Highlights
-
• Late bilinguals exhibit symmetrical translation priming effects with 50-ms primes.
-
• L1–L2 priming is more pronounced with 83-ms primes due to the slower responses for unrelated primes.
-
• Primes become largely visible with an 83-ms prime duration.
-
• Although both tasks produce significant priming, semantic categorization induces a stronger effect.
-
• Increasing L2 proficiency and exposure leads to greater L2–L1 priming.
-
• Word frequency effects vary across language directions, tasks, and prime durations.
1. Introduction
A central question in psycholinguistics concerns how bilinguals represent and process lexical entries in their mental lexicons. Despite extensive research using diverse experimental paradigms, ranging from translation priming to naming tasks (Bailey et al., Reference Bailey, Lockary and Higby2024), the findings reported to date have been largely inconsistent (Zhao et al., Reference Zhao, Li, Xiong and Xing2024), particularly with respect to the directionality and mechanisms of cross-linguistic activation. Consequently, there is still no consensus regarding how native (L1) and non-native (L2) languages interact within the bilingual mind. The nature of this interaction seems to depend on several factors, including language proficiency, language exposure, age of acquisition (AoA), as well as task demands and item-level characteristics. The qualitative approach states that any differences between L1 and L2 word processing/storage are relatively stable when an L2 is acquired after infancy (e.g., Jiang & Forster, Reference Jiang and Forster2001), as late-acquired L2 words are represented differently and processed less efficiently. The quantitative approach, however, has proposed that such differences can weaken with increased L2 proficiency and usage (e.g., Dijkstra et al., Reference Dijkstra, Wahl, Buytenhuijs, van HALEM, Al-Jibouri, Korte and Rekké2019), potentially approaching native-like processing under certain conditions. These competing perspectives reflect a broader debate about the plasticity of the bilingual lexicon and the extent to which it can approximate native-like processing in L2.
While qualitative models generally advocate for permanent differences between L1 and L2 lexicons, their specific claims provide qualitatively distinct testable predictions. For instance, the Revised Hierarchical Model (RHM; Kroll & Stewart, Reference Kroll and Stewart1994; Kroll & Tokowicz, Reference Kroll, Tokowicz and Nicol2001) posits separate L1 and L2 lexicons, which are connected to a shared conceptual system (Xia & Andrews, Reference Xia and Andrews2015). The nature of this link is different for the two lexicons: L1 words have stronger connections to semantic concepts, while L2 words have weaker connections to concepts but strong and direct links to their L1 counterparts, particularly at lower levels of L2 proficiency. As a result, L1 words are more likely to induce semantic processing, whereas L2 words (especially at lower proficiency levels) tend to activate their translation equivalents in L1 first. The Episodic L2 Hypothesis (Jiang & Forster, Reference Jiang and Forster2001) predicts even a stricter asymmetry since bilinguals have their L1 words stored in a lexical memory system in the model, as opposed to L2 words, which are stored in an entirely different memory system, namely, the episodic memory, at least during the early stages of L2 acquisition. Unlike these two models, which propose separate lexicons or memory systems, the Sense Model (Finkbeiner et al., Reference Finkbeiner, Forster, Nicol and Nakamura2004) proposes partially overlapping lexicons with graded semantic overlap across languages. In the model, words from different languages share semantic features, or senses, which determine the strength of the connections between L1 and L2 words. Due to earlier acquisition and richer exposure, L1 words have more senses than L2 words.
Unlike qualitative models, which differentiate L1 and L2 lexicons through acquisition-related differences, quantitative models argue that individual differences, such as L2 proficiency, account for processing differences between L1 and L2 words. This more flexible approach leaves room for the possibility that increasing L2 proficiency and exposure can indeed result in native-like processing for L2 words. A well-acknowledged quantitative model is Multilink (Dijkstra et al., Reference Dijkstra, Wahl, Buytenhuijs, van HALEM, Al-Jibouri, Korte and Rekké2019), which evolved from the Bilingual Interactive Activation (BIA; Dijkstra & van Heuven, Reference van Heuven, Dijkstra and Grainger1998) and the Bilingual Interactive Activation Plus (BIA+; Dijkstra & van Heuven, Reference Dijkstra and van Heuven2002). In this non-selective language access model, written input stimulates orthographic representations, which in turn activate semantic and phonological representations in a fully interactive manner. The model derives its name from the multiple links among these representations.
Multilink inherits a task/decision system from its predecessors, which adjusts word processing stages according to task demands. This feature allows flexibility; in addition to considering the activation levels of word candidates, the system checks whether the candidates meet the criteria set for the task. Multilink is an interactive model in which all representations influence each other, creating connections among orthographic, semantic, and phonological features. An input activates multiple orthographic candidates, whose activation levels are subsequently modulated at the following semantic and phonological levels. Although L2 words might initially be more difficult to activate, their activation resting thresholds are adjusted through more exposure, allowing for faster activation over time. Therefore, a core tenet of the model is that subjective frequency is an essential factor in determining word activation speed, implying that L2 words can be activated as efficiently as L1 words through sufficient exposure.
Although the models of the bilingual lexicon all acknowledge the existence of cross-linguistic activation, they have distinct views on whether this emerges due to direct lexical links or semantic mediation (Kim & Kim, Reference Kim and Kim2024). Crucially, the claims of both qualitative and quantitative models focus on the automatic processes involved in word activation rather than the use of conscious strategies in reading. Therefore, testing these claims requires an investigation of pre-conscious processing during word recognition. One experimental paradigm frequently employed to explore this is masked priming. This paradigm involves the brief presentation of a prime word immediately before a target word, making it possible to examine how properties of the prime influence target recognition (Feldman et al., Reference Feldman, Kostić, Gvozdenović, O’Connor and Moscoso del Prado Martín2012; Forster & Davis, Reference Forster and Davis1984; Rastle et al., Reference Rastle, Davis and New2004). In addition to the short duration of the prime word (usually 30–80 ms), non-letter characters (e.g., hash marks) are used just before the primes to prevent participants from consciously recognizing them, a practice called forward masking, although prime visibility may vary depending on duration and masking parameters.
Studies investigating the bilingual lexicon usually use cross-linguistic prime–target pairs, such as translation equivalents, which share semantic but not necessarily orthographic or phonological features across languages (e.g., cat in English and kedi in Turkish). Even non-cognate translation equivalents do not necessarily eliminate orthographic overlap altogether when languages share the same script, as shared letters or orthographical features might modulate word processing and result in different patterns than those observed in languages using two different scripts. Nevertheless, by minimizing the potential influence of orthography and phonology, researchers can attribute their findings primarily to semantic interaction or direct word links between languages. This is crucial for testing and differentiating among bilingual lexicon models; by isolating the semantic aspect, researchers can directly test specific claims proposed by the models without the confounding influence of other variables. For example, sport and spor are cognate pairs in English and Turkish with nearly identical orthographic and phonological features. Presenting either of these words might activate the other, not because of a genuine cross-linguistic interaction but as a result of simply tolerating minor orthographic deviations from existing lexical entries. In other words, the English word sport intended as an L2 prime for Turkish–English bilinguals might be perceived directly as the Turkish word spor, revealing an effect of identity priming (i.e., presenting the same word as the prime and target) rather than cross-linguistic priming. In addition, specific claims regarding semantic activation in the L1 and L2 words can be tested more effectively when the involvement of orthographic and phonological similarity is minimized. Altogether, translation-equivalent priming offers a robust framework for testing the specific claims of the bilingual lexicon models.
Two primary tasks that are generally employed in masked translation priming studies are the lexical decision task (LDT) and the semantic categorization task (SCT). In a standard LDT, participants must determine whether a given letter string corresponds to a real word in the target language. In an SCT, on the other hand, all stimuli are real words, and participants must decide whether the word belongs to a predetermined category (see Section 4.2 for more information). Related primes preceding targets in both tasks are translation equivalents, and unrelated primes are words with no immediate orthographic, phonological, or semantic relations to targets. Although both tasks require rapid binary responses, the underlying cognitive decisions are significantly different. The LDT calls for fast and efficient lexical access, a decision that can be based on the orthographic form of the word, while the SCT necessitates the evaluation of semantic features, thus requiring deeper conceptual processing.
Crucially, these different processing requirements allow researchers to test the claims of the bilingual lexicon models under investigation. LDTs can provide insights regarding direct lexical links in the RHM due to their form-based lexical processing requirement; if direct lexical links exist between L2 and L1 words, translation priming will emerge in both directions (and perhaps even be larger in the L2–L1 direction). SCTs, conversely, can allow an investigation of task-dependent semantic modulation in translation priming. According to the RHM, the SCT should result in reduced or absent translation priming in the L2–L1 direction due to the direct lexical links between L2 and L1 words, which force an indirect conceptual activation through L1 lexical items. In contrast, the sense model predicts that the SCT should result in similar translation priming in both directions. According to the model, giving the word category beforehand restricts the relevant senses associated with the target word, thereby making cross-linguistic activation similarly effective in both language directions due to the limited number of senses required for activation. For the Episodic L2 Hypothesis, L2–L1 translation priming is not expected to occur in an LDT since L2 words have no direct connections to the lexical memory. Finally, Multilink proposes that both tasks can produce significant translation priming in both language directions, depending on L2 proficiency and subjective frequency. The SCT can result in more translation priming due to stronger semantic processing requirements.
A recurring pattern in this literature is the asymmetry in translation priming. As discussed below, L1 primes consistently facilitate L2 target recognition, yet L2 primes often fail to produce equivalent effects for L1 targets (e.g., Chaouch-Orozco et al., Reference Chaouch-Orozco, González Alonso and Rothman2021; Jiang & Forster, Reference Jiang and Forster2001), supporting the claims of the qualitative models, albeit not uniformly across studies. This asymmetry is commonly attributed to inefficiencies in L2 lexical processing or shallower L2 word representations. Another potential factor is orthographic depth. When the native language has a shallow orthography, bilinguals who are used to direct activation patterns between letters/bigrams to lexical features/representations might struggle with L2s having deep orthographies (e.g., English).
