Working memory effects on L1 and L2 processing of ambiguous relative clauses by Korean L2 learners of English

Abstract

In this study, we report the results of two self-paced reading experiments that investigated working memory capacity effects on the processing of globally ambiguous relative clauses by advanced Korean second language (L2) learners of English. Consistent with previous monolingual literature on the processing of temporary ambiguity, we found that working memory capacity was a factor that also affected the processing of globally ambiguous relative clauses. High working memory capacity was positively correlated with a processing disadvantage reflected as slower reading times at the region where the ambiguity becomes detectable, and longer response times to decide on a correct disambiguation for the target structure. Furthermore, a similar pattern was also found in the same participants’ processing of L2 ambiguity. We conclude that for highly advanced L2 learners, the processing strategies employed for ambiguous structures are not qualitatively different between the same individual’s first language (L1) and L2, but rather differ across readers of different working memory capacities.

Keywords

global ambiguity resolution relative clause attachment second language processing working memory capacity

I Introduction

Research in the area of second language (L2) processing has focused largely on the similarities and differences between the processing strategies of native speakers and L2 learners. Recently, increased interest in the mechanisms involved in sentence processing has led to the investigation of various factors that may affect the processing strategies of these two populations. One such factor that has been the topic of extensive discussion in the monolingual literature is individual working memory capacity (WMC). Some studies have provided evidence regarding WMC effects on the post-interpretive stage of processing, i.e. which interpretation is chosen for the ambiguous relative clause (RC) structure (Kim and Christianson, 2013; Mendelsohn and Pearlmutter, 1999). Others have investigated the interpretive stage of processing. For example, individuals with high capacity may take longer to resolve syntactic ambiguities, but show shorter reading times when integrating pragmatic information (MacDonald et al., 1992).

MacDonald et al. (1992) proposed a model of sentence processing, i.e. the Capacity Constrained Parsing Model, in which WMC influences the extent to which a reader is capable of maintaining multiple representations during the processing of sentences with syntactic ambiguity. In their study, they employed a construction in which the verb was temporarily ambiguous between a main verb reading and a reduced relative reading.

(1) a. The experienced soldiers warned about the dangers before the midnight raid.

b. The experienced soldiers warned about the dangers conducted the midnight raid.

In both sentences (1a) and (1b), the role of the verb warned is temporarily ambiguous. In (1a), the sentence is disambiguated toward the main verb reading at the end of the sentence, whereas in (1b) the temporary ambiguity is resolved at the main verb conducted, which informs the reader that the reduced relative reading is correct.

The Capacity Constrained Parsing Model (MacDonald et al., 1992) predicts that only readers with sufficient cognitive resources, i.e. high capacity, will be able to maintain both possible interpretations during processing. On the other hand, readers who lack sufficient resources are likely to choose only one reading, which is usually the more preferred or more frequent, i.e. the main verb reading. This decision will result in greater processing costs for low capacity readers in cases where the temporary ambiguity is resolved toward the less preferred reading, as reanalysis is required. The results of their study confirmed these predictions. Although both groups of readers showed elevated reading times for the reduced relative reading in (1b), there was no interaction with WMC, indicating that the less preferred reading resulted in additional processing costs for both groups.¹ However, there was a significant interaction with WMC and reading time for sentences resolved toward the main verb reading, with only the high capacity readers showing an increase in reading times at the disambiguating region. MacDonald et al. (1992) claim that when the temporary ambiguity is resolved toward the preferred reading, the high capacity readers suffer from increased processing costs associated with maintaining multiple representations in working memory. Furthermore, total reading times for the ambiguous sentences did not differ significantly between groups, suggesting that the results were not due to general differences in reading speed stemming from cognitive differences.

Building on these results, a number of studies have investigated WMC effects in L2 learners’ language processing. Dussias and Piñar (2010) investigated how WMC may affect how native English speakers and Chinese L2 learners of English utilize plausibility information in the processing of long-distance wh-questions, as illustrated below.

(2) a. Who_i did the police declare t_i killed the pedestrian? (implausible filler)

b. Who_i did the police know t_i killed the pedestrian? (plausible filler)

Results showed that only the L2 learners with high WMC showed reading time patterns similar to the native English speakers, with longer reading times when the wh-word was a plausible filler for the incorrect gap position (as object of the main verb), leading to an incorrect initial analysis. The low capacity group showed an opposite pattern, with longer latencies for sentences with implausible fillers. Dussias and Piñar (2010) claim that these results indicate that native English speakers and high capacity L2 learners were capable of using plausibility information to recover from initial misparses, whereas low capacity L2 learners were not. These results extend previous findings in the monolingual processing literature showing cognitive effects on first language (L1) processing (Just and Carpenter, 1992; MacDonald et al., 1992; Traxler, 2007), and suggest that L2 learners also require sufficient cognitive resources to integrate various types of information during processing. It is only when adequate cognitive resources are available that L2 learners exhibit online processing patterns that resemble those of native speakers.

Other studies suggest that the extent to which cognitive resources play a role in L2 processing may be task-dependent (Havik et al., 2009; Williams, 2006). In Williams (2006), native and non-native readers were required to read sentences making use of plausibility information such as in (3).

(3) a. Which machine did the mechanic fix the very noisy motorbike with two weeks ago?

b. Which friend did the mechanic fix the very noisy motorbike with two weeks ago?

Longer reading times for the sentences in the plausible condition (3a), for which machine may initially be analysed as the object of fix, would be taken to suggest that the participants were able to immediately compute the plausibility of the filler-gap relation at the verb fix. The results turned out to be task dependent. When performing a ‘stop making sense’ task, which required incremental processing of the sentence, all participants showed immediate plausibility effects. However, when participants were asked to read the same sentences to perform a memory probe task, only the native English speakers and L2 learners with higher performance on the memory task used plausibility information. Furthermore, even the L2 learners with high capacity showed plausibility effects at a later point in the sentence than both high and low memory native English speakers. Williams (2006) suggests that the incrementality of L2 processing is task dependent, and that the L2 learners with low memory capacity did not have sufficient resources to perform incremental processing while at the same time undertaking a memory-demanding task.

To summarize, in contrast to previous monolingual studies, which suggest that cognitive capacity is a factor that affects sentence processing (Just and Carpenter, 1992; Kim and Christianson, 2013; MacDonald et al., 1992; Traxler, 2007), the results of studies on L2 learners have been more divergent. Studies suggest that there is a relationship between the cognitive capacity of L2 learners and L2 processing, and that L2 learners with sufficient resources may show native-like processing, but even high-capacity L2 learners may be more task dependent than native speakers (Dussias and Piñar, 2010; Havik et al., 2009; Williams, 2006).

What is crucial to the motivation of the present study is that most previous studies investigating WMC effects on ambiguity processing used temporarily ambiguous constructions (Dussias and Piñar, 2010; Havik et al., 2009; MacDonald et al., 1992; Williams, 2006). However, examining processing strategies for globally ambiguous constructions, and comparing how these strategies may differ across temporarily and globally ambiguous constructions, is a topic that merits further investigation, primarily due to differences in the final interpretation of the two different types of ambiguity. For temporarily ambiguous constructions, the ambiguity lies at a certain point in the sentence, but is ultimately resolved toward one correct final interpretation, e.g. garden-path sentences. However, for globally ambiguous constructions, the reader is not required to settle on one single correct interpretation. As both possible interpretations are valid, it is possible for readers with a more limited pool of cognitive resources, in particular, L2 learners, to settle on a good-enough interpretation (Christianson et al., 2001; Ferreira et al., 2002), or to construct a shallow representation (Clahsen and Felser, 2006). Therefore, our investigation of processing strategies in globally ambiguous structures is expected to shed more light on the mechanisms of the human sentence parser, as readers may employ processing strategies different from those used when processing a temporarily ambiguous construction.

