Explanations of asymmetrical cross–language semantic activation in multilingual individuals build on mechanisms originally developed in bilingual research. Two foundational accounts in this literature are the Revised Hierarchical Model (Kroll and Stewart, 1994) and the BIA + framework (Dijkstra and Van Heuven, 2002), both of which offer insights into how asymmetric patterns may arise during lexical access.

The Revised Hierarchical Model proposes that forward translation (L1 → L2) relies primarily on concept mediation, whereas backward translation (L2 → L1) can be accomplished through stronger lexical links. This organization predicts a commonly observed pattern in which L2 → L1 translation is faster and more automatic than L1 → L2, especially among unbalanced bilinguals (Heredia, 1997; Linck et al., 2009; Clenton, 2015; Sperl et al., 2024). The model further suggests that lexical-semantic connection strengths vary as a function of L2 proficiency and L1 dominance, thereby shaping direction–specific processing asymmetries.

Complementing the RHM, the BIA + model assumes nonselective access: words from both languages become co–activated during recognition, with subsequent top–down control mechanisms differentiating between target and non-target languages (Green, 1998; Dijkstra and Van Heuven, 2013; Schwartz and Kroll, 2006). While BIA + successfully explains cross-language co–activation and competition, it remains less explicit about the precise conditions under which directional asymmetries emerge.

Although these bilingual models provide a useful foundation, they do not fully capture the representational and control dynamics inherent in trilingual systems. Trilinguals must coordinate three lexical–semantic networks, navigate more complex competition structures, and manage activation pathways that may involve indirect or cascaded influences across languages. Crucially, one language–often the L2–may possess a special functional status due to differences in learning history, representational format, or instructional role. To address these gaps, researchers have proposed accounts such as the L2 Status Factor Hypothesis (Bardel and Falk, 2007) and the Typological Proximity Model (Rothman, 2011, 2015), both of which aim to characterize how prior linguistic systems differentially influence L3 acquisition and processing.

Bilingual models such as the RHM and BIA + offer valuable insights into cross–language activation but do not fully specify how activation flows across three languages or why particular non–native languages may hold privileged roles in trilingual processing. In L3 acquisition research, two influential frameworks—the L2 Status Factor Hypothesis and the Typological Proximity Model —provide complementary mechanisms for explaining how previously acquired languages shape activation patterns in an emerging L3. Together, these accounts generate testable predictions for semantic activation in typologically diverse systems such as Portuguese–English–Chinese.

Masked translation priming is a widely used paradigm for investigating unconscious cross–linguistic semantic activation while minimizing strategic processing effects. The technique involves presenting prime words for 50–60 ms, followed by pattern masks and target words, enabling researchers to capture automatic activation patterns that vary systematically across language directions (Forster et al., 2003). From a cognitive mechanism perspective, masked priming effects are typically interpreted as indicating that early, unconscious, automatic processing relies primarily on semantic memory systems rather than episodic memory systems, reflecting lexical-level mechanisms rather than low-level sublexical processing driven by visual similarity. This “economical” processing incorporates prime activation into the target’s processing stream upon target presentation, creating advantages for semantically related prime–target pairs (Qiao and Zhang, 2017).

Early masked priming studies used a four-level sequence–a forward mask, a briefly presented prime, a backward mask, and the target (Evett and Humphreys, 1981). Because the prime and the backward mask followed each other within a very short interval, the two sometimes fused perceptually, reducing the effectiveness of the mask. To prevent this fusion effect, later work adopted either a three-level structure (mask → prime → target) or a modified four-level structure that inserts a short blank interval between the prime and the backward mask (Forster and Davis, 1984). The present study uses this modified four-level procedure, which avoids fusion while keeping the prime below conscious awareness, allowing the paradigm to capture early, automatic cross-linguistic activation (Forster et al., 2003; Jiang, 1999; Wen and van Heuven, 2017).

