Methods

Participants

One hundred and twenty-two female Arab learners ranging in age between 18 and 25 (M = 20.91, SD = 1.82) took part in this experiment. They were all BA students at a university in Saudi Arabia where English is the main medium of instruction and were recruited from various levels of study in order to ensure variability in English proficiency levels. Nearly half of them (N = 60) were students at the preparatory-year English language program while the other half (N = 62) were senior students who were studying in various majors after meeting their English-language requirement. All participants started learning English in schools at an average age of 8.30 years (SD = 4.41), and none had spent more than 3 months in an English-speaking country.

The two groups of participants were given the 1,000 (1k) and 2,000 (2k) levels of the Updated VLT, Version A (Webb, Sasao, & Ballance, Reference Webb, Sasao and Ballance2017) as a rough estimate of proficiency. We measured the 1k and 2k levels only due to time limitations. Their average score out of 30 was 27.74 at the 1k level (SD = 2.57) and 22.77 at the 2k level (SD = 5.85). The total mean score across both levels was 50.51 (SD = 7.92). The Kuder–Richardson (Formula 21) reliability coefficient (see Bachman, Reference Bachman2004) for the VLT was found acceptable (0.89).Footnote ² The distribution of total VLT scores is presented in Figure 1. It should be noted that these were centred prior to the analysis.

Figure 1.

Distribution of VLT scores achieved by participants (Experiment 1).

Following Sonbul (Reference Sonbul2015), we included VLT scores as a continuous predictor variable in the mixed-logit model (see the Analysis section below). This step was taken to ensure a fine-tuned depiction of proficiency effects on collocation recognition.

Measures

We employed the most widely used collocation recognition measure, that is, MC. Three distractors were developed for each target item that met the following criteria:

1. 1. All distractors belonged to the most frequent 2,000 word families in English (Nation, Reference Nation2012).

2. 2. Following Nguyen and Webb (Reference Nguyen and Webb2017), the distractor combined with the noun node (in a window of -1 for adjective and -2 for verbs) had a frequency between 0 and 10 in the COCA. When the pair occurred in the COCA, the MI score was either negative or 1 maximum signaling no association. The only exception to this is the distractor *push time which has a frequency of 11 in the COCA but with a negative MI score (–3.20).

We then ensured that all distractors were semantically plausible. In a norming task, the initial version of the test was presented to 5 native speakers of English who hold MA or PhD degrees either in Linguistics or in TESOL. They were instructed to make any changes they deemed necessary in relation to the most suitable determiner in VN collocations. In addition to this, we asked them to mark any distractor that was particularly semantically marked (implausible). Any distractor that was marked as implausible by three native speakers was replaced by another one. Here are examples for the items limited ability and catch cold:

• _____________ ability

  A. narrow B. gentle C. limited D. big

• ______________ a cold

  A. ignore B. happen C. catch D. attract

Items were divided into two sections, AN and VN. The final set of 40 multiple-choice items was piloted with 5 native speakers of English (different from those who took the norming task above). They all achieved a perfect score (100%).

Scoring the test was straightforward. A response was scored 1 when it was correct and 0 if it was incorrect or missing. The reliability of the test was calculated based on the scores achieved by the 122 EFL participants and was found high (Cronbach α = 0.92).

Procedures

The participants were recruited during their normal classes.Footnote ³ They took the MC test (around 20 min) followed by the 1k and 2k levels of the Updated VLT (around 15 min), and finally the language background questionnaire.

