In the field of psycholinguistics and cognitive science, the term representation refers to “an encoding of some information, which an individual can construct, retain in memory, access, and use in various ways” (Smith, 1998, p. 391). Within the scope of bilingualism, bilingual representation in the mental lexicon is often divided into two interconnected levels: lexical representation and conceptual representation (Kroll & Groot, 2005; Kroll & Stewart, 1994). Bilingual lexical representation pertains to the specific forms of words in each language, including their orthographic and phonological properties. In contrast, bilingual conceptual representation focuses on the semantic information of the translation pairs (or interlingual semantically related word pairs) encoded in the bilingual mental lexicon.

Early bilingual representation models, such as the Word Association Model, the Concept Mediation Model (Potter et al., 1984), and the Revised Hierarchical Model (RHM; Kroll & Stewart, 1994), propose that the lexical representations of a bilingual’s two languages are language-specific, while the conceptual representation is shared. The Word Association Model posits that L2 words are directly linked to their L1 translations at the lexical level, bypassing the bilingual conceptual representation. In contrast, the Concept Mediation Model states that translations are connected via the shared conceptual representation. The RHM integrates these two views, proposing that both lexical and conceptual links exist between languages, with their strength varying as a function of translation direction and L2 proficiency. Based on these ideas, the model makes three core assumptions regarding translation performance among bilinguals (Jiang, 2023, Kroll et al., 2010, Kroll and Stewart, 1994): First, bilinguals tend to rely more on direct and strong L2 to L1 lexical connections during backward (L2 to L1) translation. In contrast, when undergoing forward translation (L1 to L2), they are more likely to use the conceptual route (L1 to concept to L2). Consequently, backward translation should be completed more rapidly than forward translation. Second, forward translation, which typically follows the semantic route, is more affected by semantic manipulation than backward translation. Third, higher-proficiency bilinguals should be influenced more by semantic manipulation than lower-proficiency ones because the link between L2 lexical representation and conceptual representation is strengthened as L2 proficiency increases. It is thus reasonable to claim that the relative strength of conceptual mediation and lexical association would differ as a function of translation direction and L2 proficiency. The third assumption, which is also called the developmental hypothesis, can also be accommodated by the theoretical framework of the BIA+ (Bilingual Interactive Activation Plus; Dijkstra and Van Heuven, 2002, van Heuven and Dijkstra, 2010). This model relies on the notion of words’ resting level of activation (RLA) for bilingual lexico-semantic processing. When a bilingual’s language proficiency is lower in L2 than L1, the RLA of L2 lexical representations is lower than that of L1, and the RLA of L2 lexical representations for bilinguals with lower-proficiency is lower than those with higher-proficiency.

Empirical evidence supporting and challenging the assumptions of the RHM has been mixed, and the validity of such assumptions remains a subject of ongoing debate. The first two assumptions, which can be framed as translation asymmetry or asymmetric links for the two translation directions, have been supported across various experimental paradigms such as translation production, translation recognition, and interlingual priming paradigms. In the translation production task, translation latencies were faster in backward than forward translation, and forward translation was more affected by semantic manipulation than backward translation (Cheung and Chen, 1998, De Groot et al., 1994, Kroll et al., 2002, Kroll and Stewart, 1994, Sholl et al., 1995). Similarly, support also came from the traditional translation recognition task, where translation recognition was faster for backward than forward translation at the behavioral level, and the N400 translation priming effect (i.e., the difference in waveforms between non-translations and translation equivalents) in this task was larger for backward than forward translation recognition at the electrophysiological level (Chen et al., 2015, Chen et al., 2020, Palmer et al., 2010, van Hell and Kroll, 2013, Wilck et al., 2019). Additionally, semantic and translation priming paradigms have documented a stronger forward priming effect compared to backward priming, which was attributed to greater conceptual mediation in forward translation (Duyck, 2005, Gollan et al., 1997, Jiang, 1999, Keatley et al., 1994, Smith et al., 2019, Witzel and Forster, 2012, Zhao et al., 2024). However, some studies failed to support these two assumptions of the RHM. For example, Heij et al. (1996) investigated the effects of semantic manipulation on translation among Dutch-English bilinguals. In their experiments, semantic manipulation was achieved by presenting either a semantically related (congruent) or unrelated (incongruent) picture alongside the target word. The results revealed that semantic manipulation had a greater (experiment 2) or equal (experiments 3 and 4) impact on backward translation compared to forward translation.

