1. Introduction
The intricate process of orally translating a word presented in one language into another language is of great interest to research. However, this question has seldom come to the limelight in the field of bilingualism because word translation is merely one of the tasks practically used to address the question in a much broader scope: the organization and processing of the two languages in the bilingual mind. To illustrate, one way to understand how the bilingual mental lexicon is organized and processed is to use translation comprehension (i.e., priming or recognition) tasks (e.g., Altarriba, Reference Altarriba1992; De Groot & Nas, Reference De Groot and Nas1991; Keatley et al., Reference Keatley, Spinks and De Gelder1994; Mott et al., Reference Mott, Midgley, Holcomb and Emmorey2020; Smith et al., Reference Smith, Walters and Prior2019; Jiang, Reference Jiang2021) or translation production tasks (e.g., Christoffels et al., Reference Christoffels, De Groot and Kroll2006; De Groot et al., Reference De Groot, Dannenburg and Vanhell1994; Herrera, Reference Herrera2019; Ibrahim et al., Reference Ibrahim, Cowell and Varley2017; Kroll & Stewart, Reference Kroll and Stewart1994; La Heij et al., Reference La Heij, Hooglander, Kerling and Van Der Velden1996; Potter et al., Reference Potter, So, Von Eckardt and Feldman1984; Van Hell & De Groot, Reference Van Hell and De Groot2008) to investigate whether connectivity between the first language (L1) and the second language (L2) in the bilingual mental memory is achieved via direct lexical associations or through a shared concept mediation. As such, the intricate process of translation per se has remained under-investigated.
1.1. Word translation production
This present study took an interest in the intricate process of word translation production per se. Translation production (but not comprehension) is favored here for the following reasons. Firstly, it provides valuable insight into the “active” translation production process in bilingual individuals compared with the “passive” translation comprehension tasks (i.e., translation priming or recognition). Secondly, translation production is of more interest due to its relevance to real-life translation, offering higher ecological validity (also cf. Christoffels et al., Reference Christoffels, Ganushchak and Koester2013). Thirdly, word translation production is particularly intriguing due to its simultaneous involvement of word recognition and production processes (also cf. Christoffels et al., Reference Christoffels, Ganushchak and Koester2013; Kroll et al., Reference Kroll, Van Hell, Tokowicz and Green2010). The process of word translation production is indeed a complex and fascinating area of study in linguistics and cognitive science. While it has received some attention, there are ongoing debates and challenges in understanding the intricacies of word translation production.
Potter et al. (Reference Potter, So, Von Eckardt and Feldman1984) outlines two potential routes through which a bilingual individual can perform word translation. The first hypothesis, known as the “word association hypothesis,” suggests that translation can occur by establishing a direct connection or association at the lexical level between the representations of the two translation equivalents. In contrast, the second hypothesis, called the “concept mediation hypothesis,” posits that translation can be accomplished by utilizing a common representation within the amodal conceptual system. Both hypotheses distinguish between two hierarchies of representation in bilingual memory (Snodgrass, Reference Snodgrass1984), namely lexical and conceptual, hence they are “hierarchical” models.
Kroll and Stewart (Reference Kroll and Stewart1994) introduced the revised hierarchical model (RHM), which effectively merges the hypotheses of word association and concept mediation in bilingual language processing. The RHM is categorized as an asymmetry model as it was originally developed to explain the observed differences in translation performance among unbalanced bilinguals. According to the RHM, L1 is assumed to have a privileged pathway to access meaning, while L2 often necessitates mediation through the L1 translation equivalent until the bilingual becomes proficient enough in the L2 to access meaning directly. Consequently, there exists an inherent asymmetry in the strength of the connections between words and concepts in the two languages of unbalanced bilinguals. In this framework, translating from L1 to L2, known as forward translation, primarily involves concept mediation. Conversely, when translating from L2 to L1, referred to as backward translation, the process relies more on direct lexical links, allowing the translation equivalent to be accessed directly without necessarily engaging the conceptual system. The RHM prescribes qualitatively distinct translation routes for forward and backward translation, leading to two key predictions: better performance in backward translation compared to forward translation, and effects of semantic manipulation only on forward but not on backward translation.
Although the RHM has offered valuable insights into the process of word translation, it has encountered significant challenges due to mixed findings in empirical research. Studies on oral word translation reported larger semantic effects on forward than on backward translation (e.g., De Groot et al., Reference De Groot, Dannenburg and Vanhell1994; Herrera, Reference Herrera2019; Ibrahim et al., Reference Ibrahim, Cowell and Varley2017; Kroll & Sholl, Reference Kroll and Sholl1992; Kroll & Stewart, Reference Kroll and Stewart1994; Sholl et al., Reference Sholl, Sankaranarayanan and Kroll1995), similar-sized effects or even effects in the opposite direction (e.g., Christoffels et al., Reference Christoffels, De Groot and Kroll2006; Christoffels et al., Reference Christoffels, Ganushchak and Koester2013; De Groot & Poot, Reference De Groot and Poot1997; Ibrahim et al., Reference Ibrahim, Cowell and Varley2017; La Heij et al., Reference La Heij, Hooglander, Kerling and Van Der Velden1996; Sakaki et al., Reference Sakaki, Hakoda and Kaminska2012; Van Hell & De Groot, Reference Van Hell and De Groot2008).
Unlike the asymmetry model which suggested qualitatively distinct routes for forward and backward translation, La Heij et al. (Reference La Heij, Hooglander, Kerling and Van Der Velden1996) put forward the idea of quantitatively different routes for two directions of translation and proposed a two-process account. According to this account, the initial process entails accessing the underlying concept linked to a given word (concept activation), and the subsequent process involves retrieving the corresponding word from the lexicon based on this conceptual information (word retrieval). For individuals with unbalanced bilingual proficiency, L2 is the weaker language. In translation from L1 to L2, the main problem is word retrieval but not concept activation; whereas in translation from L2 to L1, the primary challenge is reversed. As such, forward translation requires more effort in the word retrieval process, while backward translation encounters greater difficulties in the concept activation process. Therefore, within this framework, forward translation and backward translation vary in the relative ease of these two constituent processes.