Recent trilingual masked–priming investigations further demonstrate the paradigm’s effectiveness in revealing complex asymmetrical activation patterns. Aparicio and Lavaur (2016) found translation–priming effects only when L1 French served as the prime for L2 English and L3 Spanish targets, with no cross–activation between L2 and L3, suggesting a hierarchical organization privileging the dominant language. A similar dominance–driven pattern was observed by Li et al. (2024), who tested Chinese–English–Japanese trilinguals using triple different-script cognates. Their results showed that priming emerged only when the prime was from the participants’ strongest language (L1 Chinese)—that is, in the L1 → L2 and L1 → L3 directions—while primes from weaker languages (L2 English, L3 Japanese) produced no reliable facilitation. Importantly, Li et al.’s findings indicate that even with phonologically related cognates, cross-language activation requires sufficiently robust prime processing. Together, these studies highlight masked priming’s sensitivity to dominance-based asymmetries that are characteristic of trilingual lexical organization.

Applying masked translation priming to trilingual contexts requires careful attention to methodological factors that distinguish trilingual from bilingual research. First, the six possible language combinations in trilingual systems demand systematic investigation of all translation directions to comprehensively characterize asymmetrical patterns. Second, trilingual proficiency profiles are inherently more complex, necessitating careful control of proficiency differences across the three languages to isolate cross-linguistic activation effects from proficiency-driven differences. Third, potential cascaded activation—for example, Portuguese → English priming secondarily activating Chinese through English’s role as instructional medium—necessitates designs that can separate direct cross-linguistic effects from indirect L2-mediated pathways. Fourth, multilingual contexts may introduce additional sources of interference that warrant consideration; for instance, Toassia and Motab (2018) demonstrated L2 German interference in L3 English production among Portuguese natives, indicating that non-target languages can influence trilingual processing even when not directly manipulated.

The precise nature of these asymmetries varies substantially across studies, resulting in a complex empirical landscape that must be carefully examined to address our research questions on Portuguese-English-Chinese trilingual processing.

Evidence from hierarchical organization studies demonstrates clear L1-privileged patterns. Aparicio and Lavaur (2016) examined French-English-Spanish trilinguals using masked translation priming with 67 ms prime duration, reporting significant priming effects only when L1 French served as the prime for both L2 English and L3 Spanish targets, with no L2–L3 interaction. This suggests that L1 retains privileged access to both non-native languages, while L2 and L3 remain functionally isolated—directly relevant to RQ1 concerning asymmetrical activation patterns. Abunuwara’s (1992) Stroop experiments with Arabic-Hebrew-English trilinguals corroborated this hierarchical view, showing interference effects between L1 and both L2 and L3, but no L2–L3 interference. More recently, Ashraf (2025) confirmed this pattern in Punjabi-Urdu-English trilinguals across picture-naming, Stroop, and translation tasks, demonstrating stronger L1 activation effects compared to L2 in multiple experimental contexts.

In contrast, research supporting partial semantic integration reveals more complex activation patterns that challenge simple hierarchical models. Wang et al. (2023) conducted a systematic investigation of German-English-French trilinguals, providing comprehensive evidence for directional asymmetries directly relevant to our research questions. Using masked translation priming with a 50 ms prime duration across 8 participants, they found significant priming effects in the L2–L1, L1–L3, and L2–L3 directions, but no effects for L1–L2, L3–L1, or L3–L2. Crucially, their modified Sense Model framework showed that language dominance, rather than acquisition order, determined priming strength, with evidence of dynamic dominance shifts between L1 and L2 during lexical development.

