The good-enough processing account (GE) reasonably explains how the processor (erroneously) operates during language activities. It assumes two processing routes: an algorithmic stream, which is a structure-based, bottom-up route, and a heuristic stream, which is a usage/experience-based, top-down route (Ferreira, 2003; Christianson, 2016). While these routes apply simultaneously to interpretation, they are distinctive concerning the trade-off between accuracy and efficiency. Algorithms yield precise computations of linguistic representations but require effortful and time-consuming processing. In contrast, heuristics allow rapid and less effortful, yet sometimes underspecified, interpretation. GE maintains that heuristics can generate interpretations earlier than algorithms and that the former sometimes triumphs over the latter. This argument finds support in Ferreira (2003): monolingual English speakers occasionally misinterpret passives (e.g., The dog was bitten by the man) by incorrectly mapping thematic roles onto event participants (i.e., an agent role to the dog and a theme role to the man). This indicates a speaker’s primary commitment to simple, coarse-grained schemata from language-usage experience or world knowledge involving argument realisation in English.

The abovementioned argument is supported by various models/frameworks of sentence processing. Real-time processing places heavy demands on cognitive resources (Gibson, 1998; Lewis et al., 2006; Pozzan and Trueswell, 2015). Therefore, the processor both attempts to immediately finish interpreting input and seeks to avoid repairing misinterpretations unless urgently required (Fodor and Inoue, 1994; Piantadosi et al., 2012). Moreover, because linguistic cues are often noisy (Futrell and Levy, 2017; Gibson et al., 2019) and lossy (Christiansen and Chater, 2016), the processor often settles for options readily accessible from memory (e.g., Noun–Verb–Noun template in English; Townsend and Bever, 2001), provided that these options reasonably preserve communicative intent (Jaeger and Tily, 2011; Kleinschmidt and Jaeger, 2015). This way, the processor achieves and maintains sufficient cognitive equilibrium while minimising burdens on cognitive systems (Kool et al., 2010; Christie and Schrater, 2015; Karimi and Ferreira, 2016).

Much work revolving around GE suggests that the processor prioritises the heuristic route while selectively adopting the algorithmic route when required (Dwivedi, 2013; Kharkwal and Stromswold, 2014; Lim and Christianson, 2015; Qian et al., 2018; Tan and Foltz, 2020), aligning well with the multi-stream models of sentence processing (van Herten et al., 2006; Kuperberg, 2007). However, the literature has predominantly focused on a limited range of languages and speakers of those languages, especially (L2-)English-speaking populations, with few studies exploring beyond this scope (e.g., Russian: Stoops et al., 2014; Mandarin: Zhou et al., 2018). This perpetuates an English-centric perspective (cf. Blasi et al., 2022). Consequently, it is questionable whether previous findings on sentence-processing patterns under GE are generalisable to speakers of lesser-studied languages and usage contexts. Incorporating understudied languages and usage contexts can enrich GE by introducing diverse linguistic structures and language-specific adaptations not commonly found in major languages, thereby challenging and refining our current understanding of this framework.

The relationship between linguistic knowledge and domain-general mechanisms has been actively explored in efforts to address language behaviour (McElree, 2000; Phillips and Ehrenhofer, 2015; Pozzan and Trueswell, 2015; Christiansen and Chater, 2016; Cunnings, 2017). Of various mechanisms, working memory (WM) and inhibitory control (IC) are the major cognitive skills widely discussed in the field. WM refers to a mental workspace that retains information for a short period while concurrently conducting mental operations on this information (Baddeley and Hitch, 1974; Daneman and Merikle, 1996). Although its capacity is limited and varies by individual, it reliably predicts performance in various cognitive tasks, including language comprehension (Tokowicz et al., 2004; Huettig and Janse, 2016; Litcofsky et al., 2016). IC pertains to attentional control for natural, habitual, or dominant responses to a target stimulus that are unnecessary for this stimulus, allowing individuals to become goal-relevant (Miyake and Friedman, 2012; Diamond, 2013). There is growing evidence on the role of IC in language activities (Abutalebi and Green, 2007; Pérez et al., 2020), particularly from bilingualism research: given the parallel activation of both languages, bilinguals utilise IC to attend to one language system while suppressing the other and manage conflicts and interference between the two (Bialystok et al., 2005; Kroll et al., 2008; Wigdorowitz et al., 2023).

