Enduring cognate facilitation: evidence for form-based transfer in bilingual word recognition

Cognates are words that share meaning across languages alongside all or most of their orthography. They are generally processed faster than control words, an effect known as cognate facilitation. The current study set out to investigate whether long-term cognate facilitation in a multi-day practice paradigm is driven by shared mental representations or graded cross-linguistic similarity effects. Dutch-English participants completed a visual lexical decision paradigm over five consecutive days, with testing on the first and fifth day in one language (L1 or L2) and practice on the intervening days in the other language. Target words included identical and non-identical cognates, non-cognate translation equivalents, controls, and non-words. While results showed a robust procedural practice effect across all word types, both identical and non-identical cognates exhibited a significantly larger cross-linguistic performance transfer than translation equivalents and control words. Crucially, the transfer effects for identical and non-identical cognates did not statistically diverge, and no additional transfer was observed for purely semantic translation equivalents. These patterns suggest that shared semantic representations alone cannot account for enduring cognate facilitation. Rather than proving isolated shared structural representations, the findings demonstrate that cross-linguistic transfer in visual word recognition is heavily modulated by graded form-based similarity, reflecting tightly coupled orthographic and phonological processing pathways.

One of the skills most people can do automatically is read. We see a word and cannot not read it (cf. the Stroop effect; Stroop, 1935). People recognize letters in words easily (Rumelhart and McClelland, 1982) and subsequently access their meaning. However, for individuals who speak more than one language, recognition of letter strings as words is not restricted to the language context in which they find themselves. Indeed, it seems that a bilingual's languages are simultaneously activated and that lexical access is language non-selective (e.g., Brysbaert and Duyck, 2010; Dijkstra and van Heuven, 2002). This is demonstrated by studies employing interlingual homographs and cognates (e.g., Dijkstra et al., 1999; Van Assche et al., 2012). The former are words that share orthographic form but not meaning, whereas the latter share both. Cognates can further be divided into identical and non-identical words. An example of an identical cognate is the word school in Dutch and English, and a non-identical cognate is rivier in Dutch and its translation equivalent river in English. The overall assumption is that conceptual representations (i.e., meanings) are shared across languages in the bilingual brain (e.g., Dijkstra and van Heuven, 2002; Francis, 2005), whereas phonological representations are distinct (Peeters et al., 2013). It was the aim of the present study to shed more light on whether orthographic representations of cognates are more likely shared or distinctive.

As mentioned above, studies employing homographs and cognates have provided evidence for language non-selectivity in bilinguals. To illustrate, Dijkstra et al. (2000) asked Dutch-English participants to respond to words only if that word existed in English. In their task, they employed both interlingual homographs and control words. Results showed higher response latencies for homographs, suggesting that activation of the Dutch lexicon interfered with lexical decision for English targets. Whereas, homographs seem to interfere with lexical decision in a single language context, cognates appear to facilitate the endeavor. For instance, Caramazza and Brones (1979) asked Spanish-English subjects to classify letter strings as either English or Spanish words or non-words. Part of their stimuli were cognates, and targets were presented in both blocked (L1 or L2 only) and mixed (L1 and L2) conditions. A cognate facilitation effect was found in the L2 blocked condition and in the mixed condition, but not in the L1 condition. Although some studies have reported cognate facilitation also when processing words in L1, the effect seems to be reduced (e.g., Starreveld et al., 2014; van Hell and Dijkstra, 2002).

The cognate facilitation effect is generally attributed to cross-language activation, under the assumption that both orthographic representations of a cognate become activated and send converging activation to a shared semantic representation (e.g., Vanlangendonck et al., 2020). Theoretical accounts of word recognition, such as the Bilingual Interactive Activation Plus (BIA+) model (Dijkstra and van Heuven, 2002), put forth that visual presentation of a word leads to language non-selective co-activation of orthographic neighbors (i.e., word candidates similar to the input in different languages). The orthographic representations will subsequently activate their semantic and phonological representations (e.g., Coltheart et al., 2001; Grainger, 2008). In other words, if orthographic overlap is large across words, as is the case for cognates, the internal representations will be highly activated, which in turn will facilitate lexical decision.

