same singular number features and only the Case morphology on the embedded NP disambiguates their interpretation. In (3), the accusative Case morphology on the definite article “den” says that “den Clown” is the embedded object and the sentence is a subject RC; in (4), the presence of the nominative Case morphology on the definite article “der” says that “der Clown” is the embedded subject and thus that the sentence is an object RC. In (5) and (6), the plural article “die” introducing the plural noun “Kinder” is ambiguous between nominative and accusative Case and only the agreement morphology on the final verb disambiguates the sentence. While (5) is a subject RC, since the embedded verb and the head of the RC “die Frau” (but not the embedded NP “die Kinder”) share the same number features, (6) is an object RC, since the embedded verb and the embedded NP “die Kinder” (and not the head-NP “die Frau”) share the same number features. A number of studies with German adults have shown that object RCs disambiguated by Case are easier to process than object RCs disambiguated by agreement (Friederici 1997; Friederici et al. 1998; Meng and Bader 2000; see also Bader and Bayer 2006). Other studies with adults (Mak et al. 2002, for Dutch; Traxler et al. 2002, for English) have shown that the difficulties associated with object RC processing disappear when the head NP denotes an inanimate entity indicating that animacy is another feature that modulates comprehension. The role of animacy of the head NP in RC processing is particularly interesting in light of the fact that pragmatic plausibility facilitates object RC processing only slightly, as shown by Traxler et al. (2002). In this study, where participants were presented with object RCs with only one of the NPs denoting a pragmatically plausible agent for the event described, this facilitation did not improve comprehension (see also Gordon et al. 2001 for the role of NP-type in the comprehension of RCs; Mak et al. (2008) for the role of discourse topic). The impact of different disambiguating features on the comprehension of RCs by children has been examined with off-line tasks by Arosio et al. (2009) and Adani (2008) based on Italian, and Guasti et al. (2007) based on Greek. Arosio et al. investigated the comprehension of subject and object RCs with a Picture Selection Task in Italian-speaking children from 5 to 11 years. The experimental data from this study show that Italian subject RCs, like in (7) are generally easier than object RCs. As in Italian, there are two types of object RCs, one with a word order analogous to English, like in (8) and, another with the subject in a postverbal position, like in (9), it is possible to determine the impact of different disambiguation features. In fact, while it is the position of the embedded NP that is responsible for the object reading in (8), it is the number agreement on the embedded verb that is responsible for the object reading in (9), as in the German (6). (7) Il ragazzo che guarda i pagliacci The boy that watch3SING the clowns ‘The boy who is watching the clowns’ (8) Il ragazzo che il pagliaccio guarda The boy that the clown watch3SING
SUBJECT RC
OBJECT RC: POSITION DISAMBIGUATION
‘The boy who the clown is watching’
123
140
J Psycholinguist Res (2011) 40:137–154
(9) Il ragazzo che guardano i pagliacci The boy that watch3PL the clowns
OBJECT RC: AGREEMENT DISAMBIGUATION
‘The boy who the clowns are watching’
Arosio et al. (2009) found that the comprehension of object RCs disambiguated through the position of the embedded subject (as in (8)) is already good at the age of 5 years, but that the comprehension of object RCs disambiguated through the morphology of the embedded verb as in (9) only improves at the age of 9 years. This result holds notwithstanding the fact that Italian children are sensitive to subject–verb agreement violations and produce object RCs with a post-verbal subject at the age of 5 years (Guasti and Cardinaletti 2003). One explanation for these facts might be that while children are sensitive to verb agreement in simple sentences, they may fail to process it in RCs and this would explain their rather poor performance up to the age of 9 years (Wittek and Tomasello 2005). An important question related to this conjecture is whether the verb agreement information in (9) is accessible or not to children. In our study we want to answer this question. In fact, an alternative explanation might be that, though number information is accessible to children, it is not effective in triggering the revision of the preferred subject interpretation. According to this explanation both the position of the embedded NP and the number marking on the embedded verb convey accessible information for the revision of the preferred subject interpretation, but the former is more effective than the latter. The question is why in this case such an asymmetry should hold. One possibility is that limitations in the memory resources affect the process of reanalysis, but as a function of the nature of the disambiguating information. For the purposes of this paper, we will assume Baddeley’s influential model of the memory system (1986; Baddeley and Hitch 1974; Baddeley et al. 2009). According to this model, information is temporary stored and manipulated in a system (the Working Memory system, WM, henceforth) with a limited capacity that, among other things, includes (i) the phonological loop, consisting of a phonological cache that stores phonological information as memory traces for a few seconds and an articulatory rehearsal component that helps in reviving the memory traces, (ii) the central executive, a control mechanism that coordinates operations on the information stored in the phonological loop. According to this model, the phonological loop is a WM subsystem capable of storing limited amounts of verbal information over a short interval. The reading span task developed by Daneman and Carpenter (1980) is aimed at measuring working memory resources. In this task, subjects are asked to read lists of an increasing number of sentences and, for each sentence in a list, they are asked to make a judgment about the sentence and at the end of the list to recall the last word of each sentence. The digit span test is a measure of verbal short term memory resources, i.e., the phonological loop capability of storing limited amounts of verbal information over a short interval. It requires subjects to store a sequence of digits of increasing length in the phonological loop and to repeat it. Returning to the relation of memory and RC comprehension, experimental findings with adults (King and Just 1991) have shown that individual working memory differences, as measured by the reading span task, modulate RC processing in word by word reading tasks. In particular, the results show that high-reading-span adults are faster than low-reading-span adults at the main verb in object RCs, as in (11), while no statistically significant difference is found for subject RCs, as in (10). (10) The reporter that attacked the senator admitted the error (11) The reporter that the senator attacked admitted the error
123
SUBJECT OBJECT
J Psycholinguist Res (2011) 40:137–154
141
Using the ERP method, Bornkessel et al. (2001) and Vos et al. (2001) found different patterns of ERP responses in high-reading-span and low-reading-span German adults. While high span readers were more efficient, performed reanalysis and were not influenced by the context, low span readers did not show any sign of reanalysis and attempted to comprehend the sentences by relying on contextual information. Similarly to adults, memory differences in children modulate RC processing. Booth et al. (2000) investigated children’s comprehension of RCs and its relation to children’s memory by making use of a word by word reading task and a word by word listening task, both including a comprehension question, and two memory tasks, a reading span task that was an adaptation of Daneman and Carpenter (1980) task and a digit span test. Booth and colleagues found that digit span measures, but not reading span measures, interacted with reading and listening times and with accuracy measures; children with a high digit span had longer reading and listening times and were more accurate than children with a low digit span in the transition between the RC and the main clause. Felser et al. (2003), however, found that children’s use of different parsing strategies was modulated by their low or high working memory span as measured by a listening span test (Gaulin and Campbell 1994). Finally, Arosio et al. (2009) using a Backward Repetition Span Test (Ciccarelli 1998), where children were asked to listen to lists of an increasing number of words and to repeat them in the reverse order, found a correlation between backward repetition span measures and accuracy scores in the picture selection task mentioned above. These findings corroborate the hypothesis that processing strategies in children are motivated by limitations of computational resources like in adults and support the idea that processing differences between adults and children could depend on the different resources available to these two populations.