Cross-linguistic influence research also offers essential insights into the mechanisms underlying trilingual vocabulary processing. Mulík et al. (2019) studied Spanish-English bilinguals learning L3 Slovak using paired-associate tasks and found comparable facilitation effects for both L3–L1 and L3–L2 phonological similarity, even when L2 was not explicitly present in the task. However, participants with higher L2 proficiency showed greater benefits from L2 similarity, indicating that L2 activation strength correlates with proficiency. Extending this work, Mulík and Carrasco-Ortiz (2023) used ERP methods and revealed divergent neurocognitive mechanisms for L3 learning via L1 versus L2, showing opposite N400 effects for L1 Spanish and L2 English interlingual homophones after only 3 days of training. This points to distinct pathways of cross-linguistic influence. Otwinowska (2024) synthesized mounting evidence for cumulative cross-linguistic influence in L3 vocabulary acquisition across multiple paradigms. Analyzing L3 word processing experiments, cognate guessing tasks, and vocabulary learning studies, she demonstrated that triple cognates (shared across L1, L2, and L3) consistently provide processing advantages over double cognates, provided that L2 and L3 proficiency reaches threshold levels.

Taken together, these empirical findings establish three key patterns relevant to the present study. First, directional asymmetries are a consistent feature of trilingual processing, though their manifestations vary systematically across language combinations and proficiency profiles. Second, L2 demonstrates unique activation characteristics that distinguish it from both L1 and L3, supporting theoretical accounts of L2’s mediating role in trilingual systems. Third, typological distance significantly shapes cross-linguistic activation strength, with typologically closer pairs showing different facilitation patterns than more distant pairs.

Nevertheless, current research suffers from a critical limitation: the overwhelming focus on Indo-European language combinations restricts our understanding of how typological diversity modulates trilingual processing mechanisms. The present study addresses this gap by examining Portuguese-English-Chinese trilinguals, thereby enabling a systematic investigation of how typological distance influences asymmetrical activation patterns across all translation directions while testing theoretical predictions concerning L2-mediated cross-linguistic influence.

Fifty-nine Portuguese–English–Chinese trilinguals (36 female, 23 male; M = 30.17 years, SD = 6.95, range = 20–45 years) participated in the study. Inclusion criteria required native Portuguese speakers who had received formal instruction in both English and Chinese as foreign languages. All were right-handed, had normal or corrected-to-normal vision, and none reported any history of neurological disorders.

Language acquisition patterns showed that English (L2) was acquired significantly earlier (M = 11.31 years, SD = 2.67) than Chinese (L3; M = 19.95 years, SD = 6.05), t(58) = −10.55, p < 0.001, d = 1.37. Both foreign languages were learned primarily through classroom instruction rather than naturalistic immersion. Self-rated proficiency, measured on a 7-point Likert scale (Li et al., 2020), indicated significantly higher competence in English (M = 5.71, SD = 0.96) than in Chinese (M = 4.18, SD = 1.27) across all language modalities, t(58) = 7.85, p < 0.001, d = 1.38.

Table 1 presents detailed proficiency ratings across the four skill domains. Effect size analyses showed large differences between English and Chinese in all skills (Cohen’s d > 1.40), with English consistently rated higher. Portuguese remained the dominant language for daily communication and cognitive processing, English was primarily used in professional and academic contexts, and Chinese was mainly restricted to classroom learning and limited social interactions with Chinese speakers.

Seventy two Portuguese–English–Chinese translation triplets were selected to ensure cross-linguistic comparability. All items were concrete nouns that met three criteria: (1) high familiarity ratings across all three languages (≥ 6.0 on a 7-point scale), (2) absence of cognate relationships between Portuguese and English to avoid orthographic confounds, and (3) availability of clear translation equivalents in Chinese. Half of the stimuli denoted animate entities (animals, professions), and half represented inanimate objects (tools, furniture), conforming to binary animacy judgment requirements.

Portuguese and English words were matched for length, with Portuguese items averaging 5.97 letters (SD = 1.84) and English items averaging 5.74 letters (SD = 1.92), t(143) = 0.84, p = 0.401. Chinese stimuli were two-character compounds presented in simplified script to ensure semantic transparency and avoid the processing complexity associated with single-character presentations.