Important to note is that cognates may also serve as inhibitors of lexical access when presented under language constricting circumstances. To illustrate, Dutch-English bilinguals took longer to respond to cognates than to control words when performing a language-specific lexical decision task in English, in which all non-words were replaced by Dutch words requiring a no-response (Vanlangendonck et al., 2020). In contrast, in generalized lexical decision where languages are mixed but words in both languages require a yes-response, cognate facilitation also occurs (Lemhöfer and Dijkstra, 2004). The BIA+ model assumes that in a language-specific lexical decision task, only the stimuli belonging to the target language are bound to a yes-response. This is also the case in a generalized lexical decision task, where items from both target languages are tied to “yes.” However, in a language-specific task containing existing words from the other language, which should be classified as non-words, the cognitive system receives conflicting information when it is confronted with a cognate for which membership to both languages is affirmed. Solving this conflict takes some time, which accounts for the increased response times to cognates under these circumstances.

Although this article has employed the term “cognate” in general up till now, an important distinction should be made between identical and non-identical cognates as these lead to different processing, especially within the context of lexical decision. When Dijkstra et al. (2010) employed cognates for lexical decision with a varying degree of overlap, they found that response times to target words decreased in function of orthographic similarity (see also Lemhöfer and Dijkstra, 2004; Van Assche et al., 2009, 2011). That is, more overlap resulted in lower response latencies. In a second task, the authors asked participants to perform a language decision (i.e., identify whether a word belongs to one language or the other) and reported a reversed pattern of results, upon which they argued that larger orthographic overlap will not only lead to increased co-activation of the non-target orthographic representation, but also of the non-target language node. This will heighten response competition in language decision and cause a cognate inhibition effect. Crucially, the cross-linguistic similarity effects in this study were not just linear (Duyck et al., 2007), but discontinuous in nature, with a steep decrease/increase (depending on the task) in response latencies from near-identical to identical cognates. This may be a consequence of identical cognates sharing both meaning and complete orthography (Dijkstra et al., 1999) leading to faster processing when the task causes no competition, and to a high conflict situation when the task does impose competition.

Finally, a similar priming effect has been found in studies where non-cognate translation equivalents were used as targets. While some discussion continues to exist on the precise effect size and possible moderators of such (masked) translation priming, a 2017 meta-analysis indicated the effect can occur for both L1-L2 and L2-L1 translation, although the effect size appears notably smaller for the latter. In the case of L1-L2 translation, effect size was significantly moderated only by the interval between the offset of the prime and onset of the target, whereas the number of items per cell was the only significant moderator for L2-L1 translation (Wen and van Heuven, 2017). This effect has been shown to be robust both between similar-script and different-script language pairs—e.g., Dutch-English and Korean-English, respectively (Lee et al., 2018)—and in balanced and unbalanced bilinguals, although the asymmetry in effect size between forward (i.e., L1-L2) and backward (L2-L1) priming does diminish and sometimes disappear altogether in more balanced participant pools. These findings indicate that priming effects are not necessarily tied to orthographic similarity between words, but can also stem from semantic similarities, which will of course also play their role when the targets are identical and non-identical cognates.

Current theoretical models and predictions

While a vast body of literature confirms that form overlap yields immediate processing advantages, nearly all prior evidence stems from short-term, single-session masked priming paradigms (see Wen and van Heuven, 2017). Masked translation priming captures transient, automatic semantic or lexical co-activation between a prime and a target within a single experimental frame. Consequently, it remains unclear whether cognate facilitation merely reflects this type of transient, short-lived neighbor activation, or if it involves enduring structural modifications within the bilingual architecture. A multi-day lexical practice paradigm provides a novel testing ground to adjudicate among competing accounts by moving beyond transient priming and observing how massed cross-linguistic experience transfers across days and language switches.

