How acquirable are English articles for L2 learners? Evidence from online processing and production by L1-Mandarin and L1-Croatian speakers of L2-English

Abstract

English articles (a[n]/the) are one of the hardest features for second language (L2) learners to acquire. Theories such as the Morphological Congruency Hypothesis (MCH) predict that where articles are a unique-to-L2 feature, they will not be fully acquirable, while the Unified Competition Model (UCM) predicts full article acquisition.

To examine first language (L1) transfer in the article processing/production of L1-Mandarin and L1-Croatian learners of L2-English, we conducted two studies. To shape MCH/UCM predictions, Study 1 used a forced-choice task with 28 L1-Mandarin and 27 L1-Croatian speakers testing L1 preferences for noun premodification with anaphoric definites and referential indefinites. Study 2 tested 24 advanced L1-Mandarin/L2-English and 22 L1-Croatian/L2-English learners and 24 L1-English controls on acceptability judgment, self-paced reading (SPR), and oral production tasks.

Mixed-effects regression showed participants had explicit knowledge of (in)definite English articles, but only the L1-Mandarin/L2-English learners performed similarly to the L1-English on SPR for the indefinite article and had similar production accuracy. By contrast, the L1-Croatian/L2-English learners lacked sensitivity to omission of either article and had lower production accuracy. The L1-Croatian/L2-English results, in particular, support the MCH predictions that unique structures present persistent problems in comprehension and production, even at high proficiency.

Introduction

Much research into the second language (L2) acquisition of articles has found strong evidence of first language (L1) transfer. That is, L2 learners of English from an article-lacking L1 background (e.g., Russian, Polish, Japanese) consistently either omit or misuse articles in L2-English (Ekiert, 2004; Ionin et al., 2004; Snape, 2007; Trenkic, 2007), especially when compared with studies that have tested L2-English learners from L1 backgrounds with articles (Derkach & Alexopoulou, 2023; Murakami & Alexopoulou, 2016; Trenkic & Pongpairoj, 2013). To empirically test the role of L1 transfer in L2 article acquisition, we explored two transfer-based theories, namely, the Morphological Congruency Hypothesis (MCH) (Jiang et al., 2011) and the Unified Competition Model (UCM) (MacWhinney, 2005). Both theories predict that the acquisition of dissimilar features (differently instantiated in the L1 and L2) will remain non-target-like even at high proficiency due to unautomated activation (MCH) or competition (UCM). However, the theories differ in their predictions for unique features (existing in the L2 but not the L1), which the MCH considers as unacquirable, but the UCM considers as fully acquirable, since they neither suffer nor benefit from transfer (see Section 1.3 for details).

To test the predictions of both theories for L2-English article acquisition, in the present paper, we conducted two studies. In Study 1, 28 L1-Mandarin and 27 L1-Croatian speakers were given a forced-choice elicitation task (FCT) to determine the extent to which they preferred anaphoric definite and referential indefinite nouns to be premodified. In Study 2, we tested 24 L1-Mandarin/L2-English and 22 L1-Croatian/L2-English learners and 24 L1-English controls on their explicit knowledge (via an acceptability judgment task [AJT]), online comprehension (via a self-paced reading [SPR] task), and oral production of English articles in the same contexts as in Study 1. On the premise that articles are a unique feature for L1-Croatian/L2-English learners (no articles in the L1), we would not expect the L1-Croatian speakers to prefer noun premodification in Study 1. Moreover, if their performance on online comprehension and production is non-target-like, this would yield support for the MCH, while if the performance is target-like, this would align with the predictions of the UCM (see Section 1.3 for details).

Although Mandarin has traditionally been classed as an article-lacking language, a growing body of research has questioned this, showing that L1-Mandarin speakers strongly prefer anaphoric definite referents to be premodified by a determiner and referential indefinites by a numeral + classifier (Section 1.1), and that due to positive transfer, L1-Mandarin/L2-English learners often perform more accurately and/or similarly to L1 speakers compared with L2ers from other article-lacking L1 populations (Section 1.2). To the extent that these patterns hold, the Study 1 L1-Mandarin group is expected to prefer anaphoric/referential nouns to be premodified, which is expected to manifest as positive transfer across all tasks for the Study 2 L1-Mandarin/L2-English group. This would align with the predictions for positive L1 transfer under both the MCH and UCM.

Definiteness in English, Mandarin, and Croatian

Definiteness is a concept present in all languages and is broadly said to denote the identification of a referent within discourse, with a definite referent being one that the speaker intends to refer to and is also identifiable to the hearer (Hawkins, 1991; Trenkic, 2009). English is a language that grammaticalizes definiteness in the form of the definite article the and the indefinite article a(n) ¹ . By contrast, Mandarin and Croatian belong to languages that do not mark (in)definiteness through the use of an overt article system and, instead, infer the status of a referent from context or use word order and other determiners, which are not obligatory. Although markers of (in)definiteness appear in a wide range of contexts, for the purposes of the present paper, we focus on anaphoric definites and referential indefinites. An anaphoric referent is defined as definite if it is uniquely identifiable through previous mention, as in (1), while a referential indefinite referent, although asserted to exist in the world, is not uniquely identifiable, as in (2) (Hawkins, 1991; Ionin et al., 2022).

1. 1. Jack has a dog to keep him company. He takes the dog for long walks in the park.