Building on previous studies investigating WMC effects on temporary ambiguity, we further advance this line of research by examining WMC effects on the processing of global ambiguity by advanced Korean L2 learners of English, and whether WMC effects on processing strategies pattern differently across the L1 and L2. One of the merits of the paraphrase decision task employed in the present study is that both online reading time data and offline interpretation data can be obtained. The online reading time data reflect the interpretive stage of processing, i.e. online processing, during which participants parse the incoming sentence and construct a syntactic representation, and the paraphrase decision RT data and RC attachment preference data reflect the post-interpretive stage of processing, i.e. offline processing, where the meaning gleaned from the interpretive stage is stored and used to complete other tasks (Caplan and Waters, 1999).

In line with previous studies on L2 learners (Dussias and Piñar, 2010; Havik et al., 2009), which showed that the processing strategies of L2 learners with high WMC patterned with those of native speakers, we predict that for the advanced L2 learners in the present study, higher WMC will result in less divergence in processing strategies across their L1 and L2, despite purported cross-linguistic differences in attachment preferences.

II The present study

In the present study, we address the question of how individual cognitive ability, measured as working memory capacity (WMC), affects ambiguous RC processing, and whether these effects differ between L1 and L2 processing. We begin by describing several points critical to the design of our study. First, in contrast with numerous previous studies in the literature investigating cognitive effects on L2 processing (Dussias and Piñar, 2010; Havik et al., 2009; Juffs, 2004, 2005; Williams, 2006), which compared the performance of native speakers and L2 learners, in the present study we examine the same group of participants and compare how their processing of the ambiguous RC construction differs across their L1 and L2. This was done to provide a more direct and accurate investigation of how the processing strategies of the same group of individuals may differ according to language, which was not possible in previous studies employing two distinct populations, which were also likely to have differed in various other factors besides language. We expect that this type of direct comparison will provide us with a clearer picture of how the processing strategies employed by speakers differ between a native and second language.

A second point that merits discussion is the type of structure employed. Most previous studies investigating WMC effects in L1 and L2 processing used temporarily ambiguous constructions, disambiguated at later points in the sentence (Dussias and Piñar, 2010; Havik et al., 2009; Juffs, 2004, 2005; MacDonald et al., 1992; Williams, 2006). In these constructions, if the reader has made an initial misanalysis regarding the ambiguity, they must recover from this error and perform a reanalysis in order to arrive at the correct interpretation. In this study, participants were asked to process globally ambiguous sentences, e.g. The lawyer of the client who insulted you during the trial was intelligent. In this structure, the RC who insulted you during the trial may modify either the lawyer (High attachment, or HA) or the client (Low attachment, or LA).

This ambiguous RC structure was investigated by Cuetos and Mitchell (1988) in a study of disambiguation preferences in English and Spanish, to test the claim that LA was a universal preference, based on the structural principle of Minimal Attachment (Frazier, 1979). Attachment preferences in English and Spanish were tested with native speakers of each language. Results showed that English speakers displayed a weak LA preference, and Spanish speakers showed a strong HA preference. These data were taken as counter-evidence to language-general, universal parsing preferences, and supporting a ‘tuning hypothesis’, which proposed that the parsing preferences of speakers are ‘tuned’ by the structural frequencies in any given language (Cuetos and Mitchell, 1988).

Turning to Korean, the language of interest here, previous studies have found a HA preference for the ambiguous RC structure (Jun, 2003; Lee and Kweon, 2004). The purported HA preference in Korean has been accounted for as follows. The head-final property of Korean results in the RC preceding the complex NP that it modifies, and the lower noun (NP2), preceding the higher noun (NP1), i.e. RC–NP2–NP1. The Thematic Processing Domain (Frazier and Clifton, 1996) cannot be closed off at the first NP (NP2), which is marked with the Genitive case marker -uy, and therefore the parser must wait until it reaches the second NP (NP1), which is marked with the Accusative case marker -ul, and therefore a plausible candidate for case/theta-marking as matrix object. This enables both nouns in the complex NP to be visible to the parser, and the more salient of the two NPs would be the higher noun (NP1).

As mentioned earlier, when processing the globally ambiguous RC construction, it is possible for the reader to not attach the RC at all, thereby deriving an underspecified, or ‘good enough’, final interpretation (Swets et al., 2008). Therefore it is possible that L2 learners may employ strategies different from those used when processing a temporarily ambiguous construction. In the present study, we aim to investigate L2 learners’ processing strategies in the globally ambiguous RC construction, and compare how processing strategies may differ between L1 and L2 processing.

Regarding the effects of WMC on ambiguity processing, we predict that readers with high WMC will be more likely to maintain both possible attachment sites in memory, as they continue to look for input that will signal the best attachment. We further predict that the additional processing load associated with maintaining both interpretations in working memory will be reflected as longer reading times at the region where the ambiguity becomes detectable, in line with the predictions of the Capacity Constrained Parsing Model (MacDonald et al., 1992). Therefore, a second goal of the present study is to examine whether highly advanced L2 learners show a similar pattern of WMC effects in L1 and L2 processing of the globally ambiguous RC construction.

III Factors modulating processing difficulty

In addition to cognitive effects on ambiguity processing, another point of interest is how various factors previously reported to modulate processing load interact with WMC in the processing of the ambiguous RC construction. In the present study, we investigated two factors in particular: RC modifying position and referential load.

Previous studies investigating ambiguous RC processing have used either subject-modifying RCs (Omaki, 2005; Swets et al., 2007; Traxler, 2007) or object-modifying RCs (Felser, Marinis and Clahsen, 2003; Felser, Roberts, Marinis and Gross, 2003; Mendelsohn and Pearlmutter, 1999). However, clear differences in predicted processing difficulties between these two sentence types exist, illustrated in (4).

(4) a. Subject-modifying RC:

The agent of the star who met me at the party last night was poor.

b. Object-modifying RC:

The police arrested the agent of the star who met me at the party last night.

The Dependency Locality Theory (Gibson, 1998, 2000) predicts that the relative processing difficulty of a sentence is correlated with the integration distance between the main subject and verb. Integration distance, in turn, is measured in terms of the number of referents intervening between subject–verb integration.² Consequently, a comparison of (4a) and (4b), with the main subject italicized and the main verb underlined, clearly shows that in English, a head-initial language, integration distance is greater in (4a).

In contrast, a different pattern regarding RC processing difficulties is predicted in Korean, a head-final language. The difference between the RC construction in English and Korean is illustrated in (5).

(5) a. Subject-modifying RC:

Cayphan-cwung-ey cungin-ul moyok-han pyunhosa-uy kokayk-un

Trial-during-LOC witness-ACC insult-MOD lawyer-POSS client-TOP

ttoktttok-hayss-ta.

intelligent-PST-DECL

The lawyer of the client who insulted the witness during the trial was intelligent.’

b. Object-modifying RC:

Phansa-nun caypahn-cwung-ey cungin-ul moyok-han pyunhosa-uy

Judge-TOP trial-during-LOC witness-ACC insult-MOD lawyer-POSS

kokayk-ul thail-ess-ta.

client-ACC rebuke-PST-DCL

‘The judge rebuked the lawyer of the client who insulted the witness during the trial.’