Indeed, an increasing body of research suggests that both word association and concept mediation pathways are utilized in translation (De Groot et al., Reference De Groot, Dannenburg and Vanhell1994; Thierry & Wu, Reference Thierry and Wu2007; Wu & Thierry, Reference Wu and Thierry2010, Reference Wu and Thierry2012; Dijkstra & Rekké, Reference Dijkstra and Rekké2010; Dijkstra et al., Reference Dijkstra, Wahl, Buytenhuijs, Van Halem, Al-Jibouri, De Korte and Rekké2019; Kroll et al., Reference Kroll, Van Hell, Tokowicz and Green2010; La Heij et al., Reference La Heij, Hooglander, Kerling and Van Der Velden1996; Liu et al., Reference Liu, Qi and Di Biase2020). The degree to which each pathway is used can be influenced by multiple factors, such as word concreteness (e.g., Chaouch-Orozco et al., Reference Chaouch-Orozco, González Alonso, Duñabeitia and Rothman2023; Chen et al., Reference Chen, Liang, Cui and Dunlap2014; Van Hell & De Groot, Reference Van Hell and De Groot2008), frequency (e.g., De Groot, Reference De Groot1989; Ibrahim et al., Reference Ibrahim, Cowell and Varley2017), familiarity (e.g., Chen et al., Reference Chen, Zhang, Li and Wang2015; Cheung & Chen, Reference Cheung and Chen1998), cognate status (e.g., Ferre et al., Reference Ferre, Sanchez-Casas, Comesana and Demestre2017) and translation ambiguity (e.g., Jouravlev & Jared, Reference Jouravlev and Jared2020). Moreover, another factor contributing to the mixed findings in translation research is the variability in L2 proficiency levels among bilingual individuals. Research has indicated that as proficiency in L2 improves, there is a developmental shift from word association to concept mediation (as reviewed in Kroll & Tokowicz, Reference Kroll, Tokowicz, Froll and Groot2005; Kroll et al., Reference Kroll, Van Hell, Tokowicz and Green2010). Therefore, the two alternatives—word association and concept mediation—may correspond to different proficiency stages in L2 (Chen & Leung, Reference Chen and Leung1989; Kroll & Curley, Reference Kroll and Curley1988).
To accommodate the mixed results in research on translation, Brysbaert and Duyck (Reference Brysbaert and Duyck2010) suggested that the RHM should be replaced with connectionist models like the bilingual interactive activation plus (BIA+) model (Dijkstra & Van Heuven, Reference Dijkstra and Van Heuven2002). In response to Brysbaert and Duyck’s suggestion, Kroll et al. (Reference Kroll, Van Hell, Tokowicz and Green2010) contended that the RHM was fundamentally a model of bilingual word production while the BIA+ model is one for bilingual word recognition. It is important to highlight the complexity of word translation production in that it includes aspects of word recognition, meaning retrieval and word production. Considering such complexity, Dijkstra and Rekké (Reference Dijkstra and Rekké2010) developed a computational cognitive model, named Multilink, which aimed to simulate these more basic cognitive sub-processes involved in word translation. Of note, Multilink is a localist-connectionist model of word translation which integrates basic assumptions of both BIA+ and RHM. More recently, Dijkstra et al. (Reference Dijkstra, Wahl, Buytenhuijs, Van Halem, Al-Jibouri, De Korte and Rekké2019) expanded Multilink, providing a more thorough explanation of bilingual word recognition and translation.
Taken together, word translation production is one of the most sophisticated skills of bilingual individuals, involving several sub-processes like recognizing a word, retrieving its meaning and producing a corresponding word. Furthermore, the intricate sub-processes of word translation production become even more complicated when taking into account individual differences in L2 proficiency and variable manipulation of word characteristics including word concreteness, frequency, familiarity, cognate status and translation ambiguity. It is not surprising that empirical research on word translation has so far yielded conflicting findings, implying the dynamic and adaptable nature of translation.
1.2. The word flanker paradigm
To address the dynamics and adaptation of the translation process, we introduced a word flanker paradigm (Shaffer & LaBerge, Reference Shaffer and LaBerge1979) which was inspired by the classic Eriksen flanker paradigm (Eriksen & Eriksen, Reference Eriksen and Eriksen1974). In a typical word flanker experiment, a target word is centrally presented, flanked or surrounded by two copies of a word. Flanker words are placed adjacent to the target word, either on its left and right sides or above and below it (e.g., Guttentag et al., Reference Guttentag, Haith, Goodman and Hauch1984; Snell et al., Reference Snell, Declerck and Grainger2018). These flanker words can be identical, semantically related or unrelated to the central target word. Participants in such experiments are usually tasked with responding to the central word by semantic categorization judgment or lexical (word or non-word) decision—while trying to ignore the potentially distracting flanker words. The purpose is to observe and measure how these flanker words influence the processing or recognition of the target word. The word flanker paradigm provides a powerful tool for understanding the dynamics of linguistic processing, by varying the relationship between the target and flanker words.