WM and IC have routinely been associated with comprehension (Pérez et al., 2014; Tarchi et al., 2021). Most cognitive tasks require both skills (Diamond, 2013), which are intertwined when humans conduct these tasks (Carlson et al., 2002; Roncadin et al., 2007). Each skill, however, seems to differentially affect how language users engage in language activities (Abutalebi and Green, 2007; Linck and Weiss, 2015). This stems from their representation of varying aspects of cognitive skills (Astle et al., 2013)—even though they outwardly share some fundamental architectures (Miyake et al., 2000)—and their asymmetric changes across a lifespan (Williams et al., 1999; Robert et al., 2009).

Heritage speakers are defined as child and adult members of a linguistic minority whose home language involves limited usage experience and formal literacy education in a community and the majority language in that community is dominantly used (Rothman, 2009; Montrul, 2010). They manifest asymmetric linguistic representations influenced by various factors such as reduced home-language input, pressure on usage from the majority language, grammatical properties of a target item, and cognitive resources (Kondo–Brown, 2005; O’Grady et al., 2011; Jia and Paradis, 2015; Mikhaylova, 2018; Felser and Arslan, 2019; Karaca et al., 2024; for an in-depth overview, see Polinsky and Scontras, 2020). Previous studies have delineated distinctive attributes of heritage speakers’ morphosyntactic knowledge in comparison to monolingual or L2 speakers (Kim et al., 2009; Laleko and Polinsky, 2016; Montrul et al., 2019; Fuchs, 2022). Furthermore, research has elucidated the role of individual differences (represented as WM, IC, or overall proficiency in the heritage language) in modulating heritage speakers’ task performance (Chondrogianni and Schwartz, 2020; Bice and Kroll, 2021; Torres, 2023). Exploring the language-processing mechanisms of heritage speakers is crucial not only for unveiling the unique challenges stemming from their language-use backgrounds but also for enlightening researchers and practitioners to better understand and empower this population in preserving their heritage language.

Our study specifically focuses on Korean heritage speakers (KHSs) residing in the United States. With over 1.9 million individuals speaking Korean as a heritage or community language in the country, this demographic constitutes the fifth-largest Asian-American subgroup (U.S. Census Bureau, 2021). Despite the increasing global interest in Korean culture and language, research within the US contexts has predominantly centred on dominant heritage speaker groups such as Hispanics or Chinese (Jegerski et al., 2016; Scontras et al., 2017; Hur et al., 2020; Bice and Kroll, 2021; Torres, 2023; López Otero et al., 2024), underscoring the urgent need for scholarly attention towards KHSs. Korean, an understudied language for GE, is a Subject–Object–Verb language that maintains verb-finality, but its case-marking system allows for relatively free word order by scrambling sentential components (Sohn, 1999). We concentrate on two clausal constructions—suffixal passive and morphological causative—which contrast with respect to alignments between thematic roles and case markers as well as interpretive procedures involving verbal morphology.

In this study, we investigate how KHSs engage in sentence comprehension under GE, serving as the basic sentence-processing architecture, with a focus on the two parsing streams situated at the intersection of proficiency (as an indicator of home language usage experience), cognitive skills, and task types. Regarding task modalities, we examine two language tasks: acceptability judgement and self-paced reading. In an acceptability judgement task (AJT), a comprehender partially and holistically evaluates a sentence under few time constraints before arriving at a complete interpretation of the sentence and decides on its acceptability. A self-paced reading task (SPRT) requires a comprehender to conduct moment-by-moment, sequential, and cumulative interpretations of incoming items in real time, subject to stricter time constraints compared to AJT. Given the commonly observed disparities between non-dominant-language learners’ explicit and implicit/automatised knowledge (Jiang, 2007), some studies have explored how varying task demands affect the GE processing of L2-English sentences (translation vs. eye-tracking: Lim and Christianson, 2015; comprehension question types during SPRT: Tan and Foltz, 2020). However, research on this topic for underrepresented languages and populations remains less active.