Despite the common finding of language translation asymmetry in the literature, this observation is not universal. Some studies suggest that highly proficient bilinguals can exhibit comparable priming effects in both L1–L2 and L2–L1 directions (Nakayama et al., Reference Nakayama, Ida and Lupker2016; Perea et al., Reference Perea, Duñabeitia and Carreiras2008), depending on factors like task type, prime duration, and individual differences. Such findings align with the quantitative models and challenge the common assumption that L2 words are inherently disadvantaged in bilingual lexical processing. Instead, they point to a more nuanced interplay of linguistic and cognitive factors, raising important questions about the mechanisms underlying cross-linguistic lexical interactions. Importantly, these results also underscore the necessity of considering differences across items, tasks, and individuals when investigating the bilingual lexicon. The following literature review presents the essential findings from translation-equivalent priming research regarding these factors.
2. Literature review
A common finding in studies of translation-equivalent priming is an asymmetry between language directions. When L1 words are used as primes and L2 words as targets (i.e., the L1–L2 direction), related prime words elicit faster response times even for low-proficient bilinguals (Dimitropoulou et al., Reference Dimitropoulou, Duñabeitia and Carreiras2011a; Duyck & Warlop, Reference Duyck and Warlop2009; Jiang & Forster, Reference Jiang and Forster2001; Voga & Grainger, Reference Voga and Grainger2007), whereas L2 primes usually fail to induce priming effects for L1 targets in the L2–L1 direction (Chaouch-Orozco et al., Reference Chaouch-Orozco, González Alonso and Rothman2021; Finkbeiner et al., Reference Finkbeiner, Forster, Nicol and Nakamura2004; Jiang & Forster, Reference Jiang and Forster2001; Schoonbaert et al., Reference Schoonbaert, Duyck, Brysbaert and Hartsuiker2009; Wang & Forster, Reference Wang and Forster2015; Witzel & Forster, Reference Witzel and Forster2012). This difference, known as the language direction asymmetry, is often attributed to reduced efficiency in processing L2 primes (Dijkstra et al., Reference Dijkstra, Wahl, Buytenhuijs, van HALEM, Al-Jibouri, Korte and Rekké2019) or to shallower representations of L2 words (Jiang & Forster, Reference Jiang and Forster2001). From a quantitative perspective, L2 language proficiency can mitigate or even eliminate this asymmetry, as indicated by the similar priming effects in highly proficient bilinguals (Perea et al., Reference Perea, Duñabeitia and Carreiras2008) and increased L2–L1 priming with higher L2 proficiency (Nakayama et al., Reference Nakayama, Ida and Lupker2016). There have also been studies showing comparable translation priming in low-proficient bilinguals (Dimitropoulou et al., Reference Dimitropoulou, Duñabeitia and Carreiras2011b; Duyck & Warlop, Reference Duyck and Warlop2009), suggesting that multiple variables are at play in determining the strength of the translation priming effect, rather than a single underlying mechanism.
One commonly cited factor behind the language direction asymmetry is the onset of L2 AoA. According to the qualitative account, acquiring a language in infancy creates different lexical representations (Dimitropoulou et al., Reference Dimitropoulou, Duñabeitia and Carreiras2011b) or different connections between these representations and conceptual features (Kroll & Tokowicz, Reference Kroll, Tokowicz and Nicol2001) that are fundamentally different from those of a late-acquired L2. If this assumption holds true, increasing L2 proficiency would not result in identical lexical processing for L1 and L2 in all contexts, and the translation priming asymmetry would emerge even in highly proficient sequential bilinguals but not in simultaneous bilinguals. Supporting this AoA-based view, Duñabeitia et al. (Reference Duñabeitia, Dimitropoulou, Uribe-Etxebarria, Laka and Carreiras2010a) and Duñabeitia et al. (Reference Duñabeitia, Perea and Carreiras2010b) reported similar translation priming effects in both language directions with simultaneous bilinguals who acquired both languages in infancy. Similarly, Kahraman et al. (Reference Kahraman, De Wit and Beyersmann2024) found a significant effect of L2 AoA on the translation priming effects observed in Turkish–English bilinguals. In contrast to these findings, Nakayama et al. (Reference Nakayama, Ida and Lupker2016) analyzed the effect of L2 AoA in their data and argued that it failed to explain the translation priming effects. Rather, they argued that the null L2–L1 findings in the literature result from other factors, such as low L2 proficiency, different word lengths and frequencies across experimental lists, and low study power. While their argument is compelling, the potential role of L2 AoA cannot be easily dismissed, as the number of participants and items was also limited in Nakayama et al. (Reference Nakayama, Ida and Lupker2016).
One additional factor intertwined with the L2 AoA is the bilingual’s use of that language. Two terms can be associated with language use: language exposure and language dominance. While newborns can acquire two languages simultaneously, limited exposure to one of them (even after a certain age) can result in non-native–like acquisition or even language attrition (i.e., losing language abilities over time). For this reason, the concept of L2 exposure in some contexts, which considers both AoA and the usage of that particular language to date, can provide a more accurate measure for investigating the bilingual lexicon than merely focusing on the onset of L2 AoA (see Li et al., Reference Li, Zhang, Yu and Zhao2020). Crucially, this definition of L2 exposure differs from cultural immersion. The one discussed here measures how many years of active usage a bilingual has for a particular language, while cultural immersion refers to the degree of being immersed in an L2 context, underlining cultural features of that language and living in an L2-speaking community as well.
The second term, language dominance, refers to daily language use and preference for a language. One concern, however, arises regarding the relative proficiency of bilinguals in different languages. Some individuals can possess a basic variety of a language and engage in conversations with that limited ability. Even if the use of that language surpasses the use of their L1, such bilinguals should not be regarded as L2-dominant because of their limited ability to process L2. Therefore, the term language dominance goes beyond simple language use and refers to consistently using one language with sufficient proficiency (see Li et al., Reference Li, Zhang, Yu and Zhao2020).
In an investigation of whether language dominance would modulate translation priming, Chaouch-Orozco et al. (Reference Chaouch-Orozco, González Alonso and Rothman2021) conducted masked priming studies with Spanish–English bilinguals who had acquired their L2 after infancy (around the age of 9). Despite the abundant time the participants had spent in an L2-speaking environment before the time of the study (at least 5 years), only the L1–L2 direction induced significant priming effects, replicating the earlier findings supportive of the translation priming asymmetry (Finkbeiner et al., Reference Finkbeiner, Forster, Nicol and Nakamura2004; Jiang & Forster, Reference Jiang and Forster2001; Schoonbaert et al., Reference Schoonbaert, Duyck, Brysbaert and Hartsuiker2009; Wang & Forster, Reference Wang and Forster2015; Witzel & Forster, Reference Witzel and Forster2012). These findings contradict those of Wang (Reference Wang2013), in which bilinguals who had been using both languages dominantly showed comparable priming effects in both language directions. Note, however, that the bilinguals in Wang (Reference Wang2013) had also acquired their L2s rather early (age of 5 or earlier), making it difficult to disentangle the potential effects of L2 AoA and L2 dominance. The potential influence of language dominance was further supported in an investigation of language shift in which Basnight-Brown and Altarriba (Reference Basnight-Brown and Altarriba2007) presented compelling evidence in favor of the effect of L2 dominance on translation priming. The bilinguals in their sample who declared better L2 skills than L1 skills showed asymmetrical translation priming effects, but surprisingly, the L2–L1 direction produced more priming than the L1–L2 direction. Based on their findings, Basnight-Brown and Altarriba (Reference Basnight-Brown and Altarriba2007) argued that frequent use of an L2 can actually make it the dominant language, emphasizing the importance of considering L2 dominance in translation priming studies. One recent study supporting that latter claim is Kahraman et al. (Reference Kahraman, Casalis, Javourey-Drevet, Menut, De Wit and Beyersmann2025), in which both Turkish–English and French–English bilinguals showed greater translation priming effects with increasing proficiency and L2 exposure.
Individual differences, such as proficiency, immersion, and dominance, play a complex role in translation priming, but they show just one side of the coin. Task-related factors can also influence the emergence and strength of translation effects. Prime duration and stimulus onset asynchrony (SOA), for example, have also been reported to have an effect on translation priming patterns. Without additional configuration, prime duration is equal to SOA in the present paradigm, that is, the total time between the presentation of a prime and a target word. When a blank screen or backward mask is presented after the prime, SOA increases, and the prime duration stays the same, allowing participants to have more time to process the prime before seeing the target. Consequently, under typical masked priming conditions, increased SOA is argued to ensure more time to process the prime word without conscious recognition, as the prime word still appears for a very short duration on the screen.
The manipulation of SOA and prime duration can offer important insights. If an increase in SOA enhances the priming effects in the L2–L1 direction, it can be assumed that the prime duration was enough for participants to start processing L2 prime words, but they needed more time to finish processing. Increasing prime duration, on the other hand, can reveal whether longer prime exposure is required for efficient processing of L2 prime words. Increasing prime duration, however, might result in a pattern reflecting more conscious processing than purely automatic, pre-conscious processing. Basnight-Brown and Altarriba (Reference Basnight-Brown and Altarriba2007), for example, used 100-ms primes, which is longer than the usual prime duration used in the masked priming paradigm (30–70 ms). This, in turn, might have modulated the observed priming effects and resulted in a different pattern than would have occurred with shorter prime durations.
Investigating the potential effect of SOA on translation priming, Schoonbaert et al. (Reference Schoonbaert, Duyck, Brysbaert and Hartsuiker2009) tested Dutch–English bilinguals with two different SOAs (50-ms primes with either 100-ms or 250-ms SOAs). The increased SOA boosted priming effects in all tasks, and the only non-significant priming effect in the 100-ms SOA condition (L2–L1 semantic priming) became significant with the 250-ms SOA. In a later electroencephalogram (i.e., EEG) study, Schoonbaert et al. (Reference Schoonbaert, Holcomb, Grainger and Hartsuiker2011) used 100-ms primes with 20-ms backward masks and observed L2–L1 priming effects, but the translation asymmetry was still apparent. The authors suggest that increased prime duration and high L2 proficiency would affect translation priming similarly, both allowing for more efficient processing of L2 primes in a limited amount of time. Contrary to these findings, Wang and Forster (Reference Wang and Forster2015) did not observe any L2–L1 priming with Chinese–English bilinguals using a relatively long prime duration (80 ms). This raises the question of whether increasing the SOA above 100 ms led to the emergence of L2–L1 priming rather than prime duration itself. Supporting this latter view, Korean–English bilinguals with low L2 proficiency in Lee et al. (Reference Lee, Jang and Choi2018) showed significant L2–L1 priming effects when the SOA value was 150 ms but not when it was 60 ms, using the same prime duration of 50 ms. This further supports the idea that giving bilinguals enough processing time via increased SOA (probably above 100 ms) might help them activate L2 primes efficiently, even at lower L2 proficiency levels.