Analysis

The analysis was done using mixed-logit models for binary data with the glmer function (lme4 package) on R version 3.5.0. (R Core Team, 2018). It was conducted in both the forward and backward methods to check consistency, and the final models were identical in both directions.Footnote ⁴

The analysis included participants and items as random variables, 0/1 MC scores as the dependent measure, and the following fixed factors: item type (congruent vs. incongruent), VLT scores as a continuous variable representing estimated proficiency, and the interaction between the two main factors. We also tested for the potential effect of item-related variables on test scores. These included word class (AN vs. VN), word 1 frequency, word 2 frequency, pair frequency, and pair length. All frequencies were taken from the COCA and were log transformed to reduce skewness in data. Odds ratios (Exp[β]) transformed from log odds were used as effect size measures in the model (see Puimège & Peters, Reference Puimège and Peters2019).Footnote ⁵

Collinearity between item-related variables was checked prior to the analysis, and residuals were computed for pair length and log word 2 frequency. Both residuals correlated well with the original variables (pair length r = .92, p < .001, and log word 2 frequency r = .71, p < .001). Table 1 summarizes continuous variables.

Table 1.

Summary of continuous variables (MC test)

Note: The second column shows the range of the variables. The adjusted range after transformation and/or centring is presented in parentheses. Standard deviations and medians refer to the predictor values in the models. All variables are centred, and their means are zero.

Results and discussion for Experiment 1

Table 2 below presents the number and percentage of correct/incorrect responses in the MC test. It can be clearly seen that, overall, our Saudi EFL learners scored higher for congruent than incongruent collocations (with a 11.23 and 12.13 percentile point difference for AN and VN collocations, respectively). In contrast, the advantage for AN over VN collocations is not large with a 1.31 percentile-point difference only for congruent and 2.21 percentile-point difference for incongruent collocations.

Table 2.

Responses and percentile scores in the MC test (N = 122)

^aMax score = 1,220 (N = 122 × K = 10). ^bMax score = 2,440 (N = 122 × K = 20).

Table 3 presents the best-fit model for variables predicting MC scores. Only two main predictors were found significant. Estimated L2 proficiency as reflected by VLT scores seems to have a large effect with 12% more chances of a correct answer as the VLT score increased by 1 point. The effect of congruency seems to be even bigger with 2 times higher odds of a correct answer for congruent than incongruent collocations. Finally, none of the item-related variables (including word class) contributed significantly to the model, neither did the interaction between VLT and item type.

Table 3.

Summary of the best-fit mixed logit model for MC scores (N = 4,880, log likelihood = –2,334.10)

The model has random intercepts for participants and items. *p < .05. **p < .001.Formula:

glmer (MCQScore ~ VLTScore + ItemType + [1|ParticipantCode] + [1|ItemCode], family = “binomial”)

Thus, to answer the research questions, congruency had an overall effect on untimed collocation recognition. Estimated L2 proficiency also had a sizable effect with higher collocation scores as VLT scores increased. Finally, lack of interaction between the two main predictors seems to suggest that the effect of L1 congruency is omnipresent across the board regardless of the estimated proficiency level.

Methods

Participants

One hundred and six female Arab learners of English took part in this experiment. None of them took part in Experiment 1, but they were sampled from BA students in the same university in Saudi at two levels: preparatory-year program (N = 55) and seniors (N = 51). They ranged in age between 17 and 25 (M = 20.10, SD = 1.90). None of them had lived in an English-speaking country for more than 3 months, and they started learning English at school age (M = 6.74 years, SD = 3.66).

The average scores obtained by the two groups on the 1k and 2k levels of the Updated VLT were 28.65 (SD = 2.01) and 24.58 (SD = 4.82), respectively. The total VLT average score achieved across the two levels was 53.24 (SD = 6.52). The Kuder–Richardson (Formula 21) reliability coefficient for the total VLT score was found acceptable (0.87). Figure 2 presents the distribution of scores. Like Experiment 1, the VLT scores were centred and included in the mixed-effects model as an estimated continuous measure of proficiency.

Figure 2.

Distribution of VLT scores achieved by participants (Experiment 2).

In addition to the two groups of nonnatives, a group of 37 native speakers of English (female = 21, male= 16) were also included in this experiment to set a baseline for nativelike timed recognition. They ranged in age between 24 and 68 (M = 45.51, SD = 10.73).