The passive voice is marked across languages (Haspelmath, 1990; Siewierska, 2013), and its usage frequency in Korean is notably lower than that of the active voice (Woo, 1997; Park, 2021). The suffixal passive consists of two arguments, a nominative-marked theme subject and a dative-marked agent oblique, followed by a passivised verb. Passive morphology, signaled by one of the four allomorphic variants of verbal suffixes (-i/hi/li/ki-), serves as a key disambiguation point for identifying a sentence’s structural properties, also decreasing the verb’s valency slots (from two to one for a transitive verb). The canonical pattern (1a) follows the theme–agent–verb sequence, but the verb can be fronted via scrambling, yielding a verb-initial pattern (1b) found in colloquial speech for afterthought clarification, information amplification, or emphasis (Sohn, 1999).

The two word-order patterns exhibit contrastive characteristics regarding the timing of disambiguation. In the verb-final pattern (1a), passive morphology constitutes a late-arriving cue, compelling a comprehender to revise an initial analysis conducted before encountering this morphology. In Korean, a nominative-marked [+human] argument is likely to be interpreted as an agent, and a dative-marked [+human] argument tends to be interpreted as a recipient. These interpretations are supported by strong associations between thematic roles and case markers attested in language use (Sohn, 1999; Kim and Choi, 2004; Shin and Mun, 2023). Therefore, a plausible way of analysing (1a) prior to the verb is that the thief executes an action affecting the police. However, this analysis is incongruent with the passive-voice information conveyed by verbal morphology. Upon encountering the sentence-final verb, a comprehender must revise their initial interpretation, recalibrating the arguments’ thematic roles by mapping a theme role onto the nominative-marked entity and an agent role onto the dative-marked entity. Revision in this manner is linguistically and cognitively demanding (Rapp and Kendeou, 2007; Kendeou et al., 2013), thereby posing challenges to language activities (Kim et al., 2017; Shin, 2022). The situation differs for the verb-initial pattern (1b). The fronted verb and its morphology constitute an early-arriving cue, guiding the succeeding interpretation of the arguments and forestalling possible misinterpretations of these arguments’ thematic roles (cf. Pozzan and Trueswell, 2015). Therefore, the sentence-initial verb in (1b) is expected to inform a comprehender that the nominative-marked entity thief is not the agent but the theme and that the dative-marked entity police is not the recipient but the agent, also suppressing the typical associations between thematic roles and case markers.

The results on the two tasks performed by MSK are consistent in light of how the two parsing streams operate. Together with the higher acceptability ratings for the verb-final condition than the verb-initial condition in AJT, no processing benefit was derived from the fronted passive morphology in SPRT, as predicted. These suggest a larger role of heuristics (word-order typicality) than algorithms (interpretive procedures driven by verbal morphology) in processing the suffixal passive, aligning with previous studies (Ferreira, 2003; Dwivedi, 2013; Kharkwal and Stromswold, 2014).

The performance of KHS in the two tasks elucidates how GE operates in conjunction with various factors during comprehension. While exhibiting similar acceptability-rating behaviour to MSK, KHS demonstrated prolonged RTs at the critical regions (R2–R4) in both conditions, as predicted. This aligns with prior research that highlights a general challenge in the real-time processing of non-dominant languages (Pozzan and Trueswell, 2016; Tachihara and Goldberg, 2020; Grüter and Rohde, 2021). The fronted verb in the verb-initial condition incurred greater processing cost compared to the nominative-marked subject in the verb-final condition, whereas the post-verbal region in the verb-initial condition incurred reduced processing cost compared to that region in the verb-final condition. Importantly, KHS exhibited similar RTs when reading the verb region in both conditions, consistent with our prediction. These findings indicate a lack of evidence for the active role of early-appearing verbal morphology cues in processing verb-initial passive sentences. That is, the presumed advantage of fronted verbal morphology may not have completely surpassed the processing benefit of the canonical word order which is readily accessible from memory (and may potentially alleviate the interpretive complexity posed by passive morphology occurring at the end of a sentence). This is ascribable to multiple possibilities: heritage speakers’ susceptibility to home-language morphosyntax (Laleko and Polinsky, 2016; Kim et al., 2018; Chondrogianni and Schwartz, 2020), their reduced flexibility in dealing with scrambling due to dominant language—English in this study (Namboodiripad et al., 2018), and limited usage experience of home language (Chondrogianni and Schwartz, 2020; Hur et al., 2020; López Otero et al., 2024).