The relative contributions of SOA and prime duration are further complicated by task differences. Finkbeiner et al. (Reference Finkbeiner, Forster, Nicol and Nakamura2004) found comparable translation priming effects in both language directions with SCTs, but the L2–L1 priming effect was null in LDTs, establishing the foundations of the Sense Model presented above. The authors propose that SCTs involve category filtering, in which bilinguals focus more on the senses (conceptual representations) related to the presented categories. In an attempt to test this proposal further, Xia and Andrews (Reference Xia and Andrews2015) tested Chinese–English bilinguals using LDTs and SCTs. Although the level of priming in the SCTs was comparatively higher, the translation direction asymmetry was still apparent. As an alternative to category filtering, Xia and Andrews (Reference Xia and Andrews2015) argue that giving categories beforehand might activate related semantic categories, which in turn boost semantic feedback for lexical representations. In turn, pre-activating exemplars of a category results in more efficient L2 prime processing in that category. Since the LDT does not explicitly present any categorical cues before the presentation of items, such tasks usually fail to induce L2–L1 priming effects.
Individual differences seem to interact with task differences in translation priming. McPhedran and Lupker (Reference McPhedran and Lupker2021) replicated the finding of more translation priming in the SCT than the LDT, but, in addition, they also found that specific L2 skills affected translation priming to a greater extent in different tasks. Specifically, the LDT was more affected by reading, writing, and comprehension of spoken language, while the SCT was more affected by L2 use frequency, speaking fluency, and listening. Although the reported L2 measurements were based on subjective self-assessments by the participants, the findings nevertheless suggest that different tasks might utilize different language skills more than others. McPhedran and Lupker (Reference McPhedran and Lupker2021) argue that more exposure to language orthography through reading and writing might increase the efficiency of priming in LDTs, and cultural immersion (i.e., immersion in an L2 context) might have the same effect in SCTs.
Similar to individual differences and task-related factors, item-related differences have also been found to modulate word processing patterns. In particular, the effects of word frequency are a common finding in language studies (Brysbaert et al., Reference Brysbaert, Mandera and Keuleers2018; Chee et al., Reference Chee, Westphal, Goh, Graham and Song2003; Monaghan et al., Reference Monaghan, Chang, Welbourne and Brysbaert2017). Unfortunately, word frequency effects have been largely ignored in translation-equivalent priming studies, despite the fact that such effects have been widely reported in the L1/L2 lexical processing literature (e.g., Brysbaert et al., Reference Brysbaert, Lagrou and Stevens2017; Brysbaert et al., Reference Brysbaert, Mandera and Keuleers2018; Diependaele et al., Reference Diependaele, Lemhöfer and Brysbaert2013). With the aim of mitigating this limitation, Nakayama et al. (Reference Nakayama, Lupker and Itaguchi2018) investigated word frequency effects in translation priming among highly proficient Japanese–English bilinguals. When prime words were of lower frequency, responses in both related and unrelated conditions were slower, but translation priming remained consistent regardless of target frequency. When prime words had higher frequency values, however, the priming effect of higher target frequency was only observed in the unrelated condition. The researchers attributed this finding to the rapid activation of prime words of higher frequency, which strongly pre-activated target words even before they appeared, making their frequency values irrelevant. The effect of high prime frequency in the related condition against the priming effect of target frequency in the unrelated condition cancelled each other out, resulting in null translation priming effects under those specific conditions. Chaouch-Orozco et al. (Reference Chaouch-Orozco, González Alonso and Rothman2021) also confirmed the role of target frequency, as they found slower responses for target words with low frequency and for lower-proficiency bilinguals. As their findings regarding the frequency effects were vague, the authors encouraged more studies to investigate the influence of word frequency in translation priming.
The complexity of considering individual, task, and item differences in investigating translation priming is further enhanced by the potential factor of cross-linguistic differences. The scripts of a bilingual’s languages, for example, can either boost or hinder the observed priming effects. On one hand, same-script languages might boost translation priming, as activating the language script in one language would ease the activation of lexical items in the other due to orthographic similarity (e.g., McPhedran & Lupker, Reference McPhedran and Lupker2021). However, such similarity might also cause more lexical competition as the words in both languages use the same script, resulting in less translation priming (e.g., Gollan et al., Reference Gollan, Forster and Frost1997; Nakayama et al., Reference Nakayama, Lupker and Itaguchi2018). The effect of language script has been investigated more extensively in the L2–L1 direction, given that this direction usually produces less or no translation priming effects. Testing a sample of Japanese–English bilinguals, Nakayama et al. (Reference Nakayama, Ida and Lupker2016) found a null L2–L1 priming effect in an LDT with low-proficient participants. Xia and Andrews (Reference Xia and Andrews2015) replicated this null finding with Chinese–English bilinguals, even though the SOA value was higher (200 ms) and the proficiency of the participants was moderate to high. While these findings might be taken as support for the absence of L2–L1 priming in LDTs when language scripts are different, there are also studies showing significant translation priming for the same task and language direction. Dimitropoulou et al. (Reference Dimitropoulou, Duñabeitia and Carreiras2011b) and Lee et al. (Reference Lee, Jang and Choi2018), for example, found significant L2–L1 priming effects for Greek–English and Korean–English, respectively. The conflicting findings have raised the question of whether a bilingual’s L1 affects translation priming. It has already been shown that Korean–English bilinguals could outperform Chinese–English bilinguals in a phoneme-deletion task, and this was attributed to the relatively greater similarity between Korean and English writing systems (Wang et al., Reference Wang, Koda and Perfetti2003). Similarly, Nakayama et al. (Reference Nakayama, Lupker and Itaguchi2018) underlined that Greek and English alphabets resemble each other and that this might have been the reason behind the significant L2–L1 priming in Dimitropoulou et al. (Reference Dimitropoulou, Duñabeitia and Carreiras2011b).
The arguments regarding the same-script languages are also not conclusive, as Chaouch-Orozco et al. (Reference Chaouch-Orozco, González Alonso and Rothman2021) could not observe any L2–L1 priming with Spanish–English bilinguals. In addition, script transparency can also modulate the nature of priming effects (Li et al., Reference Li, Yang, Jia, Spinelli and Lupker2024). For instance, phonological priming emerges only when primes are written with Hiragana, a transparent script in Japanese, and not with Kanji, a logographic script (Chen et al., Reference Chen, Yamauchi, Tamaoka and Vaid2007; Forster & Yoshimura, Reference Forster and Yoshimura1994). These findings also suggest that even if the languages a bilingual knows share the same script, priming in different directions can still differ depending on the orthographic transparency of the language to which the prime belongs. Overall, the findings so far have suggested that language script and orthography can change the patterns of translation priming, but the exact pattern of this effect with different language pairs is still largely unknown.
The challenge of having a multitude of potential factors related to translation priming has resulted in the aforementioned contradictory findings. It should be noted, though, that many of the studies reported above suffered from low statistical power resulting from an insufficient number of participants, items, or a combination of both. Furthermore, participant and/or item-related variabilities make it even more difficult to directly compare the findings obtained across studies. To overcome these limitations, it is of critical importance to investigate multiple factors within the same group of participants and items, which should be adequate in number to ensure a satisfactory study power.
Against this background, we investigated in the present study the potential effects of individual differences, task type, prime duration, and word frequency on translation-equivalent priming using a large sample of Turkish–English bilinguals. Importantly, the design of the present study (see the Methods section for details) allows for a more direct test of the central claims of the major models of bilingual lexical processing introduced earlier. In order to investigate these issues, we preregistered our hypotheses on the Open Science Framework before collecting any data: https://osf.io/c3gmf. As reported in the preregistration report, additional variables were used only in the exploratory analyses (e.g., L2 AoA). Research questions, along with our hypotheses and the claims of the models, are given below.
1) Do bilinguals exhibit translation priming effects? If so, is there a language asymmetry?
Hypothesis: We expected to observe translation priming along with a language asymmetry effect in the L2–L1 direction. This asymmetrical pattern can be explained by both qualitative and quantitative models, albeit via different mechanisms. For instance, the Sense Model predicts less L2–L1 priming due to the limited number of senses associated with L2 words. The Multilink, on the other hand, would attribute the asymmetry to insufficient L2 proficiency and subjective frequency.
2) Do different tasks result in different response and translation priming patterns? Does the language direction asymmetry disappear in an SCT?
Hypothesis: We hypothesized that SCTs would elicit slower responses due to their reliance on semantic processing, but this reliance would also yield greater translation priming. Regarding language asymmetry, we had no direct assumptions. The finding of an asymmetric pattern would have supported Xia and Andrews (Reference Xia and Andrews2015), but if translation priming became similar in the SCT, it would support the Sense Model (Finkbeiner et al., Reference Finkbeiner, Forster, Nicol and Nakamura2004). The Episodic L2 Hypothesis predicts no L2–L1 translation priming in an LDT, given the absence of L2 word representations in lexical memory. Under the RHM, L1 primes should elicit priming effects with L2 targets due to strong conceptual links in an SCT, while L2 primes should produce weaker or null effects because late-acquired L2 words are expected to have weaker connections to shared conceptual nodes. The RHM would, therefore, predict robust L1–L2 priming and attenuated L2–L1 priming in the SCTs. Multilink can account for boosted translation priming in the SCT compared to the LDT, as the task/decision system can modulate word processing patterns according to task requirements. In line with the greater semantic and conceptual processing demands in an SCT, translation priming effects can also increase. The asymmetry should be consistent in both tasks according to Multilink, as cross-linguistic activation mainly depends on proficiency and frequency.
3) What is the effect of using a longer prime duration (83 ms) on translation priming?
Hypothesis: Based on the premise that L2 prime processing is often less efficient than L1 processing, we predicted that a longer prime duration would increase translation priming effects compared to a shorter prime duration, especially in the L2–L1 direction. If both prime durations tap into pre-conscious stages of word processing, this outcome would support Multilink by showing that processing time can help overcome differences in resting activation thresholds between L1 and L2 word representations. Crucially, if 83-ms primes become visible, the findings cannot be attributed to automated stages of word processing and would instead reflect conscious or strategic processing.