Currently, three distinct theoretical accounts offer competing explanations for the cognate facilitation effect, each assuming a fundamentally different mapping mechanism within the bilingual architecture. The first is a purely semantic account (Francis, 2005), which posits that facilitation is driven entirely by converging semantic activation from separate language-specific lexical entries onto a single, shared conceptual node. Under this account, formal similarity is secondary; any cross-linguistic benefit is a function of semantic overlap. The second is the cross-language lexical activation account, formalized by localist connectionist models like the BIA+ (Dijkstra and van Heuven, 2002). This framework assumes that identical and non-identical cognates possess dual, distinct structural entries in the lexicon (one for each language). When a word is presented, sub-lexical features trigger parallel feed-forward activation across these separate entries via cross-linguistic orthographic and phonological neighbor pathways. The third is the Shared Orthographic Representation Account (Lemhöfer and Dijkstra, 2004), which predicates that identical cognates do not possess dual entries at all. Instead, due to perfect formal identity, they map onto a single, language-independent structural node at the lexical level.

To evaluate these models, the current study set up a multi-day practice paradigm in which Dutch-English subjects exercised lexical decision for control words, non-cognate translation equivalents, non-identical cognates, and identical cognates across consecutive days, executing the task in either their L1 or L2. Crucially, this practice paradigm relied entirely on massed item repetition without explicit feedback, allowing us to isolate autonomous threshold changes within the network from feedback-driven task learning. Baseline scores were collected in the non-practiced language on Day 1, massed repetition practice occurred on Days 2 through 4 in the alternative language, and cross-linguistic transfer was evaluated in the original non-practiced language on Day 5.

The predictions for this multi-day repetition paradigm are explicitly tied to the underlying mechanisms. If the purely semantic account is correct, the long-term transfer of practice benefits across languages should be driven solely by conceptual overlap. Consequently, non-cognate translation equivalents should demonstrate identical practice benefits to both identical and non-identical cognates on Day 5, as all three groups share identical semantic relationships with their cross-linguistic counterparts. If the cross-language activation account (BIA+) is correct, cross-linguistic transfer relies on the gradual resting-level threshold changes of distinct neighbor representations and their mapping pathways. Because these pathways are modulated by formal proximity, savings on Day 5 should scale linearly with structural similarity: identical cognates should show a moderate advantage over non-identical cognates, which in turn should outperform translation equivalents and controls. Finally, if the shared orthographic representation account is correct, practicing an identical cognate in one language directly modifies the resting-level threshold of its single, shared lexical node. This predicts a discontinuous, disproportionately large post-practice processing advantage for identical cognates in the non-practiced language on Day 5, leaving non-identical cognates (which must rely on separate, unpracticed entries) statistically clustered with standard translation equivalents.

Participants

Sixty-nine participants took part in this study for credit or monetary compensation. All reported having Dutch and English as their L1 and L2, respectively. They had received English formal education for at least 5 years and regularly came in contact with English through textbooks and (social) media. All participants completed a language background questionnaire as well as a LexTALE test (Lemhöfer and Broersma, 2011) in both Dutch and English (see Table 1 for a summary of demographic info and LexTALE scores).

Stimulus material

The original target word list consisted of 60 cognates, 60 non-cognate translation equivalents, and 60 control words. Items were identified and selected using cross-linguistic databases and prior bilingual literature. Cognates were classified into two categories based on their formal cross-linguistic properties: 30 were categorized as semantically and orthographically identical (e.g., school), exhibiting perfect character-by-character form identity, and 30 were categorized as non-identical cognates, exhibiting semantic identity alongside high orthographic and phonological similarity (e.g., blood / bloed). Non-cognate translation equivalents (e.g., dog / hond) shared distinct conceptual meaning but possessed zero formal or phonological overlap. All 60 control words were matched to the respective target types on a trial-by-trial basis for bigram frequency, word frequency (derived from SUBTLEX-NL and SUBTLEX-UK; Keuleers et al., 2010a; van Heuven et al., 2014), word length, number of syllables, and orthographic neighborhood size. Using the WordGen stimulus generation programme (Duyck et al., 2004), non-words were generated that respected the orthographic and phonological constraints of the respective target language. See Supplementary Appendix S1 for the full list of target words and their item-level properties (including word frequencies, lengths, and form similarity indexes).