2. 2. Jack has a cat to keep him company. He wishes he also had a dog to take on walks.

In English, both of these constructions are obligatorily marked, typically by the definite article the and the indefinite article a/an, respectively.

Definiteness in Croatian and Mandarin

On the surface, both Croatian and Mandarin are classed as article-lacking languages; however, there is an ongoing debate as to whether Mandarin has begun to grammaticalize some of its markers of (in)definiteness. In Croatian, an anaphoric referent (one that has already been introduced) is predominantly bare, with optional premodification using the demonstratives this (ovo/ova/ovaj) and that (to/ta/taj) as in example (3) (present stimuli).

However, the use of this/that with anaphoric referents in Croatian is restricted only to uses in which a demonstrative would also be possible in English, namely, in immediate situations and anaphoric reference (e.g., example 3) in which the referent has already been established as definite through context/first mention and the demonstrative is used to make the definite reference even clearer (Trenkic, 2004).

In terms of referential indefinites, nouns are also most commonly bare although indefinite pronouns (no one, anybody, etc.) and the numeral one (jedan) can be used to signal indefiniteness as in example (4) (Pranjkovic, 2000, p. 347). However, just as with anaphoric referents, all of these markers are entirely optional.

In contrast to Croatian, the development of markers of (in)definiteness in Mandarin is a lot more complex. As Lopez et al. (2022) state, it would be “an oversimplification to say that Mandarin entirely lacks articles” (p. 2) because for decades, there has been evidence that in Mandarin, the numeral one (yi + classifier) and the demonstratives this/that (zhe/na) are undergoing the process of grammaticalization (Chen, 2004; Hedberg, 1996; Li & Thompson, 1981; Liu, 2010). Zhe and na have come to be obligatorily used with anaphoric referents in object (but not subject) position in Mandarin as in example (5) (Chen, 2004, p. 1153).

This is said to be the beginning of the grammaticalization process, which typically starts with demonstratives becoming common/obligatory with anaphoric referents (Chen, 2004).

However, zhe and na are also used beyond the contexts in which this/that would also be used in English, as in the anaphoric example above, ² shared general knowledge (Chen, 2004), and bridging (Crosthwaite, 2014). Nevertheless, zhe and na, although obligatory or preferred in some contexts, are still optional in a number of contexts, which would be prototypical of the definite article (e.g., general knowledge), suggesting that the process of grammaticalization is far from complete.

Similar grammaticalization claims are made of yi + classifier primarily used for marking indefinite referents in Mandarin, especially for introducing discourse-new referents as in example (6) (Li & Thompson, 1981, p. 132). This construction is frequently and preferably used for signaling new referents (Chen 2004, Crosthwaite, 2014; Liu, 2010).

Chen (2004) claims that yi is further down the grammaticalization path than zhe and na as it serves more functions where English also requires the indefinite article; however, there are still contexts in which a bare noun is preferred (e.g., with inferable referents).

Evidence for the slow grammaticalization, but also continuous optionality, of zhe and na with previously mentioned referents, and yi + classifier for newly introduced referents, comes from numerous studies, including linguistic analyses/grammar books (Hedberg, 1996; Jenks, 2018; Li & Thompson, 1981), corpora (Chen, 2004; Crosthwait, 2016; Huang, 1999; Liu, 2010), and experimental data (Crosthawaite, 2014, 2016; Ionin et al., 2022; Lopez et al., 2022). For example, when Ionin et al. (2022) and Lopez et al. (2022) asked L1-Mandarin speakers whether they preferred anaphoric definites and referential indefinites to be preceded by a demonstrative/numeral or leave the noun bare in forced-choice tasks, answers were in favor of demonstratives/numerals. Specifically, in Ionin et al. (2022), there was a 67% preference for a demonstrative to precede anaphoric definites and an 80% preference for yi + classifier to precede referential indefinites, while in Lopez et al. (2022), the preference was around 80% for both contexts.

L2-English article production and online processing

If Mandarin markers of (in)definiteness are undergoing gradual changes, the question arises as to whether these changes (in certain contexts) could facilitate positive transfer into the L2. There is a growing body of literature showing that L1-Mandarin/L2-English learners tend to perform more accurately than is expected from typical article-lacking populations (e.g., corpus studies: Derkach & Alexopoulou, 2023; Crosthwaite, 2016; metalinguistic task: Xu et al., 2016; eye-tracking Trenkic et al., 2014). Especially important for the present study is that this pattern has been observed in oral production tasks, in which L2-English learners from Slavic ³ L1 backgrounds have been found to supply articles with low accuracy (Murakami & Alexopoulou, 2016), in particular by omitting them (e.g., Avery & Radisic, 2007; Trenkic 2007). This finding contrasts with the performance of L1-Mandarin/L2-English learners, who tend to produce articles with high accuracy on oral production tasks (e.g., around 90% in Snape, 2009) and use the indefinite article to introduce non-inferable referents into discourse even at A1 (beginner) proficiency level, in contrast to article-lacking L1-Korean learners who did it much later (Crosthwaite; 2014).