4) Do individual differences (L2 proficiency, L2 exposure, and L2 dominance) modulate response times and priming effects?
Hypothesis: We expected to see faster responses and greater priming effects in the L2–L1 direction with increasing L2 proficiency, L2 dominance and L2 exposure. This finding would support Multilink, as the model attributes more efficient L2 processing to increasing L2 proficiency and exposure. The Sense Model can account for enhanced L2–L1 translation priming with improving L2 skills, as such improvement might also indicate an increase in the number of senses associated with L2 words. Nevertheless, the language direction asymmetry should persist within that model. For the RHM and the Episodic L2 Hypothesis, proficiency or exposure should have a minimal or null influence at the early processing stages of L2 primes.
5) Does word frequency modulate response times and priming effects?
Hypothesis: We expected faster responses with increasing word frequency, in line with previous research. In particular, higher frequency values for prime words were hypothesized to increase priming effects due to more efficient activation, especially for L2 primes. These patterns would have supported Multilink, in which word activation efficiency is strongly related to frequency. We also expected the possibility that word frequency might stop being a facilitator after a certain threshold, and since our items were of high frequency in the current study, its effect might be null.
3. The present study
As outlined above, there is limited evidence and no consensus regarding the effects of language direction, individual differences (L2 proficiency, exposure, and dominance), task differences (e.g., task type and prime duration), and item differences (word frequency) on translation-equivalent priming. Furthermore, the available evidence is largely based on either different-script language pairs, like Japanese–English and Korean–English, or same-script language pairs that are similar in orthography and typology (e.g., English–Spanish, Spanish–Catalan). This is especially problematic, as language diversity is widely recognized as essential in psycholinguistic studies (Norcliffe et al., Reference Norcliffe, Harris and Jaeger2015).
In the present study, we tested the cross-linguistic processing of translation equivalents by L1 Turkish speakers of L2 English. Turkish uses a Latin-script alphabet and thus has a writing system similar to English. Despite the similarity of the two scripts, Turkish, a member of the family of Turkic languages, is typologically distant from English and has a shallow orthography and highly productive morphology. As mentioned above, orthographic transparency may still substantially change observed priming effects, even when the two languages share the same script (Li et al., Reference Li, Yang, Jia, Spinelli and Lupker2024). Turkish children have been reported to acquire proficient reading in one school year (Babayiğit & Stainthorp, Reference Babayiğit and Stainthorp2007; Durgunoğlu & Öney, Reference Durgunoğlu and Öney1999; Öney & Durgunoğlu, Reference Öney and Durgunoğlu1997) and to master the use of connectives by the fifth grade (Oğuz & Özge, Reference Oğuz and Özge2020), which is faster than what is usually reported for deep orthographic languages such as English (Cain & Nash, Reference Cain and Nash2011; Seymour et al., Reference Seymour, Aro and Erskine2003). Furthermore, Turkish primary school children are sensitive to only true affixes (e.g., -er in the word driver) in the early stages of word processing (Oğuz & Kırkıcı, Reference Oğuz and Kırkıcı2023), unlike the sensitivity shown for both true and pseudo-affixes (e.g., -er in the word corner) by children who speak a language with a deep orthography (e.g., Quémart et al., Reference Quémart, Casalis and Colé2011). These marked dissimilarities between the two languages inevitably lead to the question of whether and how such typological and orthographic differences may impact the cross-linguistic activation of words, in spite of the fact that the two languages use the same script.
The phonological and orthographic dissimilarities between English and Turkish translation equivalents made it possible for us to focus more on their semantic similarities. Controlling confounding factors is often a difficult task because many translation equivalents in language pairs like Spanish and English may involve a high degree of orthographic and phonological similarity. This results in either translation-equivalent pairs with more than just semantic similarity or a rather limited number of items, which, in turn, may lead to low statistical power. Using a language pair like Turkish and English, therefore, offers more variety in terms of options for experimental word lists, thus allowing more and better-controlled items to be included in the tasks.
One of the goals of the current study was to investigate the potential impact of individual differences and L2 exposure on translation-equivalent priming. The contradictory findings in earlier studies regarding the effect of L2 proficiency call for further research with different language pairs, as is the case in the present study. In addition, the potential effects of L2 exposure and L2 dominance have also been largely neglected in earlier research. The context of the current study was well-suited for differentiating the effects of L2 exposure from L2 proficiency and dominance, as the bilinguals tested in the present study possessed high levels of L2 proficiency and had been using their L2 dominantly over the last year, despite its acquisition after infancy.
Another common controversy in translation priming research is the effect of task type and configurations. Although some studies suggest that the SCT mitigates or even eliminates the commonly observed translation priming asymmetry (Finkbeiner et al., Reference Finkbeiner, Forster, Nicol and Nakamura2004), further research is required to confirm such a pattern due to conflicting findings (Xia & Andrews, Reference Xia and Andrews2015). It is also still controversial whether increasing the prime duration and/or SOA would allow L2–L1 priming effects to emerge; longer exposure to a prime word might help bilinguals activate/process the word better, or just giving them extra time to process L2 primes might enhance the priming effect without increasing the time of exposure. By including a comparatively high number of participants and items, the current study also sought to provide further insight into these contradictory findings.
Finally, the relation between word frequency and translation priming is still largely unknown due to the limited number of studies on the issue, which necessitated further investigation of the effects of prime word and target word frequencies on response times and priming effects. Moreover, a comparison between corpus frequency values and subjective frequency values was necessary, as corpora might fail to capture the language exposure experience of bilinguals. In addition to using corpus values, we collected subjective frequency ratings to compare their efficiency in explaining the variance in participant responses.
4. Methods
4.1. Participants
We recruited 101 highly proficient Turkish–English bilinguals who were undergraduate students in the English Language Teaching DepartmentFootnote 1 at Middle East Technical University (average age: 19.5 years, SD: 1.53), an English-medium university in Türkiye. Although some of them knew additional languages, the participants had been constantly exposed to L2 English in their department and reported English as their most proficient non-native language. None of the participants reported any learning disorders or neurological conditions, and they all had normal or corrected-to-normal vision. The mean self-reported onset age of L2 acquisition was 8.30 (SD: 3.50), implying that our sample group predominantly acquired their L2 after infancyFootnote 2. The participants reported learning L2 English through formal school education and self-study. Two participants did not attend the second part of the experiment, making the final number of participants 99. In accordance with Brysbaert and Stevens (Reference Brysbaert and Stevens2018), we aimed for 1,600 observations per condition to ensure sufficient statistical powerFootnote 3, and with the final number of participants and items, the available data comfortably surpassed the suggested value (2,376 observations per condition) even after excluding incorrect or outlier responses (see the Results section below). All participants took part in the study on a voluntary basis, and ethical approval for the study was obtained from the ethics committee of the university prior to any data collection.
As several participant-level variables were included as predictors in our analyses (see the Results section), here we provide operational definitions of the constructs measured. In the present study, L2 exposure refers to the cumulative amount of active experience with the L2 and was measured using the LHQ3 (Li et al., Reference Li, Zhang, Yu and Zhao2020), which integrates the age of onset and years of regular use. While L2 dominance can sometimes be expressed as a ratio value relative to L1, we preferred the basic L2 dominance score in the LHQ3, which reflects a combination of L2 daily usage and self-reported L2 proficiency. This preference was based on the fact that a bilingual who uses both languages in a limited manner can provide a very high L2 dominance ratio, even when the usage difference is small in absolute terms. These two constructs, L2 exposure and L2 dominance, were modeled separately as they index distinct dimensions of bilingual experience that may differentially modulate lexical processing.
In order to interpret 83-ms prime duration results more accurately, an additional 34 Turkish–English bilinguals, who had similar characteristics (in terms of age, proficiency, and language background) to the main sample group, were recruited. These bilinguals completed a prime awareness task to evaluate the visibility of 83-ms primes under the present masking conditions in both language directions.
4.2. Tasks
The masked priming paradigm was combined with two different tasks: the LDT and the SCT. Using masked primes allows researchers to investigate processes involved in pre-conscious stages of word recognition, as such primes are often referred to as invisible to human consciousness but still recognized by the pre-conscious processing mechanisms (Feldman et al., Reference Feldman, Kostić, Gvozdenović, O’Connor and Moscoso del Prado Martín2012; Forster & Davis, Reference Forster and Davis1984; Rastle et al., Reference Rastle, Davis and New2004). Each task had four different versions, and with these eight tasks, each language direction and each prime duration were tested in a fully crossed design in a separate experiment. The experiments were as follows: an LDT in the L1–L2 direction with 50-ms primes, a LDT in the L2–L1 direction with 50-ms primes, an LDT in the L1–L2 direction with 83-ms primes, an LDT in the L2–L1 direction with 83-ms primes, a SCT in the L1–L2 direction with 50-ms primes, a SCT in the L2–L1 direction with 50-ms primes, a SCT in the L1–L2 direction with 83-ms primes, and a SCT in the L2–L1 direction with 83-ms primes.
In the LDTs, participants decided whether a given set of letter strings constituted a real word in the target language. To investigate the pre-conscious processes, each trial began with a set of hashtags (i.e., forward masking) followed by a brief prime word (either for 50 ms or 83 ms). Right after the prime word in lowercase letters, a target word was presented in uppercase letters, thereby avoiding a visual continuity effect. In half of the trials, the prime was the translation equivalent of the target word; in the other half, it was an unrelated word.
The same masking and priming procedures were used in the SCTs. In this task, however, participants were informed that all target words were real words and that they were required to decide whether a target word belonged to a given category. Similar to the LDTs, half of the primes were related (translation equivalents), and the other half were unrelated. Each time the category changed, a warning screen appeared, displaying the new category and two example words in that category (e.g., New Category: Animals [e.g., cat, whale]). Since tracking the category change was essential in the task, these warning screens were presented twice, and participants were required to press a key twice to avoid accidental skips. The category change screens also served as small breaks between blocks, allowing participants to rest their eyes. To give the same opportunity in the LDTs, five short break screens were added to allow participants to rest, corresponding to the times of category changes in the SCTs. With this setup, LDTs and SCTs were nearly identical, except for the criteria on which the decisions were based.