Measures

The measure employed in this experiment is a timed phrase acceptability judgment task. As evident from the review of previous research, this is the most common type of task used in collocation research to gauge timed/automatic recognition. The task was designed and administered using E-Prime 3.00 (Psychology Software Tools Inc.; https://www.pstnet.com). Participants had to decide whether each presented pair is commonly used in English using YES/NO-labelled buttons on a gamepad (IHK LLC, Shenzhen China; https://h3-grup.com). Research suggests that gamepads are more sensitive in recording RT than keyboards (see Segalowitz & Graves, Reference Segalowitz and Graves1990; Shimizu, Reference Shimizu2002).

The 40 control and 40 target items were counterbalanced into two lists (Task 1 and Task 2) so that no control/target appeared in the same list more than once. Each list included 10 AN collocations, 10 VN collocations, 10 AN control pairs, and 10 VN control pairs. In addition to these, 42 filler collocate and noncollocate pairs were included in the list (21 AN and 21 VN) to make the task as natural as possible. None of these fillers included any component word from the target collocations or control pairs. The list was divided into two parts with a break in the middle based on word class. In Task 1, the AN section preceded the one including VN pairs while in Task 2 the order was reversed. This was intended to control for the effect of word class presentation order on RT and accuracy scores.

There were 10 practice trials in the beginning to familiarize participants with the task. Thus, in each task, participants were presented with 92 pairs in total (10 practice, 20 targets, 20 controls, and 42 fillers). Half of the participants completed Task 1 (natives N = 17, nonnatives: N = 55) and the other half completed Task 2 (natives: N = 20, nonnatives: N = 51). Each trial started with a fixation screen of 12 asterisks, after 1 s, it was replaced by a blank screen for 50 ms, finally the pair appeared and stayed on the screen for 5000 ms or until a response was detected. The order of trials was randomized across subjects, and E-Prime recorded accuracy and response latency for each trial.

Procedures

The E-Prime task was conducted in a quiet computer lab and lasted 15 min. The instructions were adapted from Wolter and Gyllstad (Reference Wolter and Gyllstad2013, p. 460) and read as follows:

After completing the acceptability judgment task, the nonnatives were administered the two levels of the Updated VLT and completed them in 15 min. Finally, all participants completed a language background questionnaire.

Analysis

The analysis was conducted on R version 3.5.0. (R Core Team, 2018). Accuracy scores (0/1) for natives and nonnatives were analyzed separately in two mixed-logit models for binary data (glmer function in the lme4 package). Then, we analyzed predictors of our main dependent variable in this experiment, RT data, through two linear mixed-effects (LME) models (lmer function in the lme4 package), one for natives and the other for nonnatives. The analysis for natives was conducted to set a baseline against which the nonnative recognition can be compared.

Initially, timed-out responses were removed, and the data was checked for outliers. Responses that were faster than 500 ms were excluded from further analysis. As for the highest cutoff point, 2000 was set for natives and 4000 was set for non-natives.Footnote ⁷ This resulted in excluding 3.58% and 1.36% of data points for natives and nonnatives, respectively.

In the mixed-logit analysis for accuracy, a response was coded either as 1 (correct) or as 0 (incorrect) based on whether it matched the expected response: 1 (Yes for a collocation and No for a control pair) and 0 (No for a collocation and Yes for a control pair). The models included accuracy binary scores as the dependent variable and participant/items as random variables. The two main fixed effects were then tested: item type (congruent, incongruent, or control) and estimated proficiency (VLT scores for nonnatives only). In addition to these, the models also included the interaction between the two main factors, trial number, and five item-related variables: word class (AN vs. VN), word 1 log frequency, word 2 log frequency, log pair frequency, and pair length. Effect size was estimated based on odds ratios (Exp[β]) transformed from log odds.

We then explored variables predicting RT in LME models. Only correct responses were included in this analysis with RT as the dependent measure (after log transformation), participants/items as random variables, and the same fixed effects tested above. Effect sizes for these models are represented by marginal and conditional R ² values. The former involves only fixed effects while the latter incorporates random effects as well (see Wolter & Yamashita, Reference Wolter and Yamashita2018, for an explanation).