There is also a lively debate about the developmental hypothesis, namely the third assumption of the RHM. For instance, Talamas et al. (1999) used a variant of the translation recognition task in the backward direction to investigate this topic and supported the hypothesis. In this task, the critical trials involved non-translations that were related in form or meaning to the correct translations. They found that L2 proficiency modulated the sensitivity to lexical or semantic relatedness of translations. Less proficient bilinguals were more likely to be interfered by lexical distractors whereas the reverse was true for more proficient ones, supporting bilinguals’ developmental shift during translation recognition. Findings from brain potentials further provided convincing neurophysiological evidence that cannot be detected by traditional behavioral measures to support the developmental hypothesis (e.g., Kotz and Elston-Güttler, 2004, Liang and Chen, 2020). In a recent event-related potentials (ERPs) study by Liang and Chen (2020), participants with low- and high-proficiency performed a L2 picture-word matching task, where they had to judge whether the word presented was semantically consistent or inconsistent with the picture. The findings revealed a reliable N400 effect, indicating semantic incongruity, for both proficiency groups in the L2. Importantly, bilinguals with higher-proficiency exhibited a shorter N400 peak latency and larger mean amplitude for the difference waves, suggesting that as L2 proficiency increases, the link between the L2 word forms and concepts is strengthened as posited by RHM’s third assumption. Besides, Kotz and Elston-Güttler (2004) investigated the influence of L2 proficiency on lexical access in German-English bilinguals by examining their responses in a lexical decision task employing both associative and categorical priming paradigms. Associative priming assesses the strength of connections between related lexical items, while categorical priming evaluates the strength of links between lexical and conceptual representations. The N400 categorical priming effect reflecting conceptual mediation was observed exclusively in the highly proficient group. This suggests that while less proficient bilinguals can establish associative connections between lexical representations, only those with higher-proficiency form stronger links between L2 word forms and their conceptual representations, thereby corroborating the developmental hypothesis. However, counter evidence was reported by Sunderman and Kroll (2006) that conceptual access appeared available for less proficient bilinguals as well because they were also affected by semantic distractors of translation equivalents. Nevertheless, Kroll et al. (2010) claimed that the results of this study did not refute the developmental hypothesis in that the more critical finding from it was only lower-proficiency bilinguals showed evidence for lexical association in the translation recognition task.

As stated above, although the RHM was initially formulated to explain translation production studies, its reach has since extended into various translation paradigms (Jiang, 2023, Kroll and Tokowicz, 2001, Lijewska et al., 2018, Palmer et al., 2010). As van Hell and Kroll (2013) clarify, the model’s assumptions have been rigorously tested not only through production tasks but also via translation recognition (both the traditional one and the variant of it) and translation priming experiments, together forming a triad of critical evidence for the RHM framework. Empirical surveys further highlight this broader applicability: Brysbaert et al. (2010) reported that, between 1994 and 2009, over half of the published studies citing the RHM adopted translation-perception tasks, demonstrating the model’s influential role in research on both translation production and comprehension.

In the aforementioned models of bilingual representation and processing, representing the bilingual conceptual system as only one shared, holistic node between translation-equivalent pairs is an oversimplification. Indeed, assuming that translation equivalents exhibit full semantic equivalence is premature (Duyck & Brysbaert, 2004), especially for culturally distant language pairs like English and Chinese. De Groot (2011) emphasized that “it is a well-known fact that complete meaning equivalence of the two terms in a translation pair is a rare phenomenon” (p. 132). Liu et al. (2025) further pointed out that translation equivalents often lack full semantic equivalence, due to factors such as the abstract nature of language, variations in lexical senses, differences in connotative meanings, and mismatched referential category boundaries.