For example, Guttentag et al. (Reference Guttentag, Haith, Goodman and Hauch1984) pioneered a bilingual variant of the flanker paradigm, in which the target and flanker words were presented in different languages. The target-flanker relations were varied: 1) translations of the target words; 2) the same semantic category, but not translations; 3) different semantic category, but the same response and 4) different responses. Participants were required to make the semantic categorization of the target words verbally (Experiment 1) and manually (Experiment 2) in conditions of cross-language flankers, while making the semantic categorization judgment manually in conditions of both cross- and same-language flankers (Experiment 3). Results from all three experiments showed significant effects of cross-language flankers (i.e., the meaning of the to-be-ignored flanker words affected the processing of the target words), in favor of the parallel view of multiple word processing. With a lexical decision task, Fox (Reference Fox1996) also used semantic associates (Experiment 1) and translation equivalents (Experiment 2) as flankers to investigate the cross-language priming effects from the to-be-ignored flanker words on the processing of the target words. To revisit the parallel semantic processing in reading, Snell et al. (Reference Snell, Declerck and Grainger2018) used a flanker paradigm with a semantic categorization judgment (Experiments 2 and 3), where the target words were accompanied by either their translation equivalents or semantically unrelated words as flankers. The results showed a significant improvement in performance when the flankers were translation equivalents, indicating that readers are capable of integrating semantic information from several words simultaneously.
Therefore, these studies have employed translation equivalents as flanker words (Fox, Reference Fox1996; Guttentag et al., Reference Guttentag, Haith, Goodman and Hauch1984; Snell et al., Reference Snell, Declerck and Grainger2018), simply because translation equivalence is considered to offer the most robust semantic link between two words, thus maximizing the likelihood of detecting semantic effects (Snell et al., Reference Snell, Declerck and Grainger2018). The focus of these studies, however, has mainly concerned whether words are processed in a series or parallel during comprehension. Notably, there has not been any research specifically focusing on the cognitive process involved in word translation production using the word flanker paradigm to date.
1.3. The present study
The present study adopted the word flanker paradigm to investigate the intricate process of oral word translation per se, particularly in its dynamics of being influenced by context-dependent sub-processes. In real-life scenarios where a word is being translated, the linguistic context surrounding the to-be-translated target word may impact its translation. The word flanker paradigm provides an ideal method for examining these context-dependent sub-processes during word translation production, with the target word being translated while the flanker words serve as the context. Considering this, this research constructed a minimal context wherein adjacent words might potentially influence translation as a word is being processed for translation. To this end, a bilingual version of the flanker paradigm was implemented to examine the impact of within- and cross-language semantic effects on the production of word translation. More specifically, participants were presented with a central target stimulus word that was flanked by two identical words. The central target word could be semantically identical, related or unrelated to the flanker words. Meanwhile, the language of the central target word could be the same as or different from that of the flanker words. The participant’s task was to translate the central target word from Chinese to English (forward translation) in one session or from English to Chinese (backward translation) in the other while ignoring the flanker words. In addition, to control for the confounds from individual differences in L2 proficiency and word characteristics, we recruited unbalanced Chinese-English bilingual individuals as participants and selected nouns with relatively high frequency to be the target words for translation.
The present study addresses three main research questions: 1) Is forward translation generally slower than backward translation, in line with the RHM? 2) Are semantic effects more pronounced in forward translation than in backward translation under the within-language (i.e., target-flanker language-congruent) conditions? 3) Do such semantic effects still persist under the cross-language (i.e., target-flanker language-incongruent) conditions? According to the asymmetry model (Kroll & Stewart, Reference Kroll and Stewart1994), forward translation predominantly relies on conceptual mediation while backward translation primarily depends on word-to-word associations. As such, the two translation directions should exhibit varying degrees of sensitivity to conceptual influences: forward translation should be affected by changes in semantic or conceptual information, while backward translation should be less affected by such changes. In line with the results in the study (Experiment 3) by Kroll and Stewart (Reference Kroll and Stewart1994), we also predict 1) that forward translation would be slower than backward translation and 2) that our semantic manipulation (i.e., within-language semantic effects) would affect forward more than backward translation. However, it remains to see 3) if such a pattern of semantic effects would hold true under the cross-language conditions.
2. Methods
2.1. Participants
A total of 52 Chinese-English bilingual speakers participated in the study (mean age = 19.3, SD = 0.7, 41 females). All participants were right-handed and reported no language, hearing and neurological impairments. Their vision was either normal or corrected to normal. The study was approved by the Ethics Review Board of Chongqing Medical University, China.
All participants were native Chinese speakers with Mandarin as their mother tongue (L1). They had started learning English as their second language (L2) between the ages of 9–12 in primary education and had learned English for about 6–11 years by the time the experiment was conducted. To assess the participants’ level of language use and proficiency, an adapted form of the Language Experience and Proficiency Questionnaire (LEAP-Q; Marian et al., Reference Marian, Blumenfeld and Kaushanskaya2007) was employed. As indicated by the data in Table 1, a statistically significant discrepancy was observed between Chinese and English across all evaluated metrics (ps < .001), showing that the participants exhibited unbalanced bilingualism.