Collinearity among item-related variables was checked prior to the analysis and residuals were calculated for word 1 frequency, word 2 frequency, pair frequency, and pair length. All residuals correlated well with the original variables (log word 1 frequency r = .85, p < .001; log word 2 frequency r =.74, p < .001; pair length r = .94, p < .001, log pair frequency r = .44, p < .001). Table 4 below presents a summary of continuous variables.

Table 4.

Summary of continuous variables (acceptability judgment task)

Note: The second column shows the range of the variables. The adjusted range after transformation and/or centring, is presented in parentheses. Standard deviations and medians refer to the predictor values in the models. All variables are centred, and their means are zero.

Results and discussion of Experiment 2

Table 5 below presents RTs and accuracy scores in the acceptability judgment task for both natives and nonnatives. In general, like earlier research, natives were faster and more accurate than nonnatives in making decisions (but see Gyllstad & Wolter, Reference Gyllstad and Wolter2016, for an exception).

Table 5.

Mean RT (standard error; SE) and accuracy scores (percentage) in the acceptability judgment task

RT data seems to suggest an evident processing lag for control pairs in comparison to target collocations with an average difference of around 215 ms and 400 ms for natives and nonnatives, respectively. Comparing congruent and incongruent collocations, in contrast, does not seem to manifest a clear trend in RT for either group. The same holds true for word class differences. However, it should be noted that the RTs presented in Table 5 represent raw values before controlling for estimated proficiency and for item-related variables.

Looking now at accuracy scores, it seems that natives processed both types of collocations similarly (with a range of 95% and 98% accurate scores) but were less accurate in making decisions about control pairs (around 80% correct responses). Conversely, nonnatives showed a clear gradual increase in the percentage of correct answers from control items with around 50%, to incongruent collocations with over 70%, and finally congruent collocations with above 86%.Footnote ⁸ These percentages do not seem to be affected by word class with fairly similar scores for AN and VN collocations.

Natives

Accuracy

Natives’ best-fit mixed-logit model for accuracy scores is presented in Table 6. Only item type was found significant with almost 11 times more chances of a correct decision for both congruent and incongruent collocations in comparison to control items. In order to explore the difference between congruent and incongruent collocations, we fit the same model after setting “incongruent” as the reference level. Results showed that, as expected, natives did not differ in their decision accuracy for both types of collocations (Wald Z = 0.02, p = .98, Exp[β] = 1.01).

Table 6.

Summary of the best-fit mixed logit model for natives’ accuracy scores (N = 1,421, log likelihood = –406.8)

The model has random intercepts for participants and items. *p < .001. Formula:

glmer (JudgmentAccuracy ~ ItemType + [1|ParticipantCode) + [1|ItemCode], family = “binomial”)

Response latency

Table 7 presents the best-fit model for variables predicting natives’ response latencies. The first significant effect is corpus-derived pair frequency with shorter RT (faster processing) as the frequency increased. Then, pair length was significant with longer RT (slower processing) as the length increased. Finally, item type was significant with faster processing for both types of collocations in comparison to noncollocate pairs. Redefining the reference level as “incongruent” showed no difference between natives’ processing of congruent and incongruent collocations (t value = 0.74, p = .46). This model explains almost 60% of variance in the data as suggested by the conditional R ² value.

Table 7.

Summary of the best-fit LME model for variables predicting natives’ RT data (N = 1,263, R²marginal = 0.20, R²conditional = 0.57)

The model has random intercepts for participants and items. *p < .05. **p < .001. Formula:

lmer(LogRT ~ ResidLogPairFrequency + ResidPairLength + ItemType + [1|ParticipantCode] + [1|ItemCode])

To summarize results from natives, the analysis showed a clear advantage in both decision accuracy and processing time for target items over control pairs but almost identical processing for both types of collocations. These findings confirm our previous norming rating results (see the Items section above), goes in line with previous research findings (e.g., Wolter & Gyllstad, Reference Wolter and Gyllstad2013), and sets a baseline for comparison with the main nonnative group.

