The Revised Hierarchical Model (RHM) (Kroll and Stewart, 1994), focusing on language production, posits that late bilinguals possess distinct lexical-form representation systems for their first language (L1) and second language (L2). Crucially, it emphasizes differential strengths in the connections between these lexical forms and the conceptual system, with L1-conceptual links being stronger than L2-conceptual links. Furthermore, the model asserts stronger lexical-level (word-form to word-form) connections from L1 to L2 than from L2 to L1. Conversely, the Bilingual Interactive Activation (BIA/BIA+) model (Dijkstra and van Heuven, 2002), derived from L1 lexical acquisition models and centered on visual word recognition, proposes that L1 and L2 share a single integrated mental lexicon, with lexical recognition achieved through language-non-specific competitive activation and inhibition mechanisms (Kroll et al., 2010).

While influential, the RHM’s predictions have not been uniformly supported. The model posits a developmental shift from lexical mediation (L2→L1 translation via L1 lexical links) to conceptual mediation (direct L2-concept access) with increasing L2 proficiency. However, empirical studies report conflicting patterns. Some support the predicted asymmetry in early learners (e.g., faster L2→L1 translation; Kroll and Stewart, 1994; Duyck and de Houwer, 2008; Duyck and Warlop, 2009), while others find faster L1→L2 translation even for novel or low-proficiency words (e.g., Brysbaert et al., 2014; Hu, 2020), or a lack of asymmetry altogether (e.g., Tseng et al., 2018; Mohsen and Almudawis, 2021; Rice and Tokowicz, 2020). A key limitation is that the RHM does not formally account for lexical competition processes during access. More importantly, it treats bilinguals as a relatively homogeneous group, classifying them primarily by global L2 proficiency. This overlooks stable, underlying individual differences in cognitive processing that may systematically influence how learners rely on lexical versus conceptual pathways, potentially explaining these discrepant results. For instance, individuals with a strong verbal representational preference might develop and rely on direct L2-concept verbal associations differently than those with an imagistic preference, even at similar proficiency levels.

The BIA/BIA+ model, while robust in accounting for cross-linguistic competition in recognition, faces its own empirical challenges when applied to the late bilingual population central to this study. The model’s architecture, emphasizing integrated lexicons and language-non-specific access, does not readily account for the asymmetry in bidirectional lexical activation (i.e., L1-to-L2 vs. L2-to-L1 strength) characteristic of late (or non-balanced) bilinguals (Wei, 2013). Furthermore, it does not explicitly address developmental changes in how L2 lexical forms activate conceptual representations (Yang and Wang, 2023), a core concern in L2 learning research. Studies on conceptual processing often yield mixed findings; some show evidence for shared conceptual stores from early stages (e.g., Duyck and de Houwer, 2008), while others suggest qualitative differences in conceptual activation between L1 and L2 (e.g., Tseng et al., 2018). These discrepancies may arise because the model, in its standard form, does not incorporate individual variation in the quality or modality of conceptual representations (e.g., the relative strength of verbal vs. imagistic conceptual codes), which could moderate the efficiency and pathway of conceptual access for L2 words.

The contradictory findings within both theoretical frameworks suggest a critical missing variable: systematic individual differences in cognitive representation. The Dual Coding Theory (DCT) (Paivio, 1980) posits two functionally independent but interconnected systems for information representation: a verbal system and a non-verbal (imagistic) system. Furthermore, Hayes et al. (2020) demonstrated that during the information acquisition phase, regardless of whether words or images of the objects they denote are presented, verbal-preferential individuals consistently represent information verbally, while imagistic-preferential individuals favor imagistic representation. This underscores the stability of representational preference in information processing. Individuals exhibit stable representational preferences, tending to favor one modality over the other during encoding and retrieval (Riding et al., 2011; Sun, 2018; Kang et al., 2013).