A comparison of the RC construction in Korean (5a, 5b) and in English (4a, 4b) shows a clear contrast. The difference in predicted processing difficulty between the two languages is due to opposite head-directionality: English is a head-initial language, and Korean is a head-final language, which results in differences in canonical word order (SVO vs. SOV). Therefore, the Dependency Locality Theory (Gibson 1998, 2000) predicts that, in English, readers should have more difficulty when processing the subject-modifying RC, while in Korean, more difficulty is predicted when processing the object-modifying RC.

Referential load is also a factor predicted to modulate predicted processing difficulty. The Givenness Hierarchy (Warren and Gibson, 2002) predicts that the processing of 1st and 2nd pronouns will consume less resources than definite descriptions of the form ‘the + noun’, i.e. ‘full NPs’. Therefore, processing a sentence with a full NP is predicted to require more resources than a sentence with a 1st/2nd person pronoun in the same position, thereby leaving a more limited pool of cognitive resources with which to process the rest of the sentence.

The sentences in (6) contrast in the referential load of the RC embedded object. In (6a), the embedded object is a full NP, and therefore predicted to be higher in processing load than (6b), in which the embedded object is a 1st/2nd person pronoun. In the present study, we manipulate the factors of RC modifying position and referential load to vary the degree of predicted processing difficulty. In particular, we aim to investigate how the factor of processing difficulty interacts with WMC effects in the processing of the ambiguous RC construction. Our prediction is that there will be an interaction of processing difficulty and WMC, so that the difference in processing patterns affected by WMC will become more pronounced in sentences with heavier processing load.

(6) a. High referential load:

The lawyer of the client who insulted the witness during the trial was intelligent.

b. Low referential load:

The lawyer of the client who insulted me/you during the trial was intelligent.

To summarize, in this article we predict that the claims of the Capacity Constrained Parsing Model (MacDonald et al., 1992) will hold for the processing of global ambiguity as well as temporary ambiguity, although readers are not pressed to settle on a final interpretation when processing sentences with global ambiguity. Readers with sufficient cognitive resources will be able to maintain both possible readings of a globally ambiguous sentence in working memory, and the effects of this additional load will be manifested as longer reading times on the critical region. The strain on the high capacity readers’ cognitive resources from retaining both interpretations will be elevated when processing sentences with heavy processing load, resulting in a larger gap in reading times between the high and low capacity readers. In the following, we provide an outline of two experiments that tested these predictions in Korean and English.

IV Experiment 1

This experiment investigated whether the predictions of the Capacity Constrained Parsing Model (MacDonald et al., 1992) would be borne out in the processing of globally ambiguous RCs in Korean, and addressed the following questions:

whether readers differing in WMC employ different processing strategies; and

whether effects of WMC are modulated by hypothesized processing difficulties.

In order to avoid a possible confound in the measurement of individual WMC with L2 proficiency, the participants’ WMC was measured in Korean (L1), and this measure was used in both experiments.

1 Method

a Participants

Thirty-four native Korean L2 learners of English who were students at the University of Illinois at Urbana-Champaign, USA, participated in the study. All participants had normal or corrected to normal vision and were naive to the purpose of the study. Participants were highly advanced speakers of English and had resided in the USA for a minimum of five years at the time of the experiment. In order to prevent language transfer effects from languages other than English and Korean, participants fluent in a third language were excluded from the experiment (Cuetos et al., 1996).³ Demographic and language background information of the participants is presented in Table 1.

Table 1.

Language background and demographic information for participants.

	Mean	Range
Age	24.5	19–32
Age of arrival	15.53	13–22
Self-rated reading*	7.78	5–9
Self-rated writing*	6.9	5–9
Self-rated speaking*	8.44	5–10
Self-rated listening*	8.5	5–10
Cloze test scores**	29.31	23–38
WMC scores***	32.81	21–41

Notes. * Self-ratings for English proficiency on a scale of 1 (none) to 10 (native-like); ** Measuring English proficiency (maximum score = 40). ***Administered in Korean; maximum score: 42.

b Materials

The experimental stimuli were 40 sentences with globally ambiguous RCs. The relation between the two nouns was the family/kinship relation or the functional relation, which has been shown to have the least bias toward either of the two NPs (Gilboy et al., 1995). In order to investigate whether processing strategies were affected by processing load, the experimental items differed in levels of predicted processing difficulty. RC modifying position (Gibson, 1998, 2000) and referential load (Sanford et al., 2005; Warren and Gibson, 2002) were chosen as independent factors, resulting in a 2 × 2 design. A sample sentence in each condition is presented in (7).

(7) a. Subject-modifying RC, Heavy referential load:

Cayphan-cwung-ey cungin-ul moyok-han pyunhosa-uy (NP2) kokayk-un (NP1) ttoktttok-hayss-ta. (V)

Trial-during-LOC witness-ACC insult-MOD lawyer-POSS client-TOP intelligent-PST-DECL

‘The lawyer of the client who insulted the witness during the trial was intelligent.’

b. Subject-modifying RC, Low referential load:

Cayphan-cwung-ey na-lul moyok-han pyunhosa-uy kokayk-un ttoktttok-hayss-ta.

Trial-during-LOC me-ACC insult-MOD lawyer-POSS client-TOP intelligent-PST-DECL

‘The lawyer of the client who insulted me during the trial was intelligent.’

c. Object-modifying RC, Heavy referential load:

Phansa-nun caypahn-cwung-ey cungin-ul moyok-han pyunhosa-uy kokayk-ul thail-ess-ta.

Judge-TOP trial-during-LOC witness-ACC insult-MOD lawyer-POSS client-ACC rebuke-PST-DCL

‘The judge rebuked the lawyer of the client who insulted the witness during the trial.’

d. Object-modifying RC, Low referential load:

Phansa-nun caypahn-cwung-ey na-lul moyok-han pyunhosa-uy kokayk-ul thail-ess-ta.

Judge-TOP trial-during-LOC me-ACC insult-MOD lawyer-POSS client-ACC rebuke-PST-DCL

‘The judge rebuked the lawyer of the client who insulted me during the trial.’

The critical regions were the two nouns comprising the complex NP, segments NP2 and NP1, which is where the global ambiguity becomes detectable, and the matrix verb, segment V, where the reader must perform the integration between the matrix subject and verb; compare (7a).

Sentences were distributed across four lists in a Latin square design so that each participant saw each experimental item in only one of the four conditions. Each list contained 88 filler sentences not relevant to the target structure. Twenty-four of the fillers were globally ambiguous, e.g. The big circles and squares were blue, and the remaining were not. One fourth of the fillers were designed to be ambiguous due to the concern that the participants might notice the relation between the RC construction and the potential ambiguity, resulting in conscious attention to the target sentences. The paraphrase decision task for the filler sentences did not probe the ambiguity of the target sentence.

c Procedure

The same participants took part in both Experiment 1 and Experiment 2, in order to compare WMC effects across L1 and L2 processing. In order to mitigate possible language transfer or priming effects, the two sessions were scheduled at least three months apart for each participant. In addition, the order in which the two sessions were administered was counterbalanced across participants. Participants were assigned to a different list for each session. Each session was conducted entirely in the language that was tested by research assistants who were native speakers of Korean or English, in order to prevent participants from functioning in ‘bilingual mode’ (Grosjean, 2001). For the same reason, the cloze test (designed to measure L2 proficiency) and the language background questionnaire were also administered in the English session.

d Reading span task

A Korean version of the reading span task used in Conway et al. (2005) was used to measure participants’ WMC. The task consisted of 12 sets of sentences, ranging from two to five sentences per set, with 42 sentences total. The sets were randomized so that participants did not necessarily see sets increasing in number of sentences. The participants read each sentence aloud, made a judgment regarding the semantic plausibility of the sentence, and then read aloud and memorized a random letter at the end of each sentence. When the end of a set was reached, participants were instructed to write down all letters in that set in the order in which they had appeared. One point was received for every letter in the correct position, with a maximum possible score of 42 points.

e RC paraphrase decision task

Participants read each sentence in a self-paced word-by-word, non-cumulative, moving window reading paradigm, programmed in E-Prime (Professional, v2.0). Every item was followed by a paraphrase decision task (see example in (8)) (Kim, 2008; Kim and Christianson, 2013; Zhou and Christianson, 2015) in which participants had to determine whether a two-sentence paraphrase of the previous sentence was accurate or not, by pressing Y or N on an E-Prime button box. For the experimental stimuli, the paraphrase was one of the two possible interpretations for the ambiguous RC, i.e. HA or LA. A sample test item is presented in (8).