Baseline, practice, and post-practice test

The experiment comprised three distinct experimental sessions: (1) the baseline tast on Day 1 (one lexical decision task), (2) the lexical practice paradigm spanning Days 2 through 4 (consisting of nine lexical decision tasks in total, broken down into three blocks per day), and (3) the post-practice test on Day 5 (one final lexical decision task). To ensure the paradigm measured cumulative lexical exposure and threshold adaptation rather than feedback-driven strategic learning, no feedback regarding speed or accuracy was provided to participants at any point during the practice blocks.

Within each individual block, the presentation of the letter strings was randomized across participants, ensuring that each string was shown only once per block run. On the first day of the experiment, participants filled out an online questionnaire and received a Pavlovia link to the baseline test. This session consisted of 180 target words (60 cognates, 60 non-cognate translation equivalents, 60 controls) either in English or Dutch (language was randomly assigned) and 180 non-words. For the practice phase, participants encountered the exact same 180 target stimuli across all nine subsequent lexical decision tasks. This was done to allow massed practice and drive potential long-term threshold changes for those items. Conversely, non-words were entirely unique and never repeated across any blocks of trials or sessions to strictly avoid the Hebbian learning of non-existing lexical items. The entire experiment was programmed in PsychoPy (Peirce et al., 2022).

The practice phase consisting of nine lexical decision tasks was spread over three days. Each day, participants were granted access to the next session, consisting of three blocks of lexical decision tasks, each of which contained 180 words (60 cognates, 60 non-cognate translation equivalents, 60 controls) and 180 non-words. Languages were switched as compared with the baseline test, i.e., if the baseline test contained L1 Dutch words, practice sessions contained L2 English words, and vice versa. Each trial started with the presentation of a fixation cross for 700 ms, followed by an interval of 300 ms. The letter string was subsequently displayed until a response was provided, for a maximum of 2,500 ms. There was another interval of 800 ms before the next trial started.

The post-practice test entailed completing another lexical decision task of 180 words and 180 non-words, but this time in the same language as the baseline test. Participants were not informed about the changes in language until right before each session, and were therefore not aware of differences between the first and last session on the one hand, and the sessions in between on the other, until after completing them.

Procedure

The experiment was completely carried out online, with links to the appropriate Pavlovia webpage provided to participants through SONA systems and email. Participants were asked to perform the experiment on a desktop or laptop computer in a quiet space where they would not be disturbed. Before signing up, participants also received written instructions to complete their tasks individually, and these were repeated before the start of every session alongside instructions to press the right arrow key if they believed the presented string of letters was an existing word in either Dutch or English (depending on the group and session), and pressing the left arrow key when they believed it was not. The specific language to be considered was mentioned at the start of every session. Emphasis was placed on both accuracy and speed. Presentation of the letter strings was randomized across participants, and each string was shown only once per session. A preparation session of 10 trials preceded the baseline test.

Demographics

In order to see whether groups (Group 1: testing in L1, practice in L2; Group 2: testing in L2, practice in L1) were comparable, t -tests were performed on the variables of age [ t (67) = −0.123, p = 0.903], L2 age of acquisition [ t (67) = 1.590, p = 0.116], LexTALE scores for Dutch [ t (67) = −0.888, p = 0.378], and LexTALE scores for English [ t (67) = −1.708, p = 0.092]. No significant differences were found. Additionally, no differences were detected for gender ratio [χ 2 (1) = 1.477, p = 0.224].