Our Study Arosio et al. (2009) study shows that, for children, RCs disambiguated by number agreement (as in (9)) are harder than those disambiguated by the position of the embedded subject (as in (8)). However, from that study we do not know what happens when children encounter the disambiguating information, i.e., the number agreement morphology on the verb. Do they ignore it or do they process it, but fail to revise the preferred analysis? In order to gain some insight into these questions, we carried out a Self-Paced Listening Task (Booth et al. 2000; Felser et al. 2003; Ferreira et al. 1996; Ferreira and Anes 1994), in which children self administered, segment by segment, RCs disambiguated by number agreement morphology on the verb. The core assumption of self-paced listening techniques is that, listening times of phrases presented one by one reflect how long it takes to integrate them into an ongoing structural representation, and they are longer when this integration is more difficult. Though we are aware of the fact that by making use of this technique we address processing routines of rather unnatural linguistic input, since sentences are broken into segments and their prosodic frame is disrupted, this technique allows us to measure the integration cost of a morphosyntactic manipulation of segmented input in terms of different listening times. We believe that this may help us in understanding how this manipulation modulates RC comprehension processes in children. Additionally, in our study, we aimed at understanding how different features impact on comprehension processes and therefore we investigated the role of animacy, assuming that
123
142
J Psycholinguist Res (2011) 40:137–154
animacy is a feature, like gender, attached to a noun in the lexicon.1 We saw that animacy is an influential factor in adults’ RC processing (Mak et al. 2002; Traxler et al. 2002). We know that animacy information plays an important role in the acquisition of the distinction between raising and control verbs in children as young as 3 or 4 (Becker 2006). Then, we would expect it to influence RC comprehension in our 9 year old children as well. A third question concerns the role of memory in sentence comprehension, as the availability of memory resources is often advocated to explain why children display worse performance than adults on certain tasks and as the impact of different features is likely to depend on the memory resources that their processing requires. Participants, Procedure and Materials In our study, we tested a group of 51 monolingual Italian-speaking children, mean age 9;33, SD = 0.5, in (i) a self-paced listening experiment with a final comprehension question in order to investigate children’s processing and off-line comprehension of RCs, (ii) a digit span test in order to investigate the relationship between verbal short-term memory resources and the comprehension of RCs and (iii) a listening span test (Palladino 2005; this is an adaptation for Italian of Daneman and Carpenter (1980) task) to study the relationship between working memory resources and RC comprehension. Children were recruited from schools in Milan and Como (Italy) and were tested individually in a quiet room at their schools during two sessions in the morning; in one session we administered the self-paced listening task and in the second we administered the memory tests. The second session took place two/three days after the first and the children’s general behaviour and attention to the tasks was consistent in the two sessions. Self-Paced Listening Experiment The self-paced listening experiment is modelled after Mak et al. (2002) study with adults and has a 2 × 2 design with factors Sentence Type (subject RC vs. object RC) and Animacy (animacy vs. inanimacy of the object NP) with digit span test measures or listening span test measures as covariates. We constructed 24 sets of four sentences like in (12). As we can see, in condition (12a) we have a subject RC with an animate embedded object; in condition (12b) we have a subject RC with an inanimate embedded object; in condition (12c) we have an object RC with an animate object (the head-NP) and in condition (12d) we have an object RC with an inanimate object (the head-NP). All sentences contain an embedded post-verbal NP (either an object or a subject) and are disambiguated towards an object or a subject RC by the number morphology on the embedded auxiliary that agrees with the head NP, but not with 1 Evidence for the hypothesis that animacy of the NP might be a grammatical feature visible to morphosyn-
tactic operations comes from languages in which agreement between the verb and its subject and/or its interval argument are modulated by animacy. For example, in the Georgian examples below the verb agrees with its subject when this denotes an animate entity, as in (i), but not when this denotes an inanimate entity as in (ii) (sentences are from Harris 1981:149). (i)
Knutebi goraven Kittens roll−3PLUR ‘The kittens are rolling’
(ii)
Burtebi goravs Balls roll−3SING ‘The balls are rolling’
123
J Psycholinguist Res (2011) 40:137–154
143