(8) a. Target sentence:

Cayphan-cwung-ey cungin-ul moyok-han pyunhosa-uy (NP2) kokayk-un (NP1) hyunmyung-hayss-ta.

Trial-during-LOC witness-ACC insult-MOD lawyer-POSS client-TOP wise-PST-DECL

‘The lawyer of the client who insulted the witness during the trial was wise.’

b. HA paraphrase version:

Pyunhosa-uy kokayk-un hyunmyung-hayss-ta. Pyunhosa-nun cayphan-cwung-ey cungin-ul moyok-hayss-ta.

‘The lawyer of the client was wise. The lawyer insulted the witness during the trial.’

c. LA paraphrase version:

Pyunhosa-uy kokayk-un hyunmyung-hayss-ta. Kokayk-un cayphan-cwung-ey cungin-ul moyok-hayss-ta.

‘The lawyer of the client was wise. The client insulted the witness during the trial.’

The two paraphrase versions were counterbalanced across target items and lists so that each participant saw only one paraphrase version per sentence, and an equal number of HA and LA paraphrases per list. Both the answers (YES/NO) and response times (RTs) to the paraphrase decision task were recorded. Recall that both versions of the paraphrase are ‘correct’, and consequently all answers to the paraphrase decision task for the target construction should technically be ‘YES’. For the filler sentences, the paraphrase decision task tested whether the participant had correctly understood the content of the sentence, and was designed to draw the participants’ attention away from the ambiguity of the target sentences and focus their attention on normal comprehension. The task took about 35–40 minutes to complete.

2 Data analysis

Data from the self-paced reading and paraphrase decision task were analysed. The dependent variables for the self-paced reading task were the reading times at the two segments comprising the complex NP modified by the RC, and the following verb: NP2, NP1, and V. The complex NP is the region where the global ambiguity of the RC construction becomes detectable. Although the reader must reach the higher noun, NP1, in order to become fully aware of the ambiguity and consider both NPs as candidates for the modifying RC, the possessive marker -uy on the lower noun, NP2, serves as a cue that another NP will follow. Therefore, readers with sufficient cognitive resources are predicted to be more sensitive to these cues, and be able to pick up on the potential ambiguity of the construction as early as NP2. The dependent variables for the paraphrase decision task were response time (RT), i.e. how long the participant took to decide whether a paraphrase was correct or not, along with RC attachment preference (see below). We predicted that higher WMC will lead to increased sensitivity to the global ambiguity, resulting in both longer reading times at the critical region and paraphrase decision RTs.

The reading time data were analysed using linear mixed effects models (Baayen et al., 2008), which allow for simultaneous consideration of all factors in structuring experimental data. The factors included in our analysis were RC modifying position, referential load, and individual WMC. For all models, random effect structure was fitted using likelihood ratio tests, while the fixed effect structure was fitted using a stepwise model selection procedure, with only predictors and interactions that were significant (p < .05) retained in the model and reported below. All final models had random intercepts and slopes for participants and items. The p-values for fixed effects were obtained using Markov-chain Monte Carlo sampling of random-intercept-only models, as removal of random slopes did not change the significance or pattern of any results. Paraphrase decision task RTs were also analysed according to the procedures above.

3 Results

Three measures are reported: reading times, paraphrase decision RTs, and RC attachment preference. Comprehension accuracy for the filler sentences was used as a cutoff measure to exclude data from participants who did not properly engage in the task. One participant with accuracy below 85% in both Experiment 1 and 2 and one participant below 85% accuracy in Experiment 2 were excluded from the remaining analyses, leaving 32 participants.

a Reading time data

Reading times shorter than 80 ms and greater than 3 SD from the mean for each condition were eliminated, resulting in a loss of less than 4% of the data. All reading times were log-transformed and centered, i.e. mean-centered around 0 to remove colinearity and allow the polarity of the betas to be easily interpreted (Baayen et al., 2008). Subject and Item were included as random variables. Reading times for the three critical segments NP2, NP1, and V are presented in Table 2. The results of the linear mixed model for each reading time variable are presented in Table 3.

Table 2.

Mean (SD) reading times at NP2, NP1, and V in Korean in milliseconds.

Conditions	NP 2	NP1	V	RT
SRC, H	542.11 (246.00)	693.11 (444.90)	737.62 (497.29)	4341.29 (1107.99)
SRC, M	464.43 (139.68)	658.19 (395.19)	720.30 (402.24)	4151.99 (1050.09)
ORC, H	587.76 (317.60)	637.91 (356.65)	764.67 (510.67)	5020.16 (1184.54)
ORC, M	532.93 (330.32)	671.65 (416.92)	806.73 (563.91)	4820.22 (1087.47)

Notes. RT = response time for paraphrase task; SRC = subject-modifying RC (relative clause); ORC = object-modifying RC; H = heavy referential load; M = minimal referential load.

Table 3.

Fixed effects in the linear mixed model of reading times at NP2, NP1, and RTs in Korean.

Predictor	Coefficient	SE	T	P
NP2 (R² = .24 ):
Intercept	355.90	62.02	5.74	< .0005
WMC	0.13	2.56	4.51	< .0005
WMC × RC	0.07	47.52	2.63	< .01
NP1 (R² = .11):
Intercept	559.15	181.12	3.09	< .005
WMC	0.11	3.91	3.81	< .0005
V (R² = .18):
Intercept	557.81	102.41	5.45	< .0005
WMC	0.09	4.22	3.05	< .005
RT (R² = .18):
Intercept	3496.50	476.16	7.34	< .0005
RC	0.16	191.33	5.85	< .0005
WMC	0.06	10.27	2.09	< .05

Notes. RC = relative clause modifying position; RT = paraphrase task reaction times; WMC = working memory capacity.

At NP2, reading times showed an effect of WMC (t = 4.52, p < .0005), indicating an increase in reading times as WMC increased. In addition, reading times for NP2 showed a significant interaction (t = 2.63, p < .01) between WMC and RC modifying position. This interaction indicates that the reading times for the subject-modifying RC were affected to a greater extent by WMC, compared to the subject-modifying RC. A main effect of WMC was also found for segments NP1 (t = 3.81, p < .0005) and V (t = 3.05, p < .005), with longer reading times positively correlated with an increase in WMC. An analysis of reading times prior to the critical region did not reveal significant effects of WMC for any of the segments (ps > .62).

c Paraphrase decision reaction times

RTs shorter than 1000 ms and greater than 3 SD from the mean were eliminated, resulting in a loss of less than 4% of the data. All RTs were log-transformed and centered. Subject and Item were included as random variables. Mean RTs are presented in Table 2. The results of the linear mixed model for each RT variable are presented in Table 3.