Examples of translation equivalents across different language pairs that are not identical in meaning are widespread and common. Consider the English word “colony” and its Chinese counterpart “殖民地 /zhí mín dì/.” While “colony” emphasizes settlement in its meaning, “殖民地” carries a connotation of exploitation, illustrating how the shades of meanings of translation equivalents can diverge (Dong et al., 2005). Even among cognates—loanwords that share similar pronunciations and meanings across languages—semantic discrepancies can arise. For instance, the Chinese term “沙发 /shā fā/,” borrowed from the English “sofa,” originally described a seat for two or more people. However, in Chinese, “沙发” can also refer to a seat for just one person, reflecting a loss of the original connotation (Jiang, 2023). Such discrepancies are not confined to noun translation pairs. English speakers conceptualize the adjective “narrow” as limited width, whereas Japanese speakers interpret its equivalent “狭い /semai/” as both narrow and small in overall area (Wolter et al., 2020). In contrast, the Chinese equivalent adjective “窄 /zhǎi/” aligns more closely with “narrow” in meaning, as both exclusively denote limited width; on the other hand, the Japanese term “狭い” conveys an additional sense of compactness or reduced overall area (Wolter et al., 2022). Moreover, the English adjective “impressed” typically carries a positive connotation (Bradley & Lang, 1999), but its Italian counterpart “impressionato” often has a more negative connotation (Fairfield et al., 2017). In some cases, a single source word may have multiple corresponding translations in the target language to fully convey its meaning. For example, the English word “blue” requires Russian equivalents of both “cиний” (dark blue) and “гoлyбoй” (light blue) (Berlin & Kay, 1991). In addition, the homonymous word like “bank” corresponds to Chinese translations “银行 /yín háng/” for its financial institution meaning, and “河岸 /hé àn/”, “河畔 /hé pàn/”, “岸边 /àn biān/”, or “岸 /àn/” for the riverside meaning. As a result, subsequent theoretical models have considered the degree of semantic overlap of translation pairs in the bilingual mental lexicon.

Because assuming a shared or language-independent semantic system as posited by early models is premature, during the 1990s and beyond, subsequent models of bilingual representation and processing entertained the notion that concepts in both languages do not fully overlap. The Distributed Conceptual Feature Model (DCFM; De Groot, 1992a, De Groot, 1992b, Van Hell and De Groot, 1998) was the first bilingual model that has taken into account the overlapping nature of bilingual conceptual representation by adopting a feature-based conceptual knowledge account from monolingual studies (Collins & Quillian, 1969). In this theoretical model, the source word and its translation in the target language are not linked to a single, localist semantic unit; rather, translation pairs differ in the extent to which they activate the same semantic features within the bilingual semantic memory.2 It posits that the degree of semantic overlap between words and their translation equivalents in the bilingual mind depends on the number of shared semantic features. According to the DCFM, concrete translation pairs share more semantic features than abstract ones, leading to a greater degree of semantic overlap in the bilingual mental lexicon. Thus, during lexical access, the increased semantic overlap between concrete pairs leads to greater spreading activation and facilitates translation. Empirical evidence from behavioral and electrophysiological experiments (Chaouch-Orozco et al., 2024, Chen et al., 2022, De Groot, 1992b, Ferré et al., 2017, Jin, 1990, Schoonbaert et al., 2009, Van Hell and De Groot, 1998) supports the existence of a concreteness effect or a larger translation priming effect for concrete pairs, thereby indicating the distributed, overlapping nature of the bilingual conceptual representation.

Subsequently, the Shared Distributed Asymmetrical Model (SDAM; Dong et al., 2005) also highlights the varying degree of semantic overlap in bilingual conceptual representation. It posits a shared conceptual store that contains features common to both languages, alongside two language-specific conceptual stores, each encoding culture- and language-specific conceptual information. This model aligns with the DCFM by emphasizing shared and distinct semantic features between the two languages. Similarly, the Sense Model (Finkbeiner et al. 2004) predicts that bilingual conceptual representation consists of distinct “bundles” of features or semantic senses. Shared senses between translation equivalents correspond to the overlap seen in the DCFM, while language-specific senses reflect unique semantic features inherent to each language. For example, the sense of “black” in English (color) may overlap with the Japanese “kuroi,” but other senses (e.g., “black humor” in English vs. “darkness” in Japanese) are language-specific (Finkbeiner et al. 2004). Lastly, the Modified Hierarchical Model (Pavlenko, 2009) similarly suggests that the bilingual conceptual store is not unified. It maintains the view that the bilingual conceptual representation would be fully shared, partially overlapping or entirely language-specific, based on research into investigating the subtle differences in cross-linguistic category boundaries for translation equivalents.