Main analyses

Analyses were performed on response time (RT; in ms) Table 2 and accuracy (ACC; in %) Table 3 data from the Baseline Test on Day 1 and the Post-practice Test on Day 5 on the one hand, and from the first and last practice session on Day 2 and Day 4 on the other. Two participants were excluded from analyses due to extremely low latencies in the final practice session (minimums of 5 and 12 milliseconds, first quantiles at 155 and 263 ms). Because these response latencies cannot physically reflect cognitive lexical processing, it indicates a post-hoc performance failure to comply with the task instructions. Excluding these participants ensures that the final modeling represents valid lexical decision performance, though it slightly restricts the sample to highly compliant participants.

Analysis was carried out in R (R Core Team, 2022) by means of the lme4 -package (Bates et al., 2015), which was used to fit a (Generalized) Linear Mixed-Effects Model for both the reaction times and the accuracy scores of the four sessions mentioned above. The reaction time model used a linear mixed model (LMM) with a default Gaussian distribution, and the accuracy model used a generalized linear mixed model (GLMM) with a binomial distribution and a logit link function. To maximize statistical power and minimize Type I error inflation, a maximal random effects structure was specified including random slopes for participants across sessions. However, this full structure resulted in severe convergence issues due to model overparameterisation. Hence, random slopes were abandoned and the final architectures restricted to random intercepts. While removing random slopes can slightly increase the risk of Type I errors for the within-subject fixed effects, it represents a necessary concession to achieve model identification with complex, item-heavy psycholinguistic designs. Furthermore, p - values and degrees of freedom were estimated using the lmer -test package. Only correct replies were considered in the response time analyses and the data were also trimmed by calculating an individual mean across all trials and excluding any response 2.5 SD removed from the mean. This procedure eliminated 3.0% of the data.

Reaction times

For both the Baseline-Post-practice test comparison and the comparison between the first and final practice session, we started from an intercept-only model with Participant and Item as random effects. Predictor variables were included one by one to determine the main effects, followed by an ANOVA test comparing the newer model with the best previous model. These variables included Word Type (Identical Cognate vs. Non-Identical Cognate vs. Control vs. Translation Equivalent vs. Non-word), Session (Baseline vs. Post-practice and First Practice vs. Last Practice), Language (Practice in L1 vs. Practice in L2 and Testing in L1 vs. Testing in L2), and LexTALE scores. The model was then further refined by including interaction effects in the same stepwise manner.

For the Practice data, this yielded a model with main effects from Session, with words being processed more quickly in the Last Practice [ t (43, 652.915) = −9.008, p < 0.001], and Word Type, with Non-words being processed more slowly overall [ t (900.846) = 19.286, p < 0.001]. Interaction effects were found between Session and Word Type on the one hand, and Word Type and Language on the other. Non-words benefited more from practice than Controls [ t (36, 502.836) = −14.404, p < 0.001], and there was no difference in processing speed between L1 and L2 as was the case for control words and other types [ t (735.552) = −2.274, p < 0.023]. An attempt to include random slopes for each participant over sessions resulted in convergence issues and was therefore abandoned. Likewise, including participants' Group as control variable resulted in a rank-deficient model, and was also renounced, providing us our final model. Variance Inflation Factors, calculated by means of the car package (Fox and Weisberg, 2019), ranged from 1.08 to 2.22, indicating low correlation. When compared with an intercept-only model, the fitted model performed significantly better (p < 0.001), with AICs of 609,904 and 608,125, respectively.

The same procedure was employed for the Test data (Baseline vs. Post-practice). Main effects were found for Session, with faster RTs in Post-practice as opposed to Baseline [ t (43, 145.257) = −11.963, p < 0.001]. Additionally, there were effects of Word Type, with both Non-words being processed more slowly [ t (1, 000.729) = 16.851, p < 0.001] and Identical Cognates more quickly [ t (922.554) = 2.038, p = 0.042]. No main effects were found for any of the other word types. With regard to Language, slower RTs were found for L2 words [ t (111.829) = 2.879, p = 0.005]. Interestingly, there was an interaction between Session and Identical Cognates [ t (43, 156.080) = −5.895, p < 0.001], Non-identical Cognates [ t (43, 164.347) = −3.108, p = 0.002] and Non-words [ t (3, 185.980) = −16.248, p < 0.001], indicating Practice had a stronger speeded processing effect for these words than for Controls. No such effect was found for Translation Equivalents. Crucially, however, the effect was similar for Identical and Non-identical Cognates (p = 0.233). Finally, an interaction effect was observed between Word Type and Language, with Non-words [ t (951.438) = 2.966, p = 0.003] suffering more and Identical Cognates [ t (1, 594.640) = −3.350, p < 0.001] suffering less from the L2 condition's slowing influence. There were no three-way interactions. For this model too, Variance Inflation Factors were calculated by means of the car package (Fox and Weisberg, 2019). These ranged from 1.13 to 1.93, indicating low correlation here as well. When compared with an intercept-only model, the fitted model performed significantly better (p < 0.001), with AICs of 602,513 and 600,709, respectively.