Prior to the experiment, two Portuguese-English-Chinese trilinguals who did not participate in the main study independently rated the familiarity of candidate words on a 7-point scale (1 = very unfamiliar, 7 = very familiar). Words were included in the final stimulus set only if both raters gave a score of 6 or above. The mean familiarity ratings of the final word set were 6.45 (SD = 0.38) for Portuguese, 6.52 (SD = 0.35) for English, and 6.48 (SD = 0.41) for Chinese, ensuring high familiarity and translation accuracy across all three languages.

A within-subjects design was used to examine masked translation priming across six translation directions: Portuguese↔English (PT → EN, EN → PT), Portuguese↔Chinese (PT → ZH, ZH → PT), and English↔Chinese (EN → ZH, ZH → EN). This comprehensive design allowed systematic examination of bidirectional priming across all language pairs, essential for detecting asymmetrical cross-linguistic semantic activation.

Testing took place individually in acoustically isolated rooms, with E-Prime 3.0 used for precise timing control. Each trial followed a standardized masked priming sequence: forward mask (######, 500 ms) → prime (100 ms) → blank interval (50 ms) → backward mask ($$$$$$, 50 ms) → target (displayed until response or maximum 2,500 ms). The 200 ms stimulus onset asynchrony (SOA) was chosen based on prior findings that this interval captures automatic semantic processing while minimizing conscious translation strategies (Forster and Davis, 1984).

Stimuli appeared in uppercase against a white background. Portuguese and English words were displayed in 24-pt Arial, and Chinese characters in 28-pt SimHei, to ensure comparable visual salience across scripts. Participants performed animacy judgments using designated response keys (“A” = animate, “L” = inanimate), with key-mapping counterbalanced across participants to control for motor biases. The animacy judgment task was selected to ensure semantic processing of targets without explicitly encouraging translation strategies, providing an appropriate measure of unconscious cross-linguistic activation. Reaction times were recorded from the onset of the target display until the participant’s keypress response (Figure 1).

After six practice trials with feedback, participants completed 144 randomized experimental trials presented without feedback to avoid learning effects. No breaks were given to maintain consistent testing conditions. At the end of the session, participants completed a comprehensive language history questionnaire (Li et al., 2020). Each session lasted approximately 30 min. All participants gave written informed consent.

To justify the adequacy of the experimental design, we followed established recommendations in masked priming research indicating that statistical power in within-subject RT paradigms is primarily determined by the number of trials rather than the number of participants (Brysbaert and Stevens, 2018). Each participant completed 144 critical trials (24 per translation direction), yielding a dense trial structure that increases sensitivity to small priming effects in generalized linear mixed-effects models.

According to best-practice guidelines for reaction-time experiments (Brysbaert and Stevens, 2018), designs with ≥ 1,600 observations per fixed-effect comparison typically achieve ≥ 80% power to detect small effects (d ≈ 0.15–0.20). In the present study, each direction-specific GLMM included between 1,416 and 1,472 usable RT observations after trimming, closely approaching the recommended threshold. Together, the within-subject manipulation, the high number of item-level observations, and the use of mixed-effects modeling provide adequate statistical power to detect masked priming effects of the magnitude commonly reported in bilingual and trilingual studies.

Reaction times were analyzed using generalized linear mixed-effects models (GLMMs) implemented in the lme4 package in R (Bates et al., 2015). Because RT distributions are positively skewed, models were fitted with a Gamma distribution and a log link, which provides a better distributional fit than Gaussian models in masked priming paradigms. For each language pair, the following model was specified: RT ∼ PrimingType∗Direction+(1∣Participant) + (1∣Item).

PrimingType (related vs. unrelated) was contrast-coded (−0.5 = related, +0.5 = unrelated).

We first fitted a unified mixed-effects model to reaction times, including Language Pair, Direction, and Relatedness as fixed effects, as well as all corresponding interactions. Random intercepts were specified for participants and items, with by-participant random slopes for Relatedness. The model was fitted using a Gamma distribution with a log link function and converged successfully without singular fits (N = 610).