These theoretical, verbal models have acknowledged that the extent of semantic overlap in the bilingual conceptual representation is not the same. However, it is important to note that such hand-coded models concerning bilingual conceptual representation fall short of capturing the nuanced, continuous nature of semantic overlap for various translation pairs in the bilingual lexicon. That is, these models lack the granularity needed to measure subtle differences in semantic equivalence of translation pairs. For instance, models such as the DCFM simply makes distinctions in semantic overlap based on dichotomous word types such that concrete translation pairs exhibit greater semantic overlap than abstract ones, and that cognate pairs have a higher degree of semantic overlap compared to non-cognate pairs. However, it has been shown that each pair of translation equivalent has its own unique, gradient level of semantic alignment (Liu et al., 2025, Thompson et al., 2020), and this alignment exists along a continuous spectrum, rather than only fitting into a rigid dichotomy based on word types that the model proposes. Besides, feature-based models like the DCFM and SDAM face several other significant challenges in capturing the nuances and complexities of bilingual conceptual representation. One major issue is determining what information is encoded by the features themselves. For example, when distinguishing between objects like different dog breeds, it becomes unclear how specific features like “four legs” or “short legs” should be defined and applied into the conceptual representation system, particularly when such features vary in degrees (e.g., short legs in a Corgi vs. long legs in a Great Dane) (Jared et al., 2013), and the determination of semantic features becomes even more elusive when comparing translation equivalents between languages. Furthermore, it is difficult to define how detailed these features should be, and manually coding the infinite number of potential features for both the source word and its target translation is both labor-intensive and impractical. For these reasons, there is only one published study that delved into the bilingual conceptual representation using a feature-based approach (Matsuki et al., 2021).

Fig. 1 illustrates the traditional, feature-based hypothesis that indicates different degree of semantic overlap for translation pairs in the bilingual conceptual representation (De Groot, 1992a, De Groot, 1992b, Dong et al., 2005, Van Hell and De Groot, 1998), and our context-based, bilingual distributional hypothesis that cross-linguistic context consistency drives the variability of semantic overlap between translation equivalents in the bilingual mental lexicon. The core idea behind feature-based models is that translation pairs are represented in bilingual semantic memory as a collection of conceptual/semantic binary features (e.g., birds have wings, birds can fly, whereas cars do not and cannot), and the correlation or overlap of these numbers of common features determines the extent of semantic overlap for translation pairs in the bilingual conceptual representation.

To our knowledge, all current computational models in bilingualism research take a simple, localist view of the bilingual conceptual representation (e.g., Dijkstra et al., 2019, Dijkstra and Van Heuven, 2002, Zhao and Li, 2013). Models like BIA+, Multilink (Dijkstra et al., 2019), and SOM-based (Self Organizing Map) bilingual model (Zhao & Li, 2013) all adopted one shared, holistic node to represent the bilingual conceptual system. Consequently, these models are unable to capture and estimate the degree of semantic relatedness between words within a single language or the degree of semantic equivalence between translation equivalents across languages. In fact, Dijkstra et al. (2019) have acknowledged that their localist approach oversimplifies the complexity of bilingual lexical semantics and advocate researchers for incorporating distributed representations at the conceptual level in the future.