Notably, KHS’s performance was modulated by cognitive skills and proficiency. This interplay was more complex in SPRT than AJT: the KHS’s acceptability ratings were proportionate to their scores on the three measures (digit span, Flanker, proficiency), but the contributions of these measures to their RT patterns varied at different regions and conditions, which deviated from our predictions. KHS spent less time reading R2 in the verb-final condition with expanding IC capacities, but this trend was missing from the verb-initial condition. This difference is ascribed to an increased degree of interpretive challenge involving verb-initiality. That is, the fronted verb—manifesting atypical word order and inviting (re)calibrations of the mapping between thematic roles and case markers early on—may have substantially canceled out processing support from IC when KHS handled the early-appearing verb/morphology cue.

KHS spent less time reading R4 in the verb-initial condition with increasing proficiency. This can be interpreted in two ways. First, it suggests KHS’s larger space for, and more engagement in, handling passive morphology (and algorithmic parsing tied to that morphology) in the verb-final condition as proficiency increased. Second, it implies KHS’s enhanced efficiency in processing a dative-marked agent in the verb-initial condition as proficiency increased. Meanwhile, the expected role of proficiency over the entire structure did not emerge. This finding contradicts previous research demonstrating the facilitative role of general language proficiency in a non-dominant language for achieving target-like processing of clausal constructions (Jackson, 2008; Rah and Adone, 2010; Kaan, 2014; Hopp, 2017). Rather, this finding lends support to the idea that proficiency in the target language (as one factor of individual variability) selectively influences sentence processing within that language contingent upon task types (Roberts, 2012).

KHS spent more time reading R5 (spill-over involving clausal integration for complete interpretation) in the verb-final condition as their WM capacities expanded although the effect was weak. This finding is the reverse of our prediction, implying KHS’s increased capacity for, and commitment to, the integration procedures involving the canonical word-order condition proportionate to their WM skills. Such capacity and commitment, in turn, enable a comprehender to reserve more space for coping with previous and current inputs at this region.

Nevertheless, it is premature to draw firm conclusions about the operational characteristics of the two parsing routes solely from Study 1. The fact that each group spent comparable RTs across the two verb regions (as shown in Figure 2) might also imply the possibility that heuristic parsing was counterbalanced by algorithmic parsing. Hence, additional evidence is required to convincingly elucidate the interplay between KHS’s sentence-processing behaviour and various factors within GE, particularly concerning the heuristic-before-algorithm strategy. We thus conducted another experiment, focusing on the morphological causative manifesting interpretive procedures involving verbal morphology distinctive from those of the suffixal passive.

The morphological causative consists of three arguments: a subject (causer), an indirect object (causee), and a direct object (theme), as in (3a). The verb carries one of the seven allomorphic variants of verbal suffixes (-i/hi/li/ki/wu/kwu/chwu-), thereby increasing the verb’s valency slots. The verb can move to the sentence-initial position as in (3b).

The interpretation of the arguments’ thematic roles hinges upon causative morphology, but this process does not invoke substantial challenges to the extent that passive morphology does. To illustrate, in (3a), the nominative-marked [+human] argument Mia-ka is understood as a causer (as an extension of an agent, sharing the concept of a volitional actor). The dative-marked [+human] argument Pola-eykey is understood as a causee (as an extension of a recipient); the dative marker ensures these extensions by sharing the same semantic component—GOAL (Sohn, 1999). Causative morphology does not invite the same kind of recalibration of the mapping between thematic roles and case markers as that needed in passive morphology. Therefore, the degree of cognitive burdens that verbal morphology poses to processing the morphological causative is not enormous.