The model revealed significant baseline differences in reaction times across language pairs, as well as a significant overall asymmetry between forward and backward translation directions. The main effect of Relatedness did not reach significance, and the LanguagePair × Relatedness interaction was also not significant. Fixed effects from the unified model are summarized in Table 2.

Given the study’s theoretical focus on directional asymmetries in trilingual lexical access, we next conducted planned follow-up analyses examining simple effects of Relatedness within each translation direction. Fixed effects from the unified model are summarized in Table 2.

Data were preprocessed following standard procedures for masked priming studies. The goal of this section is to provide an overview of the descriptive patterns that motivate the subsequent mixed-effects analyses. We report mean reaction times (RTs) and priming effects for each translation direction to illustrate the overall asymmetry pattern before turning to the inferential statistics. These descriptive results allow readers to understand the direction and magnitude of the effects observed in the GLMMs.

No participants were excluded based on overall accuracy, as all participants responded correctly on more than 80% of trials. Trials with incorrect animacy judgments were removed (6.8% of trials). Reaction times below 200 ms or exceeding±2.5 standard deviations from each participant’s mean were further excluded (2.1% of remaining trials). In total, 91.1% of the original dataset was retained for analysis. Table 2 presents mean RTs and priming effects (unrelated RT–related RT, where positive values indicate facilitation) for each translation direction. Overall latencies varied systematically across target languages, with Chinese as target producing longer RTs (Portuguese → Chinese: 811 ms; English → Chinese: 870 ms) compared to Portuguese or English targets (694–758 ms). Priming effects revealed marked directional asymmetries, ranging from slight inhibition in English → Portuguese (−4.62 ms) to substantial facilitation in English → Chinese (+117.72 ms).

Data were analyzed using generalized linear mixed-effects models (GLMMs) with a Gamma distribution and log link function to accommodate positively skewed RT distributions. Models were implemented with the lme4 and lmerTest packages in R (Bates et al., 2015; Kuznetsova et al., 2017). Following conventions in trilingual priming research, separate models were fit for each language pair to facilitate clearer interpretation of directional effects. This approach avoids conflating cross-pair variability arising from differences in script, baseline RT, or distributional properties across languages. It also allows each pair to be modeled with an appropriate variance structure, improving interpretability while maintaining comparability across models.

Each model included fixed effects of Priming Type (related vs. unrelated, contrast-coded), Direction (forward vs. backward, contrast-coded), and their interaction. Random intercepts were specified for participants and items. Model parameters were estimated using maximum likelihood, and significance was assessed with likelihood ratio tests (Table 3).

We first examined the Portuguese–English pair to establish whether masked semantic priming occurs between the participants’ dominant language (L1) and their highly proficient L2. Results showed a significant main effect of Priming Type, χ²(1) = 4.14, p = 0.042, indicating that targets preceded by translation-related primes were judged faster than those preceded by unrelated primes, consistent with automatic semantic facilitation in masked priming. Neither Direction, χ²(1) = 0.22, p = 0.638, nor the interaction, χ²(1) = 2.50, p = 0.114, reached significance. Simple effects analysis indicated significant facilitation for Portuguese→English (+13.85 ms), z = −2.04, p = 0.042, Cohen’s d = 0.06, although the effect size was negligible. By contrast, English → Portuguese showed no reliable effect (−4.62 ms), z = −0.01, p = 0.991, d = −0.02.

We next turned to the Portuguese–Chinese pair to assess whether priming emerges between the typologically most distant languages in the trilingual system. The analysis revealed a significant main effect of Direction, χ²(1) = 6.12, p = 0.013, reflecting overall RT differences between directions. Neither Priming Type, χ²(1) = 0.25, p = 0.617, nor the interaction, χ²(1) = 0.15, p = 0.696, reached significance. Simple effects showed no reliable priming in either direction: Portuguese → Chinese (+10.11 ms), p = 0.617, d = 0.03, or Chinese → Portuguese (+32.44 ms), p = 0.292, d = 0.16. Although the numerical difference was slightly larger in the Chinese → Portuguese direction, the effect did not reach significance.