Another limitation of such computational bilingualism model is their reliance on an algorithmic implementation of a series of pre-defined “rules.” While many regard the practice of predefining assumptions based on prior theoretical frameworks or experimental conclusions and adjusting parameters to simulate behavioral effects observed in experiments as a merit, it can also be seen as a potential limitation. This approach makes the models highly dependent on the choices made by human programmers, rendering them subject to the biases and decisions of those designing them. For example, BIA + and Multilink incorporate several adjustable parameters that influence the model’s performance. One key parameter is the RLA assigned to each word form, which is correlated with its frequency in the language. Dijkstra et al. (2019) suggest that differences in processing between early and late bilinguals can be modeled in Multilink by varying the RLA, assuming that late bilinguals encounter L2 words less frequently, resulting in lower resting activation for L2 words. Similarly, the SOM-based bilingual model also depends on the modeler’s decisions. The model (Zhao & Li, 2013) introduces new lateral connections within each layer to the original DevLex-II model (e.g., Zhao & Li, 2010) to support the theoretical framework of non-selective bilingual lexical access and to model interactions between two lexicons during L2 acquisition that are presumed to occur in its study (Zhao & Li, 2013). These lateral connections are added using the Hebbian learning rule (“neurons that fire together, wire together”) to model the dynamic interactions and strengthening of cross-linguistic links during the progress of L2 acquisition. Hence, the reliance on a priori assumptions and the manipulation of parameters in these models represents one of the most influential “wildcards” available to the modeler (Hummel and Holyoak, 2003, Jones et al., 2006), yet it speaks least to the ecological validity of cognition and human behavior (Schwieter & Festman, 2023).

Cognitively-oriented computational models based on distributional semantics can provide valuable insights into language processes, as their architectures and underlying theoretical underpinning offer cognitively plausible and psychologically valid accounts of meaning representation in psycholinguistics and cognitive science (e.g., Latent Semantic Analysis: LSA; Landauer & Dumais, 1997). In monolingual context, distributional semantic models (DSMs) have been demonstrated to be instrumental in shedding light on cognitive processes, particularly in understanding how meaning is learned and represented in the human mind. DSMs capitalize on the distributional hypothesis (Firth, 1957, Harris, 1954), which posits that words with similar meanings tend to occur in similar contexts. Grounded in this hypothesis, these models align with associative learning mechanisms affecting natural language acquisition, where repeated exposure to patterns of linguistic co-occurrence helps language learners infer semantic relationships between words. These models also offer a framework for understanding the gradual acquisition of meaning, where language learners incrementally build semantic representations through exposure to diverse linguistic contexts, with models simulating this process by progressively refining word vectors as more co-occurrence data is processed, reflecting the gradual development of semantic knowledge in the human mind (Günther et al., 2019). Therefore, DSMs have been demonstrated to align well with human performance in psycholinguistic and cognitive tasks such as semantic similarity judgments, lexical decision, and word association (Mandera et al., 2017), and these models can currently be considered state-of-the-art models of semantic representation (Kumar, 2021).

While DSMs build on cognitively-plausible mechanisms for how concepts may be learned and represented from natural language experience, a major limitation lies in their application to (predominantly) monolingual environment. Recently, inspired by the notion of the distributional hypothesis, Liu et al. (2025) developed a Semantic Alignment Model (SAM) that can serve as a potential candidate for bilingual conceptual representation. The architecture of SAM is illustrated in Fig. 2. By measuring the extent to which the cross-linguistic contextual usage patterns for a particular translation-equivalent pair (e.g., “dragon | 龙 /lóng/”) are consistent, Liu et al. (2025) successfully quantified, in a large-scale, data-driven manner, the computed semantic alignment Rc values of translation-equivalent pairs. In other words, by comparing the context of use between English and Chinese languages, the similarities and differences in the semantic profiles of a given pair of translation equivalent can be derived. Liu et al. (2025) validated this metric induced from SAM by showing that semantic alignment derived across different sets of DSMs is consistent. Besides, semantic alignment corresponds to human-based explicit assessment of translation pairs, such that it not only reflects human intuitions about semantic similarity of translation pairs but also captures bilinguals’ usage frequency of translations. Furthermore, the study shows that the relationships between semantic alignment and various psycholinguistic factors (i.e., translation ambiguity, word sense, concreteness, imageability) mirror those between semantic similarity ratings and these measures. However, while human-based assessment of words are considered the traditional benchmark for evaluating the reliability of distributional semantic measures (Marelli & Baroni, 2015), the extent to which this metric can be indicative of human behavior at the implicit level is still unknown, and remains an open, empirical question.