Table 1. Language background of participants: age of acquisition (AoA), scores of self-rated proficiency, language exposure and language use
2.2. Materials and Procedure
All experimental stimuli (520 words) were commonly used concrete nouns, which were selected from English textbooks for the seventh-to-ninth Graders in schools where English is taught as a foreign language starting from the third grade. These words were specifically chosen for their relatively high frequency and relative ease of translation, ensuring that they were easily understandable and translatable for the unbalanced bilingual participants. We initially selected 186 semantically related word pairs as candidate stimuli. An independent cohort of 37 undergraduate students then rated the semantic relatedness of these word pairs with a five-point scale (1 = highly unrelated, 5 = highly related). Based on their rating, the top 100 pairs with the highest relatedness (average score = 3.94, SD = 0.63) were categorized as semantically related lists (Sublist 1). For the semantically unrelated list (Sublist 2), another 100-word pairs were selected, some of which were created by reorganizing the remaining 86 semantically related pairs into unrelated ones. These 100-word pairs were then rated by the same cohort of undergraduates as semantically unrelated (average score = 1.66, SD = 0.42). There was a significant difference in semantic relatedness between Sublist 1 and Sublist 2 (p < .001). In addition, another set was composed of an additional 100 words categorized as the semantically repeated list (Sublist 3). Therefore, the master list comprised 500 words and was organized into 300 sets of target-flanker triplets (i.e., a target word and two flanking words), which were equally divided into three sublists (see Appendix; also see Figure 1 for examples). Finally, an extra 20 words were picked for practice trials, which were later incorporated into the main experiment as warm-up trials. Nevertheless, selecting appropriate words from the total vocabulary of 2,200 words from English textbooks for junior middle school students posed a challenge. Given our primary research question concerning semantic effects on word translation, we prioritized the semantic relatedness factor. The objective ratings of target word concreteness (Brysbaert et al., Reference Brysbaert, Warriner and Kuperman2014), frequency (Van Heuven et al., Reference Van Heuven, Mandera, Keuleers and Brysbaert2014) and length of Chinese characters and English syllables were presented across three semantic-relatedness conditions in Table 2. There were no significant differences in concreteness between the related and unrelated conditions, nor between the unrelated and identical conditions (ps > .05). However, a significant difference was observed between the related and identical conditions (p < .05). Similarly, no significant differences in frequency were found between the related and unrelated conditions or the unrelated and identical conditions (ps > .05), but a significant difference existed between the related and identical conditions (p < .05). In terms of the length of Chinese characters, no significant difference was noted between the identical and unrelated conditions (p > .05), but differences were found between the related and identical conditions and between the related and unrelated conditions (ps < .001). Finally, there were no significant differences in the length of English syllables between the related and identical conditions or the unrelated and identical conditions (ps > .05), yet a significant difference was evident between the related and unrelated conditions (p < .01). We will return to this point when we discuss limitations of the study.
Figure 1. Left panel: Flanker-target relations as a function of target-flanker language congruency (congruent vs. incongruent) and semantic relatedness (identical vs. related vs. unrelated). Right panel: Examples of six conditions involving flanker-target relations with target-flanker triplets as stimuli are provided here (only the English-to-Chinese translation is shown). Participants are asked to orally translate the central target word into the other language while ignoring the flanker words. 花朵 = flower, 桔子 = orange, 钢笔 = pen.
Table 2. Ratings of target word concreteness, frequency and length of Chinese characters and English syllables
We employed a bilingual version of the flanker paradigm (Guttentag et al., Reference Guttentag, Haith, Goodman and Hauch1984) in which a central target word was flanked by two flanking words on either side. The target-flanker relations were manipulated as a function of semantic relatedness and language congruency. More specifically, depending on the sublist they originated from, the target word could either be semantically related (e.g., “chair desk chair” from Sublist 1), unrelated (e.g., “roof shark roof” from Sublist 2) or identical (e.g., “gate gate gate” from Sublist 3) to its flanking words. Meanwhile, the language of the target word could be the same as or different from that of the flanking words, leading to either congruent (within-language) or incongruent (cross-language) conditions between the target and flankers. As such, the target-flanker relations yielded a total of six distinct conditions (see Figure 1: left panel). From the master list of 500 words, 300 sets of target-flanker triplets (i.e., a target word and two flanking words) were constructed, with 50 sets for each of the six conditions. A third factor considered was the direction of translation. There were two translation sessions: in one, the target words were Chinese, requiring a Chinese-to-English (or forward) translation, while in the other, they were English, necessitating an English-to-Chinese (or backward) translation.
The study followed a 2 × 2 × 3 design, with translation direction (Chinese-to-English vs. English-to-Chinese), target-flanker language congruency (congruent vs. incongruent) and target-flanker semantic relatedness (identical vs. related vs. unrelated) serving as within-subjects variables. Overall, the study included 600 sets of triplets (or trials), with half of these sets consisting of English target words and the other half comprising Chinese target words. There were two sessions of translation tasks: one involving Chinese-to-English translation and the other involving English-to-Chinese translation. The allocation of tasks was counterbalanced across participants. There was a ten-minute break between sessions. Each of the two sessions consisted of 300 trials across six different conditions. The 300 trials were pseudorandomized, ensuring that trials of the same condition were separated by a minimum of three trials. Each session was composed of four blocks, each containing 77 trials, with the initial two trials serving as warm-up trials. There was a 1.5-minute break time between blocks.
Each trial (see Figure 1: right panel) started with a fixation cross lasting 300 milliseconds (ms). Following this, a central target word flanked by two identical words on either side appeared for a duration of 2,700 ms. Once a response was registered or when the maximum time of 2,700 ms had elapsed, a blank screen was then displayed for 1,500 ms. A long stimulus duration was employed here to mimic real-world translation scenarios, where the target word for translation is always accompanied by adjacent words. The longest target word subtended a horizontal visual angle of 4.76 °, while the longest potential triplet spanned 7.04 ° horizontally. The trials were presented using E-Prime 3.0 software (Psychology Software Tools, Pittsburgh, PA). Participants viewed the screen from an approximate distance of 60 cm, with the screen operating at a refresh rate of 60 Hz and a resolution of 1,024 × 768. Response times (RTs) were recorded using the Chronos microphone device (Psychology Software Tools, Pittsburgh, PA), and measured from the onset of the stimulus. A digital voice recorder was used to record vocal responses for later verification of incorrect responses.
In the pilot experiment, the translation task proved to be difficult for most participants, who exhibited notably delayed responses (failing to meet the required 2,700 ms timeframe per trial), incorrect responses or an inability to translate altogether. Therefore, we implemented a familiarization procedure to help the participants improve the correct responses. This procedure might have led to individual differences due to learning effects (see Discussion). In the days preceding the experiment, participants were requested to complete language background questionnaires through email, which served as an initial tool to screen and select participants. Afterward, participants familiarized themselves with a set of 520 words (500 for the main experiment and 20 for practice), arranged randomly in an emailed word text file containing both English and Chinese versions. On the spot before the experiment, participants were asked to translate the 520 stimulus words in a booklet from Chinese to English and then vice versa. Correct translations were provided if participants produced an incorrect response. Following this, a practice block of 48 trials was conducted to acquaint participants with the task procedure. The instructions and procedures during the practice session were identical to those in the actual experimental sessions. The main experiment was conducted in a dimly lit sound-proof booth. Participants were required to translate the designated word into the other language as accurately and quickly as possible while ignoring the flanker words. Meanwhile, outside the soundproof booth, the experimenter documented errors using a wireless omnidirectional microphone system. After the experiment, participants received a debriefing.