Finally, we analyzed the English–Chinese pair to determine whether L2 facilitates access to L3, as predicted by accounts proposing a mediating role of the non-native language. This language pair produced the strongest semantic facilitation effects, with markedly faster responses for targets preceded by translation-related primes. There was a robust main effect of Priming Type, χ²(1) = 11.86, p < 0.001, and a marginal Priming Type × Direction interaction, χ²(1) = 3.77, p = 0.052. Simple effects revealed robust facilitation in English → Chinese (+117.72 ms), z = −3.44, p < 0.001, Cohen’s d = 0.30, representing a medium effect size. By contrast, Chinese → English showed no reliable priming (+17.19 ms), z = −0.70, p = 0.484, d = 0.09 (Table 4).

This section synthesizes the results across all six translation directions and highlights the overarching pattern of cross-linguistic semantic priming observed in the dataset. Across all six translation directions, only two showed significant priming: English → Chinese (p < 0.001) and Portuguese → English (p = 0.042). All other directions failed to reach significance. Here, “priming” refers specifically to faster animacy judgments for targets following translation-related primes, reflecting early automatic semantic activation rather than explicit translation or language-level activation.

As shown in Figure 2, the heatmap of priming effects (ms) illustrates the asymmetrical pattern, with robust facilitation in the L2 → L3 (English → Chinese) direction and weaker or null effects elsewhere. Figure 3 presents a forest plot of effect sizes (Cohen’s d with 95% confidence intervals). Both figures converge in demonstrating reliable priming for English → Chinese and Portuguese → English, but not in other directions.

The present study revealed a clear asymmetrical pattern of cross-linguistic semantic activation among trilingual speakers, with significant priming effects observed exclusively in the English → Chinese and Portuguese → English directions, while other language combinations failed to produce reliable effects. This directional asymmetry directly addresses RQ1 and provides important insights into the organizational principles of trilingual semantic representation.

The Portuguese → English effect aligns with the predictions of the Revised Hierarchical Model (RHM), which posits that L1 lexical items maintain direct conceptual connections, whereas L2 access is mediated through L1 (Kroll and Stewart, 1994). This explains why Portuguese primes facilitated English targets, while the reverse direction did not. Extensive bilingual research corroborates this asymmetry, consistently demonstrating robust L1 → L2 priming but weak or absent L2 → L1 effects across both similar and distinct writing systems (De Groot and Nas, 1991; Voga and Grainger, 2007; Tomaz et al., 2024).

The Bilingual Interactive Activation Plus (BIA+) model (Dijkstra and Van Heuven, 2002) provides a complementary account by emphasizing differential resting activation levels. Portuguese as L1 maintains higher baseline activation, enabling rapid conceptual activation and spreading to L2. By contrast, English as L2 operates at a lower resting activation level, limiting its ability to facilitate L1 targets under masked conditions.

The significant English → Chinese effect is particularly noteworthy, as it challenges traditional assumptions of L1 dominance. It aligns with evidence showing that L2 can function as a gateway to L3 when proficiency is high or when L2 assumes functional dominance (Chen, 2020). Pathak et al. (2025) similarly reported that dominant L2s facilitated faster L3 processing in multilinguals. These observations are consistent with the Shared Distributed Asymmetry Model, which proposes that L2 exerts stronger influence on L3 than vice versa, reflecting acquisition sequence rather than strict L1 dominance.

The present findings demonstrate both convergence and divergence with prior trilingual research. The Portuguese → English asymmetry shows strong continuity with bilingual traditions, replicating forward priming effects documented across multiple language pairs (Bailey et al., 2024; Dimitropoulou et al., 2011). However, notable differences emerge in trilingual contexts.