This paper presents SAM, a computational model that quantifies the consistency of cross-linguistic usage patterns in translation pairs to induce the semantic relevance of them. SAM achieves this by applying a simple linear association to measure the semantic distance between related contextual words and the translation pair within their respective English and Chinese semantic spaces in a bottom-up fashion, thereby capturing the computed semantic alignment of this translation pair. The model builds on previous research line by analyzing the statistical distribution of words within their textual context to infer semantic relationships between them. SAM adheres to this tradition of semantic representation that aims to model word meanings without relying on predefined conceptual or perceptual primitives or human knowledge, aside from the associative learning mechanism itself (Griffiths et al., 2007, Landauer and Dumais, 1997, Lund and Burgess, 1996). Our study contributes to distributional semantic studies by being the first to extend this lineage to bilingualism by demonstrating that the pairwise cross-linguistic semantic associations—comparing the semantic relationships between English contextual words with a source word and the corresponding Chinese contextual words with its translation equivalent—underlies the latent semantic overlap stored in bilingual semantic memory, as illustrated in Fig. 1B.

Liu et al. (2025) have shown that semantic alignment correlated well with human-rated similarity judgment as a traditional benchmark and aligned with explicit assessment of translation pairs; however, this does not necessarily guarantee that this distributional measure would be an effective predictor of implicit language processing tasks (Baayen, 2013, Marelli and Baroni, 2015). Indeed, Gatti et al. (2024) stated that explicit tasks by human participants focus on conscious, deliberate judgment, potentially overlooking the subtle, automatic associations that influence decision-making at a chronometric level. Therefore, this study aims to test our novel hypothesis using SAM through three complementary analytical lenses:•

First, we evaluate whether semantic alignment can replicate the behavioral effects associated with bilingual conceptual representation, as observed in the implicit task. Previous literature (e.g., Chaouch-Orozco et al., 2024) has explained the asymmetric effect of concreteness on translation priming across opposing translation directions through Paivio, 1991, Paivio, 2010 Bilingual Dual Coding Theory (BDCT). According to BDCT, sensorimotor information linked to concrete L1 words activates their corresponding L2 counterparts more strongly than the reverse direction. Conversely, abstract L1 and L2 words activate their counterparts in similar extents, as both rely primarily on linguistic rather than sensorimotor information to establish meaning. As a result, an asymmetric advantage of concreteness emerges during translation priming. Our study is the first to test whether this phenomenon, namely the asymmetric concreteness advantage on translation priming, can be captured solely by language itself—without relying on any grounding information—through a computational simulation.

•

Second, we assess semantic alignment’s predictive power for both implicit translation recognition process and explicit semantic similarity judgment. We expect that higher semantic alignment should predict greater implicit processing efficiency and higher explicit similarity ratings of translation equivalents. We then compare its predictability with that of concreteness, namely our baseline and the traditional proxy of semantic overlap.

•

Third, we examine the consistency of semantic alignment predictions with the three assumptions of the RHM, and compare these prediction patterns against the baseline concreteness effect. Specifically, we hypothesize that forward translation recognition, where conceptual mediation tends to occur, is more likely to be influenced by the facilitatory effect of semantic alignment (RHM’s first assumption). On the other hand, backward translation recognition, where the lexical link is strong, should be unaffected or affected to a lesser extent by semantic alignment (RHM’s second assumption). We also hypothesize that more fluent bilinguals should be more sensitive to the meaning alignment of translation equivalents because rising L2 proficiency strengthens the conceptual mediation (RHM’s third assumption).