Nonnatives

Accuracy

Table 8 below presents the best-fit model for variables predicting accuracy scores. VLT scores and item type contributed significantly to the model with no interaction between the two. The odds of a correct answer in the acceptability judgment task was 6% bigger as the VLT increased by one unit. As for congruency, both congruent and incongruent collocations led to more correct answers in comparison to the control reference level (almost 8 and three times bigger, respectively). Finally, a model where “incongruent” was the reference level revealed significantly more correct responses for congruent than incongruent collocations (Wald Z = 3.30, p < .001, Exp[β] = 2.55).

Table 8.

Summary of the best-fit mixed logit model for non-natives’ accuracy scores (N = 4,162, log likelihood = –2,220.2)

The model has random intercepts for participants and items. *p < .001. Formula:

glmer(JudgmentAccuracy ~ VLTScore + ItemType + [1|ParticipantCode] + [1|ItemCode], family = “binomial”)

Response latency

The best-fit LME model for nonnatives’ RT data is presented in Table 9 below. As indicated by the conditional R ² value of the model, it explains 47% of variance in the data. trial number, log pair frequency, and pair length had significant effects with longer RTs as the experiment went on (showing a fatigue effect for nonnatives) and as the length of the pair increased, but shorter RTs as the frequency increased. Word class and the frequency of individual words, however, did not add any significant value to the model.

Table 9.

Summary of the best-fit LME model for variables predicting non-natives’ RT data (N = 2,845, R²marginal = 0.19, R²conditional = 0.47)

The model has random intercepts for participants and items. *p < .05. **p < .01. ***p < .001. Formula:

lmer(LogRT ~ Trial + ResidLogPairFrequency + ResidPairLength + VLTScore*ItemType + [1|ParticipantCode] + [1|ItemCode]).

As for the main fixed factors, item type had a main effect with slower responses for control items in comparison to both congruent and incongruent collocations. Upon redefining the refence level as “incongruent,” we found that the two types of collocations did not differ in RT in the final model (t value = –0.46, p = .65). Finally, VLT scores did not show up as a significant main effect but rather significantly interacted with item type.

Figure 3 depicts the interaction between the two variables along with the effect size for each item type revealing quite interesting results. Looking at RTs to the three types of items (congruent, incongruent, and control) and how they changed as the VLT scores increased, one can note the following points. First, RTs for all items decreased as estimated proficiency increased. Second, nonnatives achieving low scores in the VLT seem to process congruent collocations normally with a big RT advantage over noncollocate pairs. However, incongruent pairs do not seem to have gained their collocation status yet as their processing is rather similar to baseline pairs. Third, while both congruent and baseline items seem to show a gradual decrease in processing time, incongruent collocations demonstrate sharp and substantial changes (with the biggest effect size). This led to more nativelike processing (see above) as estimated proficiency increased, that is, smaller differences between congruent and incongruent collocations and larger differences between them and control pairs. ^{Footnote 9} This finding is further supported by the statistical figures revealed upon redefining the reference level for item type to “incongruent.” In that model, the interaction with VLT scores was still significant (t = 2.60, p = .009). This stands in sharp contrast with the lack of variance between congruent and incongruent collocations (as a main effect) reported above. Thus, one might conclude that there was not an inherent difference between congruent and incongruent collocations. Rather, they only showed a difference in recognition time as estimated proficiency (VLT scores) decreased.

Figure 3.

Interaction between VLT score (centred) and item type for the nonnative RT data (Experiment 2).

Thus, unlike untimed recognition (see Experiment 1) where congruency and proficiency effects are omnipresent, timed processing manifests a proficiency-modulated advantage for congruent over incongruent collocations. The L1 effect seems to be strong for lower level EFL learners, but it fades away gradually as estimated proficiency increases approaching nativelike collocation processing.

Other determinants of L2 collocation recognition

Although our study focused on two main factors of L2 collocation recognition (i.e., congruency and proficiency), previous research has pointed out to several other important effects including word class, collocation frequency, and the frequency of individual component words. These were included in our mixed-effects models and showed interesting findings in both experiments.