the embedded postverbal NP in the case of subject RCs, or that agrees with the embedded postverbal NP, but not with the head NP in the case of object RCs. (12) a. SUBJECT RCs, ANIMATE OBJECT (SA) Il pasticcere osserva i gatti che stanno rincorrendo il topolino The baker watches the cats that are chasing the mouse ‘The baker watches the cats that are chasing the mouse’ b. SUBJECT RCs INANIMATE OBJECT (SI) Il pasticcere osserva i gatti che stanno rincorrendo il gomitolo The baker watches the cats that are chasing the ball-of-yarn ‘The baker watches the cats that are chasing the ball of yarn’ c. OBJECT RCs ANIMATE OBJECT (OA) Il pasticcere osserva il topolino che stanno rincorrendo i gatti The baker watches the mouse that are chasing the cats ‘The baker watches the mouse that the cats are chasing’ d. OBJECT RCs INANIMATE OBJECT (OI) Il pasticcere osserva il gomitolo che stanno rincorrendo i gatti The baker watches the ball-of-yarn that are chasing the cats ‘The baker watches the ball of yarn that the cats are chasing’ The 4 sentences of each set were divided into four lists such that each list included 6 items for each of the 4 conditions. Items in the lists were randomly ordered and interspaced by 24 filler sentences and six familiarization sentences were added at the beginning of each list. Consequently, each list included 6 familiarization sentences, 24 filler sentences and 24 experimental items, for a total of 54 trials. In our study, we had to administer the self-paced listening task in one session, (we were permitted to test children for only two sessions and in one of them we administered the memory tests), and therefore we could not add more items for condition in our design, since this would have made the task longer and inappropriate to the short lasting attention resources of children. Sentences were segmented and segments were acoustically recorded independently, so that prosodic or coarticulatory cues were absent, and then reassembled. An example of sentence segmentation is given below where the slashes indicate segmentation boundaries: /
S1 NP+VP
/
S2
/ S3 / S4
NP
that AUX
/
S5 PP
/
S6
/
NP
(13) / Il pasticcere osserva / il topolino / che / stanno / rincorrendo / i gatti / The baker watches the mouse that are chasing the cats ‘The baker watches the mouse that the cats are chasing’ Children self administered sentences in the auditory modality, segment after segment, by pressing the space bar of the keyboard of a personal computer administering the stimuli (see Ferreira et al. 1996). Stimuli were delivered through loudspeakers connected to the personal computer. In order to prevent subjects from pressing the response key before they had heard the whole segment, the next segment was administered only when the response key was pressed after the end of the previous segment. If the subjects had pressed the response key before the end of the segment, they had to press it again at the end of the segment. Listening times (henceforth, LTs) were recorded from the end of the segments. LTs for segments
123
144
J Psycholinguist Res (2011) 40:137–154
where the subjects had previously pressed the response key before the end of a segment were tagged and substituted in the analysis (see below). We wrote our experiment in Matlab, using Psychophysics Toolbox extensions (Brainard 1997; Pelli 1997). Participants were instructed to listen carefully to each sentence segment and to move from one sentence segment to the next one by pressing the space bar of the keyboard and to answer a comprehension question asking who was the patient/theme of the described event; the end of the sentence was signalled by an acoustic tone. Since in Italian wh-questions may be ambiguous between subject and object interpretation, the comprehension question was asked by an interrogative passive unambiguous sentence. A comprehension question for sentences (13) is given below (14) Chi viene rincorso? ‘Who is chased?’ In order to facilitate the task, immediately after having heard the comprehension question, children saw two pictures on the computer monitor representing the intended patient/theme and agent of the described event. Children answered the comprehension question by naming the patient/theme of the described event or by pointing to the picture representing it. While children went through the self-paced task, the computer monitor was blank. Digit Span Test and Listening Span Test A standard auditory digit span test was administered to children to measure verbal short-term memory resources. Series of digits of increasing length, from 3 to 9 digits, were randomly ordered in lists. These were digitally recorded in audio files by an Italian native speaker and played through loudspeakers connected to a personal computer administering the test at a rate of about one digit per second. Children were asked to listen carefully to the lists of digits and repeat them aloud in the correct order. The test was stopped when children missed 3 out of 3 trials within one level. A maximum of one minute was allowed for the recall of a list. We also administered a listening span test aimed at measuring working memory resources. In this test, children were presented with a series of an increasing number of sentences previously recorded by a female voice. Sentences were presented through loudspeakers connected to a computer and presentation was controlled by the experimenter. Children had to answer a question about the sentence heard and to recall as many sentence final words as possible. The test was dismissed when children missed 3 out of 3 trials within one level.
Results We conducted two types of analyses: analysis of off-line accuracy of responses to the comprehension question presented after each sentence and analysis of listening times. In both cases, we employed a mixed model analysis (Little et al. 1996) because it allowed us to carry out a statistical analysis equivalent to a within-subject ANOVA, suited for our experimental design, with the added possibility of having a covariate that changed within participants. We performed both subject, F1, and item analyses, F2. Accuracy Results Children’s responses as a function of their digit span (henceforth d-span) are depicted in Fig. 1. As we can see from this figure, subject RCs are responded to very accurately, regardless of children’s d-span (where d-span ranged from 4 to 6 with a M = 5.24 and SD = 0.62).