Accuracy

Analyses for accuracy were performed on the same four sessions as for reaction times, and the same two participants were excluded because of exceedingly low scores in the final practice session (50% and 54% accuracy), both as compared with their other scores and to their peers (overall mean = 92%), providing further indication that they did not carry out the experiment as required. The model in this case consisted of a Generalized Linear Mixed Effects Model, constructed in R by means of the lme4 package (Bates et al., 2015), with Score (Correct vs. Incorrect) as binary dependent variable. Predictor variables under consideration included Word Type (Identical Cognate vs. Non-Identical Cognate vs. Control vs. Translation Equivalent vs. Non-word), Session (Baseline vs. Post-practice and First Practice vs. Last Practice), Language (Practice in L1 vs. Practice in L2 and Testing in L1 vs. Testing in L2), and LexTALE scores for both Dutch and English, and just as in the reaction time analysis Participant and Item were included as random effects.

In the analysis of the Practice data, significant main effects were found for Session and Language, showing higher ACC during the last practice (z = −3.223, p = 0.001) and for L1 Dutch words (z = −3.467, p < 0.001). In addition, ACC was significantly lower for Non-words (z = −7.533, p < 0.001) than for Controls. Furthermore, there was an interaction between Word Type and Session, with an increased benefit of practice for Non-words (z = 6.320, p < 0.001). There was also an interaction between WordType and Language, defined as a smaller disadvantage of the L2 condition when the stimuli were either Non-words (z = 2.069, p = 0.039) or Identical Cognates (z = 2.441, p = 0.015). The fitted model significantly outperformed an intercept-only model (p < 0.001), with AICs equal to 24,099 and 24,244, respectively. Variance Inflation Factors ranged from 1.38 to 2.58, indicating low correlation.

The same procedure was followed to analyse the data from the Baseline and Post-practice Test, resulting in the same final model, albeit with slightly different effects. There was a main effect of Session, with higher ACC during Post-practice (z = −5.658, p < 0.001) as well as a main effect of Word Type, with lower ACC for Non-words (z = −6.397, p < 0.001) and higher ACC for Identical Cognate (z = −2.977, p = 0.003) and Non-identical Cognates (z = −2.923, p = 0.003) than for Controls. No such effect was found for Translation Equivalents. There was also an effect of Language, with higher ACC in L1 Dutch (z = −2.188, p = 0.029). Furthermore, the analyses showed an interaction between Language and Word Type, with Non-words (z = −2.695, p = 0.007) being significantly less accurate in L2 English than the other words, and accuracy scores for Identical Cognates (z = 3.209, p = 0.001) being identical for both language conditions. Finally, interaction effects were also found for Session and Word Type, with Non-words (z = 8.524, p < 0.001), Identical Cognates (z = 6.679, p < 0.001), and Non-identical Cognates (z = 5.447, p < 0.001) benefitting from practice, whereas Translation Equivalents did not. Again, the fitted model significantly outperformed an intercept-only model (p < 0.001), with AICs of 22,374 and 22,658, respectively. VIFs ranged from 1.31 to 2.13, indicating low correlation.