On the untimed side of processing, in contrast with Nguyen and Webb (Reference Nguyen and Webb2017) where word and collocation frequency were significant determinants of MC test scores, Experiment 1 in the present study did not find such effects. This might be due to the fact that Nguyen and Webb experimentally manipulated the frequency of individual words and collocations (at three distinct levels) allowing such factors to show up. Our study did not specifically manipulate frequency either at the word or at the collocation level leading to lack of variation. As for word class (VN vs. AN), earlier corpus-based studies (Laufer & Waldman, Reference Laufer and Waldman2011; Nesselhauf, Reference Nesselhauf2003) reported higher difficulty in the production of VN than AN collocations. Our results, however, agree with those reported by Nguyen and Webb (Reference Nguyen and Webb2017), showing no significant difference between the two types of collocations. It seems that nonnatives process them similarly, at least at the recognition level.

Looking now at the acceptability judgment task, our results exhibited the collocation-frequency effects found in previous studies that gauged timed processing (e.g., Wolter & Gyllstad, Reference Wolter and Gyllstad2013; Wolter & Yamashita, Reference Wolter and Yamashita2018; see also Sonbul, Reference Sonbul2015). This finding supports usage-based models of language development, which claim that experience with the language is a major predictor of language processing and development (Bybee, Reference Bybee2006) both in the L1 and L2. However, our findings go against Wray’s (Reference Wray2005) claim that natives and nonnatives process formulaic language differently. Experiment 2 results seem to show that as estimated proficiency increases, nonnatives process collocations on par with natives (see Gyllstad & Wolter, Reference Gyllstad and Wolter2016, for similar findings). Along the same lines, Groom (Reference Groom, Barfield and Gyllstad2009) showed that, with more immersion in the L2 context, nonnatives seem to exhibit more nativelike collocation patterns in written production.

When it comes to word-level frequency effects, though, insignificant effects in the present study stand in contrast with Wolter and Yamashita (Reference Wolter and Yamashita2018). As reported above, this might be related to a ceiling effect as all component words in the collocations were highly frequent. Finally, as for word class effects on real-time processing, our results show that, like Wolter and Yamashita (Reference Wolter and Yamashita2015), nonnatives do not differ in their timed processing of AN and VN collocations. This result is interesting in our context as the word order in Arabic is the same for VN collocations but is the opposite for AN collocations. Despite this difference in word order, it seems that Saudi EFL learners in the present study processed them similarly. This seems be due to the long experience, an average of about 13 years (from 6.74 years to 20.10 years, see the Methods section in Experiment 2), they had with English. This might have allowed them to overcome this mismatch in word order between the two languages. It would be interesting to explore the effect of word class configuration with beginner Arab EFL learners.

Theoretical implications

Although the current study did not originally aim to address various models of bilingual language processing, it seems to have certain tentative implications. Previous research employing timed collocation measures (e.g., Wolter & Gyllstad, Reference Wolter and Gyllstad2013; Yamashita & Jiang, Reference Yamashita and Jiang2010, see the Background section) has attempted to extend theoretical models of bilingual word processing to collocations. Yamashita and Jiang (Reference Yamashita and Jiang2010), for example, took their findings as support of the revised hierarchical model (RHM; Kroll & Stewart, Reference Kroll and Stewart1994), which assumes a nonselective view of processing in the early stages of L2 learning whereby both languages are activated making it difficult to suppress one over the other (i.e., L2 is accessed through the L1). As L2 proficiency increases, a more direct, faster route to the L2 lexicon is possible.

The present study adds to Yamashita and Jiang’s (Reference Yamashita and Jiang2010) findings through (a) including EFL participants only, thus controlling for L2 context, and (b) treating proficiency as a continuum to allow for a more fine-grained depiction of its effect. Our results seem to lend support to the RHM as the nonnatives with low estimated levels of proficiency showed dependence on the L1, but the ones at higher levels seemed to follow a more direct route to lexical access exhibiting more nativelike processing (see Figure 3 above).