3. Results
In the assessment of RTs, only the valid trials were subjected to analysis (All data and scripts are available at https://osf.io/w3cnm/). Trials were categorized as erroneous, with a percentage of 5.7% for forward translation and 4.0% for backward translation, if the participants committed either: (i) an incorrect translation or (ii) vocalized errors, including false starts or verbal hesitations. Additionally, trials were excluded from the analysis under the circumstances where the participants failed to respond within the 2,700-ms timeframe (7.9% for forward translation and 9.9% for backward translation) or if their RTs deviated more than three standard deviations from the group mean (1.9% for forward translation and 2.0% for backward translation).
3.1. Results of RTs
For the analysis by subjects, a three-way repeated-measures analysis of variance (ANOVA) was conducted on the RTs of word translation production as a function of translation direction (Chinese-to-English vs. English-to-Chinese), target-flanker language congruency (congruent vs. incongruent) and semantic relatedness (identical vs. related vs. unrelated). The results are shown in Figure 2, Tables 3 and 4. A main effect was observed in translation direction [F(1, 51) = 5.65, p = .021], revealing that RTs were faster by 44 ms for forward translation (Chinese-to-English) as compared to backward translation (English-to-Chinese), with mean RTs of 1,118 ms and 1,162 ms, respectively. A significant main effect emerged for target-flanker language congruency [F(1, 51) = 66.12, p < .001], indicating that RTs were slower by 40 ms in language-congruent compared to language-incongruent conditions (1,160 ms vs. 1,120 ms, respectively). Furthermore, semantic relatedness also showed a significant main effect [F(2, 50) = 272.13, p < .001], with mean RTs of 1,066 ms, 1,162 ms and 1,193 ms for semantically identical, related and unrelated conditions, respectively. Post hoc comparisons, corrected using Bonferroni adjustments (as applied similarly hereinafter), demonstrated significant differences (ps < .001) among all three semantic conditions, with RTs being 31 ms faster for semantically related compared to unrelated conditions, and 96 ms faster for semantically identical compared to related conditions.
Figure 2. A by-subject repeated-measures ANOVA of word translation production on mean response times (RTs) and error rates (%) as a function of translation direction (Chinese-to-English vs. English-to-Chinese), target-flanker language congruency (congruent vs. incongruent) and target-flanker semantic relatedness (related vs. unrelated vs. identical). *p < .05, **p < .01, ***p < .001.
Table 3. By-subject and by-item analyses of response times (RTs) in milliseconds (ms) and error rates in percentage (%) as a function of translation direction (Chinese-to-English vs. English-to-Chinese), target-flanker language congruency (congruent vs. incongruent) and target-flanker semantic relatedness (related vs. unrelated vs. identical)
Table 4. A by-subject repeated-measures ANOVA performed on response times (RTs) and error rates of word translation production as a function of translation direction (tra.dir), target-flanker language congruency (lan.con) and semantic relatedness (sem.rel). *p < .05, **p < .01, ***p < .001
A significant interaction was obtained between translation direction and target-flanker language congruency [F(1, 51) = 9.01, p = .004]. A post hoc analysis revealed that the RTs in the Chinese-to-English translation direction were slower when the target and flanker languages were incongruent compared to when they were congruent [F(1, 51) = 9.16, p = .004]. Conversely, in the Chinese-to-English translation direction, there were no notable differences in response RTs between congruent and incongruent conditions [F(1, 51) = 2.68, p > .05]. Furthermore, a significant interaction effect between translation direction and target-flanker semantic relatedness was found [F(2, 50) = 14.40, p < .001]. A post hoc analysis revealed an absence of a difference between semantically related and unrelated conditions in the Chinese-to-English direction (p > .05), contrasted by a significant semantic facilitation effect in the English-to-Chinese direction (p < .001). The semantic repetition effect (i.e., identical vs. related conditions) was also significant across both translation directions (ps < .001). Additionally, an interaction between target-flanker language congruency and semantic relatedness was observed [F(2, 50) = 235.08, p < .001]. A post hoc analysis demonstrated no difference in RTs between semantically related and unrelated conditions when language congruency was incongruent (p > .05). However, a significant semantic facilitation effect was observed under language congruent conditions (p < .001). Moreover, the semantic repetition effect was significant in both language-congruent (p = .008) and language-incongruent (p < .001) conditions.
Additionally, a three-way interaction among translation direction, target-flanker language congruency, and semantic relatedness also emerged [F(2, 50) = 6.97, p = .002]. For the Chinese-to-English translation, a post hoc analysis showed that in language congruent conditions, RTs were faster for semantically related compared to unrelated conditions (p = .049) and for semantically identical compared to unrelated conditions (p < .001), although no difference was evident between semantically identical and related conditions (p > .05). In language incongruent conditions, RTs were significantly faster for semantically identical conditions compared to both related and unrelated conditions (ps < .001), with no significant difference between semantically related and unrelated conditions (p > .05). Conversely, for the English-to-Chinese translation, a post hoc analysis demonstrated that in language congruent conditions, RTs were faster for semantically related conditions as compared to both unrelated and identical conditions (ps < .001). Yet, no significant difference emerged between semantically identical and unrelated conditions (p > .05). In language incongruent conditions, the translation performance showed similar patterns in the English-to-Chinese translation as those observed in the Chinese-to-English translation.