Paraphrase decision RTs showed a main effect of WMC (t = 2.09, p < .05), indicating that the time required for paraphrase decision increased in proportion to WMC. In addition, RTs were longer for sentences with object-modifying RCs, resulting in a main effect of RC modifying position (t = 5.85, p < .0005). RTs were significantly longer when the paraphrase was in the HA version (t = 5.05, p < .0001), suggesting that participants were more reluctant to accept the HA interpretation. Paraphrase RTs for the filler items did not show significant effects for WMC (p = .24)

d Relative clause attachment preference

RC attachment preferences were coded according to the following coding scheme. For the HA paraphrases, attachment preference was coded as HA if the participant answered ‘YES’ to the paraphrase decision task, and LA if the participant answered ‘NO’. The opposite applied to the LA paraphrases, so that ‘YES’ answers were coded as LA, and ‘NO’ answers as HA. In the resulting coded data, HA was recorded as 1, and LA was recorded as 0, in line with notation procedures in previous research on RC attachment (Brysbaert and Mitchell, 1996; Cuetos and Mitchell, 1988; Konieczny and Hemforth, 2000; Pynte, 1998). Arcsine square root transformations were performed to eliminate spurious effects that may arise from performing ANOVAs on categorical data (Hogg and Craig, 1995), and the transformed data were submitted to a 2 × 2 repeated measures ANOVA with the two factors of RC modifying position and referential load as within-participants variables and list as a between-participants variable (Pollatsek and Well, 1995). Attachment preference data is reported in Table 4.

Table 4.

Relative clause (RC) attachment preference data in Korean by condition.

	High	Low
SRC	0.34	0.41
ORC	0.36	0.40

Notes. High = high referential load; Low = low referential load; SRC = Subject-modifying RC (relative clause); ORC = Object-modifying RC.

The analysis of the attachment preference data showed a significant effect of referential load (F₁(3, 31) = 8.27, p < .01; F₂(3, 39) = 3.82, p < .05), indicating a higher proportion of LA preference in the high referential load condition. The factor of RC modifying position, and the interaction between the two factors was not significant (ps > .41). The mean attachment preference across all participants and conditions was .38 (LA), which was significantly different from chance (p < .0001, t = 5.60).

4 Discussion

Experiment 1 investigated WMC effects on the processing of the globally ambiguous RC construction in Korean. In addition, the interaction of WMC effects with factors modulating processing difficulty, i.e. RC modifying position and referential load was also investigated. We will begin by discussing WMC effects on the RC processing of L1 ambiguity.

The analysis of reading times showed a significant effect of WMC at segments NP2, NP1, and V. Because Korean is a head-final language, the word order for the ambiguous RC construction is the opposite of English, so that the modifying RC precedes the complex NP, and the lower noun (NP2) precedes the higher noun (NP1): RC–NP2–NP1. Although the global ambiguity of this structure becomes clear at NP1, we predicted that readers with sufficient cognitive resources would show sensitivity to the cue provided by the possessive marker -uy attached to NP2 and be able to pick up on the global ambiguity of the RC construction as early as the lower noun, NP2.

Results showed that WMC significantly affected reading times in these two regions, NP2 and NP1, bearing out our predictions. WMC effects were also obtained for the region following the complex NP, the main verb V. Recall that the main verb of the sentence is a point of high processing load, where the integration of the main subject of the sentence, held in working memory, with the incoming verb must be performed. Therefore, the significant effects of WMC at these three segments indicate that an increase in WMC is correlated with longer reading times required to process the critical region where the global ambiguity of the RC structure becomes apparent, and the following region where the subject–verb integration must be performed. Furthermore, the factor of WMC did not significantly affect any of the segments elsewhere in the sentence, which suggests that the significant effects of WMC obtained for the critical regions were not due to a tendency for high capacity readers to spend more time reading. These results correctly bear out our predictions and those of the Capacity Constrained Parsing Model (MacDonald et al., 1992). We argue that the results of Experiment 1 not only support the predictions of the Capacity Constrained Parsing Model (MacDonald et al., 1992), but also extend these predictions to the processing of globally ambiguous structures.

This interpretation of the results is further supported by the paraphrase decision RT data, which are consistent with the results obtained in the reading time data, by showing a significant effect of WMC. To elaborate, the time required to decide whether a paraphrase was correct or not increased in proportion to WMC. We interpret these results to suggest that high WMC enabled readers to be sensitive to the global ambiguity of the target construction, so that they retained both interpretations in working memory, and therefore required additional time to choose between the two.

Note that in contrast to the temporarily ambiguous construction used in MacDonald et al. (1992), for which there is ultimately only one correct interpretation, there is no single correct interpretation for the globally ambiguous RC construction in the present study. Therefore, in no situation is the low capacity reader faced with a penalty, regardless of which interpretation is chosen. In the case of the main verb/reduced relative ambiguity used in MacDonald et al. (1992), the low capacity reader suffered a time disadvantage by being forced to make a reanalysis when the temporary ambiguity was disambiguated toward the less favored reading. In the present study, lower WMC was correlated with shorter reading times and paraphrase decision task RTs regardless of which interpretation was favored.

Taken together, the online reading time data, in parallel with the offline paraphrase decision task RT data, suggest that in both the interpretive and post-interpretive stages of processing, awareness of the global ambiguity of the RC construction increased in proportion to WMC. A lack of sufficient cognitive resources required to consider both HA and LA interpretations led to the selection of only one, more preferred, or frequently used interpretation. Not being burdened by the additional processing load of maintaining both possible interpretations in working memory was reflected by shorter reading times and paraphrase decision RTs.

The next point of discussion is how predicted processing difficulty interacted with WMC in the processing of the RC construction. Crucial to our research question, a significant interaction between RC modifying position and WMC was found at segment NP2. To elaborate, a larger effect of WMC at NP2 was found when the RC modified the object of the sentence. Recall that in Korean, the object-modifying RC construction is predicted to pose a heavier processing load, due to greater distance between subject and verb integration (Gibson 1998, 2000), compared to the subject-modifying RC. WMC effects were predicted to be more salient for sentences with higher processing difficulty, as more cognitive resources will be drained when retaining multiple representations while simultaneously processing a sentence with extreme processing load. The interaction found between RC modifying position and WMC at segment NP2 correctly bears out these predictions.

Finally, WMC did not affect offline RC attachment preference, which resulted in a significant LA preference of 0.38, though higher WMC was associated with longer paraphrase RTs for the HA paraphrase version, suggesting more reluctance to accept the HA interpretation. We propose that this LA preference, which is inconsistent with some previous studies of Korean RC attachment (Jun, 2003; Lee and Kwon, 2004), was due to the high degree of syntactic complexity and processing difficulty in the experimental materials used, which is likely to have swayed participants toward adopting a Recency strategy. The experimental materials used in our study used transitive verbs in the embedded RC, in contrast to previous studies on RC attachment.⁴ Furthermore, RC attachment preference was modulated by the factor of referential load, with a higher percentage of LA preference in the heavy referential load condition, supporting our claim that readers tend to resort to a Recency strategy when faced with heavy processing load. These results are also consistent with previous studies arguing that purported cross-linguistic differences in RC attachment preferences may be outweighed by other factors such as syntactic complexity or cognitive resources (Kim and Christianson, 2013; Swets et al., 2008; Traxler, 2007).

Experiment 2 investigated how the same participants in Experiment 1 process the ambiguous RC structure in their L2 (English), and how the pattern of WMC effects on L2 processing compare to those observed for L1 processing.