Thus, despite recent calls for leaving the RHM behind due to evidence that it fails to account for various aspects of L2 lexical processing (see, for example, Brysbaert & Duyck, Reference Brysbaert and Duyck2010), our results seem to suggest that predictions of the RHM still hold value. It should be noted, however, that results of the present study should be treated rather cautiously as proficiency was only estimated based on a short vocabulary test rather than a standardized measure. In addition, our results stop at the level of lexical access/processing. Models of bilingual lexical processing also make assumptions about knowledge representation (see Kroll & Tokowicz, Reference Kroll, Tokowicz, Kroll and De Groot2005 for an overview), but this was not the focus of the present study.

In contrast with the RHM, proponents of the age of acquisition/order of acquisition effects on bilingual lexical processing claim that the most important determinant of word recognition (both time and accuracy) is the order in which the words were learned (see, for example, Izura et al., Reference Izura, Pérez, Agallou, Wright, Marín, Stadthagen-González and Ellis2011). Wolter and Gyllstad’s (Reference Wolter and Gyllstad2013) and Wolter and Yamashita’s (Reference Wolter and Yamashita2018) findings are interpreted within these effects whereby congruent collocations, which are plausible in the L1, are learned first and will always show a processing advantage over incongruent collocations. However, our results seem to suggest that this explanation might not be precise and might be concealed by the narrow range of proficiency in these earlier studies. Our findings seem to suggest that, unlike frequency effects, which uphold even at the highest levels of L2 proficiency, congruency effects seem to fade away as proficiency increases.

Another similar explanation of the congruency effect is Jiang’s (Reference Jiang2000) model of L2 lexical fossilization (see Wolter & Gyllstad, Reference Wolter and Gyllstad2011). Jiang divides lexical knowledge into two levels: lexeme (phonology, orthography, and morphology) and lemma (syntax and semantics), claiming that the latter is amenable to influence from the L1 even at very advanced L2 levels and might result in a fossilized L2 state. The model also claims that only exceptional nonnative cases might be able to escape these L1-lemma effects to achieve nativelike performance. Looking at Figure 3 and the systematic trend for more nativelike recognition as estimated proficiency increases, one might possibly refute such claims of resistant L1 effect on L2 processing. Our results seem to suggest that with enough L2 exposure, nonnative learners might most likely be able to overcome influences from the L1, achieving more nativelike recognition (cf. Wolter & Gyllstad, Reference Wolter and Gyllstad2011, Reference Wolter and Gyllstad2013).

Pedagogical and research implications

The study has important implications for L2 teaching, testing, and research practice. Since the L1 effect seems to diminish gradually as proficiency increases, L2 learners can be pushed out of the fossilized state of lexical development (Jiang, Reference Jiang2000; see above) through more exposure to language. This can eventually lead to dealing with the L2 input through a nativelike eye removing the L1 lenses that might overshadow processing in the early stages of L2 learning. However, the question remains: how best can L2 learners be provided with such an exposure especially in the EFL context where contact with English is often limited to classroom time? One feasible option is extensive reading programs, which, as argued by Nation (Reference Nation2011), have been underapplied in EFL classrooms in spite of their well-documented value. These programs can expand vocabulary knowledge and help learners develop lexical aspects that are not usually amenable to teaching such as collocations. This is not to say that collocations should not be taught explicitly. Previous research has shown direct collocation teaching as an effective tool in developing knowledge collocations (e.g., Laufer & Girsai, Reference Laufer and Girsai2008), but teachers need to be careful about which collocations to teach and how best to teach them.