A closely related line of evidence comes from Li et al. (2024), who examined masked translation priming in Chinese–English–Japanese trilinguals using triple different-script cognates. Their results revealed a strong dominance-driven asymmetry: significant priming emerged only when primes originated from the participants’ strongest language (L1 Chinese), whereas L2 and L3 primes produced no reliable effects. Importantly, even with phonologically related cognates—which typically yield larger priming effects—the weaker languages (English and Japanese) failed to prime the stronger one, underscoring the necessity of robust prime activation for cross-language facilitation. This pattern contrasts with the present study in two respects. First, our use of non-cognate translation pairs isolates semantic activation from phonological similarity, allowing us to show that the same dominance-driven asymmetry arises even without cognate facilitation. Second, unlike Li et al.’s demonstration of strong L1 → L2/L3 effects, our Portuguese → Chinese and Chinese → Portuguese pairs did not produce priming, highlighting the additional constraining role of typological distance rather than dominance alone.

Aparicio and Lavaur (2016) examined unbalanced French (L1)–English (L2)–Spanish (L3) trilinguals and found that significant priming occurred only when primes originated from the dominant language (L1). Specifically, L1 primes significantly accelerated recognition of both L2 and L3 targets, suggesting that multilingual mental lexicons may be organized around L1, with limited direct interaction between non-dominant languages (L2 and L3). By contrast, the absence of Portuguese → Chinese and Chinese → Portuguese priming in the present study diverges from reports of observable L1–L3 activation in some trilingual populations.

Wang et al. (2023) identified significant L1–L3 priming in German–English–French trilinguals, whereas Tytus (2017), using semantic categorization tasks, reported bidirectional L1–L2 priming but no L1–L3 effects. Such cross-study patterns suggest that L3-to-L1 priming is generally weak, particularly when proficiency is low or typological distance is large.

Nevertheless, some studies have reported different representational patterns. Chen and Zhang (2012), for example, found that Cantonese–Mandarin–English trilinguals shared a common semantic representation across all three languages, with strong conceptual connections between each pair. This stands in sharp contrast to the directional asymmetries observed here. However, several features of their sample warrant caution: Cantonese is a linguistic variant of Mandarin with high similarity; participants were raised in Mandarin-speaking environments with formal education; and all had passed the College English Test Band 6. These factors render them closer to highly proficient bilinguals than prototypical trilinguals, which may explain their symmetric representation pattern.

Similarly, Dilraba (2012) investigated Uyghur trilinguals using cross-linguistic long-term repetition priming in category judgment tasks and found that when L3 proficiency was high, L3 could effectively prime L2. By contrast, in the present study, L3 (Chinese) was the last-acquired, least-proficient, and least-used language, which likely contributed to the absence of significant L3 → L2 priming effects. Task paradigms may also account for discrepancies: Wang (2016) showed that semantic categorization tasks are more likely to elicit cross-linguistic priming, whereas lexical decision tasks often produce weaker or null L3 → L2 effects.

Taken together, these comparative findings highlight that typological distance, proficiency, and task paradigm jointly shape cross-linguistic activation in trilinguals. They also suggest that Portuguese–English–Chinese trilinguals represent a unique configuration that both aligns with and challenges models developed primarily from Indo-European language combinations.

Typological distance and language status are widely recognized as key determinants of cross-linguistic influence in third language acquisition, with effects documented across lexical, morphosyntactic, semantic, and phonological domains (Hammarberg, 2001; Chiswick and Miller, 2005). Typological distance encompasses both objective structural differences and psychotypology, learners’ subjective perceptions of similarity (Jaekel et al., 2024). Research shows that learners often rely primarily on lexical similarity when judging relatedness, rather than phonological or syntactic overlap (Rothman, 2015). When subjective perceptions align with objective typology, cross-linguistic transfer and integration are more likely to occur (Ringbom, 2007; Singleton and Ó Laoire, 2006).