In recent years, L2 article acquisition has also been investigated using online processing methods such as eye-tracking (Trenkic et al., 2014) and SPR (Cho, 2022; Ionin et al., 2021, 2022). In their 2021 and 2022 studies, Ionin et al. employed a combination of a grammaticality judgment task (GJT) and SPR and FC tasks (see Section 1.1 for discussion of the FCT from the 2022 study) to test L1-Mandarin and L1-Korean learners of L2-English (both groups considered article-lacking by the authors) on anaphoric and bridging definites and (non)referential indefinites. Across both studies, the L1-English controls were found to have both explicit and implicit knowledge of article omission and substitution in tested contexts, as were the L1-Mandarin/L2-English learners. This finding for the L1-Mandarin groups could be predicted based on the premise that Mandarin prefers anaphoric definites and referential indefinites to be premodified, which is being positively transferred into English. However, in the 2022 study, the L1-Korean/L2-English learners, whose L1 is arguably fully article-lacking, were also sensitive to article omission in both indefinite and definite contexts on the SPR task. These findings are in line with Cho (2022), who found L1-Korean participants to be sensitive to article substitution in anaphoric and bridging (where referent is not identifiable through previous mention but shared knowledge) definite and non-unique indefinite contexts on an SPR task. However, in contrast to Ionin et al. (2022), whose participants had explicit awareness of the GJT, the participants in Cho’s (2022) study did not demonstrate explicit knowledge of all constructions, discrepancies which the authors ascribed to task differences.

MCH or UCM as sources of variable L2-English article acquisition?

Although there are many transfer-based models in the field of Second Language Acquisition (SLA) that propose to account for variable acquisition of (morpho)syntax, in the current paper, we explore two approaches (i.e., the MCH and UCM, described further below), which both predict positive transfer for similar cross-language features, such as English articles in anaphoric and referential contexts for L1-Mandarin/L2-English learners. Similarly, both the MCH and UCM predict negative transfer for dissimilar features, which exist in both the L1 and L2 but are instantiated differently, ultimately resulting in non-target-like acquisition. However, the MCH and UCM differ in their account of unique-to-L2 features, such as English articles for the L1-Croatian/L2-English learners, with the MCH predicting the same non-target-like acquisition as for dissimilar features (Jiang et al., 2017), while the UCM proposes that unique features neither compete with nor transfer, leading to target-like processing and production of that feature (cf. van Hell & Tokowicz, 2010).

The MCH posits that the acquirability of a particular morphosyntactic feature is dependent on whether the L2 variant is congruent or incongruent with the L1 variant (Jiang et al., 2011, 2017). A feature is considered congruent if it is morphologically marked in both the L1 and the L2 but incongruent if only marked in the L1. Successful L2 performance is predicated upon morpheme meanings being automatically activated in a speaker’s mind, but since morphological meaning differs between languages, this can lead to incongruency. The automatic activation of incongruent morphemes has to be acquired over time, and the process is more laborious and often incomplete, leading to delays and errors in both online comprehension and production (Jiang et al., 2017). The MCH postulates that L1-L2 incongruent morphemes will never be acquired to a native-like level regardless of proficiency. The effects of morphological (in)congruency have been reported with English articles (Chrabaszcz & Jiang, 2014) but also with other structures such as the plural -s (Jiang et al., 2011). Since the MCH does not predict offline performance problems, especially at higher proficiency (Jiang et al., 2011), we expect all Study 2 participants to demonstrate explicit knowledge of English articles on the AJT. In online performance (comprehension and spontaneous production), which the MCH conceptualizes as tapping into implicit knowledge reflecting acquisition, we predict that the L1-Croatian/L2-English group, for whom articles are incongruent, will show non-target-like processing and performance. Although Mandarin does not have a grammaticalized article system per se, we predict that for anaphoric definites and referential indefinites, the frequently preferred premodification of such nouns will positively transfer into L2-English sentence comprehension and production.

The UCM, while similar to the MCH in its proposals for positive transfer, further distinguishes dissimilar features from unique ones. Under the UCM, dissimilar features are highly susceptible to negative transfer as they give rise to online competition and often result in ungrammatical utterances if speakers rely more on the entrenched L1 system, which is more strongly activated. By contrast, unique-to-L2 features—those present in the L2 but not in the L1—should neither benefit from positive transfer nor suffer from negative transfer (Tolentino & Tokowicz, 2011). Additionally, the UCM predicts that if L2 cues are widely available and reliable, target-like acquisition should eventually be possible. However, while studies generally support the presence of either positive or negative transfer as predicted by the UCM, the findings regarding unique-to-L2 features are less consistent, with some structures being acquirable and others not (e.g., for a review, see Tolentino & Tokowicz, 2011). Since the UCM predicts similar (successful or unsuccessful) performance across tasks, we expect both L2 groups in Study 2 to perform in accordance with what is afforded by their L1 on all three experimental tasks. Specifically, for the L1-Croatian/L2-English group, articles are a unique-to-L2 feature, which should result in target-like acquisition and performance on both offline and online tasks. We make similar predictions for the Study 2 L1-Mandarin/L2-English group, but for different reasons: the influence of positive L1 transfer rather than feature uniqueness.