The inclusion of the SCT in addition to the LDT was essential for testing the theoretical aims of the present study. The LDT is known to primarily engage orthographic and lexical-level processes; the SCT, on the other hand, requires participants to evaluate conceptual category membership and places stronger demands on semantic activation pathways. This allowed us to test whether translation priming reflects automatic lexical activation or whether it is modulated by conceptually driven filtering mechanisms. Therefore, the comparison of the LDT and SCT was seen as a principled test of the theoretical distinction between form-driven and meaning-driven accounts of translation priming.
A prime awareness experiment was created to evaluate the visibility of 83-ms primes in the L1–L2 and L2–L1 language directions and to assess the extent to which processing remained automatic or conscious. Since 50-ms primes are standard in the masked priming literature to ensure pre-conscious processing (e.g., Forster et al., Reference Forster, Mohan, Hector, Kinoshita and Lupker2003), we did not assess prime awareness for this duration. The same four experimental lists were used without non-words or filler words, and participants were told explicitly to pay attention to prime words and decide whether they were translation equivalents of the target words using the buttons of a Logitech F-510 gamepad (objective awareness evaluation). After this decision, the participants rated their awareness of the prime word, using the numbers (1–4) on the gamepad (subjective awareness evaluation). The two tasks (L1–L2 and L2–L1 awareness) were pseudo-randomized to avoid order effects.
4.3. Word lists
In the current design, we used translation equivalents as related primes and targets, which had little to no orthographic or phonological similarity. For the unrelated conditions, target words were preceded by prime words with no apparent orthographic, phonological, or semantic relations. Since we planned to use the same experimental words for the SCTs and LDTs, we chose 184 translation-equivalent pairs from six different categories: animals (e.g., cat-kedi), human body (e.g., leg-bacak), food (e.g., peach-şeftali), household items (e.g., table-masa), nature (e.g., tree-ağaç), and construction (e.g., wall-duvar). By using a Latin-square design, we created four different experimental lists. All items had three to eight letters. Two corpora, the corpus of contemporary American English (COCA; Davies, Reference Davies2008) and TS Corpus v2 (Sezer & Sever Sezer, Reference Sezer and Sever Sezer2013), were used to check word frequencies. The final 4 lists did not differ in their L1 word length, L2 word length, L1 word frequency, and L2 word frequency (p > .05). Although neighborhood density can modulate word processing patterns (Andrews & Hersch, Reference Andrews and Hersch2010), the lists in the current study did not control this variable, as there were no Turkish corpora or tools available at the time to check the orthographic neighborhood values of Turkish words at the time. Nevertheless, the lists were pseudo-randomized across conditions, and the presentation order of lists was not a significant factor in the exploratory analyses.
Unrelated primes were taken from the related primes of other lists, making related and unrelated primes similar in terms of word length, frequency, and category membership. ANOVA results showed that unrelated primes and related primes had similar word lengths and number of syllables across lists (all p values > .05). Although we carefully distributed unrelated primes to avoid semantic relatedness with target words, this was further tested using a Likert scale to measure the relatedness of related and unrelated primes to target words. In the scale, each prime word was rated on a 1–5 scale in relation to its degree of relatedness to a given target word (1 was completely unrelated, and 5 was completely related). In addition, the scale also asked whether the category memberships of target items were accurate. Using Google Classroom, the scale was administered to 18 participants, who did not participate in the main study but were similar to our sample group in terms of L1 and L2 language skills, age, and L2 AoA. The results showed a significant difference between the relatedness ratings of related primes (4.89 out of 5) and unrelated primes (1.30 out of 5) to target words.
ANOVAs showed that orthographic similarity between primes and targets did not differ between the lists (Appendix A), based on three different orthographic similarity measures: the Normalized Levenshtein values (F(1,732) = .25, p = .620), the Levenshtein Distance values (F(1, 732) = .05, p = .829), and the Orthographic Similarity values (F(1,732) = 2.52, p = .113). Related primes had more orthographic similarity to their target words compared to unrelated primes (p < .001), but the effect size for each measure was small (Eta2 values for the Normalized Levenshtein, the Levenshtein Distance, and the Orthographic Similarity were 0.03, 0.03, and 0.04, respectively). Nevertheless, we tested whether this difference had any influence on the results by including the orthographic similarity measures in the main analysis. Since all these measures were non-significant (no main effects or interactions), we removed them and concluded that the difference between related and unrelated primes in terms of orthographic similarity had no effect on priming effects or response times. We included the orthographic similarity measures in the main analyses to see if they had a main effect or interacted with priming effects.
As the two different tasks required different targets for the “No” answers, we created non-words for the LDTs and found filler words for the SCTs. A total of 184 non-wordsFootnote 4 for each language direction were created using the Wuggy software (Keuleers & Brysbaert, Reference Keuleers and Brysbaert2010); for each list, non-words were generated based on the words in another list that was not used to extract unrelated primes. For instance, if unrelated word primes for List 1 were taken from the words in List 2, non-words for List 1 were generated using the words in List 3. For the filler items, we found 184 L2 (English) words with similar lengths, frequencies, and concreteness values to the experimental words using the LexOPS package (Taylor et al., Reference Taylor, Beith and Sereno2020), and the translation equivalents of these words were used as Turkish filler words.
4.4. Individual difference measurements
4.4.1. LexTALE
The Lexical Test for Advanced Learners of English, (LexTALE) (Lemhöfer & Broersma, Reference Lemhöfer and Broersma2012) was used to estimate the L2 proficiency levels of the participants in an objective manner. This measurement includes a brief LDT and considers the correct answer rate and response times to predict language proficiency.
4.4.2. Language history questionnaire
The Language History Questionnaire Version 3 (LHQ 3.0; Li et al., Reference Li, Zhang, Yu and Zhao2020) was used to measure the L2 proficiency, L2 exposure, and L2 dominance scores of the participants. Comparing the subjective L2 measurement of the questionnaire with the objective measurement of LexTALE, we also aimed to investigate which method could provide a more accurate L2 measurement that can better explain the potential interaction between L2 proficiency and translation priming. L2 exposure in this questionnaire is measured by considering the age of onset for L2 acquisition and the years of L2 use after that initial acquisition, while L2 dominance scores are based on L2 proficiency and daily use of different L2 language skills. Critically, L2 dominance was calculated using only L2-related measures and does not indicate dominance relative to L1.
4.4.3. Subjective word frequency scale
Since the bilinguals in the current study mainly acquired their L2 in a classroom setting, corpus values might not have been fully representative of their L2 experiences. Despite the finding of Chen and Dong (Reference Chen and Dong2019) that refers to L2 English corpus values as accurate for Chinese–English bilinguals, the study’s small sample size and potential cros-slinguistic effects necessitate further confirmation of this proposal. Therefore, the subjective measurement of word frequency was conducted via a 7-point Likert scale (Balota et al., Reference Balota, Pilotti and Cortese2001), in which ratings ranged between never seeing the word and seeing it a few times a day. The scale was administered online using Google Forms after the main experiments had been completed. Each participant rated all experimental words in terms of their perceived frequency.
4.5. Procedure
We carried out two pilot studies with five and four participants, respectively. The participants were studying English at the School of Foreign Languages at that time and had intermediate English levels (determined by their school). In line with their feedback, we changed 4 experimental words and 2 non-words, changed the category of 9 non-exemplar words, and finally changed the names of two categories created for the SCTs.
Each participant used the Appointlet website to assign their names to available times in the laboratory. Open Sesame (ver. 3.3.12; Mathôt et al., Reference Mathôt, Schreij and Theeuwes2012) was used to run the experiment on a 15-inch Lenovo laptop (60-Hz display refresh rate), and the participants used a Logitech F310 gamepad to give their answers. Each participant was given a participant ID, which helped with pseudorandomization by determining the order of their tasks. We told participants to give accurate responses as fast as possible by using their dominant-hand button for the correct answers and the non-dominant-hand button for the wrong answers. Also, we informed them that they could take short breaks during the screens, labeled “short break” or “new category.” The participants completed four different practice sessions, each for a different experiment (L1–L2 LDT, L2–L1 LDT, L1–L2 SCT, and L2–L1 SCT). After the practice, no further explanation or help was provided to the participants.
In the first data collection period, 50 bilinguals completed the 83-ms versions of all experiments, while 51 of them completed the 50-ms versions (two LDTs and two SCTs in different language directions). Once a list was used in an experiment, it was eliminated from subsequent tasks within that session, ensuring that participants saw a word or its translation equivalent only once across four experiments on that day. Each session (four experiments) took about 30 minutes, and each participant completed the LexTALE using their own mobile phone after the first session. Each day, we uploaded the collected data to the Open Science Framework (OSF) under a folder named after the date. The same participants completed the same 4 experiments with a different prime duration (either 50 ms or 83 ms, depending on their first session) after at least 25 days (30.6 days on average), as the test – retest effect was shown to be non-significant for LDTs after around a month (Pluck, Reference Pluck2020). Upon completing the second session, we told the participants to complete the Subjective Word Frequency Scale and LHQ v3 at their own time. Only two participants quit the study after the first session, making the final number of participants 99. Although each participant saw an item only once in a session and extreme precautions were taken for pseudo-randomization, potential order effects were also examined in the main analyses by considering item, task, and session orders. The significant effects and interactions were the same; therefore, these were not included in the final models.
5. Results
We used R 4.2.3 (R Core Team, 2023) for data processing and analysis. The prime awareness task revealed that 83-ms primes were mostly visible. The participants could correctly identify prime–target relationships 91% of the time in the L1–L2 direction and 88% of the time in the L2–L1 direction. Subjective awareness ratings were also high, averaging 3.41 for L1 primes and 3.21 for L2 primes (out of 4 points). Therefore, the two prime durations indexed differences between pre-conscious processing and conscious processing. For the main experiment, the experimental items consisted of 36800 data points in total after excluding non-words and fillers. Two participants were excluded due to their absence in the second part of the experiment, and the items with high error rates (over 40%) in any task were excluded from the data of all tasks (9 items in total) to ensure a reliable comparison across different tasks. Upon examining the response time distribution, responses that were too fast (under 200 ms) or too slow (above 2000 ms) were removed (329 responses in total). The final experimental data had 33,193 data points. The descriptive statistics for the final data are given in Table 1.
Descriptive statistics of the tasks and prime conditions
Table 1. Long description
Descriptive statistics of the tasks and prime conditions (LDT: Lexical decision task, SCT: Semantic Categorization task).
Note. LDT, lexical decision task; SCT, semantic categorization task.
Bold values indicate the priming effect size for each task and language direction.