Among late bilinguals, the majority utilize both imagistic and verbal modalities for representation, classifying them as dual-preferential users (Yao et al., 2010; Alfred et al., 2020). However, approximately 35% of all individuals prefer a single representational modality, either imagistic or verbal (Sanjanaashree et al., 2014). Research involving native Chinese speakers indicates that an individual’s representational preference influences the effectiveness of their L1 abstract word memorization. Specifically, significant differences in explicit memory performance for L1 abstract words exist between imagistic-preferential and verbal-preferential individuals. This discrepancy arises because imagistic-preferential individuals process abstract words, which rely more heavily on verbal representation, less efficiently than their verbal-preferential counterparts (Niu and Zhao, 2005).

For late bilinguals, whose L2 lexical-conceptual links are initially weak, this preference may fundamentally shape the early mediation pathways. An imagistic-preferential learner, when acquiring a new L2 word, might be compelled to rely on L1 lexical mediation to first access a conceptual image, whereas a verbal-preferential learner could establish a more direct L2 verbal-conceptual link. This aligns with the Encoding-Retrieval Compatibility principle (Eysenck and Keane, 2010; Xiao et al., 2020; Kang et al., 2013), suggesting these differential encoding pathways will recur during retrieval, affecting lexical access efficiency. Therefore, representational preference is not merely a supplementary factor; it is a candidate cognitive construct that can elucidate why the predictions of RHM and BIA show such variability across studies and participants. By explicitly modeling this preference, we can move beyond classifying bilinguals solely by proficiency and begin to explain divergent access patterns within proficiency groups.

Currently, L2 lexical access research has rarely accounted for representational preference distinctions among participants, indicating that existing studies have largely overlooked the influence of individual cognitive traits on L2 lexical access. Due to the absence of direct links between the L2 lexical system and the conceptual system in late bilinguals, coupled with the weaker associative networks within the L2 lexical system itself, imagistic-preferential individuals might be compelled, particularly during initial L2 learning or the acquisition of novel L2 words, to initially rely on mediation through the L1 lexical system to activate the imagistic system for conceptual access subsequently. In contrast, verbal-preferential individuals can complete this process without activating the imagistic system. The Revised Hierarchical Model (RHM) proposes that the connection between L2 lexical form and concept develops from a “lexical mediation” stage via L1 to a “conceptual mediation” stage as L2 proficiency increases. Consequently, the lexical access and retrieval differences between late bilinguals with different representational preferences should be more pronounced for L2 than L1 vocabulary.

For imagistic-preferential individuals, the mediation pathway via the L1 lexical system conflicts with their representational preference. Therefore, the semantic access pathway for L2 words may undergo a developmental shift: from reliance on the L1 lexical system pathway when processing unfamiliar words, toward an imagistic system pathway analogous to L1 lexical access when processing familiar words. However, unbalanced verbal-preferential bilinguals, being adept at L1 lexical representation, do not necessitate this “pathway shift.” For dual-preferential users, overcoming competition between the two representational modalities is consistent throughout the representation of unfamiliar and familiar L1 words.

Representational preference constitutes a significant individual cognitive factor influencing the process of L2 lexical access, as evidenced in related research. To examine the hypotheses above and integrate the factor of individual representational preference into existing lexical access theories—thereby advancing the understanding of the cognitive mechanisms underlying lexical learning—this study conducts an experiment that employs statistical methods with a lexical translation judgment task, aiming to investigate whether classifying participants based on representational preference constitutes a more effective cognitive classification variable for understanding L2 semantic access, compared to the conventionally used variables, such as cognitive style and L2 proficiency, etc. in prior research (Hu, 2020). Then it investigates potential differences in task performance across L2reference constitutes a more effective cognitive classification variable for understanding L2 semantic access, istic, dual), examining whether the three participant groups exhibit fundamentally distinct lexical access patterns.

A priori power analysis was conducted using G*Power 3.1 (Faul et al., 2009) for a between-subjects ANOVA (three groups) with an alpha of.05, power of.80, and a medium effect size (f = 0.25) (Cohen, 1992). The analysis indicated a required total sample size of 78 participants (i.e., 26 participants per group). To ensure robust group comparisons and account for potential data loss, we recruited a larger sample of 160 participants. The 160 participants were undergraduates (59 males), aged 19–25 years (M = 21.13, SD = 2.46), with Mandarin as their first language and approximately 10 years of English learning experience (M = 8.71, SD = 1.60). All participants volunteered for the experiment, completed questionnaires, and received monetary compensation upon completion. Following the assessment of representational preference (see section 2.3), 35 participants were randomly selected from each of the three preference groups (verbal, imagistic, dual) for the subsequent translation judgment task, meeting and exceeding the power requirement.