Study 1: forced-choice task

We first present the results of Study 1, an FCT with L1 speakers of Mandarin and Croatian, which was conducted for two main purposes: (1) to provide evidence on the extent to which Mandarin and Croatian speakers prefer anaphoric definites and referential indefinites to be premodified by a demonstrative and numeral, respectively, and (2) to inform the MCH and UCM predictions for Study 2. Crucially, Study 1 was conducted after Study 2 (below) and with different participants, but it is presented first due to its importance for shaping Study 2 predictions. ⁴

Task and participants

The FCT was based on Ionin et al. (2022) and consisted of stimuli that were used in the SPR task and AJT, which were translated into Mandarin and Croatian, respectively, by professional translators. For the FCT, administered via Qualtrics, participants were asked whether each target sentence (based on the context) is best preceded by a demonstrative (this/that), numeral one (+ classifier for Mandarin), or nothing (bare). The stimuli were presented in two lists to avoid participants judging the same (or similar) context more than once. Study 1 participants were 28 China-based, L1-Mandarin speakers (list 1 = 11; list 2 = 17; mean age: 27; 22–41) and 27 Croatia-based, L1-Croatian speakers (list 1 = 17, list 2 = 10; mean age: 33; 19–55). For Croats, item 18 in the indefinite context was removed from the analysis because the Croatian translation was giving a definite reading, which is not the case in English. Item 22 was removed from all Study 1 and 2 analyses for both groups because both contexts gave a definite reading.

Results and discussion

The results for the anaphoric definite condition in Study 1 revealed that 83% of the L1-Mandarin speakers and 74% of the L1-Croatian speakers showed a preference for a demonstrative rather than a bare noun (Table 1). The L1-Mandarin results are similar to preference rates from previous studies of 81.82% (Lopez et al., 2022) and 67% (Ionin et al., 2022). While the L1-Croatian higher preference rate contrasts with the lower 55% preference by the L1-Koreans in Ionin et al. (2022), both of which are article-lacking, it is not entirely surprising as demonstratives are commonly used in Croatian with anaphoric referents. However, this is entirely optional and restricted to anaphoric referents, which can be marked with a demonstrative (instead of an article) even in English.

Table 1. Results of the Study 1 FCT task by group and context

With referential indefinites, the L1-Mandarin speakers showed a 62% preference for a numeral (Table 1), which is somewhat lower than the 80% in Ionin et al. (2022) and 87% in Lopez et al. (2022). However, an additional 12% of L1-Mandarin speakers also chose a demonstrative, which results in an overall 84% preference for premodification compared to a 26% preference for bare nouns. By contrast, the L1-Croatian speakers showed only a 35% preference for numeral premodification of referential indefinites.

Next, we use the findings of Study 1 to inform our predictions for Study 2.

Study 2: online comprehension and production

Study 2 investigated how advanced L1-Mandarin/L2-English and L1-Croatian/L2-English learners process and produce English articles in real time. The L2 participants and the L1-English control group undertook an AJT, SPR task, and an oral production task with the aim of testing explicit knowledge, online processing, and production of English articles, respectively. The study aimed to answer the following research questions:

1. 1. To what extent do L1-English, L1-Mandarin/L2-English, and L1-Croatian/L2-English speakers possess explicit knowledge of the English article system?

2. 2. To what extent are L1-English, L1-Mandarin/L2-English, and L1-Croatian/L2-English speakers sensitive to violations of English articles in online processing?

3. 3. How accurately do L1-English, L1-Mandarin/L2-English, and L1-Croatian/L2-English speakers produce the indefinite and definite articles in obligatory contexts in spontaneous oral production?

Based on theory and the results of Study 1, we further refined the MCH and UCM predictions for Study 2:

For RQ1, we predict that all participants will demonstrate explicit knowledge of the English article system. Under the MCH, offline performance at advanced proficiency is likely to be unaffected by any potential procedural issues, and under the UCM, both L2 groups should be able to acquire articles (due to uniqueness in the case of L1-Croatian participants and L1 transfer for the L1-Mandarin participants).

For RQ2 and RQ3, both theories predict target-like performance by the L1-Mandarin group on both tasks (i.e., positive transfer), which is supported by the presented Study 1 data, which (despite some optionality) on the whole show a strong preference for the anaphoric definite and referential indefinite nouns to be premodified. Under the MCH, the L1-Croatian group would be expected to show non-target-like performance due to morphological incongruency in the meaning of definiteness, while under the UCM, target-like performance would be expected because unique features are considered to be fully acquirable. In Study 1, the L1-Croatian participants preferred a demonstrative before an anaphoric referent (similar to the L1-Mandarin group), so if this is available for positive transfer, it should facilitate both online comprehension and production in the definite context. We would not expect a similar effect for the indefinite context, as the majority of L1-Croatian participants in Study 1 preferred the noun to be bare.

Participants

Study 2 tested a total of 70 participants: 24 L1-Mandarin/L2-English (mean age: 24; 21–26), 22 L1-Croatian/L2-English learners (mean age: 34; 24–49), and a control group of 24 L1-English speakers (mean age: 22; 18-28). The L2 participants were of advanced proficiency at the time of testing as confirmed by the Oxford Quick Placement Test (QPT) (L1-Mandarin/L2-English score M = 49, SD = 1.41; L1-Croatian/L2-English score M = 53, SD = 2.19). In addition, L2ers were late bilinguals who had learned English in formal education and had not lived in an English-speaking country before the age of 18. The L1-Mandarin/L2-English participants, but not the L1-Croatian/L2-English participants, had lived in the UK for about three months prior to testing (studying at university), but there were no observable effects of this brief immersion either in the participants’ QPT scores or performance on the Study 2 experiments. The present study was part of a larger project, which also investigated tense-aspect acquisition. In those results, we observed a different pattern in that Mandarin participants, whose L1 does not overtly mark tense, performed less accurately on the oral production task than Croatian participants, whose L1 has overt tense marking, that is, a pattern opposite to that found with articles. Also, there is no evidence to suggest that a three-month immersion would make a substantial difference to the L1-Mandarin/L2-English group’s language skills. In fact, Trenkic and Warmington (2019) show that in a similar population, there was little improvement in language skills even after eight months of immersion on a UK university course.