Portuguese and Chinese, as cross-family languages, are typologically distant. Portuguese, a Romance language within the Indo-European family, is characterized by rich inflectional morphology, whereas Chinese, a Sino-Tibetan language, relies on word order and particles rather than inflectional morphology. Their orthographic systems further accentuate this divide: Portuguese employs an alphabetic script, whereas Chinese uses morphosyllabic characters. Jiang’s (2000) psycholinguistic model suggests that typological distance constrains lexical integration. When languages are typologically close, L2 vocabulary is more likely to achieve full semantic integration; when distant, lexical representations often remain partially integrated, maintaining relative separation. This framework accounts for the absence of Portuguese ↔ Chinese priming in the present study. From a psychotypological perspective, Portuguese speakers may perceive Chinese as highly incompatible with Romance languages, thereby raising activation thresholds and reducing the likelihood of cross-language priming under masked conditions.

By contrast, Portuguese and English, as Indo-European languages, share abundant cognates and structural similarities. These features facilitate cross-linguistic activation and account for the reliable priming effects observed in the Portuguese–English direction. The findings thus underscore typological distance as a decisive factor modulating semantic representation in trilinguals.

Beyond distance, the mediating role of L2 is equally critical in trilingual processing. Research has long demonstrated that L3 output is not only influenced by L1 but often more strongly shaped by L2 (De Angelis et al., 2009). Green’s (1998) inhibitory control model predicts that because L1 has the highest baseline activation, it requires stronger suppression to minimize interference with L2 and L3. When L3 activation is relatively weak, L2 emerges as the more salient competitor. Williams and Hammarberg (1998) termed this the “foreign language effect,” arguing that L2, as a non-native language, is cognitively closer to L3 and thus more likely to intrude during L3 production. Bardel and Falk’s (2007) “L2 status factor” hypothesis further refined this account, proposing that L2 and L3 share learning strategies and knowledge structures, granting L2 a privileged role in trilingual acquisition.

A growing body of evidence substantiates this framework. Bardel and Lindqvist (2007) showed that in Swedish learners of Italian (L3), L2 influence persisted consistently, whereas L1 played only a minor role. Bono and Melo-Pfeifer (2011) reported that over 60% of unintentional lexical intrusions in French speakers of Spanish (L3) originated from English L2. More recently, Stoehr et al. (2024), using a rapid trilingual switching paradigm with Spanish–Basque–English participants, found that L3 interfered more strongly with L2 than with L1, highlighting the robustness of L2–L3 connectivity.

The present findings provide direct evidence for this L2 mediation effect. The robust English → Chinese priming indicates that English (L2) functioned as a mediating pathway for activating Chinese (L3), while the absence of Portuguese → Chinese priming reflects the combined constraints of typological distance and acquisition sequence on L1–L3 interactions. In other words, although Portuguese is the native language, its substantial typological distance from Chinese limited direct semantic activation. By contrast, English, with higher proficiency, its role as the instructional medium for learning Chinese, and greater similarity to L3 in representational and learning strategies, conforms to the predictions of the L2 status factor and exerts a stronger mediating role in the trilingual semantic system.

These findings both support and challenge the Typological Proximity Model. On one hand, the absence of Portuguese ↔ Chinese priming validates TPM’s prediction that greater typological distance reduces cross-linguistic facilitation. On the other hand, the strong English → Chinese effect highlights an underexplored dimension: the privileged mediating role of L2 in linking typologically distant languages. Future models of trilingual processing must therefore integrate typological distance, psychotypology, and L2 status as dynamic factors shaping multilingual semantic representation.

In sum, typological distance and L2 mediation jointly shape the architecture of trilingual lexical representation. While distance constrains the probability and strength of cross-linguistic activation, L2 assumes a distinctive role in bridging distant systems. These findings deepen our understanding of trilingual lexical processing mechanisms.