Previous research into cognate facilitation has universally established that the effect is driven by language non-selective cross-language activation. Within interactive-activation models such as the BIA+ (Dijkstra and van Heuven, 2002), this cross-linguistic activation can theoretically emerge from two distinct architecture tracks: either through parallel feed-forward activation spreading across distinct but highly similar orthographic neighbor representations, or through the direct retrieval of a single shared structural representation stored across languages (e.g., Dijkstra et al., 1999; Lemhöfer and Dijkstra, 2004). The main objective of the current study was to refine understanding of these cross-language activation mechanics by evaluating how multi-day lexical practice modifies these pathways over time. By subjecting Dutch-English bilinguals to a lexical practice paradigm exclusively in one language and evaluating performance shifts in the other, the study aimed to contrast an activation account based on distinct neighbor co-activation against a structural account based on a shared lexical processing node.

Evaluating the three competing theoretical accounts, the current study yielded three different predictions. Under a purely semantic interpretation, non-cognate translation equivalents should demonstrate identical cross-linguistic practice transfer to both identical and non-identical cognates due to their identical shared conceptual relationships. Under the account of cross-language activation, transfer relies on separate representations activated via formal neighbor pathways and processing advantages should therefore scale linearly with structural proximity, meaning identical cognates will statistically outperform non-identical cognates. Finally, under the shared orthographic representation theory, identical cognates uniquely share a single lexical node, predicting a discontinuous, disproportionately large performance leap in the non-practiced language that leaves non-identical cognates (which possess separate entries) far behind.

Looking at the first vs. the final practice session, it was confirmed that training indeed occurred for all word types, with faster reaction times overall during the final practice session. However, no significant differences emerged between word types, except for non-words, which were invariably processed less swiftly. The uniform acceleration observed during this intervening practice phase likely reflects a combination of general procedural task learning alongside short-term activation tracks, such as episodic or echoic memory traces resulting from the dense, massed repetition of the item list. Because these memory traces operate uniformly across all repeated real-word targets during exposure, they result in general, undifferentiated speed-ups. The non-words ultimately benefited most from practice, as subjects demonstrated reaction times during the last practice session that were more in line with the other word types.

When analysing the test phases before and after practice, non-words were still found to be processed more slowly and identical cognates more quickly. Non-words were also identified more slowly overall in the L2, and significantly more so than other word types, whereas identical cognates were also classified more slowly in L2, although the effect of language was markedly smaller than for the other word types. Additionally, practice in one language appeared to offer a general processing advantage for both types of cognates as well as for non-words, which were all processed significantly faster during the final test. This finding indicates a substantial role for task learning itself; because non-words were never repeated across sessions, the ability to classify strings as existing or non-existing represents a procedural skill that can be acquired and improved with practice (see also Keuleers et al., 2010b). Intriguingly, control words and translation equivalents did not benefit from this general task learning to the same degree as cognates and non-words. With regard to any post-practice transfer effects resulting from completing lexical decisions in the other language, no between-group differences were observed that were not already explained by the different language conditions themselves.

Pertaining accuracy, the results were highly consistent with the latency patterns concerning the effects of session, language, and word type. Scores were higher in the final practice session compared with the first, and in post-practice (test phase) compared with baseline. Overall, subjects performed better in the L1 and made more errors on non-word trials. Concurrently, higher proportions of correct responses were observed for both classes of cognates during the test phase. Mirroring the reaction time data, non-words exhibited a disproportionately larger practice benefit compared with real-word types. Furthermore, identical cognates and non-words appeared less susceptible to the processing costs typically associated with the L2 condition. Specifically, during the testing phase, accuracy rates for identical cognates remained comparable across both L1 and L2, whereas all other word types demonstrated a significant attenuation in performance, with non-words experiencing the most pronounced decrement.