The L1-English and L1-Mandarin/L2-English participants were recruited from a UK university by sending a call for participants through different departments and snowball sampling. The L1-Croatian/L2-English participants were recruited through a foreign languages school in Croatia. All participants were offered a small reward for their time.

Design

Oral production task

The oral production task in the present study sought to investigate how accurately the participants produced English articles, more specifically, singular countable nouns introduced into the discourse as either first-mention (indefinite) or subsequent (definite) referents. The participants were asked to watch a short, animated Pixar film, “Partly Cloudy,” in which a cloud makes different baby animals that are then taken to their respective parents by a dedicated stork. This task required the participants to keep track of (in)definite referents (i.e., stork, cloud, and baby animals) in discourse in real time by narrating the story as it unfolded. Participants were not allowed any preparation time (a) to make the task tap into implicit knowledge and (b) to avoid ceiling effects, given their advanced proficiency.

Self-paced reading task and acceptability judgment task

For the SPR experiment, 48 items were designed to test the participants’ sensitivity to violations of English articles: 24 items in the definite context and 24 items in the indefinite context. The items consisted of a brief context scenario that introduced a referent into the discourse as either definite or indefinite, followed by a sentence which appeared one word at a time in the middle of the computer screen and contained the critical regions for analysis. Each trial concluded with a follow-up sentence. All nouns used as the critical region were concrete singular nouns in direct object position. The critical region sentence contained a noun phrase (NP) preceded by (a) a correct article, (b) a substituted article (indefinite where definite is required, and vice-versa), or (c) no article, as shown in examples 7 and 8 below.

Definite context

7. On her trip to Mexico, Sally bought a hat from a market. It’s very colorful and interesting. (scenario)

1. a. She hasn’t worn the hat since her return three weeks ago. She hasn’t had the chance. (match)

2. b. She hasn’t worn a hat since her return three weeks ago. She hasn’t had the chance. (substitution)

3. c. She hasn’t worn hat since her return three weeks ago. She hasn’t had the chance. (omission)

Indefinite context

8. On her trip to Mexico, Sally bought a scarf from a market. It’s very colorful and interesting. (scenario)

1. d. She wishes she’d bought a hat that perfectly matched the new scarf. Sadly, she didn’t find one. (match)

2. e. She wishes she’d bought the hat that perfectly matched the new scarf. Sadly, she didn’t find one. (substitution)

3. f. She wishes she’d bought hat that perfectly matched the new scarf. Sadly, she didn’t find one. (omission)

A total of 120 items were used in the study: 48 experimental (article) items and 96 filler items made up of 48 items used as stimuli in another study and 24 fillers. None of the filler items tested structures similar to articles. All items were pseudorandomized into six presentation lists, and each participant was assigned to one list, seeing a total of 96 items: 24 experimental items (one experimental item in one condition, for example, article match, substitution, or omission in either the indefinite or definite context) and 48 fillers (i.e., 24 of the items serving as stimuli for another study and the 24 fillers shared by both studies). A third of all experimental items were followed by a comprehension question to ensure the participants were paying attention. Each comprehension question (e.g., “Did Frank buy a ring for his girlfriend?”) required a “yes” or “no” response. All participants answered the comprehension questions with accuracy above the 80% criterion, except one participant whose data were removed from the study and replaced with another participant who met this accuracy cut-off.

For the acceptability judgment task (AJT), the exact same 96 items were used as in the SPR tasks, and participants were asked to indicate how grammatically acceptable they found each sentence on a 6-point Likert scale (1 = not grammatical at all, 6 = fully grammatical).

Tasks and procedures

The tasks were presented in the order from most implicit to most explicit task (i.e., oral production, SPR, AJT) to prevent the participants from discovering which grammatical structures were being investigated. Participants were tested in a dedicated room on an individual basis and were seated in front of a laptop. The SPR task was designed and run in E-prime, which presented the participants with trials starting with a brief scenario appearing at the top of the screen. Next, the critical item and follow-up sentences appeared one word at a time in the middle of the screen, and the participants were told to press the SPACE key on the laptop when they were ready to move from one word to the next. The final AJT task was run in Qualtrics.

Analysis and statistical approach

All stimuli, data sets, and R scripts can be accessed via this OSF link: https://osf.io/wzj9s/. The data in the present study were analyzed using mixed-effects linear regressions in R (R Core Team, 2022) as mixed-effects models are a more robust and flexible approach to analysis of variance in cases when the assumption of independence has been violated due to multiple data points being collected from the same participants (Winter, 2020).