In line with our pre-registration report, we tried different methods to normalize the data and selected the Order Norm transformation method (Peterson & Cavanaugh, Reference Peterson and Cavanaugh2020) applied by the BestNormalize package (Peterson, Reference Peterson2021), as this transformation provided the most appropriate normalization based on histograms and Q–Q plots. Missing data were imputed using the K-NN method in the caret package (Kuhn, Reference Kuhn2008) for prime subjective frequency ratings (1.30% missing), target subjective frequency ratings (0.98% missing), L2 dominance score (1.01% missing), and L2 immersion (5.05% missing).
In order to determine whether subjective or objective measures better predicted response times better, an initial model selection phase was carried out with potential combinations of these variables. In addition, we tested models with composite scores based on both objective and subjective measurements. The model selection procedures are summarized in Supplementary Table 1. A total of six different full models (linear mixed-effect models) were created using subjective and objective measurements for word frequency and L2 proficiency, along with other independent variables and potential interactions. Despite the differences in word frequency and L2 proficiency measures (subjective vs. objective), all six models had normalized response times as their dependent variable and participant and word IDs as their random effects. The independent variables were prime type, task type, language direction, prime duration, L2 proficiency (subjective or objective), L2 dominance, L2 exposure, target frequency (subjective or objective), and prime frequency (subjective or objective). The main interaction was prime type × task type × language direction × prime duration, and in order to investigate further interactions, each additional variable was examined in an additional interaction added sequentially to this base formula (e.g., prime type × task type × language direction × prime duration × L2 proficiency).
Based on AIC (lowest), BIC (lowest), and marginalized R 2 (highest) values, the model including objective measurements for both indices was selected. In the next step, the selected full model was refined through automated backward elimination (lmerTest package, step function; Kuznetsova et al., Reference Kuznetsova, Brockhoff and Christensen2017), in line with our pre-registration report. This procedure was used to remove non-significant variables and improve model fit. Using Satterthwaite’s approximation, non-significant variables were removed iteratively based on the highest p-value until all the remaining variables were statistically significant. Random slopes were also considered in this process where appropriate. A likelihood ratio test showed that the refined model did not significantly differ from the full model (df = 61, Chiq = 64.118, p = .37). In addition, AIC and BIC values were also lower for the refined model (81,863 vs. 81,921 and 82,191 vs. 82,762, respectively); therefore, we selected this reduced model. The refined model showed six significant main effects, nine significant two-way interactions, four significant three-way interactions, and one significant four-way interaction for the response time data (Table 2).
The final linear mixed-effects model used for the discussions
Table 2. Long description
The final linear mixed-effects model used for the discussions.
Bold values indicate statistical significance.
In order to test whether the results were due to overfitting (considering high-order interactions), a simpler model with core predictors (prime type, task type, language direction, and prime duration) was also created. The significance of main effects and most interactions was consistent across the models; while task type × language direction interaction became significant in the simpler model (p = .023), it was more parsimoniously captured in the final model as a three-way interaction among task type, language direction, and target frequency (p = .005). Furthermore, the final model provided lower AIC and BIC values (81,863 vs. 82,032 and 82,191 vs. 82,200), along with a higher proportion of explained variance (Marginal R 2 = .08 vs .06). Overall, this confirmatory analysis indicated that the findings in the final model were not due to overfitting and that the inclusion of higher-order interactions was necessary.
Several main effects had significant influences on response times (prime type, task, language, prime duration, L2 exposure, and prime corpus frequency). The participants gave faster answers to target words after related primes, indicating a significant priming effect (Table 2). In addition, responses were faster in their L1 than in L2, in the LDTs, when their L2 exposure scores were higher, when prime duration was shorter (50 ms), and, finally, when prime frequency was higher.
The interaction among prime type, language, and task revealed comparable priming effects in both language directions and tasks when the prime duration was 50 ms (Figure 1). The SCTs elicited more priming effects compared to the LDTs, but differences between language directions were not statistically significant in both tasks with 50-ms primes.
Response times in different language directions and tasks with 50-ms primes. Diff values at the bottom of the graphs show the response time difference between the related and unrelated conditions (Focused on the 300- to 1,500-ms interval for better visuals).
Figure 1. Long description
Response times in different language directions and tasks with 50-ms primes. Diff values at the bottom of graphs show the response time difference between the related and unrelated conditions (Focused on the 300–1500 ms interval for better visuals).
Individual differences had a significant influence on the observed priming effects. Response times decreased with increasing L2 proficiency (based on LexTALE scores) in both language directions (Table 2), and this effect was more apparent in the SCTs. Although higher L2 proficiency elicited faster responses in both language directions, the difference between related and unrelated responses increased in the L2–L1 direction with L2 proficiency, indicating increased L2–L1 priming effects (Figure 2). In the L1–L2 direction, increasing language proficiency decreased the response time gap after related and unrelated primes, decreasing the observed priming effects in that direction.
The interaction between L2 proficiency (LexTALE) scores and prime type in the L1–L2 (left) and L2–L1 (right) directions.
Figure 2. Long description
The interaction between L2 proficiency (LexTALE) scores and Prime Type in the L1-L2 (left) and L2-L1 (right) directions.
Both L2 exposure and L2 dominance decreased response times, with the latter providing stronger effect with 50-ms primes (Supplementary Material, Figure 6). In addition to decreasing response times in general, L2 exposure also interacted with prime type, increasing translation priming effects significantly (Figure 3).
The interaction between L2 exposure scores and prime type in the L1–L2 (left) and L2–L1 (right) directions.
Figure 3. Long description
Response times in different language directions and tasks with 83-ms primes. Diff values at the bottom of graphs show the response time difference between the related and unrelated conditions. (Focused on the 300–1500 ms interval for better visuals).
Response times in different language directions and tasks with 83-ms primes. Different values at the bottom of the graphs show the response time difference between the related and unrelated conditions (focused on the 300- to 1,500-ms interval for better visuals).
Figure 4. Long description
The interaction between L2 Exposure scores and Prime Type in the L1-L2 (left) and L2-L1 (right) directions.
When prime duration was 83 ms, a language asymmetry effect emerged since the L1–L2 direction produced more priming effects compared to the L2–L1 direction. The asymmetry was associated with an inhibitory effect of the longer prime duration in the unrelated conditions, which was more prominent in the L1–L2 direction (Table 1). Importantly, the observed differences between language directions were similar across the two tasks (Figure 4). Prime Duration also showed a significant interaction with L2 exposure and L2 dominance. The effect of L2 exposure scores was higher in the 83-ms prime condition, while L2 dominance scores decreased response times more in the 50-ms prime condition, showing opposite influences.
Finally, both prime frequency and target frequency were significant variables in the model. Although Target Corpus Frequency did not interact with prime type and thus failed to modulate priming effects, it had a significant three-way interaction with Task and Language (Supplementary Material, Figure 7). In the L1–L2 direction, higher frequency values for target words decreased response times, indicating a facilitative effect. However, in the L2–L1 direction, response times became slower with increasing target frequency, showing an inhibitory effect.
Prime Corpus Frequency decreased response times significantly in the related conditions, indicating increased priming effects. Furthermore, Prime Corpus Frequency significantly interacted with task, prime duration, and prime time: Although the higher frequency of primes mostly decreased response times in the related conditions, for the SCTs with a longer prime duration (83 ms), the effect for the related condition was less prominent, but it also emerged in the unrelated condition as well. Notably, the effect of decreasing response times with increasing Prime Corpus Frequency was more powerful in the unrelated condition of 83-ms SCTs compared to the related conditions, decreasing the magnitude of priming effects under these conditions (Figure 5).
Interaction among prime corpus frequency, task, prime type, and prime duration.
Figure 5. Long description
Interaction among Prime Corpus Frequency, Task, Prime Type, and Prime Duration.
Overall, findings are summarized as follows by referring to each research question. Regarding the potential effect of language direction (RQ 1), bilinguals showed symmetrical translation priming effects in both language directions at pre-conscious processing stages (50-ms primes). In terms of potential task differences (RQ 2), SCTs elicited more priming effects than LTDs, with symmetrical patterns between language directions across both tasks. The effect of prime duration (RQ 3) also reflected some of the differences between pre-conscious and conscious word processing, as 83-ms primes were mostly visible. The prolonged prime duration, tapping into more conscious processing, elicited more priming effects, and this was due to delayed responses in the unrelated conditions, especially in the SCT. Notably, the symmetrical translation priming pattern also changed with 83-ms primes, as the responses in the L2–L1 directions were slower, creating a language asymmetry. As for individual differences (RQ 4), higher L2 proficiency, exposure, and dominance predicted faster responses, while higher L2 proficiency and exposure boosted L2–L1 priming effects. Finally, with regard to word frequency (RQ 5), higher prime frequency increased responses and priming effects. Target frequency, however, resulted in divergent effects across language directions; higher target frequency elicited faster responses in the L1–L2 direction but slower responses in the L2–L1 directions, especially with the SCT.
6. Discussion
In this study, we investigated the potential interaction between the languages of a bilingual in the early stages of word processing. Using the LDTs and SCTs in opposite language directions (L1–L2 and L2–L1) with two prime durations (50 ms and 83 ms), the current study showed that highly proficient Turkish–English bilinguals exhibit robust translation priming effects in different language directions, tasks, and prime durations, broadly consistent with our main hypotheses. Moreover, individual differences and word frequency modulated response times and priming effects. The findings are more readily accounted for within the quantitative view, as priming effects were observed in both language directions and higher L2 proficiency and exposure boosted the observed priming effects. In the discussion below, we refer back to the research questions presented at the end of the Literature Review by number (e.g., RQ 1) when necessary.
Despite mostly acquiring their L2 after infancy, participants showed significant and identical priming effects in both language directions with 50-ms primes (RQ 1), posing challenges to the claims of qualitative models. According to the RHM, L1 words activate L2 words through shared semantic representations, but L2 words can activate L1 words through direct links (Kroll & Tokowicz, Reference Kroll, Tokowicz and Nicol2001). Although this would suggest larger priming effects in the L2–L1 direction, Xia and Andrews (Reference Xia and Andrews2015) argue that L2 words might not actually activate their equivalents in L1 due to weaker associations between L2 words and semantic representations, proposing a revision to the model based on language asymmetry findings in the literature (e.g., Chaouch-Orozco et al., Reference Chaouch-Orozco, González Alonso and Rothman2021; Finkbeiner et al., Reference Finkbeiner, Forster, Nicol and Nakamura2004; Jiang & Forster, Reference Jiang and Forster2001). The current results are difficult to reconcile with the old and new versions of the model; L2 word primes indeed activated L1 target words, and the size of that priming was similar to that of the L1–L2 direction. This finding is not readily compatible with the Episodic L2 Hypothesis either, as the model eliminates the possibility of L2–L1 priming altogether by positing that L2 words are in an entirely different memory system that is not sensitive to lexical decisions (Jiang & Forster, Reference Jiang and Forster2001).