For the analysis by items, a univariate ANCOVA was performed on RTs for translation direction, language congruency and semantic relatedness with frequency and concreteness as covariates. The results are shown in Figure 3, Tables 3 and 5. A main effect was observed in translation direction [F(1, 599) = 22.56, p < .001], revealing that RTs were faster by 46 ms for forward translation (Chinese-to-English, 1,109 ms) as compared to backward translation (English-to-Chinese, 1,155). A significant main effect emerged for target-flanker language congruency [F(1, 599) = 13.59, p < .001], indicating that RTs were slower by 35 ms in language-congruent compared to language-incongruent conditions (1,149 ms vs. 1,114 ms, respectively). Furthermore, semantic relatedness also showed a significant main effect [F(2, 598) = 59.98, p < .001], with mean RTs of 1,058 ms, 1,158 ms and 1,179 ms for semantically identical, related and unrelated conditions, respectively. Post hoc comparisons demonstrated significant differences (ps < .001) between semantically identical and related conditions and between semantically identical and unrelated conditions, and marginally significant differences (p = .73) between semantically related and unrelated conditions. An interaction between target-flanker language congruency and semantic relatedness was observed [F(2, 598) = 50.96, p < .001]. A post hoc analysis demonstrated significant differences (ps < .001) in RTs between semantically identical and related conditions and between semantically identical and unrelated conditions only when language congruency was incongruent. No other interactions were observed. In addition, the effect of word frequency as a covariate was significant [F(1, 599) = 33.08, p < .001] while the effect of word concreteness as a covariate was non-significant [F < 1].
Figure 3. A by-item univariate analysis of covariance (ANCOVA) of word translation production on mean response times (RTs) and error rates (%) as a function of translation direction (Chinese-to-English vs. English-to-Chinese), target-flanker language congruency (congruent vs. incongruent) and target-flanker semantic relatedness (related vs. unrelated vs. identical), with word frequency and concreteness as covariates. *p < .05, **p < .01, ***p < .001.
Table 5. A by-item univariate ANCOVA performed on response times (RTs) and error rates of word translation production as a function of translation direction (tra.dir), target-flanker language congruency (lan.con) and semantic relatedness (sem.rel), with word frequency (wor.fre) and concreteness (wor.con) as covariates. *p < .05, **p < .01, ***p < .001
3.2. Results of error rates
For the analysis of subjects, a parallel repeated-measures ANOVA was performed on the percentages of errors committed (see Figure 2, Tables 3 and 4). A significant main effect of translation direction was observed [F(1, 51) = 14.23, p < .001], indicating a 1.6% higher rate of errors for forward translation (Chinese-to-English) relative to backward translation (English-to-Chinese), with respective to the average error percentages of 5.7% and 4.0%. Additionally, a statistically significant main effect was found for the variable of target-flanker language congruency [F(1, 51) = 20.09, p < .001], revealing a 1.1% increase in errors in language-congruent conditions compared to language-incongruent conditions, with corresponding mean error percentages of 5.4% and 4.3%. Moreover, a significant main effect of target-flanker semantic relatedness was observed [F(2, 50) = 4.96, p = .011]. The mean error percentages were 5.0%, 5.3% and 4.2% for semantically related, unrelated and identical conditions, respectively. Subsequent post hoc comparisons showed significantly fewer errors in semantically identical conditions as compared to both semantically related (p = .011) and unrelated (p = .011) conditions. Conversely, no statistically significant differences were found between semantically related and unrelated conditions (p > .05). An interaction between target-flanker language congruency and semantic relatedness was observed [F(2, 50) = 27.12, p < .001]. Under conditions of target-flanker language incongruence, a post hoc analysis revealed an absence of statistically significant differences in the error rates between semantically identical and unrelated conditions (p > .05). However, both a semantic facilitation effect (p = .041) and a semantic repetition effect (p < .001) were observed to be statistically significant. In contrast, when the target-flanker language was congruent, a post hoc analysis indicated significantly reduced error rates in semantically identical conditions as compared to both semantically related and unrelated conditions (ps < .001). Conversely, no statistically significant differences were noted between semantically related and unrelated conditions (p > .05). No other main effects or interactions reached statistical significance.
For the analysis by items, a parallel univariate ANCOVA was performed on the percentages of errors for translation direction, language congruency and semantic relatedness with frequency and concreteness as covariates (see Figure 3, Tables 3 and 5). A significant main effect of translation direction was observed [F(1, 599) = 17.80, p < .001], indicating a 1.5% higher rate of errors for forward translation (Chinese-to-English) relative to backward translation (English-to-Chinese), with respective to the average error percentages of 5.7% and 4.1%. Additionally, a marginally significant main effect was found for the variable of target-flanker language congruency [F(1, 599) = 3.74, p = .054], revealing a 0.7% increase in errors in language-congruent conditions compared to language-incongruent conditions, with corresponding mean error percentages of 5.3% and 4.5%. Moreover, a significant main effect of target-flanker semantic relatedness was observed [F(2, 598) = 7.75, p < .001]. The mean error percentages were 5.2%, 5.6% and 3.9% for semantically related, unrelated and identical conditions, respectively. Subsequent post hoc comparisons showed significantly fewer errors in semantically identical conditions as compared to both semantically related (p = .007) and unrelated (p < .001) conditions. Conversely, no statistically significant differences were found between semantically related and unrelated conditions (p > .05). An interaction between target-flanker language congruency and semantic relatedness was observed [F(2, 50) = 21.76, p < .001]. A post hoc analysis indicated significantly reduced error rates in semantically identical conditions as compared to both semantically related (congruent: p = .016, incongruent: p < .001) and unrelated conditions (congruent: p = .011, incongruent: p < .001). However, no statistically significant differences were noted between semantically related and unrelated conditions in either the target-flanker language congruent or incongruent conditions (ps > .05). No other interactions reached significance. In addition, an effect of word frequency as a covariate was significant [F(1, 599) = 13.29, p < .001] while an effect of word concreteness as a covariate was non-significant [F(1, 599) = 2.47, p > .05].