Cognitive style: Cognitive style was assessed using the Santa Barbara Learning Style Questionnaire (SBCSQ) (Mayer and Massa, 2003). This self-report instrument consists of six scenario-based questions where participants indicate their preference for verbal or visual information processing on a 7-point Likert scale.

Verbal working memory capacity: Verbal Working Memory Capacity was measured using the Digit Span subtest from the Wechsler Adult Intelligence Scale-Revised in China (WAIS-RC). Both the Forward and Backward recall tasks were administered, and the score was used as the index of working memory capacity.

Vocabulary size: An online assessment tool¹ was used to estimate participants’ receptive English vocabulary size (Fokin et al., 2025; Keuleers et al., 2015; Stoeckel et al., 2019).

Household income level: A 7-point scale (lowest-income households, low-income households, lower-middle-income households, middle-income households, upper-middle-income households, high-income households, highest-income households) based on the household income classification system used in the China Family Panel Studies (i.e., CFPS)² was administered to measure participants’ family economic status.

Subjective socioeconomic status (SES): The Subjective SES Scale (Adler et al., 2000) was used to measure participants’ perceived family economic standing.

L2 experience and aptitude: The Language History Questionnaire 2.0 Simplified Chinese version (LHQ 2.0; Li et al., 2014)³ was administered to assess L2 experience and self-reported learning aptitude.

Objective L2 proficiency: Participants reported their most recent College English Test Band 4 (CET-4) score as an objective measure of L2 proficiency.

Subjective L2 proficiency: Participants rated their overall L2 proficiency on a scale from 1 (not proficient) to 10 (highly proficient).

Complete description and coding scheme of Control Variables are listed in Table 1.

Stimuli: Adopting the Verbal Attentional Task (VAT) paradigm from Alfred et al. (2020), word-picture paired stimuli were created. The pictorial stimuli consisted of spade, club, and diamond symbols. Three congruent word-picture pairs (e.g., “黑桃” [spade] + ♠◻) and six incongruent pairs (e.g., “梅花” [club] + ♦◻) were designed. Vertical positions (word above/below image) were counterbalanced, resulting in six congruent and twelve incongruent stimulus pairs (see Table 2).

Procedure: Stimuli were presented via E-Prime 2.0. Each trial began with a central fixation cross (“+”) for 2000 ms, followed by an 80-ms presentation of the stimulus pair. A blank screen then appeared for 2000 ms, during which participants pressed “J” for spade, “K” for club, and “L” for diamond on the keyboard. Trials advanced automatically after a keypress or a timeout (2000 ms).

The formal experiment included 144 congruent trials (75%) and 48 incongruent trials (25%), totaling 192 trials (Alfred et al., 2020; Hayes et al., 2020). Practice trials (congruent-only; 20 trials) required > 80% accuracy to proceed to the formal experiment. Participants were not informed about incongruent trials.

Scoring: The VAT score was calculated based on the 48 incongruent trials only, as these trials force a choice between verbal and imagistic information. For each incongruent trial, the response could be congruent with the word or with the picture. The score was calculated using the formula: (Number of word-congruent responses - Number of image-congruent responses)/Total valid incongruent trials. This calculation yields a score ranging from −1 to +1. Scores approaching +1 indicate a strong preference for verbal representation, while scores approaching −1 indicate a strong preference for imagistic representation, and scores near 0 indicate no strong preference (dual-preference).

Measurement outcomes of representational preference are presented in Figure 1. Participants were then grouped according to established VAT score thresholds, i.e., Verbal-preferential group (VAT scores ≥ +0.75; n = 35); Imagistic-preferential group (VAT scores ≤ −0.75; n = 39); Dual-preferential group (n = 86). For subsequent experimental procedures, the Dual-preferential group was further refined to include only those with VAT scores between -0.4 and +0.4 (as per Alfred et al., 2020; Hayes et al., 2020).