The oral production data were transcribed, and only relevant NPs were extracted for analysis, that is, target NPs referring to the main (gray) cloud, the stork, or the baby animals that appear during their interactions. The chosen NPs were all concrete countable nouns (as in the SPR task) appearing in both the indefinite and definite contexts. The lead author then coded and scored a total of 70 scripts of oral production data and checked the intrarater reliability by recoding a portion of the scripts two months after the initial scoring. To check interrater reliability, a second marker (L1-English speaker) was recruited and trained to score a subset of the data, which showed 99% agreement (477/480 decisions). Overall accuracy of supplied articles with the target NPs for each participant was determined using the target-like use measure, which takes into account article overuse, namely, how many times the article was supplied correctly but also how many times it was supplied in nonobligatory contexts. To analyze the data, we ran a mixed-effect model with score (accuracy %) as the outcome and the fixed-effect predictors context (indefinite/definite) and group (L1-English, L1-Mandarin/L2-English, and L1-Croatian/L2-English) modeled as an interaction. The participant variable was modeled as a random effect.

The SPR data were cleaned for analysis, which meant removing all reading times (RTs) below 100 and above 2000 milliseconds (Keating & Jegerski, 2015) affecting 0.025% and 0.18% trials, respectively. The data were then log transformed due to the data being positively skewed. To analyze the SPR data, we ran a total of six linear mixed-effects models (one model per group for the indefinite article and one model per group for the definite article) using the lmer function in the lme4 package (Bates et al., 2019). Log RTs were entered as a continuous variable, while two fixed-effect predictors, condition (match, substitution, omission) and segment (0, 1, 2, 3; see Table 2 for sentence segmentation), were entered as an interaction. The R package buildmer (Voeten, 2022) was used to obtain a maximal feasible formula.

Table 2. Example of distribution of segments in a target sentence. Full sentence: She hasn’t worn the hat since her return three weeks ago

The AJT data were also analyzed using linear mixed-effects models using the lmer function in the lme4 package (Bates et al., 2019). There was one model for the indefinite article and one for the definite article. AJT ratings (Likert scale 1–6, 1 = not acceptable, 6 = fully acceptable) were entered as the outcome and an interaction between condition (match, substitution, omission) and group (L1-English, L1-Mandarin/L2-English, L1-Croatian/L2-English) as predictors. Participant and item were modeled as random effects with maximal slopes specified, and the R package buildmer (Voeten, 2022) was used to obtain a maximal feasible formula.

For all results, we report model estimates, p values, 95% confidence intervals, standard error, degrees of freedom, and t values (as appropriate).

Results: oral production task

All three participant groups produced the indefinite articles less accurately than the definite article (Table 3). However, the L1-English and L1-Mandarin/L2-English groups produced the indefinite and definite articles with around 80% and 90% accuracy, respectively. By contrast, the L1-Croatian/L2-English group produced the indefinite article accurately only 55% of the time and the definite article 81% of the time.

Table 3. Accuracy of indefinite and definite article suppliance by participant group on the Study 2 oral production task

We ran a linear mixed-effects regression to check whether the L2 groups differed in their performance from the L1 group in the indefinite and definite contexts. The fixed and random effects explained 71% of the variance in the model (conditional R ² = 0.713). There was no significant difference between the performance of the L1-Mandarin/L2-English and L1-English groups in either context (indefinite: estimate = –1.542 [–9.818, 6.735], t = –0.365, p = 0.716, definite: estimate = –2.259 [–10.119, 5.619], t = –0.560, p = 0.576), while there was a significant difference between the L1-Croatian/L2-English group and the L1-English (indefinite: estimate = –18.170 [–26.633, –9.707], t = –4.208, p = <.001, definite: estimate = –15.049 [–23.095, –7.002], t = –3.666, p = <.001).

Below are examples of the participants’ oral production (roughly the same section of the story has been selected for all samples, and the target article is underlined, or if it’s missing, an x has been placed):

L1-English participant 409

The stork’s now returned but it’s got loads of the porcupine’s spines stuck in his head. The grey cloud pulled the spines out of the stork’s head and some of its feathers came off, and then it gave it a hug. It appears to have produced…

L1-Mandarin/L2-English participant 101

And the cloud think the bird is so miserable with all the pins on his head so the cloud get rid of these pins away from his head. And hugged the bird. And the purple cloud think of something else, he think he can send to the bird.

L1-Croatian/L2-English participant 213

A bird again. x Bird is very very damaged as I can see. The cloud is in the process of making something but I don’t know what. Oh it’s a shark, x sea shark. A bird if frightened and she flew away. On another cloud. Now the cloud, x big cloud is very angry and….he’s mean. Now he’s sad…the cloud.

Results: SPR task

This section reports the results of the mixed-effects regression analyses for the SPR data for all three participant groups. In the indefinite context, the fixed and random effects together explained 69% of the variance in the L1-English model, 1.4% in the L1-Mandarin/L2-English model, and 64% in the L1-Croatian/L2-English model (conditional R ² = 0.694, 0.014 ⁵ , 0.640). The L1-English groups were sensitive to omission errors on the first segment following the noun (estimate = 0.195 [0.106, 0.283], t = 4.329, p = <.001). The L1-Mandarin/L2-English groups were also sensitive to omission errors but at a later segment (estimate = 0.131 [0.011, 0.251], t = 2.145, p = 0.032). By contrast, the L1-Croatian/L2-English group did not show sensitivity to omission errors in the indefinite context. Finally, all three groups showed no sensitivity to substitution errors.