123
J Psycholinguist Res (2011) 40:137–154
145
Fig. 1 Response accuracy for each type of sentence (SA subject RC with animate object; SI subject RC with inanimate object; OA object RC with animate object; OI object RC with inanimate object) calculated over subjects for the three levels of d-span (Span 4 N = 14; Span 5 N = 22; Span 6 N = 15). Vertical bars indicate the standard error at the 0.95% level of confidence
This is not the case for object RCs. Therefore, an effect of Sentence Type is expected. Object RCs with inanimate objects are comprehended better than object RCs with animate objects. Hence, an animacy effect is also expected. Finally, children with d-span equal to 4 perform less accurately than the others, both in object RCs with animate objects and in object RCs with inanimate objects, although they are more accurate in the latter case than in the former. Therefore, we expect an effect of d-span due to children with d-span 4 versus the other two groups. These findings are supported by the statistical analyses. First, we compared the response accuracy among the four lists. As no difference was found, F1(1, 3) = 0.91, Ms = 0.08, p = 0.43, F2(1, 5) = 0.52, Ms = 0.15, p = 0.67, we conflated all the data together for further analyses. Then, we analyzed the factors Sentence Type (subject RC and object RC) and Animacy (animate head NP and Inanimate head NP) with d-span as a covariate and we found a main effect of Sentence Type significant both in the subject, F1(1, 50) = 65.42, Ms = 0.148, p < 0.001, and in the item analysis, F2(1, 5) = 29.92, Ms = 0.32, p < .001, with subject RCs being more accurately answered than object RCs. We also found an animacy effect significant both in the subject, F1(1, 50) = 77, 76, Ms = 0.062 p < 0.001, and in the item analysis, F2(1, 5) = 29, 92, Ms = 0.32, p < 0.01, with inanimate RCs being answered more accurately than animate RCs, and a d-span effect, significant both in the subject, F1(1, 49) = 12.69, Ms = 0.141, p < 0.001, and in the item analysis, F2(1, 1204) = 26.48, Ms = 0.06, p < 0.001, likely due to children with d-span 4 who responded less accurately to object RCs. We also found an interaction between the factors sentence and animacy significant both in the subject, F1(1, 50) = 71.89, Ms = 0.063 p < 0.001, and in the item analysis, F2(1, 5) = 25.26, Ms = 0.18, p < 0.01. This interaction is clearly due to object RCs with inanimate heads which are answered more accurately than object RCs with animate heads. To unpack the interaction,we analyzed animate and inanimate RCs separately with d-span as a covariate. While no effect was found in the case of subject RCs, we found an effect of animacy in the case of object RCs that was significant both in the subject, F1(1, 50) = 77.84, Ms = 0.12, p < 0.001, and in the item analysis, F2(1, 5) = 27.14, Ms = 0.34, p < 0.01. We also found an effect of d-span that was significant both in the subject, F1(1, 49) = 13.50, Ms = 0.24, p < 0.001, and in the item analysis, F2(1, 599) = 26.99, Ms = 0.12, p < 0.001. To further analyze the effect of d-span, we compared the performance of the three groups of children in object RCs with animate and inanimate heads with t tests. We found a significant difference between children with span 4 and children with span 5 or 6, both in object RCs with animate heads
123
146
J Psycholinguist Res (2011) 40:137–154
t (1, 214) = −2.16, p < 0.001, t (1, 172) = 3.29, p < 0.001, respectively, and in object RCs with inanimate heads, t (1, 214) = −4.18, p < 0.001, t (1, 172) = 3.96, p < 0.001, respectively. No difference was found between children with span 5 and 6. We performed the same analyses with listening span test measures (w-span) in order to investigate the role of working memory in children’s RC processing, but we never found any effect of w-span. In summary, we found an effect of Sentence Type, with subject RCs being responded to more accurately than object RCs. We found an effect of animacy due to the fact that object RCs with animate objects were responded to less accurately than object RCs with inanimate objects. We found an effect of d-span due to the fact that children with span 4 performed less well that the other two groups of children in the comprehension of object RCs (with inanimate or animate heads). No effect of w-span was found. Listening Time Analysis Based on the distribution of listening times (LTs) for each segment, we individuated extreme values at the right of the distribution (typically the distribution of RTs has a long tail to the right) and replaced them with the median of the segment in the given condition. In addition, LTs less than 100 ms were substituted with the median as before (0.94% of the total LTs), based on the observation that in adults the simple RT to a stimulus is greater than 180 ms (the accepted figures for mean reaction times for college-age individuals being 190 ms for light stimuli and about 160 ms for sound stimuli; Luce 1986) and this must be all the more so for children who have slower RTs. These LTs were randomly distributed across conditions. We found that 22/7,344 = 0.29% of the LTs were those of segments for which subjects previously pushed the response button before the end of the segment; given the low number of these LTs and the fact that they are randomly distributed across subjects, conditions and segments, they were also substituted with the median. In this way, 122/7,344 = 1.6% of LTs were replaced. As the length of some segments differed and LTs might depend on segment length, for each subject we regressed LTs on segment length to find the slope of the interpolating regression line. Afterwards, we rotated the scatterplot around its gravity centre (i.e., the point whose coordinates are the means of the segment length and listening time values). In this way, we removed the variability of LTs due to the length of the segments. The effect of this transformation of the data is represented in Fig. 2. Notice that with this rotation, the overall means (i.e., the centre of gravity of the scattergram) do not change. We examined LTs for all sentences, those that were answered correctly and those that were answered incorrectly. We made this choice because, in contrast to adults, children’s accuracy is poor, at least in some conditions and this is shown in the offline question, but they may be nevertheless attempting to understand. Before starting to analyze the