4. Discussion
This study investigated within- and cross-language semantic effects on the production of word translation with a bilingual version of the flanker paradigm. The findings are summarized as follows: 1) forward translation was faster than backward translation; 2) within-language semantic facilitation effects were observed in both forward and backward translation, but such effects were more pronounced in backward than in forward translation; 3) no cross-language semantic facilitation effects were found; 4) there were substantial cross-language semantic repetition effects. Below we restrict our discussion to the literature which involves explicit oral word translation.
4.1. (Within-language) semantic effects on oral word translation
The first finding showed shorter translation latencies from L1 to L2 (forward translation) than from L2 to L1 (backward translation), which contradicts the first prediction of the RHM. Moreover, semantic facilitation effects in this study were observed in both forward and backward translation only under the target-flanker language-congruent (i.e., within-language) conditions when the target word and the flanker words were semantically related. This finding was also opposite to the second prediction of the RHM. Our first two findings, albeit contrary to the RHM’s two predictions, were consistent with a number of subsequent studies. For example, De Groot et al. (Reference De Groot, Dannenburg and Vanhell1994) manipulated semantic properties (e.g., concreteness, imageability) of to-be-translated words and found different translation speeds in directions between concrete and abstract words. More subsequent studies manipulated semantic properties of target words to examine the semantic effects on translation asymmetry, such as frequency and imageability (e.g., De Groot & Poot, Reference De Groot and Poot1997), concreteness and context availability (e.g., Van Hell & De Groot, Reference Van Hell and De Groot2008), cognate status and frequency (e.g., Christoffels et al., Reference Christoffels, De Groot and Kroll2006). Results from these studies demonstrated that semantic effects were present in both forward and backward directions. La Heij et al. (Reference La Heij, Hooglander, Kerling and Van Der Velden1996) used a picture-word interference paradigm and showed similar semantic effects in both translation directions. These previously reported results, together with our findings, presented a challenge to the RHM, suggesting quantitatively different routes for two directions of translation.
Of note, however, a similar size of behavioral findings have also aligned with the RHM’s proposed asymmetry (e.g., Herrera, Reference Herrera2019; Ibrahim et al., Reference Ibrahim, Cowell and Varley2017; Kroll & Sholl, Reference Kroll and Sholl1992; Sholl et al., Reference Sholl, Sankaranarayanan and Kroll1995), demonstrating that when translating from L1 to L2 (forward translation), semantics were more likely to be involved compared to translating from L2 to L1 (backward translation). These results appear to prescribe qualitatively distinct translation routes for forward and backward translation as proposed by the asymmetry model (Kroll & Stewart, Reference Kroll and Stewart1994).
A key factor for the mixed results in translation asymmetry may lie in the to-be-translated word characteristics. Variables such as cognate status (e.g., Christoffels et al., Reference Christoffels, De Groot and Kroll2006; Ferre et al., Reference Ferre, Sanchez-Casas, Comesana and Demestre2017), word concreteness (e.g., Chaouch-Orozco et al., Reference Chaouch-Orozco, González Alonso, Duñabeitia and Rothman2023; Chen et al., Reference Chen, Liang, Cui and Dunlap2014; Van Hell & De Groot, Reference Van Hell and De Groot2008), frequency (e.g., De Groot, Reference De Groot1989; Ibrahim et al., Reference Ibrahim, Cowell and Varley2017) and familiarity (e.g., Chen et al., Reference Chen, Zhang, Li and Wang2015; Cheung & Chen, Reference Cheung and Chen1998) have been found to influence two directions of word translation. Considering those possible confounds, the present study controlled for word characteristics in several ways. Firstly, different scripts (i.e., Chinese and English) in this study, as compared to the same script (i.e., Dutch and English) in the study (Experiment 3) by Kroll and Stewart (Reference Kroll and Stewart1994), were used to avoid the orthographical and/or phonological overlap effect as observed for cognates. Secondly, this study selected relatively high-frequency target words, whereas the frequency of target words in the study by Kroll and Stewart (Reference Kroll and Stewart1994) varied considerably. This discrepancy is evident in the differing error rates observed: 4.8% in this study compared to 46% in their study (Experiment 3).
Therefore, the contradictory translation asymmetry obtained between the current study and Kroll and Stewart (Reference Kroll and Stewart1994) may be attributed to the selection of relatively high-frequency target words in this study (also see Türker, Reference Türker2018). Previous studies showed that translation latencies were reduced as frequency increased (Christoffels et al., Reference Christoffels, De Groot and Kroll2006; De Groot, Reference De Groot1992; Jescheniak & Levelt, Reference Jescheniak and Levelt1994). As a consequence, the frequency of to-be-translated words is a determinant of translation performance (De Groot, Reference De Groot1992; De Groot et al., Reference De Groot, Dannenburg and Vanhell1994). Kroll et al. (Reference Kroll, Van Hell, Tokowicz and Green2010) also suggested that those apparently contradictory findings in the literature may be attributed to the low frequency of the items used in their experiment (Experiment 3; Kroll & Stewart, Reference Kroll and Stewart1994) compared with other studies. In Multilink, a computational model of bilingual processing, frequency is simulated as a critical factor which can explain a large percentage of the variance in translation performance (Dijkstra & Rekké, Reference Dijkstra and Rekké2010; Dijkstra et al., Reference Dijkstra, Wahl, Buytenhuijs, Van Halem, Al-Jibouri, De Korte and Rekké2019). Increasing data showed that the frequency of word use modulates the translation asymmetry (e.g., Christoffels et al., Reference Christoffels, De Groot and Kroll2006; De Groot, Reference De Groot1992; Ibrahim et al., Reference Ibrahim, Cowell and Varley2017; Jescheniak & Levelt, Reference Jescheniak and Levelt1994; Soo & Monahan, Reference Soo and Monahan2023).