In the definite context, the fixed and random effects together explained 68% of the variance in the L1-English model, 49% in the L1-Mandarin/L2-English model, and 68% in the L1-Croatian/L2-English model (conditional R ² = 0.680, 0.488, 0.684). The L1-English group showed a sensitivity to omission errors also in the definite context on the first segment following the noun (estimate = 0.093 [0.004, 0.181], t = 2.055, p = 0.040). The L1-Mandarin/L2-English and L1-Croatian/L2-English groups were not sensitive to omission errors, and all three groups showed no sensitivity to substitution errors. Full results for both contexts are available in Tables 4 and 5 as well as Figure 1.

Table 4. Summary of fixed-effects for the SPR analysis for the definite article by Study 2 participant group

Notes: Sub = substitution, Om = omission, Seg = Segment, df = degree of freedom, CI = confidence interval, SE = standard error.

Table 5. Summary of fixed-effects for the SPR analysis for the indefinite article by Study 2 participant group

Notes: Sub = substitution, Om = omission, Seg = Segment, df = degree of freedom, CI = confidence interval, SE = standard error.

Figure 1. Definite (first) and indefinite article SPR task log RTs in Study 2 across four segments: match = green, substitution = blue, omission = red.

Results: AJT

Figure 2 shows that all three groups rated the sentences in which the article matched as more acceptable than sentences in which the article was substituted or omitted. However, the ratings are lower for omission than substitution for all groups. Also, L1-English ratings for omission are the lowest of all groups, whereas for the L1-Croatian/L2-English group, they are the highest and relatively similar to substitution ratings.

Figure 2. Bar charts showing mean AJT scores for the indefinite and definite articles by Study 2 participant groups.

In the indefinite and definite contexts, the models explained 36% (conditional R ² = 0.364) and 47% (conditional R ² = 0.470) of the variance, respectively. We found that all three groups rated correct sentences as significantly more acceptable than either omission or substitution in both contexts, except the L1-Mandarin/L2-English group, who did not make such a distinction with substitution in the indefinite context (Table 6). It is worth noting that even for the L1-Croatian/L2-English group, the effect for substitution was not as strong as for omission errors.

Table 6. Summary of fixed-effects for the AJT analysis for the indefinite and definite article by Study 2 participant group

Notes: sub = substitution, om = omission, CI = confidence interval, SE = standard error.

Discussion

The present study investigated how advanced L1-Mandarin/L2-English and L1-Croatian/L2-English learners and L1-English speakers process and produce English articles. RQ1 sought to find out whether the three participant groups possessed explicit knowledge of the English article system, while RQ2 asked whether the three participant groups were sensitive to violations of English articles in online processing. The final RQ3 elucidated how accurately the three participant groups produced the indefinite and definite articles in obligatory contexts in spontaneous oral production. We first discuss the results by context, focusing on AJT/SPR omission and production data, while SPR substitution is discussed separately due to methodological implications. Finally, theoretical implications are considered.

Indefinite context

In Study 2, we found high levels of explicit knowledge of the indefinite article in the referential indefinite context in all groups. However, comprehension and production results show that only the L1-English and L1-Mandarin/L2-English groups were sensitive to omission of the indefinite article on the SPR and produced it with high accuracy (88% and 85%, respectively). By contrast, the L1-Croatian/L2-English groups were not sensitive to indefinite omission and produced the indefinite article accurately only 55% of the time. The L1-Mandarin/L2-English results point to positive L1 transfer, which facilitated comprehension and production, most likely arising from more frequent premodification of referential indefinites in Mandarin with yi + classifier, which has been well-documented in the literature (Section 1.1) and is supported by Study 1 results. The L1-Croatian/L2-English results from Study 2 align with the results from Study 1, showing that in Croatian, referential indefinites are preferred bare, which arguably has a detrimental effect on the Study 2 L1-Croatian/L2-English group’s comprehension and production of the indefinite article, as predicted by the MCH. It seems that whatever explicit knowledge Croats have available, it is not accessible implicitly/automatically in L2-English comprehension and production.

Definite context

Similarly to the indefinite context, we found that all three participant groups demonstrated explicit knowledge that omission of the definite article in anaphoric contexts was ungrammatical. Also, the L1-English groups were sensitive to omission of the definite article on the SPR task and produced it with high accuracy (96%). However, the L1-Mandarin/L2-English groups were not sensitive to definite omission but nevertheless produced the article with high accuracy (94%). This is particularly surprising given the evidence from previous research (see Section 1.1) as well as our Study 1, which showed that L1-Mandarin speakers prefer anaphoric referents to be premodified by a demonstrative. There is some evidence of positive transfer, as L1-Mandarin/L2-English learners were able to effectively use the definite article in production but did not seem to be able to do the same in comprehension (Study 2). We suggest that either (a) we were not able to capture an SPR effect due to our sample size (in Figure 1, there seems to be a small slow down for omission of the first segment following the noun, but in the present data, it was not found to be statistically significant), or (b) the grammaticalization of demonstratives with anaphoric referents has not progressed to the same extent as yi + classifier for referential indefinites (Chen, 2004). As for the Study 2 L1-Croatian/L2-English group, their performance remained similar to that in the indefinite context, namely, that they were not sensitive to definite omission on the SPR and produced the definite article significantly less accurately (81%) ⁶ than either the L1-English or L1-Mandarin/L2-English speakers. Therefore, similar to the indefinite context, our results align with the predictions of the MCH. Study 1 did show that Croats somewhat preferred anaphoric referents to be premodified by a demonstrative, but this did not seem consistently available for transfer in online comprehension or production, in contradiction with the UCM notion that “whatever can transfer will” (MacWhinney, 2005; p. 17).