The two tasks in the study, the LDT and SCT, resulted in distinct response patterns in terms of accuracy and speed, despite their nearly identical procedures except for the response criteria (categorical vs. lexical decisions). Participants were less accurate and slower with the SCT, especially in the L1–L2 direction (Table 1), suggesting a potential difficulty in determining categorical features compared to making a lexical decision. The SCTs also elicited larger priming effects than the LDTs (Figure 1), consistent with our second hypothesis (RQ 2). In spite of these differences, however, the same priming effect pattern was observed for the LDTs and SCTs in the current study; both tasks showed similar priming effects across language directions with 50-ms primes. This is not consistent with the Sense Model, which predicts comparable priming effects only in the SCT (Finkbeiner et al., Reference Finkbeiner, Forster, Nicol and Nakamura2004). The model posits that L2 lexical representations are weaker and can only be activated with short prime durations through the pre-activation of categorical cues (Xia & Andrews, Reference Xia and Andrews2015). While such pre-activation might account for the greater semantic priming in the SCTs compared to the LDTs, the sense model cannot predict identical L2–L1 and L1–L2 priming effects in sequential bilinguals, thus providing limited support for the weak-L2-representations view. Overall, these findings present further challenges for qualitative models by showing that sequential bilinguals can indeed exhibit L2–L1 priming as efficiently as L1–L2 priming, even with a very brief prime duration (50 ms).
Multilink can explain the aforementioned findings through its task–decision system. The identical priming effects with 50-ms primes indicate that longer prime durations or SOA values are not necessary to elicit similar L2–L1 and L1–L2 priming effects in sequential bilinguals, in contrast to the 100-ms prime duration in Schoonbaert et al. (Reference Schoonbaert, Holcomb, Grainger and Hartsuiker2011) and 150-ms SOA in Lee et al. (Reference Lee, Jang and Choi2018). If L2 words had weaker representations or less efficient connections to semantic representations, as qualitative models have suggested (Finkbeiner et al., Reference Finkbeiner, Forster, Nicol and Nakamura2004; Jiang & Forster, Reference Jiang and Forster2001; Kroll & Tokowicz, Reference Kroll, Tokowicz and Nicol2001), this similarity between language directions would not have been observed under these conditions. Within Multilink, both tasks can show comparable priming in opposite language directions since the emergence of translation priming depends on the efficiency of the task–decision system and resting activation levels. Given the numerous semantic features shared between translation equivalents, the task–decision system can adjust word processing patterns and induce more translation priming during more semantics-focused tasks. This also provides an explanation for why the SCT elicits more priming than the LDT.
Further support for Multilink, particularly for its resting activation levels, comes from the influence of individual differences on response times. In the current study, L2 dominance also decreased response times, with the effect being more prominent with 50-ms primes, possibly because this measurement might be more related to pre-conscious processes. More crucially, both L2 proficiency and L2 exposure scores positively correlated with faster answers in general (Table 2) and more priming effects, in line with Kahraman et al. (Reference Kahraman, Casalis, Javourey-Drevet, Menut, De Wit and Beyersmann2025) and our initial hypothesis (RQ 4). This aligns closely with Multilink, in which higher L2 proficiency and exposure can adjust the activation levels for orthographic, phonological, and semantic features for L2 words. In the model, more exposure leads to higher proficiency, which in turn results in faster word activation. Furthermore, both higher L2 proficiency and exposure scores increased L2–L1 priming. Since L2 proficiency gradually improves as a result of L2 exposure, both scores can indicate better processing of brief L2 primes and the activation of relevant features in the pre-conscious processing, suggesting modulation of more automatic stages (Figures 2 and 3). The way the task–decision system performs on different tasks can also depend on different language skills, given that particular L2 skills seem to be more prominent in modulating translation priming in different tasks (McPhedran & Lupker, Reference McPhedran and Lupker2021).
Overall, the findings paint a coherent picture of whether early acquisition or L2 proficiency and exposure determine L2–L1 translation priming. Most participants in the current study acquired their second language after the age of 7, long after the L1 vocabulary system had been established (the average onset of L2 AoA was 8.30). Despite their rather late start for L2 acquisition, the participants appear to have closed the gap to a considerable extent through massive L2 exposure, adjusting the activation levels of lexicon features and increasing overall L2 proficiency gradually. In line with this, the total amount of L2 exposure, rather than the onset of L2 AoA, appears to play a stronger role in the emergence and strength of L2–L1 translation priming. An exploratory analysis confirmed this argument, as the onset of L2 AoA had no significant influence on response times or translation priming effects. The results then challenge the idea that early acquisition of an L2 is required for native-like word representations and processing (e.g., Dimitropoulou et al., Reference Dimitropoulou, Duñabeitia and Carreiras2011b; Jiang & Forster, Reference Jiang and Forster2001; Kroll & Tokowicz, Reference Kroll, Tokowicz and Nicol2001). Our participants were English Language Teaching students who had been using their L2 (English) extensively for both comprehension and production, and identical priming effects in both language directions suggest that our sample of bilinguals might have used their L2 at such a frequency that they may approximate a state of balanced bilingualism. Taken all together, L2 proficiency and L2 exposure can explain the current results better and support the central premises of Multilink.
One controversial issue regarding translation priming is the potential effect of language orthography. Both Turkish and English use the Latin script, but they have distinct orthographic depths; the former possesses a shallow orthography, while the latter possesses a deep orthography. The similarity in the scripts might have helped the bilinguals process L2 primes with more efficiency (Nakayama, Reference Nakayama, Lupker and Itaguchi2018), allowing identical priming effects in both language directions with 50-ms primes. However, Chaouch-Orozco et al. (Reference Chaouch-Orozco, González Alonso and Rothman2021) failed to find significant L2–L1 priming effects in bilinguals speaking Spanish and English, a language pair that also uses the same script and has a similar contrast in orthographic depth, underscoring the necessity of investigating the potential effect of orthography and script in future studies. Nevertheless, the emergence of strong priming effects in both language directions shows that differences in language orthography do not necessarily prevent solid L2–L1 priming effects from emerging when the languages share the same script.
A distinct pattern was obtained in the experiments using a longer prime duration (83 ms). In both tasks, experiments with 83-ms primes elicited more translation priming, which is partially consistent with our initial hypothesis (RQ 3). Contrary to our expectations, however, this was the result of slower responses after unrelated primes rather than more efficient prime processing indicated by faster answers after related primes. This inhibitory effect was significant and robust in both language directions in the unrelated condition, but it was more pronounced in the SCTs (Table 1 and Figure 4). Crucially, the second experiment measuring prime awareness showed that 83-ms primes were mostly visible to the bilinguals, indicating a shift toward more conscious than pre-conscious processing stages. Therefore, the longer prime duration (83 ms) in the current study should be interpreted within the framework of conscious and visible word processing rather than early pre-conscious processing. These findings suggest that language direction asymmetry may be absent at pre-conscious stages but can emerge when processing becomes conscious.
Processing prime words in conscious stages is likely to activate more competitors, which in turn delays the responses after unrelated primes. More competition under different conditions is often associated with an inhibitory effect in word processing. For instance, ambiguous primes with multiple potential meanings are known to require additional cognitive effort compared to unambiguous primes (Bilenko et al., Reference Bilenko, Grindrod, Myers and Blumstein2008), possibly due to the activation of more competitors. Another known factor is the number of orthographic neighbors for a real word, a phenomenon called the prime lexicality effect (e.g., Davis & Lupker, Reference Davis and Lupker2006; Kida et al., Reference Kida, Barcroft and Sommers2022). When primes are orthographic neighbors to targets with no semantic connections, real-word primes delay the responses while non-word primes have no effect. Importantly, the degree of this inhibition is increased by the number of orthographic neighbors between the prime and target, underscoring the detrimental effect of more activated competitors. Similar to how inhibition arises in such situations, delayed responses in the unrelated conditions of the 83-ms experiments may also stem from increased lexical competition at conscious, rather than pre-conscious, word processing stages. Although primes and targets shared little to no orthographic similarity in the current experiment, the boost for orthographic and phonological features through a prolonged prime duration and conscious processing might have activated more unrelated competitors, disrupting semantic activation.
The reason why this inhibitory effect mainly emerged in unrelated conditions might be potentially related to a reset mechanism in word processing (Holcomb & Grainger, Reference Holcomb and Grainger2007). Semantic activation follows orthographic activation in the brain (e.g., Holcomb & Grainger, Reference Holcomb and Grainger2006), and this may cause semantic processing to continue even after the disappearance of a prime word (Nakayama et al., Reference Nakayama, Lupker and Itaguchi2018). It has been proposed that the processing system suppresses orthographic and phonological information to leave only semantic information, which is crucial in comprehending sentences in many contexts (Schoonbaert et al., Reference Schoonbaert, Holcomb, Grainger and Hartsuiker2011). Such suppression might be even faster when the following word presents an apparent orthographic mismatch (Nakayama et al., Reference Nakayama, Lupker and Itaguchi2018), similar to the orthographic dissimilarity between primes and targets in our experiments. When primes are related, semantic activation might suppress the irrelevant competitors quickly, resetting everything except for the shared semantic meaning. With unrelated primes, however, this process can become more challenging as no obvious candidate with semantic similarity to the target appears, and a more strategic analysis might be required to reject all lexical competitors by an inhibitory process (e.g., Green, Reference Green1998). Given that automatic priming and strategic priming activate different patterns in the brain, and the latter being more cognitively demanding (Rossell et al., Reference Rossell, Bullmore, Williams and David2001), unrelated primes might have delayed responses more in the longer prime duration when prime processing is mostly conscious.