effects, it is worth pointing out a general consideration on our data, which will be useful to appreciate their coherence, especially in view of the fact that few experiments have been conducted measuring listening times in children. Figure 3a displays the mean LTs of the six segments without distinguishing among the four conditions. Three curves are represented, corresponding to the three levels of our participants’ dspans. First we may notice that LTs of d-span 4 participants are generally slightly greater than those of the other participants’ d-spans, although this difference is not significant with the exception of segment 2, as is clear from the figure. Second, the first two segments are different from the others. The LTs of the first segment are longer than the others because the segment, itself is longer than the others, because it includes a subject and a verb. In addition, it is likely that participants take longer to listen to the first segment as they are at
123
J Psycholinguist Res (2011) 40:137–154
147
Fig. 2 Transformation of the raw data in rotated data for a typical participant. a Shows the dependence of listening times on segment length. The dotted line represents the regression line and ϕ is the angle of the slope. b Displays the data after rotation of an angle ϕ around their centre of gravity. In this way, for each subject, the variability due to the segment length was removed
Fig. 3 a Listening times (LT) in seconds for the 6 segments of the sentences as a function of the d-span of participants. Each point is based on the listening time for the given segment in the four conditions (Sentence × Animacy). b Listening times (LT) in seconds of the 24 sentences. Each point is based on the listening times for a given sentence in the four conditions
123
148 Table 1 Mean listening times in seconds as a function of animacy and clause type SA subject RC with animate object, SI subject RC with inanimate object, OA object RC with animate object, OI object RC with inanimate object
J Psycholinguist Res (2011) 40:137–154
S1 S2 S3 S4 S5 S6
SA
OA
SI
OI
1.515 0.340 0.879 0.995 1.123 1.179
1.528 0.389 0.921 1.033 1.188 1.197
1.540 0.364 0.890 1.029 1.162 1.146
1.556 0.378 0.910 1.030 1.170 1.164
the beginning of a new sentence. The LTs of the second segment (including the direct object) are the shortest, likely because the previous segment made the expectation of a direct object very high and thus sped up the following processing. Finally, the remaining segments, which include the beginning of the relative clause (segment 3), the disambiguating segment, the spill over segment (segments 4 and 5) and the final segment (segment 6) are listened to more or less equally. These consistent trends suggest that the experiment is accurate. Figure 3b displays the mean LTs for the 24 sets of sentences. As we can see from the figure no sentence appears to be different from the others in terms of LTs (see analyses below). Mean listening times in seconds for individual segments of each sentence type tested are reported in Table 1. For each segment, a mixed model analysis performed on LTs with 2 factors (Sentence Type and Animacy with 2 levels each) and a covariate (d-span) with 51 observations per cell showed no effect or interaction on segments 1, 2 and 6. At segment 3, i.e., the relative pronoun, we found an effect of Sentence Type that was significant in the subject, (F1(1, 49) = 4.85, Ms = 0.054, p < 0.05), but not in the item analysis, (F2(1, 5) = 1.79, Ms = 0.15, p = 0.23). At this position subject RCs were listened to faster than object RCs. A similar result was found by Mak et al. (2002) with adults, but at the previous segment, i.e., at the NP. We think that in our experiment the effect at the relative pronoun is a spill over effect arising at the previous NP. As in Mak et al. (2002) we attribute it to the fact that at segment 2, different sets of nouns were compared (animate vs inanimate and singular vs plural ones). Although at segment 6, the same two sets of nouns were compared, no difference was found at that position, again, as in Mak et al. (2002). At the disambiguating segment (4), the auxiliary stanno (are), no effect was found. However, at segment 5, the gerundive verb, we found a main effect of Sentence Type that was significant in the subject, (F1(1, 49)=8.34, Ms = 0.048, p < 0.01), but not in the item analysis, (F2(1, 5) = 1.13, Ms = 0.35, p = 0.33). This was qualified by an interaction between Sentence Type and animacy that was significant in the subject, (F1(1, 49) = 5.79, Ms = 0.04, p = 0.01), but not in the item analysis, (F2(1, 5) = 2.37, Ms = 0.10, p = 0.18). As shown in Fig. 4, the two RCs with inanimate objects do not differ, while a difference is observed between subject RCs and object RCs with animate objects. This is confirmed by the statistical analysis. In the animate condition, a statistically significant difference is found between subject and object RCs in the subject, F1(1, 50) = 15.97, Ms = 0.04, p < 0.001, but not in the item analysis, F1(1, 5) = 2.21, Ms = 0,29, p = 0.19. Thus, the effect of Sentence Type is essentially due to object RCs with animate objects, which are listened to longer than subject RCs. Remember that we did not discard trials in which the comprehension question was responded to incorrectly. Thus, children slow down in object RCs with animate objects regardless of the answer they provide to the comprehension question. To see whether the type of answer to the comprehension question influences LTs on the critical segment (S5), we compared LTs that were associated with a correct answer to those that were associated with an incorrect answer, but we did not find any difference, in the subject, F1(1, 84) = 0.35, Ms = 0.16, p = 0.55) or in the item analysis F1(1, 5) = 0.47,
123
J Psycholinguist Res (2011) 40:137–154
149
Fig. 4 Interaction between sentence and animacy. The figure clearly shows that the effect of sentence is only due to the level animate (white disk) of the factor Animacy. Vertical bars indicate the standard error at the 0.95% level of confidence
Ms = 0.12, p = 0.51). Thus, children listened longer to object RCs with animate objects regardless of their offline answer. We performed the same mixed model analysis with 2 factors (Sentence Type and Animacy with 2 levels each) and with w-span as covariate, but we did not find any main effect or interaction involving w-span.