This research is structured as follows: The first experiment uses the standard translation recognition paradigm to investigate the non-holistic nature of bilingual conceptual representation, as indicated by concreteness (e.g., Chaouch-Orozco et al., 2024, Ferré et al., 2017, Van Hell and De Groot, 1998). In addition, this experiment revisits RHM’s three key theoretical hypotheses. Specifically, it examines: 1) whether backward translation is associated with stronger lexical connections; 2) whether forward translation is more sensitive to semantic manipulation than backward translation, given that forward translation is primarily conceptually mediated; and 3) whether bilinguals with higher-proficiency are more affected by semantic manipulation, due to stronger conceptual activation in L2 for them (Jiang, 2023). It is worth noting that the first experiment does not include computational estimates. In the subsequent simulation experiment, we validate the computational system against the effects observed in the first behavioral experiment, specifically focusing on effects related to bilingual conceptual representation. The rationale for this simulation is that a credible cognitive computational model should be able to mirror the patterns of effects found in behavioral studies of language processing (Caliskan et al., 2017, Marelli et al., 2017). Concreteness is dichotomized into concrete versus abstract categories in the first two experiment as it enables us to use computed semantic alignment to model response times for concrete and abstract translation and non-translation pairs, that is, the very conditions that we used in the translation recognition experiment, and test whether the performance of SAM mirrors bilinguals’ response times for these word types at the group level (following Marelli et al., 2017; and Marelli & Baroni, 2015 in the Priming Effects at Different SOAs section). Finally, in the third experiment, we conduct two ad-hoc analyses to assess how well the output generated from SAM predicts bilinguals’ implicit processing performance and their explicit rating intuition. Specifically, we evaluate whether computed semantic alignment outweighs human-rated concreteness, the traditional proxy of semantic overlap, in predicting bilinguals’ chronometric performance on the implicit processing task and their semantic similarity performance on the explicit judgment task. Additionally, we compare the consistency of our direct metric of semantic overlap (semantic alignment) and the indirect proxy of it (concreteness) with RHM’s three assumptions. Because the third experiment aims to evaluate and compare the predictive power of semantic alignment and concreteness as measures for semantic overlap in bilingual conceptual representation, it is suitable to treat them into continuous variables and see how they influence the implicit and explicit performance of bilinguals.

It is necessary to explicitly state the connection and underlying logic of the three experiments. As stated above, the objectives of the first experiment are twofold: 1) testing the non-holistic nature of the bilingual conceptual representation, laying the testing ground for simulating relevant effects for it using SAM in experiment 2; 2) examining RHM’s three assumptions that remain intensely debated, setting the foundation for testing whether predictions from SAM-derived metric are consistent with these assumptions in experiment 3. Given the expectation that bilingual conceptual representation is not entirely shared, the first aim of the behavioral study lays the groundwork for applying a cognitive computational model, seeking to induce the varying and continuous estimates of semantic overlap between translation equivalents through language statistics. Accordingly, since the computational system is specifically designed as a distributional model of bilingual conceptual representation (Liu et al., 2025), the second experiment focuses solely on mirroring behavioral effects pertinent to bilingual conceptual representation in its simulation. Evaluating the model’s cognitive plausibility then enables us to utilize its output—semantic alignment—as a direct metric for semantic overlap to examine bilingual lexico-semantic organization, mirroring how concreteness has been employed as an indirect proxy of it for this research line (De Groot, 1992a, De Groot, 1992b, De Groot and Hoeks, 1995, Ferré et al., 2006). The second aim of the first experiment sets the stage for employing the semantic alignment metric and concreteness proxy for semantic overlap as predictors of chronometric performance in translation recognition where conceptual mediation is likely involved, such as the forward translation direction and the task performed by more proficient bilinguals. The third experiment hence utilizes the alignment index and concreteness proxy for semantic overlap to assess their prediction consistency with the three assumptions of the RHM and compare their predictive power of bilingual language processing efficiency recorded in the first experiment. It is hypothesized that conceptual-mediation structure will be influenced (more) by semantic manipulation, whereas lexical-association connection will remain unaffected (or affected to a lesser extent). As a final step, we examine and compare their predictability of bilinguals’ explicit judgment of semantic similarity for translation equivalents. From a computational psycholinguistic standpoint, the second experiment demonstrates SAM’s ability to account for existing behavioral phenomena, while the third experiment contributes to the predictive power from its output.