Although the gains obtained from Experiment 1 and Experiment 2 might not be directly comparable (correct/incorrect responses in Experiment 1 vs. RT for correct responses only in Experiment 2), some broad implications related to L2 collocation testing can be drawn. Previous research has mostly used untimed paper-and-pencil tests and has made claims regarding congruency (e.g., Bahns & Eldaw, Reference Bahns and Eldaw1993; Biskup, Reference Biskup, Arnaud and Béjoint1992) and the way it interacts with learning conditions (e.g., Peters, Reference Peters2016). Our results showed that congruency has differential effects on untimed and timed L2 collocation processing. It seems that nonnative users always exhibit L1 effects when they are prompted to answer untimed tests. However, during time-pressured processing, the L1 effect seems to disappear as estimated proficiency increases. Based on these findings, one might tentatively propose that language teachers need to interpret results of paper-and-pencil tests cautiously as they do not necessarily translate into automatic processing (see Sonbul & Schmitt, Reference Sonbul and Schmitt2013, for similar claims). Moreover, research on L2 collocation learning and processing should start to incorporate both types of measures in order to capture congruency effects more closely. To the best of our knowledge, only one study (Szudarski & Conklin, Reference Szudarski and Conklin2014) attempted to explore the role that L1 plays in L2 collocation learning using a timed acceptability judgment task.

One last implication of the present study is that the semantic deviation of control items in previous research (see the Background section above) does not necessarily refute earlier findings on collocation processing. Even when control items were made semantically plausible, both natives and nonnatives processed them more slowly than real collocations.

Limitations and suggestions for future research

Our study is limited in several ways. First, English language proficiency was assessed using knowledge of the most frequent 2,000 word families only. Including higher levels might have resulted in expanding the variation in proficiency levels, leading to cleaner depiction of the interaction with congruency. The study also did not explore the effect of transparency which has been shown to highly correlate with congruency (see Gyllstad & Wolter, Reference Gyllstad and Wolter2016; Yamashita, Reference Yamashita2018).

Second, we only tested recognition of two types of collocations (VN and AN). A recent study by Siyanova-Chanturia and Janssen (Reference Siyanova-Chanturia and Janssen2018) employed a timed recall measure (phrase elicitation task) in exploring the processing of formulaic language (i.e., binomials). Future research on determinants of L2 collocation processing and/or leaning may need to start assessing recall using such psycholinguistic techniques to tap into spontaneous productive language use. Moreover, findings for VN and AN collocations might need to be expanded through investigating other configurations (e.g., verb + adverb, verb + preposition, adverb + adjective) as the learning/processing burden may not be the same for all word classes.

A third limitation of the present study is related to the nature of the tasks used. Our timed task required an explicit judgment. Although such a task has been common in previous research, other psycholinguistic measures can more directly tap into natural language processing (e.g., self-paced reading, primed LDT, eye tracking, and event-related potentials). Although several studies have explored frequency effects on formulaic language using such sensitive tasks (see, for example, Siyanova-Chanturia, Conklin, & van Heuven, Reference Siyanova-Chanturia, Conklin and van Heuven2011; Sonbul, Reference Sonbul2015), very few studies (Carrol & Conklin, Reference Carrol and Conklin2014, Reference Carrol and Conklin2017; Carrol et al., Reference Carrol, Conklin and Gyllstad2016; Wolter & Gyllstad, Reference Wolter and Gyllstad2011) have explored congruency effects on L2 processing of formulaic language online (i.e., in real time). Employing such sensitive measures might lead to differential effects of congruency and proficiency and the interaction between them. Turning now to the untimed MC test, it is limited in that it only gauged collocation form without considering meaning. Nguyen and Webb (Reference Nguyen and Webb2017) suggested that future research may need to focus on predictors of meaning recall/recognition.

Fourth and finally, we were only able to recruit female learners. Future research may need to include male learners as well to evaluate the generalizability of results to both male and female Arab EFL learners.

Conclusion

This study aimed at exploring the effect of L1 on L2 collocation recognition and how this might be influenced by estimated levels of proficiency. In two experiments, we assessed both untimed and timed L2 collocation recognition in the underresearched Arabic EFL context. Results showed that the L1 influenced the ability to answer MC items regardless of proficiency level while the effect on timed recognition faded away as L2 competence increased. These results go in line with Yamashita and Jiang’s (Reference Yamashita and Jiang2010) findings and point out to an important distinction to be considered in future collocation research between untimed collocation processing and the ability to use collocations fluently in real time.