Discussion The goals of our study were to investigate whether children were sensitive to number agreement morphology in agreement disambiguated object RCs, whether different features affect the comprehension of RCs and whether individual memory differences modulate the comprehension of these structures. Our results confirm previous findings that subject RCs are easier to comprehend and to process than object RCs. However, we found that these difficulties are modulated by animacy, as in adults, and by d-span. Accuracy measures and LT measures show that object RCs with animate objects are more difficult to comprehend than subject RCs with animate objects. Accuracy measures also show that d-span modulates comprehension of object RCs (subject RCs being almost at ceiling regardless of the d-span). Children with d-span 4 comprehended object RCs less well than children with d-span 5 and 6 and this was true for both types of object RCs, although in this group a difference between the two types of object RCs was observed. The d-span difference did not emerge in the LT analysis; nor did the difference between object RCs with animate and inanimate objects. This discrepancy may depend on the low number of children for each d-span group or on the nature of the integration processes that are measured by LTs in a self-paced listening task, that are different from global comprehension as measured by accuracy. Finally, neither a main effect, nor an interaction was found in the case of w-span. This finding replicates what was found by Booth et al. (2000) in an experiment where centre embedded RCs, i.e., structures similar to ours, were tested with children of a more or less similar age (8–11 years). Our results are at odds with Felser et al. (2003) where an effect of w-span was found in structures involving attachment of an RC to a low or high NP; this discrepancy may be attributed to the fact that Felser et al. investigated structures involving attachment preference, while we studied filler gap dependencies, as Booth and colleagues did. Our findings are also at odds with a number of studies on adults, where effects of reading-span, as measured by Daneman and Carpenter’s task, were found (Bornkessel et al. 2001; King and Just 1991) but are in agreement with
123
150
J Psycholinguist Res (2011) 40:137–154
an adult study on a patient with selective short term memory damage who performed less accurately than controls on complex sentences involving a filler-gap dependency, but not on long sentences lacking structural complexity, such as coordination of two propositions (see Papagno et al. 2007; see also Waters and Caplan 1996a,b for a different approach to memory). Although we cannot solve the problem of which memory resources (i.e., phonological short term storage or working memory) are more important and at which age, the fact that d-span effects were found in studies involving RC processing suggests that for filler-gap dependencies phonological short term storage resources are particularly important, perhaps for children more than for adults. The advantage of subject RCs over object RCs is often explained by syntactic based processing theories in terms of economy principles of gap prediction, such that a gap is posited as soon as a relative pronoun is encountered, i.e., in subject position (Minimal Chain Principle, De Vincenzi 1991 or the Active Filler Hypothesis, Frazier and Flores d’Arcais 1989). Alternatively, the Syntactic Prediction Locality Theory (Gibson 1998) explains the subject advantage in terms of the number of categories that must be predicted to complete an ongoing analysis. However, as Mak et al. (2002) pointed out, these theories do not anticipate an animacy effect. An alternative that Mak and colleagues suggest is that animacy can modulate the assignment of the grammatical functions: if the head NP is inanimate, adults do not take any decision and wait for the next piece of information; when the head NP is animate, they engage in a subject RC analysis that they have to revise later. This hypothesis is incompatible with our finding that children with d-span 4 do commit to a subject RC analysis, as revealed by their higher error rate in the comprehension of object RCs with inanimate objects with respect to subject RCs. We conjecture that the source of these errors is their initial engagement in a subject RC analysis, which sometimes fails to be abandoned. Adults and children with higher memory resources may either not engage in a subject RC analysis, when they hear an inanimate object and wait until more information is available (as suggested by Mak et al. (2002)) or they do engage, but the revision, not being costly, never fails. This latter hypothesis is compatible with the assumption that adults and children share the same processing routines and what changes may be the functioning of the memory system; the former hypothesis entails that children with lower spans use parsing strategies (subject RC analysis regardless of the animacy of the head NP) that are different from those used by children with more memory resources and by adults (when the relative head NP is inanimate). The fact that no d-span effect is found on LTs is compatible with both possibilities. Object RCs with animate objects were particularly taxing for all children and this was evident in both LT measures and accuracy measures. Children answered less accurately and, at the gerundive verb, listened longer to object RCs than to subject RCs with animate objects. This difference was observed at the word following the disambiguating auxiliary, i.e., at the gerundive verb. This is not uncommon and spill-over effects are often observed in self-paced reading studies on adult processing as well (Mak et al. 2002 and references cited there). As neither an effect of d-span nor any interaction involving d-span was observed, the slowing down at the gerundive verbs is common to all children. In addition, we have seen that this slowing down is observed regardless of whether the answer to the comprehension question was correct or not. This means that all children engage in a subject RC analysis. When they encounter the number morphology on the auxiliary (or shortly after), they realize that the subject and the verb do not agree, i.e., that the number morphology is ungrammatical with respect to the ongoing analysis. Some of the children revise their initial commitment and turn the structure into an object RC and thus answer the comprehension question correctly; other children do notice the ungrammaticality, but do not revise the initial analysis and thus answer the comprehension question incorrectly. This means that the slowing down signals
123
J Psycholinguist Res (2011) 40:137–154
151