For MSK, their RTs (particularly those involving the verb regions) and acceptability ratings for the morphological causative were almost identical to those for the suffixal passive, demonstrating no early-arriving-cue advantage and late-arriving-cue disadvantage. This corroborates the heuristic-before-algorithm strategy for sentence processing by monolingual speakers (Ferreira, 2003; Dwivedi, 2013; Kharkwal and Stromswold, 2014; Lee et al., 2023), underscoring the prominent role of heuristic parsing (prioritising canonicity/typicality involving word order and case-marking facts) over algorithmic parsing (involving interpretive procedures driven by verbal morphology) in sentence processing.

For KHS, the verb-final condition was deemed more acceptable than the verb-initial condition in AJT, and more RT was spent at R2 and R3 in both conditions compared to MSK, aligning largely with our predictions. The RT difference between the two conditions in SPRT was substantial at R5 (with the verb-final condition more time-consuming than the verb-initial condition) and at the verb regions (with R2 in the verb-initial condition more time-consuming than R4 in the verb-final condition). The insignificant by-condition difference at R2 may imply the emergence of an early-arriving-cue benefit, but this should be interpreted with caution as this insignificance seems to originate from the considerably increased RTs at this region in the verb-final condition when compared to Study 1, the reason of which is unclear. Importantly, the fact that KHS exhibited notably longer RTs when reading the verb in the verb-initial condition compared to the verb-final condition, as predicted, suggests a larger role of heuristic parsing than algorithmic parsing in sentence processing. In addition, the notable decrease in RTs at R4 in the verb-final condition of Study 2 in contrast to Study 1 suggests that, despite the cognitive demands associated with clausal integration (as indicated by the significant RT gap at R5), KHS may have leveraged the interpretive procedures involving causative morphology—presumed to be less taxing than those involving passive morphology—to some extent when coping with this region.

Pertaining to proficiency and cognitive skills, only the Flanker scores meaningfully influenced the KHS’s acceptability ratings. This differs from Study 1, and implies that construction-specific properties (e.g., alignments between thematic roles and case markers, interpretive procedures driven by verbal morphology) selectively adjust the activation of these factors and to different extents in this construction. The similar kind of trend was also found in SPRT. The proficiency and digit-span scores influenced KHS’s RT patterns in the same manner as that found in Study 1, but the Flanker scores incurred more RTs at R5 in the verb-final condition (although the effect was weak). This is inconsistent with our prediction; note that the IC effects on KHS’s RT patterns emerged at R2 in Study 1. The role of proficiency found here was the same as that in Study 1, corroborating the argument that general proficiency in a non-dominant language contributes selectively to sentence processing in that language.

The locus of this asymmetry concerning IC is ascribable to the properties of the two construction types in this study. Compared to the interpretive procedures involving the suffixal passive, those involving the morphological causative are less costly (see Sections 2.1 and 3.1). In other words, the algorithmic stream applied to each construction type differs qualitatively. This less demanding nature in the morphological causative may have allowed KHS to allocate more cognitive resources when conducting clausal integration at R5 towards a full interpretation, resulting in more RTs spent at this region (cf. Kaiser, 2014). This also aligns with why KHS spent more time at R5 in the verb-final condition proportionate to the WM scores in Study 1 (see Section 2.4).

In this respect, the null Flanker effects found at R2 in the verb-initial condition in Study 2, as well as at R2 in the verb-initial condition and at R5 in the verb-final condition in Study 1, point to the same potential mechanism underlying the non-dominant-language mind: its capacity and willingness to conduct (algorithmic) parsing induced by linguistic cues proportional to the computational complexity involving these cues. To illustrate, scrambled word order is more challenging to compute than canonical word order because scrambling is less frequently used and invites contextual/discoursal effects on interpretation. This may have overridden the presumed early-arriving-cue advantage in the verb-initial conditions of both constructions substantially. Passive morphology is more taxing than causative morphology with respect to the interpretive procedures that it drives, which may have reduced the space for the non-dominant-language mind to control information that is irrelevant to the target knowledge via IC. If this reasoning is valid, it also provides additional support for the idea that IC and WM, although interconnected, function separately and differently during sentence comprehension (Abutalebi and Green, 2007; Linck and Weiss, 2015).