SPR substitution errors

A surprising finding in Study 2 was that neither the L1-English speakers nor the L2 groups were sensitive to article substitution in either context, despite mostly demonstrating explicit awareness that substitution is ungrammatical (albeit somewhat less so than omission). These findings are not straightforward to explain, especially since Cho (2022) and Ionin et al. (2021) found L1-English, L1-Mandarin/L2-English, and L1-Korean/L2-English participants to be sensitive to substitution. For one, it needs to be considered to what extent the substitution stimuli are conclusively ungrammatical ⁷ in the present study, especially in the definite context. For instance, in example (7) (see Section 3.2.2), the intended interpretation of the substitution sentence in the definite context is a violation of a specific referent with an indefinite article. However, a possible substitution reading of “any hat” cannot be fully excluded. All stimuli were piloted with L1-English speakers before being included in the study, and our AJT results confirm that (apart from the L1-Mandarin indefinite substitution) all participants found substitution to be ungrammatical (albeit less than omission). Nevertheless, in future research, it needs to be ensured that all stimuli elicit a conclusive ungrammatical reading for substitution.

Coupled with this potential design issue, another possibility is that as long as an (i.e., any) article is there, during speedy processing, most participants can quickly repair the error, which is semantic rather than syntactic, and move on with the rest of the sentence. Ionin et al. (2021) also found evidence that substitution violations pose a lesser processing load than omission, which, in the current study, the SPR task may not have been able to reliably pick up.

Theoretical implications

Our results from Study 1 and Study 2 provide two major theoretical implications. First, there seems to be evidence of positive L1 transfer for the L1-Mandarin/L2-English group in Study 2, which facilitated their online comprehension in the indefinite context and production of both articles. This is in line with the positive transfer predictions of both the MCH and UCM and supports previous evidence for the slow grammaticalization of yi + classifier with referential indefinites. Positive L1 transfer in the definite context was observed on the production task but not in comprehension, which is more likely due to the limited statistical power of our study than it is evidence against the grammaticalization of zhe and na in Mandarin. Second, for L1-Croatian/L2-English learners, articles are a unique-to-L2 structure, which the MCH would consider a source of persistent problems, but the UCM would consider as fully acquirable, since such structures should neither hinder nor aid acquisition (i.e., the MCH and UCM differ in their predictions for these learners). Our Study 2 L1-Croatian/L2-English results were consistently in line with the MCH’s prediction that a morphological incongruency will lead to persistent problems in comprehension and production, regardless of high proficiency.

Nevertheless, these findings are in direct contrast with Cho (2022) and Ionin et al. (2022), who found article-lacking L1-Korean/L2-English learners sensitive to omission/substitution of both anaphoric definites and referential indefinites in SPR, against the predictions of the MCH. It seems that whatever acquisition was observed in these studies, it does not extend to the current study’s article-lacking population (Croatian). In fact, L1-Korean/L2-English learners in other studies tend to show non-target-like performance compared to L1-Mandarin/L2-English learners, demonstrating both lower accuracy and later acquisition of articles in both written and spoken production (Crosthwaite, 2014, 2016). This is in line with the oral production results of our Study 2 in that (a) the L1-Mandarin/L2-English learners produced articles with high accuracy and (b) an article-lacking group (Croatian) was significantly less accurate on the production task.

Therefore, in the studies above, the cause of the L1-Korean/L2-English participants’ target-like performance on SPR is not entirely clear. First, it is possible that these studies contain task features that are tapping into different types of knowledge and in different ways (e.g., SPR vs. production). Second, differences in task design, administration, and data analysis could have led to different results, and close replications of ideally Ionin et al. (2021, 2022) and the present study with different L1 populations are important next steps.

Limitations

Our study is not without limitations. First, the number of participants per group is on the low side, which could have potentially affected the power to detect all effects. In particular, a replication of the substitution data with a larger sample is needed to uncover any uncaptured effects. Second, the AJT in the form we presented it was arguably not the most suitable method of eliciting explicit knowledge; an AJT in which the target structure was underlined (as in Ionin et al., 2021) may have been more reliable. Finally, we focused only on two specific contexts (anaphoric and referential referents), thus not capturing the full spectrum of article use.

Conclusions

From the results of the present study, we draw several important conclusions. First, the L1-Mandarin/L2-English and L1-Croatian/L2-English participants did not process and produce L2-English articles in similar ways; the L1-Mandarin/L2-English participants’ patterns of processing and production were somewhat similar to those of the L1-English participants. Our findings suggest positive transfer for the L1-Mandarin/L2-English group in referential indefinite and anaphoric definite contexts, and as such, we argue that this group should not necessarily be treated as fully article-lacking in future research, especially when these contexts are examined. For the L1-Croatian/L2-English group, we observed no positive transfer from either their explicit knowledge or L1 preferences, and their performance on the SPR and oral production tasks remained non-target-like. This finding aligns with the predictions of the MCH.