the detection of the ungrammaticality with respect to the ongoing analysis and, for some children, also the revision of the preferred analysis. If this is correct, it is not surprising that no d-span effect is found at this segment, as all children, regardless of their d-span, slow down, i.e., detect the ungrammaticality. What differs is the action that is performed when this ungrammaticality is detected. These findings can be explained under Fodor and Inoue (2000) diagnosis and repair model, which was proposed to account for adults’ processing. According to these authors, the informativeness of a feature determines the ease of reanalysis or lack thereof. To see how this model works in our case, we need to broaden our perspective and consider our results and those reported in Arosio et al. (2009) that we discussed in session 1, where object RCs with the embedded subject located in the preverbal position were also considered. Consider the object RC disambiguated by the preverbal position of the embedded subject below (15) Il ragazzo che il pagliaccio sta guardando The boy that the clown is3SING watching OBJECT RC: POSITION DISAMBIGUATION ‘The boy who the clown is watching’ Children, as well as adults, start to hypothesize a subject RC analysis at the relative pronoun and posit a gap in subject position connected to the NP head of the RC. When the embedded NP is encountered, they know that the subject RC analysis must be abandoned and that the new NP is the subject of the RC. This is a positive symptom: the diagnosis provides information about how to repair the structure. The adjustment is local and an NP replaces the postulated subject gap. In addition, one grammatical function was assigned before the reanalysis and one is still assigned after it, the subject function in either case. The reanalysis in this case is relatively straightforward and it is not surprising that Italian speaking children are already good at understanding this structure at the age of 5 years (Arosio et al. 2009). Consider now (16), that exemplifies the kind of object RC that was tested in our selfpaced listening experiment. Here the information that triggers the detection of the temporary ungrammaticality is the number agreement morphology on the embedded auxiliary. (16) Il ragazzo che stanno guardando i pagliacci The boy that are3PL watching the clowns OBJECT RC: AGREEMENT DISAMBIGUATION ‘The boy who the clowns are watching’ This piece of information is a negative symptom as it only indicates that the parsed structure is not correct: the head of the RC and its gap do not agree with the inflected auxiliary. However, it does not indicate the way in which the structure should be repaired. Under this model, all children can detect the ungrammaticality, and thus slow down, but they may solve the ungrammaticality in different ways, as the symptom does not instruct them unambiguously as to how to continue. One path they could take is to stick to the ongoing analysis (subject RC) and correct the number feature on the auxiliary to adjust it to the current analysis, a solution that is inexpensive and is not taxing memory resources. In that case, children will answer incorrectly to the comprehension question. A second option, the correct one, is to undo the indexing between the relative pronoun and the subject trace (and the head NP), keep the head NP in memory, look for a new gap and for a new subject. Given that Italian is a null subject language; children could fill the preverbal subject position with a null pronominal subject, pro, licensed by the agreement morphology. However, this subject would fail to be identified, as there is no antecedent for it in the sentence. Another possibility is to wait for the postverbal subject and connect it to the preverbal pro, under the assumption
123
152
J Psycholinguist Res (2011) 40:137–154
that a postverbal subject forms a chain with a preverbal expletive pro (Rizzi 1982). In this last case, when reanalysis is started, a second chain is going to be built that will be closed and interpreted when the postverbal subject is found (see De Vincenzi 1991). It is clear that reanalysis in object RCs disambiguated by number agreement is more complex and requires previous input to be stored and retrieved in order to achieve a correct reanalysis. Moreover, two chains are simultaneously being built and it is not surprising that only children with enough memory resources can perform it and answer the comprehension question correctly. Nor is it surprising that children can correctly analyze (15) before (16) and that (16) starts to be understood around the age of 9–11 years, as shown by Arosio et al. (2009) and confirmed by the present experiment. On the one hand, in the case of (15) the diagnosis gives a hint as to how the structure should be repaired and this way of repairing is straightforward; on the other, in the case of (16), the diagnosis leaves two options open and the correct one is highly demanding in terms of memory resources. The Diagnosis and Repair model can also be extended to object RCs with inanimate objects. Assume that children engage in a subject RC analysis in this case as well, at least this seems to be so for those with d-span 4. Our results suggest that children find it easier to abandon an analysis (subject RC) that is not highly plausible when the incoming information does not confirm it, as happens when they see the number agreement on the auxiliary and/or the verb. The detection of the mismatch gives a hint as to how the structure has to be repaired: inanimate entities are typically objects, not subjects, and the number agreement or the lexical verb itself favours this diagnosis. This is different from the case of object relatives with animate heads, discussed above. Although the number agreement on the auxiliary and the head of the relative do not agree, the features on the NP do not give any hint as to how the sentence could be repaired. In summary, in spite of the fact that accuracy measures show that children err in the comprehension of object RCs disambiguated by number agreement, LT measures reveal that children do process number agreement. However, this is not a guarantee that the structure will be correctly understood. Number agreement on verbs is a negative symptom, and the ungrammaticality induced by it can be repaired in different ways. The correct one, turning the RC into an object RC, requires memory resources to access previous input and to build chains that are not available to all children in our group. In contrast, animacy of the relative head greatly improves the comprehension of number agreement disambiguated object RCs. Animacy is a positive symptom and reanalysis, if it occurs, is straightforward; the head of the relative is the object and an object gap is postulated after the verb.
References Adani, F. (2008). The role of features in relative clause comprehension: A study of typical and atypical development. PhD. Dissertation, University of Milano-Bicocca. Arosio, F., Adani, F., & Guasti M. T. (2009) Grammatical features in the comprehension of Italian Relative Clauses by children. In A. Gavarrò, et al. (Eds.), Merging features: Computation, interpretation and acquisition. Oxford University Press Baddeley, A. D. (1986). Working memory. Oxford: Clarendon Press. Baddeley, A. D., Eysenck, M. W., & Anderson, M. C. (2009). Memory. Hove: Psychology Press. Baddeley, A. D., & Hitch, G. J. (1974). Working memory. In G. A. Bower (Ed.), The psychology of learning and motivation: Vol. 8 Advances in research and theory. San Diego: Academic Press. Bader, M., & Bayer, J. (2006). Case and linking in language comprehension: Evidence from German. Berlin: Springer. Becker, M. (2006). There began to be a learnability puzzle. Linguistic Inquiry, 37, 441–456. Brainard, D. H. (1997). The Psychophysics Toolbox. Spatial Vision, 10, 433–436.
123
J Psycholinguist Res (2011) 40:137–154
153