Material: High- and low-frequency L2 lexical stimuli were selected from the Snodgrass and Vanderwart standardized picture set (Snodgrass and Yuditsky, 1996). Before the main experiment, 20 university L2 English learners (who did not participate in the main experiment) rated the familiarity and complexity of the pictures and their corresponding English words on 7-point scales (1 = completely unfamiliar, 7 = highly familiar; 1 = extremely low complexity, 7 = extremely high complexity). Based on these ratings, 10 English words with familiarity ratings ≥ 6.5 were selected as Familiar Words; 10 English words with familiarity ratings between 2.3 and 4.5 were selected as Novel Words (Li et al., 2011). The complete stimulus set was created with corresponding pictures (complete materials in Supplementary Appendix).

Procedure: Stimuli were presented using E-Prime 2.0. The experimental process for novel words involved: (1) A learning phase (2 repetitions) with 1,000 ms fixation followed by 6,000 ms simultaneous presentation of English words, Chinese translations, and pictures (Bartolotti and Marian, 2014; Bartolotti and Marian, 2017; Bai et al., 2011); (2) A 5-min serial subtraction by three interference task (Kole and Healy, 2013; Dikmans et al., 2020); (3) A test phase where 1,000 ms fixation preceded L2 English words, followed by 200 ms blanks and translation prompts requiring speeded correct/incorrect judgments (disappearing upon response or after 2,000 ms; see Figure 2). Familiar words tasks appeared only in the test phase without prior notification of familiarity status. Each word generated correct/incorrect trials (incorrect trials used random lexical distractors from the material words), presented 3 times in randomized order (60 trials per direction; 120 trials in total), preceded by 20 practice trials (see Figure 3).

Data was processed and analyzed using SPSS 25.0. Data processing entailed the exclusion of 1 verbal-preferential participant whose accuracy fell below 75% across familiar word sequences, alongside 23 participants (8 dual-preferential, 8 imagistic-preferential, 7 verbal-preferential) failing to meet the 75% accuracy threshold in novel word sequences. Reaction time (RT) data from correct trials were subsequently analyzed, with outliers exceeding ± 2 standard deviations removed; these extreme values constituted less than 5% of the dataset.

Standardized residual histograms and normal P-P plots were generated for RT data across four task conditions: novel words (L2→L1; L1→L2) and familiar words (L2→L1; L1→L2). Each to verify normality of error distributions. Residual independence was assessed via Durbin-Watson tests, yielding values of 1.319 (novel L2→L1), 1.091 (novel L1→L2), 2.053 (familiar L2→L1), and 1.887 (familiar L1→L2). Multiple linear regression analyses were then conducted for each task condition, with representational preference (dual/imagistic/verbal), L2 proficiency, and cognitive style (visual/verbal) as predictors. The significance of regression coefficients (evaluated via t-tests) determined predictive differences among these independent variables on task-specific RTs (Yuan, 2019). Control variables included age, gender, language aptitude, L2 learning experience, objective English proficiency, usage frequency, subjective socioeconomic status, household income, working memory, and vocabulary size. Following multicollinearity diagnostics using Variance Inflation Factors (VIF) and Tolerance indices, two variables exhibiting high collinearity were removed: L2 learning experience (VIF = 8.016, Tolerance = 0.125) and subjective socioeconomic status (VIF = 8.483, Tolerance = 0.118). All retained variables demonstrated VIF < 2 and Tolerance > 0.5, confirming acceptable multicollinearity levels (Conversano et al., 2020).

To test our main hypotheses, we conducted: (1) Multiple linear regressions for each translation direction (L1→L2, L2→L1) and word type (familiar, novel) with representational preference, L2 proficiency, and cognitive style as predictors; (2) Separate mixed-design ANOVAs on RTs for familiar and novel words, with Representational Preference (three groups) as a between-subjects factor and Translation Direction (L1→L2, L2→L1) as a within-subjects factor.