V Experiment 2

1 Method

a Participants

The same participants in Experiment 1 also participated in Experiment 2, so that each participant took part in two sessions: Korean and English. The order of the two sessions was counterbalanced across participants.

b Materials

The experimental items were English translations of the Korean sentences in Experiment 1. The same sentences were used in both sessions to reduce any bias in the data that may result from differences in the words used in the ambiguous RC construction (Gilboy et al., 1995). Two native English speakers reviewed the experimental material to ensure naturalness of the sentences. Recall that predictions regarding the relative processing difficulty of the target sentences differing in RC modifying position in English are contrastive with those for Korean. In English, the subject-modifying RC is predicted to be more difficult to process than the object-modifying RC (Gibson’s Dependency Locality Theory; Gibson, 1998, 2000).

c Procedure and data analysis

Experimental procedures and data analysis were identical to those for Experiment 1. As mentioned previously, at the end of the English session, participants were given a cloze test to measure English proficiency.⁵ Two native speakers of English graded the cloze tests, and the maximum score possible was 40. After the cloze test, participants filled out a language background questionnaire which asked for basic demographic information, self-rated proficiency in English reading, writing, listening and speaking, and AOA (age of arrival).

2 Results

The factors included in the final, best-fit model were RC modifying position, referential load, and WMC. None of the variables reflecting L2 proficiency, i.e. cloze test scores and self-rated fluency ratings, improved model fit when entered into the regression equation (likely due to the restricted range), and were removed from the final model. Three measures are reported: reading times, paraphrase decision RTs, and RC attachment preference. Two participants whose comprehension accuracy failed to reach 85% (including one participant already excluded from Experiment 1) were excluded from the data analyses.

a Reading time data

The head-initial property of English results in a word order for the ambiguous RC construction (NP1–of–NP2–RC) which is the opposite of Korean (RC–NP2–NP1). Consequently, the critical segments where participants were predicted to show sensitivity to the global ambiguity differed from the Korean stimuli. Reading times were measured at the segments comprising the embedded RC, i.e. the relative pronoun who, the RC embedded verb RCV, and the RC embedded object RCO. These critical segments are the regions where participants with a sufficiently large pool of cognitive resources were predicted to show sensitivity to the global ambiguity of the RC construction. Reading times shorter than 80 ms and greater than 3 SD from the mean were eliminated, resulting in a loss of less than 5% of the data.

Mean reading times for the three critical segments who, RCV, and RCO are presented in Table 5. The results of the linear mixed model for each reading time variable are presented in Table 6.

Table 5.

Mean (standard deviation) reading times at ‘who’, ‘RCV’, and ‘RCO’ in English in milliseconds.

Conditions	who	RCV	RCO	RT
SRC, H	548.71 (324.33)	664.51 (500.21)	1301.02 (868.79)	5178.52 (1139.51)
SRC, M	638.67 (583.77)	578.18 (175.39)	594.95 (405.87)	4998.39 (898.52)
ORC, H	624.57 (373.31)	604.15 (402.04)	1009.98 (300.01)	5517.33(894.21)
ORC, M	566.95 (263.00)	545.54 (238.52)	484.15 (380.12)	5439.85 (925.10)

Notes. RT = response time for paraphrase task; RCV = embedded verb in modifying relative clause (RC); RCO = embedded object in modifying RC; SRC = subject-modifying RC; ORC = object-modifying RC; H = heavy referential load; M = minimal referential load.

Table 6.

Fixed effects in the linear mixed model of reading times at ‘who’, ‘RCV’, ‘RCO’, and RTs in English.

Linear Mixed Model of Reading Times
Predictor	Coefficient	SE	t	P
who (R² = .07):
Intercept	272.80	140.11	1.95	< .05
WMC	0.07	4.20	2.34	< .05
RCV (R² = .07):
Intercept	99.27	217.96	0.46	.65
WMC	0.07	6.54	2.48	< .05
RCO (R² = .32):
Intercept	1513.20	198.59	7.62	< .0005
WMC	0.09	4.75	2.09	< .0005
Ref	− 0.29	56.43	−10.88	< .0005
RC	− 0.09	56.43	− 3.54	< .0005
WMC × RC	− 0.15	2.32	− 3.74	<.0005
RT (R² = .13):
Intercept	6734.61	413.30	17.24	< .0005
WMC	0.92	30.95	2.34	< .05
Ref	− 0.06	260.21	− 2.25	< .05

Notes. WMC = working memory capacity; RC = relative clause modifying position; Ref = referential load; RT = paraphrase task response times; RCV = embedded verb in modifying RC; RCO = embedded object in modifying RC.

A significant effect of WMC was found for reading times at the relative pronoun who (t = 2.34, p < .05), indicating that an increase in reading times was associated with an increase in WMC. Main effects of WMC were also observed for segments, RCV (t = 2.48, p < .05), and RCO (t = 2.09, p < .0005). Reading times were longer at segment RCO when it was a full NP, as opposed to when it was a 1st/2nd person pronoun, resulting in a significant effect of referential load (t = −10.88, p < .0005). This effect is likely due in part to length and frequency differences between the full NPs and pronouns, however, and will not be discussed further. Reading times were also significantly longer at segment RCO when the RC modified the matrix subject, resulting in a main effect of RC modifying position (t = −3.54, p < .0005). In addition, a significant interaction of WMC and RC modifying position was found for this region (t = −3.74, p < .0005). This interaction indicates that WMC effects increased when the RC modified the subject, which was predicted to be higher in terms of processing difficulty than the object-modifying RC (Gibson, 1998, 2000). WMC was not a significant factor in the model for reading times in any of the segments preceding the critical region (ps > .37).

b Paraphrase decision reaction times

RTs shorter than 1000 ms and greater than 3 SD from the mean were eliminated, resulting in a loss of less than 5% of the data. The remaining data were analysed using the same procedures as Experiment 1.⁶ Mean RTs for the experimental sentences are presented in Table 5. Results of the linear mixed model for each RT variable are presented in Table 6. Paraphrase decision RTs showed a main effect of WMC (t = 2.34, p < .05), indicating longer decision times with an increase in WMC. In addition, RTs were longer for sentences in which the embedded object in the RC was a full NP, resulting in a main effect of referential load (t = −2.25, p < .05). Paraphrase RTs for filler items did not results in any significant effects of WMC (p = .36).

c Relative clause attachment preference

Attachment preferences were coded and analysed as in Experiment 1. The mean attachment preference across all participants and conditions was .54 (weak HA), and was not significantly different from chance (p = .26, t = 1.14). The analysis of the attachment preference data did not show any significant main effects or interactions (ps > .43). Mean attachment preferences by condition are presented in Table 7.

Table 7.

RC attachment preference data in English by condition.

	High	Low
SRC	0.51	0.57
ORC	0.55	0.53

Notes. High = high referential load; Low = low referential load; SRC = Subject-modifying relative clause (RC); ORC = Object-modifying RC.

3 Discussion

In Experiment 2, we investigated WMC effects on the processing of the globally ambiguous RC construction in English and its interactions with predicted processing difficulty. We also examined whether the WMC effects on ambiguity processing we found in Experiment 1 carried over to L2 processing strategies. We will begin by discussing WMC effects on English RC processing.

In English, the word order for the ambiguous RC construction is NP1–NP2–RC, i.e. the servant of the actress who was on the balcony. Therefore, the critical region where the global ambiguity of the structure becomes apparent is the modifying RC, i.e. segments who, RCV, and RCO. Results showed a significant positive-going relationship between WMC and reading times at these three segments, indicating that higher WMC resulted in longer reading times, suggesting sensitivity to the ambiguity of the construction. As in Experiment 1, the factor of WMC did not significantly affect any of the segments elsewhere in the sentence, which suggests that the significant WMC effects obtained for the critical regions were not due to a tendency for high capacity readers to spend more time reading.