Booth, J. R., Mac Whinney, B., & Harasaki, Y. (2000). Developmental differences in visual and Auditory processing of complex sentences. Child Development, 71(4), 981–1003. Bornkessel, I. D., Fiebach, C. J., & Friederici, A. D. (2001). On the cost of syntactic ambiguity in human language comprehension: an individual differences approach. Cognitive Brain Research, 21, 11–21. Ciccarelli, L. (1998). Comprensione del linguaggio, dei processi di elaborazione e memoria di lavoro: uno studio in età prescolare. PhD dissertation, University of Padua, Italy. Daneman, M., & Carpenter, P. A. (1980). Individual differences in working memory and reading. Journal of Verbal Learning and Verbal Behaviour, 19, 450–466. De Villiers, J., Tager-Flusberg, H., Hakuta, K., & Cohen, M. (1979). Children’s comprehension of relative clauses. Journal of Psycholinguistics Research, 8, 499–518. De Vincenzi, M. (1991). Syntactic parsing strategies in Italian. Dordrecht: Kluwer. Dixon, P. (2008). Models of accuracy in repeated-measures designs. Journal of Memory and Language. doi:10.1016/j.jml.2007.11.004. Felser, C., Marinis, T., & Clahsen, H. (2003). Children’s processing of ambiguous sentences: A study of relative clause attachment. Language Acquisition, 11, 127–163. Ferreira, F., & Anes, M. (1994). Why study spoken language?. In M. A. Gernsbacher (Ed.), Handbook of psycholinguistics (pp. 33–56). San Diego: Academic Press. Ferreira, F., Henderson, J. M., Anes, M. D., Weeks, P. A., Jr., & McFarlane, D. K. (1996). Effects of lexical frequency and syntactic complexity in spoken language comprehension: Evidence from the auditory moving window technique. Journal of Experimental Psychology: Learning, Memory, and Cognition, 22, 324–335. Fodor, D. J., & Inoue, A. (2000). Syntactic features in reanalysis: Positive and negative symptoms. Journal of Psycholinguistic Research, 29(1), 25–36. Frauenfelder, U., Segui, J., & Mehler, J. (1980). Monitoring around the relative clause. Journal of Verbal Learning and Verbal Behaviour, 19, 328–337. Frazier, L., & Flores d’Arcais, G. (1989). Filler-driven parsing: A study of gap-filling in Dutch. Journal of Memory and Language, 28, 331–344. Friederici, A. D. (1997). Diagnosis and reanalysis: Two processing steps the brain may differentiate. In J. Fodor & F. Ferreira (Eds.), Reanalysis in sentence processing. Dordrecht: Kluwer. Friederici, A. D., Steinhauer, K., Mecklinger, A., & Meyer, M. (1998). Working memory constraints on syntactic ambiguity resolution as revealed by electrical brain responses. Biological Psychology, 47, 193–221. Friedmann, N., & Novogrodsky, R. (2004). The acquisition of relative clause comprehension in Hebrew: A study of SLI and normal development. Journal of Child Language, 31, 661–681. Gaulin, C., & Campbell, T. (1994). Procedure for assessing verbal working memory in normal school-age children: Some preliminary data. Perceptual and Motor Skills, 79, 55–64. Gibson, E. (1998). Linguistic complexity: Locality of syntactic dependencies. Cognition, 68, 1–76. Gordon, P. C., Hendrick, R., & Johnson, M. (2001). Memory interference during language processing. Journal of Experimental Psychology: Learning, Memory, and Cognition„ 27, 1411–1423. Guasti, M. T., & Cardinaletti, A. (2003). Relative clause formation in romance child production. Probus, 15, 47–88. Guasti, M. T., Stavrakaki, S., & Arosio, F. (2007). Number and Case in the comprehension of relative clauses: Evidence from Italian and Greek. In A. Gavarrò & M. J. Freitas (Eds.), Language acquisition and development: Proceedings of GALA 2007.. Newcastle: Cambridge Scholar Press. Harris, A. C. (1981). Georgian syntax. A study in relational grammar. Cambridge, UK: Cambridge University Press. King, J., & Just, M. A. (1991). Individual differences in syntactic processing: The role of working memory. Journal of Memory and Language, 30, 580–602. King, J. W., & Kutas, M. (1995). Who did what and when? Using word- and causal- level ERPs to monitor working memory usage in reading. Journal of Cognitive Neuroscience, 7(3), 376–395. Little, R. C., Milliken, G. A., Stroup, W. W., & Wolfinger, R. D. (1996). SAS system for mixed models. North Carolina: SAS Publishing. Love, T. E. (2007). The processing of non-canonically ordered constituents in long distance dependencies by pre-school children: A real-time investigation. Journal of Psycholinguistic Research, 36, 191–206. Luce, R. D. (1986). Response times: Their role in inferring elementary mental organization. New York: Oxford University Press. MacWhinney, B., & Pléh, C. (1988). The processing of restrictive relative clauses in Hungarian. Cognition, 29, 95–141. Mak, W. M., Vonk, W., & Schriefers, H. J. (2002). The influence of animacy on relative clause processing. Journal of Memory and Language, 47(1), 50–68.
123
154
J Psycholinguist Res (2011) 40:137–154
Mak, W. M., Vonk, W., & Schriefers, H. J. (2008). Discourse structure and relative clause processing. Memory & Cognition (Vol. 36, pp. 170–181). Utrecht: Ms. University of Utrecht. McKee, C., Nicol, J., & McDaniel, D. (1993). Children’s application of binding during sentence processing. Language and Cognitive Processes, 8(3), 265–290. Meng, M., & Bader, M. (2000). Mode of disambiguation and garden path strength: An investigation of subject–object ambiguities in German. Language and Speech, 43, 43–74. Palladino, P. (2005). Uno strumento per esaminare la memoria di lavoro verbale in bambini di scuola elementare: taratura e validità. Psicologia Clinica Dello Sviluppo, 1, 129–150. Papagno, C., Cecchetto, C., Reati, F., & Bello, L. (2007). Processing of syntactically complex sentences relies on verbal short-term memory. Evidence from a STM patient. Cognitive Neuropsychology, 24, 292–311. Pelli, D. G. (1997). The VideoToolbox software for visual psychophysics: Transforming numbers into movies. Spatial Vision, 10, 437–442. Rizzi, L. (1982). Issues in Italian syntax. Dordrecht: Foris Publications. Roberts, L., Marinis, T., Felser, C., & Clahsen, H. (2007). Antecedent priming at trace positions in children’s sentence processing. Journal of Psycholinguistic Research, 36, 175–188. Schriefers, H., Friederici, A. D., & Kühn, K. (1995). The processing of locally ambiguous relative clauses in German. Journal of Memory and Language, 8, 499–520. Sekerina, I., Stromswold, K., & Hestvik, A. (2004). How do adults and children process referentially ambiguous pronouns? Journal of Child Language, 31, 123–152 Traxler, M. J., Morris, R. K., & Seely, R. E. (2002). Processing subject and object relative clauses: Evidence from eye-movements. Journal of Memory and Language„ 47, 69–90. Vos, S. H., Gunter, T. C., Schriefers, H., & Friederici, A. D. (2001). Syntactic parsing and working memory: The effects of syntactic complexity, reading span, and concurrent load. Language and Cognitive Processes, 16, 65–103. Waters, G. S., & Caplan, D. (1996). Processing resource capacity and the comprehension of garden path sentences. Memory and Cognition, 24, 342–355. Waters, G. S., & Caplan, D. (1996). The measurement of verbal working memory capacity and its relation to reading comprehension. Quarterly Journal of Experimental Psychology, 49, 51–79. Wittek, A., & Tomasello, M. (2005). German-speaking children’s productivity with syntactic constructions and case morphology: Local cues help locally. First Language, 25, 103–125.
123