Crucially, these findings indicate that while practice in either language improved lexical decision performance in the other—as evidenced by baseline shifts in control words and substantial gains in non-word processing—the magnitude of this cross-linguistic transfer was more pronounced for words sharing formal overlap. Accuracy scores, which were already near ceiling, remained largely unaffected by practice. Additionally, while all word types demonstrated some level of improvement due to general task practice, the transfer effect across languages was markedly stronger in both identical and non-identical cognates, whereas no such additional benefits were observed for non-cognate translation equivalents. Importantly, while short-term echoic or episodic memory traces explain the symmetrical speed-up during the training days, they cannot account for this long-term asymmetry emerging on Day 5. The absence of a distinct transfer effect for translation equivalents suggests that shared semantic representation in itself may be insufficient to generate long-term cross-linguistic practice benefits of this scale, given that equivalents should theoretically activate the same semantic nodes. Instead, these patterns highlight the critical role of form-based similarity, pointing toward a mechanism governed by either cross-language activation or shared lexical tracks. For non-identical cognates, which by definition do not share identical spelling across languages, the observed facilitation is highly compatible with models of cross-language activation, where sub-lexical orthographic similarity or the concurrent activation of distinct cross-linguistic orthographic neighbors drives the transfer (e.g., Dijkstra and van Heuven, 2002). Conversely, for identical cognates, the substantial cross-linguistic savings could theoretically be accounted for by a shared orthographic representation mechanism, where a single, language-independent lexical node is structurally modified by practice (e.g., Lemhöfer and Dijkstra, 2004).

Rather than providing definitive proof for a structurally shared orthographic representation of identical cognates, however, these findings are best understood as reflecting graded cross-linguistic similarity effects operating alongside general task learning. Since the post-practice contrasts between identical and non-identical cognates were not statistically distinct in the current paradigm, the behavioral data cannot fully adjudicate between a single shared localist representation (e.g., Lemhöfer and Dijkstra, 2004) on the one hand and a highly integrated bilingual lexicon operating via intense cross-linguistic activation of formal neighbors (e.g., Dijkstra and van Heuven, 2002) on the other. Nevertheless, the fact that form-overlapping words (cognates) benefited significantly more than purely semantic equivalents highlights that orthographic overlap is a primary locus for long-term cross-language practice transfer in visual lexical decision. This distinct longer-term pattern, however, seemingly contrasts with the short-lived translation priming effects reported in the meta-analysis by Wen and van Heuven (2017). It suggests that while purely semantic priming effects may be transient, multi-day practice paradigms likely rely more heavily on enduring changes along formal orthographic processing pathways. Ultimately, while the present data do not map the exact structural architecture of the lexicon, the clear divergence between cognates and standard translation equivalents indicates that the degree of orthographic similarity plays a leading role in how cross-language experience modifies lexical processing.

A critical limitation of the current study, however, particularly within the context of the BIA+ framework (Dijkstra and van Heuven, 2002), is its inability to disentangle the relative contributions of orthographic and phonological representations. The BIA+ model posits that visual word recognition triggers rapid parallel activation of both orthographic and phonological codes across both languages. As Dutch and English share a similar script, cognates in this language pair systematically exhibit a high correlation between form-based similarity and sound-based similarity. Consequently, the cross-linguistic transfer effects observed here, which were primarily discussed in terms of orthographic similarity, may equally reflect the role of shared or overlapping phonological representations, or an ongoing interactive feedback loop between the two domains (Dijkstra et al., 1999). The non-negligible impact of phonology in visual lexical decision tasks has been thoroughly documented, demonstrating that phonological overlap independently modulates cognate facilitation (e.g., Frances et al., 2022; Lemhöfer and Dijkstra, 2004). This is underscored by robust evidence from different-script bilinguals who demonstrate substantial cognate facilitation effects despite an absolute absence of orthographic overlap (e.g., Wang and Li, 2026). Because phonological similarity was not explicitly controlled or manipulated in the current stimulus selection, structural conclusions should be moderated. The observed long-term practice benefits possibly may not be attributed solely to the orthographic level; instead, they might reflect a broader form-based transfer mechanism wherein phonological pathways either co-contribute to or reinforce the enduring modifications within the bilingual lexicon. Future research utilizing cross-script paradigms or systematically varying orthographic-phonological consistency during practice is required to isolate these processing tracks.