The reading time patterns found in Experiment 2 are strikingly similar to the results of Experiment 1, suggesting that the disambiguation strategies employed by the participants were not qualitatively different between processing in their L1 and L2, but rather qualitatively different across readers of different WMCs. These results provide evidence in support of the Capacity Constrained Parsing Model (1992), and also show that the model’s predictions can be extended to highly advanced L2 learners, in addition to native speakers. A sufficiently large pool of working memory increases ability to retain both possible interpretations of a globally ambiguous structure in working memory, even when processing a non-native language. These results are consistent with previous studies showing that high capacity L2 learners employ qualitatively different processing strategies from low capacity L2 learners during L2 processing (Dussias and Piñar, 2010; Havik et al., 2009).

The main effect of RC modifying position was also consistent with our predictions, with longer reading times at the end of the modifying RC, i.e. segment RCO, for the subject-modifying RC, which was predicted to be higher in terms of processing difficulty compared to the object-modifying RC; see (4a, 4b).

Crucial to our research question, a significant interaction was found between RC modifying position and WMC at the embedded object of the RC. This segment (RCO) marks the end of the modifying RC and is a point of high processing load, where the reader integrates the modifying RC with one of the two attachment sites. Furthermore, the interaction was in the opposite direction of the Korean data, with a larger effect of WMC in the subject-modifying RC condition. Therefore, the interaction showing greater WMC effects in the subject-modifying RC is consistent with our predictions that sentences with heavier processing load will further strain the cognitive resources of high capacity readers, as they are already bearing the additional load of retaining two possible interpretations. In sum, effects of WMC and processing difficulty and the interaction between the two factors found in the reading time data for Experiment 2 are consistent with the pattern of results found in Experiment 1. Taken together, they suggest that participants were employing similar strategies across both languages.

The analysis of the RT data for the paraphrase decision task also revealed a similar pattern between L1 and L2 processing, with longer paraphrase decision RTs correlated with an increase in WMC. Consequently, results suggest that participants are consistent in the processing strategies they employ in the interpretive and post-interpretive stages of processing across their L1 and L2, and the particular processing strategies employed are affected by WMC. Only advanced L2 learners with a sufficiently large pool of WMC were sensitive to the potential ambiguity of the RC construction and capable of retaining both interpretations in working memory, whether processing in the L1 or the L2.

A minor difference in the paraphrase RT data that contrasted with the Korean data was that while effects of referential load observed in the reading times carried over to the paraphrase RT data, effects of RC modifying position did not. We suggest that even for highly advanced L2 learners, the additional cognitive load associated with processing a non-native language, in addition to the burden of retaining two interpretations simultaneously, may have resulted in a ceiling effect for sensitivity to processing difficulty in the post-interpretive stage. Consequently, while larger effects of WMC were found the subject-modifying RC during online processing, RC modifying position did not affect difficulty in deciding whether a particular interpretation was correct or not.

The final point of discussion concerns participants’ final RC attachment preference, which did not show a significant HA or LA preference. These results are not surprising when considering various previous studies which have reported no fixed attachment preference for L2 learners when processing the ambiguous RC construction in the L2 (Dussias, 2003; Papadopoulou and Clahsen, 2003).

To summarize, our predictions regarding WMC effects on L2 ambiguous RC processing were borne out in the data. WMC effects were observed at the region where the global ambiguity of the RC construction became detectable, with an increase in WMC leading to longer reading times. These data were supported by longer RTs for higher WMC for the paraphrase decision task. These results support the predictions of the Capacity Constrained Parsing Model (MacDonald et al., 1992), suggesting that a sufficiently large pool of WMC enabled sensitivity to the potential global ambiguity and ability to retain both interpretations in working memory. They are also strikingly similar to the pattern of results obtained for L1 (Korean) processing in Experiment 1.

VI General discussion and conclusions

The present study investigated WMC effects on the processing of the ambiguous RC construction by Korean L2 learners of English, in both their L1 and L2. The research questions investigated were:

Does WMC affect the processing strategies chosen for the globally ambiguous RC construction? If so, are these WMC effects consistent with the predictions of the Capacity Constrained Parsing Model?

Do highly proficient L2 learners of English exhibit similar effects of WMC during both L1 and L2 processing?

Does hypothesized processing difficulty affect processing strategies?

In order to answer the questions above, we examined the interpretive and post-interpretive stages of sentence processing by means of a combined self-paced reading and paraphrase decision task.

Results showed that WMC did affect the processing strategies chosen for the globally ambiguous construction. An increase in WMC led to inflated reading times for the critical region where the global ambiguity of the RC construction becomes detectable, consistent with the predictions of the Capacity Constrained Parsing Model (MacDonald et al., 1992). Furthermore, the participants in this study showed strikingly similar patterns of WMC across L1 and L2 processing, with the slow-down in reading times at the critical region carrying over to the post-interpretive stage of processing, as longer paraphrase decision times. These results extend and strengthen the claims of the Capacity Constrained Parsing Model by showing that its predictions extend to the processing of globally ambiguous structures, and highly advanced L2 learners.

The final research question concerned the interaction of WMC and hypothesized processing difficulty. Korean and English are good test languages to examine this question, as they contrast in head-directionality, which leads to different predictions regarding processing difficulty according to RC modifying position (Gibson, 1998, 2000).

Reading times for the critical region in both Experiments 1 and 2 showed a significant interaction between RC modifying position and WMC, with stronger WMC effects when processing difficulty is hypothesized to be higher, i.e. object-modifying RC for Korean, and subject-modifying RC for English. These results are consistent with our predictions that the difference in reading times due to WMC would be greater when processing sentences with high processing difficulty, as the strain on cognitive resources should add to the already heavy load of retaining two possible interpretations in working memory for participants with higher capacity. In sum, the participants in the present study showed a strikingly similar pattern of WMC effects on the online processing of global ambiguity in both their L1 and L2.

The final point of discussion concerns RC attachment preference. As noted above, purported cross-linguistic differences in RC attachment preferences have been the topic of extensive previous studies (Cuetos and Mitchell, 1988; see also Brysbaert and Mitchell, 1996; Jun, 2003; Kamide and Mitchell, 1999; Konieczny and Hemforth, 2000; Papadopoulou and Clahsen, 2003; Pynte, 1998). However, more recent studies have suggested that factors such as sentence complexity and WMC may affect RC attachment preferences substantially (Kim and Christianson, 2013; Swets et al., 2007; Traxler, 2007). Kim and Christianson (2013) reported a LA preference in Korean for sentences with high syntactic complexity, in contrast to the purported HA preference in Korean reported in previous studies (Jun, 2003; Lee and Kweon, 2004). Swets et al. (2007) reported a range of attachment preferences mediated by WMC in English larger than previously reported differences in attachment preferences between English, a LA language, and other HA languages. The results of the present study, showing a significant LA preference in Korean and an absence of a significant attachment preference in English, support the claims of these recent studies, suggesting that while cross-linguistic differences in RC attachment may exist, factors such as sentence complexity and WMC may override purported cross-linguistic differences.

In conclusion, the present study revealed that a sufficiently large pool of cognitive resources increases the possibility that highly proficient L2 learners will be sensitive to the global ambiguity of the RC construction in Korean and English. The pattern of WMC effects in the participants’ L2 English processing was strikingly similar to the WMC effects obtained for their L1 processing.

The present study contributes to the field of research in language processing by extending the findings of previous studies on WMC effects on monolingual processing to L2 processing. The experimental methodology used in the present study allowed us to obtain data from both the interpretive and post-interpretive stages of sentence processing, and consequently provided us with a clearer picture of the similarities and differences between L1 and L2 processing. Furthermore, in contrast with the majority of previous studies comparing the processing of L2 learners with a native speaker control group, we investigated how the processing strategies for the same group of highly proficient L2 learners differed in the processing of the same ambiguous RC construction in their first and second languages. This type of direct comparison provided us with deeper insight regarding cognitive effects in L2 processing, and also shed light on the strikingly similar pattern of processing strategies that readers adopt across L1 and L2 processing.