The stronger inhibitory effect in the SCTs compared to the LDTs is strongly related to the nature of this task. Whether primes provide inhibitory or facilitative influence for responses, task requirements determine the strength of priming effects (e.g., De Wit & Kinoshita, Reference De Wit and Kinoshita2015; Zhang et al., Reference Zhang, Wu and Wang2025). When making a lexical decision, deciding whether the given letters constitute a word does not require any semantic or categorical decisions; therefore, the activation of more competitors, which are all legitimate words, might be less inhibitory in the LDT. In the SCT, where semantic processing is crucial (Finkbeiner et al., Reference Finkbeiner, Forster, Nicol and Nakamura2004), this competition might become more costly. With a short prime duration tapping into pre-conscious processing, bilinguals might activate the semantic features that have the strongest associations with the target word in the SCT, which usually indicates a single category. The word bird, for example, is usually associated with animal species. A brief exposure to this word can activate the same category for both L1 and L2 speakers. When the prime is presented for longer (e.g., 80 ms or more) and prime processing becomes more conscious, the semantic features that have weaker connections to the target word might also be activated. This argument is supported by the findings that a long SOA value fully activates the semantic features of a prime (Hofmann et al., Reference Hofmann, Kleemann, Roelke-Wellmann, Vorstius and Radach2022), increasing the semantic richness of the word along with its challenge in a semantic-focused task (Yap et al., Reference Yap, Tan, Pexman and Hargreaves2011). The same word, bird, might also refer to a piece of equipment in badminton, and thus conscious processing of this word can activate the category of sports. When the prime is unrelated, having multiple unrelated categories activated by the prime might slow down the responses substantially with a longer prime duration.
Another unexpected pattern was the emergence of a language asymmetry effect with the longer prime duration. Although we predicted a language asymmetry effect (RQ 1), we certainly did not expect it to appear with only 83-ms primes. Nevertheless, 83-ms primes were largely visible and investigated patterns involving more conscious word processing stages, as revealed by the prime awareness task. The results showed that the unrelated conditions were similar across language directions, and we found that the asymmetry appears to be driven by the different patterns for the related conditions across language directions. In particular, the related conditions of both tasks elicited similar response times in the L1–L2 direction across different prime durations, but in the L2–L1 direction, the responses were slowed down by an average of 9 ms with the 83-ms primes. In fact, these slower responses were the primary factor for the language direction asymmetry, and without them, the priming sizes would have been similar across language directions in both tasks, showing comparable translation priming.
Contrary to the common argument in the literature regarding inefficient L2 prime processing (e.g., Nakayama et al., Reference Nakayama, Lupker and Itaguchi2018), the current results show that bilinguals can exhibit similar translation priming in opposite language directions even with 50-ms primes, thus arguing against inefficient L2 prime processing with 50-ms primes. The reason why this inhibitory effect only emerged in the L2–L1 direction for related tasks can also be associated with the involvement of conscious processing, which can activate more unrelated competitors as discussed above. The suppression of more competitors or more-conscious processes might be easier for L1 primes than L2 primes in the conscious word processing stages. Apparently, a longer prime duration does not necessarily increase the efficiency of processing L2 primes (e.g., Wang & Forster, Reference Wang and Forster2015), and the boost in L2 processing can be observed by giving bilinguals more time to process rather than exposing them to the prime longer (Lee et al., Reference Lee, Jang and Choi2018). The findings also contradict the claim of Schoonbaert et al. (Reference Schoonbaert, Holcomb, Grainger and Hartsuiker2011), who argue that higher L2 proficiency and increased prime duration would both allow more efficient L2 prime processing in a similar way. When the prime duration is long enough to tap into conscious word processing stages, the patterns of translation priming can change substantially, showing opposite patterns to those observed in the pre-conscious processing stages. This finding is especially important considering that primes of approximately 80 ms or longer are used in translation priming research (e.g., Wang & Forster, Reference Wang and Forster2015).
The study’s final focus was the potential role of target and prime frequencies. The effect of word frequency had been mostly ignored in the translation-priming studies (Chaouch-Orozco et al., Reference Chaouch-Orozco, González Alonso and Rothman2021), leaving several questions unanswered. In the current study, target and prime frequencies showed significant effects on response times. Contrary to Nakayama (Reference Nakayama, Lupker and Itaguchi2018), in which prime frequency had no effect on response times, higher prime frequency values predicted faster answers for the related primes in both 50-ms and 83-ms prime conditions, boosting observed priming effects (RQ 5). In the 83-ms prime duration version of the SCTs, higher prime frequencies were associated with faster answers for both related and unrelated primes (Figure 5). The longer prime duration and the involvement of conscious prime processing might have allowed the activation of additional unrelated semantic features in the unrelated condition, and prime frequency might have mitigated this inhibitory effect to some extent by prioritizing the most common semantic features attached to the prime word, making the categorical decision easier.
Target frequency played completely opposite roles in different language directions. In the LDTs, higher target frequency predicted faster responses for L2 target words, in line with previous research (e.g., Diependaele et al., Reference Diependaele, Lemhöfer and Brysbaert2013), whereas in the L2–L1 direction, it slowed down the responses for L1 targets. Interestingly, this pattern was more prominent in the SCTs. Although we expected faster answers with increasing frequency for L2 words, we did not hypothesize that the same value would have a detrimental effect on L1. The bilinguals may have activated L2 words of higher frequency more easily, thus comparing them with the features activated by prime words faster. For L1 target words, however, highly frequent L1 words might have activated more competitors, some of them conflicting with the features of activated L2 primes. In the SCT, this might also activate some unrelated categories, further slowing down responses. This inhibitory effect of word frequency for L1 words might, in the present data, only be apparent for target words, which stay on the screen much longer than prime words and can be consciously seen.
The findings of the current study are more consistent with quantitative models, particularly Multilink, which propose that the differences between L1 and L2 word processing diminish as L2 skills improve. First, the enhanced translation priming observed in the SCT compared to the LDT can be associated with Multilink’s task–decision system. Since this system modulates word processing based on specific task criteria, and the SCT requires a semantic evaluation of the target, the shared semantic features between translation equivalents become more essential in processing, enhancing the observed translation priming effects. In addition, identical priming effects with 50-ms primes in both language directions align well with Multilink’s argument regarding resting activation levels. The results regarding 83-ms primes, on the other hand, should be interpreted within a conscious processing framework, rather than the pre-conscious and automatic paths and interactions modeled by Multilink. Although our sample consisted of sequential bilinguals, these participants had experienced extensive L2 exposure through their academic programs (English Language Teaching). Under Multilink, this massive exposure adjusts the resting activation levels of L2 words, making their activation as efficient as L1 words. This finding also challenges the qualitative view that early acquisition is required for native-like processing. Crucially, the finding that L2 proficiency and L2 exposure enhanced L2–L1 translation priming further supports this mechanism in the current study. Finally, the observed prime frequency effects provide additional evidence for Multilink. Because subjective frequency is an essential factor modifying word activation levels in the model, the finding that higher prime frequency enhanced translation priming fits well within Multilink’s architecture. However, given the limitations discussed below, the current findings alone cannot be taken as definitive evidence confirming the validity of Multilink, or any other bilingual lexicon model for that matter.
7. Limitations and recommendations for further research
The current study focused only on highly proficient bilinguals, which is likely to reveal a different picture compared to other proficiency levels. Further studies should investigate translation priming using the same language pair and less-proficient bilinguals. This way, the development of the interaction between the L1 and L2 lexicons could be better understood.
A second limitation is related to how L2 exposure and dominance information were collected. Although L2 proficiency was also measured through a subjective tool (LHQ), the models with LexTALE scores as an indicator of L2 proficiency had lower AIC scores, indicating better fit, confirming the necessity of using objective measurements to assess L2 proficiency skills (McPhedran & Lupker, Reference McPhedran and Lupker2021). While there are more objective ways to measure L2 exposure and dominance, further studies might carry out interviews with participants, which might allow the collection of more accurate data due to the availability of making explanations regarding concepts.
Although using longer prime durations than the conventional 50-ms duration is common in translation priming research, the current study showed that a prime duration around 80 ms can actually tap into conscious stages of word processing. This raises a critical issue, as most bilingual lexicon models have claims regarding pre-conscious word processing stages. Future research using long prime durations, such as 83 ms, must be cautious in interpreting their findings within the framework of pre-conscious word processing. Therefore, it is critical to investigate safe prime durations and masking procedures (e.g., using backward masks) to prevent prime visibility. Furthermore, although many studies, including this one, assume that 50-ms primes reliably tap into pre-conscious word processing stages, future work should investigate whether participants are aware of 50-ms primes in masked priming experiments.
Supplementary material
The supplementary material for this article can be found at http://doi.org/10.1017/S1366728926101679.
Data availability statement
The data that supports the findings of this study are openly available in OSF at https://osf.io/c3gmf/overview.
Acknowledgments
We thank our colleagues and reviewers for their valuable suggestions.
Funding statement
This research received no specific grant from any funding agency, commercial, or not-for-profit sectors.
Competing interests
The authors declare no competing interests.
AI contributions to research
Gemini was used for grammar and spell check.
Appendices
A. Appendix. Orthographic similarities between items
The orthographic similarity between targets and primes was similar in all lists, as shown by the Normalized Levenshtein values (F(1,732) = 8.50e-03, p = .927; Eta2 (partial) = 1.16e-05, 95% CI [0.00, 1.00]), the Levenshtein Distance values (F(1, 732) =21.46, p < .001; Eta2 (partial) = 0.03, 95% CI [0.01, 1.00]), and the Orthographic Similarity values (F(1,732) = 8.50e-03, p = .927; Eta2 (partial) = 1.16e-05, 95% CI [0.00, 1.00]). Overall, related primes had slightly greater orthographic similarity to target words than unrelated primes since translation equivalents were used as related primes, while unrelated primes were handpicked. The effect sizes for these differences were tiny for each measure (0.04, 0.03, and 0.04 for the Normalized Levenshtein, the Levenshtein Distance, and the Orthographic Similarity, respectively), and the orthographic similarity scores were not significant in any analyses. Therefore, we excluded them from the final analysis and decided that this little variation did not affect the response patterns.
B. Appendix. Post hoc power analysis
Post hoc Analysis with Raw Estimates, NCP-based Power, and Labels (Significant Predictors Only).
Figure 4. Long description
The interaction between L2 Exposure scores and Prime Type in the L1-L2 (left) and L2-L1 (right) directions.
Note. Power labels are as follows: low (<0.50), medium (0.50–0.79), and high (≥0.80).
Open access