Footnotes

Acknowledgements

This work was supported by Hankuk University of Foreign Studies Research Fund.

Declaration of Conflicting Interest

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) received no financial support for the research, authorship, and/or publication of this article.

Notes

References

Baayen

Davidson

Bates

(2008) Mixed-effects modeling with crossed random effects for subjects and items. Journal of Memory and Language 59: 390–412.

Brysbaert

Mitchell

(1996) Modifier attachment in sentence parsing: Evidence from Dutch. The Quarterly Journal of Experimental Psychology 49A: 664–95.

Caplan

Waters

(1999) Verbal working memory and sentence comprehension. Behavioral and Brain Sciences 22: 77–126.

Christianson

Hollingworth

Halliwell

Ferreira

(2001) Thematic-roles assigned along the garden path linger. Cognitive Psychology 42: 368–407.

Clahsen

Felser

(2006) Grammatical processing in language learners. Applied Psycholinguistics 27: 3–42.

Conway

ARA

Kane

Bunting

Hambrick

Wilhelm

Engle

(2005) Working memory span tasks: A methodological review and user’s guide. Psychonomic Bulletin and Review 12: 769–86.

Cuetos

Mitchell

(1988) Cross-linguistic differences in parsing: Restrictions on the late-closure strategy in Spanish. Cognition 30: 73–105.

Cuetos

Mitchell

Corely

MMB

(1996) Parsing in different languages. In: Carreiras

Garcia-Albea

Sebastian-Galles

(eds) Language processing in Spanish. Hillsdale, NJ: Lawrence Erlbaum, 145–87.

Dussias

(2003) Syntactic ambiguity resolution in L2 learners: some effects of bilinguality on L1 and L2 processing strategies. Studies in Second Language Acquisition 25: 529–57.

10.

Dussias

Piñar

(2010) Effects of reading span and plausibility in the reanalysis of wh-gaps by Chinese–English. Second Language Research 21: 121–51.

11.

Felser

Marinis

Clahsen

(2003) Children’s processing of ambiguous sentences: A study of relative clause attachment. Language Acquisition: A Journal of Developmental Linguistics 11: 127–63.

12.

Felser

Roberts

Marinis

Gross

(2003) The processing of ambiguous sentences by first and second language learners of English. Applied Psycholinguistics 24: 453–89.

13.

Ferreira

Bailey

KGD

Ferraro

(2002) Good-enough representations in language comprehension. Current Directions in Psychological Science 11: 11–15.

14.

Frazier

(1979) On comprehending sentences: Syntactic parsing strategies. Unpublished PhD dissertation, University of Connecticut, CT, USA.

15.

Frazier

Clifton

Jr (1996) Construal. Cambridge, MA: MIT Press.

16.

Gibson

(1998) Linguistic complexity: Locality of syntactic dependency. Cognition 68: 1–76.

17.

Gibson

(2000) The dependency locality theory: A distance-based theory of linguistic complexity. In: Miyashita

Marantz

O’Neil

(eds) Image, language, brain. Cambridge, MA: MIT Press, 95–126.

18.

Gilboy

Sopena

Clifton

Frazier

(1995) Argument structure and association preferences in Spanish and English complex NPs. Cognition 54: 131–67.

19.

Grosjean

(2001) The bilingual’s language modes. In: Nicol

(ed.) One mind, two languages: Bilingual language processing. Malden, MA, 1–22.

20.

Havik

Roberts

van Hout

Schreuder

Haverkort

(2009) Processing subject–object ambiguities in L2 Dutch: A self-paced reading study with German L2 learners of Dutch. Language Learning 59: 73–112.

21.

Hogg

Craig

(2005) Introduction to mathematical statistics. 5th edition. New York: Macmillan.

22.

Juffs

(2004) Representation, processing and working memory in a second language. Transactions of the Philological Society 102: 199–225.

23.

Juffs

(2005) The influence of first language on the processing of wh-movement in English as a second language. Second Language Research 21: 121–51.

24.

Jun

S-A

(2003) Prosodic phrasing and attachment preferences. Journal of Psycholinguistic Research 32: 219–49.

25.

Just

Carpenter

(1992) A capacity theory of comprehension: Individual differences in working memory. Psychological Review 99: 122–49.

26.

Kamide

Mitchell

(1997) Relative clause attachment: nondeterminism in Japanese parsing. Journal of Psycholinguistic Research 26: 247–54.

27.

Kim

(2008) Working memory effects on bilingual sentence processing. Unpublished doctoral dissertation. Department of Linguistics, University of Illinois at Urbana-Champaign, IL, USA.

28.

Kim

Christianson

(2013) Sentence complexity and working memory effects in ambiguity resolution. Journal of Psycholinguistics Research 42: 393–411.

29.

Konieczny

Hemforth

(2000) Modifier attachment in German: Relative clauses and prepositional phrases. In: Kennedy

Pynte

(eds) Reading as a Perceptual Process. Amsterdam: Elsevier, 517–28.

30.

Lee

Kweon

(2004) A sentence processing study of relative clause in Korean with two attachment sites. Discourse and Cognition 11: 126–41.

31.

MacDonald

Just

Carpenter

(1992) Working memory constraints on the processing of syntactic ambiguity. Cognitive Psychology 24: 56–98.

32.

Mendelsohn

Pearlmutter

(1999). Individual differences in relative clause attachment ambiguities. Poster presented at the 12th Annual CUNY Conference on Human Sentence Processing, City University of New York, USA.

33.

Omaki

(2005) Working memory and relative clause attachment in first and second language processing. Unpublished MA thesis, Department of Second Language Studies, University of Hawaii, HI, USA.

34.

Papadopoulou

Clahsen

(2003) Parsing strategies in L1 and L2 sentence processing: A study of relative clause attachment in Greek. Studies in Second Language Acquisition 25: 501–28.

35.

Pollatsek

Well

(1995) On the use of counterbalanced designs in cognitive research: A suggestion for a better and more powerful analysis. Journal of Experimental Psychology, Learning, Memory and Cognition 21: 785–94.

36.

Pynte

(1998) The time course of attachment decisions: Evidence from French. In: Hillert

(ed.) Sentence processing: A cross-linguistic perspective. San Diego: Academic Press, 227–45.

37.

Sanford

AJS

Sanford

Filik

Molle

(2005) Depth of lexical-semantic processing and sentential load. Journal of Memory and Language 53: 378–96.

38.

Swets

Desmet

Clifton

Jr Ferreira

(2008) Underspecification of syntactic ambiguities: Evidence from self-paced reading. Memory and Cognition 36: 201–16.

39.

Swets

Desmet

Hambrick

Ferreira

(2007) The role of working memory in syntactic ambiguity resolution: A psychometric approach. Journal of Experimental Psychology 136: 64–81.

40.

Traxler

(2007) Working memory contributions to relative clause attachment processing: A hierarchical linear modeling analysis. Memory and Cognition 35: 1107–21.

41.

Warren

Gibson

(2002) The influence of referential processing on sentence complexity. Cognition 85: 79–112.

42.

Williams

(2006) Incremental interpretation in second language sentence processing. Bilingualism: Language and Cognition 9: 71–88.

43.

Zhou

Christianson

(2015) I ‘hear’ what you’re ‘saying’: Auditory perceptual simulation during silent reading. Quarterly Journal of Experimental Psychology. DOI: 10.1080/17470218.2015.1018282.