It is important to note, however, that the frequency of target words in this study was not adequately controlled, despite their relatively high frequency. Consequently, this lack of control affected the semantic priming and repetition effects, leading to a relative reduction in these effects in the by-item analysis compared to the by-subject analysis. In addition, word concreteness across the three semantic relatedness conditions was not well controlled in this study. However, the by-item analysis with concreteness as a covariate showed no significant effect. We will return to these points again when we discuss the limitations of the study.
4.2. Cross-language semantic effects on oral word translation
In addition, we observed substantial cross-language semantic repetition effects, with the translation of a target word being markedly facilitated by its flanking translation equivalents. The improved performance with translation-equivalent flankers was also obtained in studies on bilingual processing with comprehension tasks (e.g., Hoversten & Martin, Reference Hoversten and Martin2023; Snell et al., Reference Snell, Declerck and Grainger2018). Interestingly, no cross-language semantic facilitation effects were found under target-flanker language incongruent conditions. This finding provided the first evidence for the absence of cross-language semantic effects on oral word translation, suggesting that the translation of a target word presented in one language may not be subject to the semantic contexts (i.e., the flanker words) in the other language. However, some studies with the flanker paradigm have demonstrated cross-language effects during lexical decision or semantic categorization (e.g., Fox, Reference Fox1996; Guttentag et al., Reference Guttentag, Haith, Goodman and Hauch1984; Hoversten & Martin, Reference Hoversten and Martin2023; Snell et al., Reference Snell, Declerck and Grainger2018). One possibility for this discrepancy may lie in the task difference, with our study adopting a production task while other studies adopting comprehension tasks.
The two-process account offers a plausible explanation for the present or absence of within- or cross-language semantic relatedness (or repetition) priming effects in this study. In the process of concept activation, the target word and its semantically related flanking words may co-activate their semantic category, leading to speeded performance for translation. However, this facilitation is sensitive to within-language contexts, whereas it seems to be immune to cross-language contexts simply because the cognitive effort consumed for the translation of a target word leaves little room for the translation of the flanker words. In this case, the cross-language semantic effects of the flankers appeared to be little to no. In the process of word retrieval, the facilitation effect was present only in cross-language contexts but not in within-language contexts, simply because the translation of a target word was exactly its flanking translation equivalents.
4.3. Dynamic processes at the lexical and conceptual levels in oral word translation
Now, it is evident that both forward and backward translation in this study benefit from the semantic facilitation effects, despite that the semantic effects seemed to be weaker in forward translation than in backward translation. As such, there appears to be quantitative different translation routes for forward and backward translation (La Heij et al., Reference La Heij, Hooglander, Kerling and Van Der Velden1996), rather than qualitatively distinct translation routes as suggested by the RHM. If this holds true, it could be argued that both forward and backward translation are conceptually mediated. However, the semantic manipulation previously used seems to exert elusively varying effects on forward or/and backward translation, casting doubt on the single-route/process (i.e., either word association or concept mediation) account of word translation.
Instead, word translation may involve processes at the lexical and conceptual levels simultaneously. To demonstrate with three commonly investigated variables in translation studies, proficiency, frequency and familiarity may reflect a point in common: a dynamic lexical-conceptual interplay at word translation. Language proficiency can be explained by the relative strength of the links between words and concepts in bilingual memory structure (De Groot, Reference De Groot1992; De Groot et al., Reference De Groot, Dannenburg and Vanhell1994; Dijkstra & Rekké, Reference Dijkstra and Rekké2010; Dijkstra et al., Reference Dijkstra, Wahl, Buytenhuijs, Van Halem, Al-Jibouri, De Korte and Rekké2019; Kroll & Stewart, Reference Kroll and Stewart1994; Kroll et al., Reference Kroll, Van Hell, Tokowicz and Green2010). It is true of word frequency and familiarity. Take word frequency for example, frequent use of a word in a bilingual’s L1 strengthens the link between the L1 word and its concept. Likewise, it applies to his/her L2 as the frequency of L2 use increases. In a sense, frequent use of a word leads to more familiarity and thus higher proficiency, be it L1 or/and L2.
A more plausible hypothesis is that word translation may involve lexical and conceptual processes asynchronously rather than simultaneously. Following this idea, we may break down the translation process into two (or more) sub-processes in which word translation involves dynamic equilibrium between lexical and conceptual activation in bilingual memory. Actually, La Heij et al. (Reference La Heij, Hooglander, Kerling and Van Der Velden1996) proposed “a two-process account” for the variance in the literature of word translation. This account posits that word translation consists of two serially occurred processes, including a) concept activation: retrieving the meaning of the to-be-translated word and b) word retrieval: retrieving its translation equivalent on the basis of retrieved concept information. In this view, both forward and backward translation are conceptually mediated, but different in the relative ease or difficulty of these two processes. For less proficient bilinguals, L2 processing is relatively difficult. As such, L2 word retrieval is the main problem for translating from L1 into L2 (forward translation), whereas concept activation of L2 is the main problem for translating from L2 into L1 (backward translation).
The two-process account can well explain our data. In this study, unbalanced bilinguals have relative difficulty in English processing although the stimulus words were of relatively high frequency. Backward (English-to-Chinese) translation was more sensitive to semantics because concept activation of English words was the major problem while word retrieval of Chinese translation equivalents was the minor problem. As a consequence, the co-activation of semantically related items directly facilitated the target for concept activation, thus yielding speeded responses. In contrast, forward (Chinese-to-English) translation seemed less sensitive to semantics because concept activation of Chinese words was the minor problem while word retrieval of English translation equivalents was the major problem. Therefore, the dynamic interplay of word retrieval and concept activation could provide insight into the extent to which both directions of translation are affected by semantic effects. The two-process account can also provide an explanation for the translation asymmetry in the study (Experiment 3) of Kroll and Stewart (Reference Kroll and Stewart1994) in which words with a variance of frequency were selected. In forward translation, word retrieval of low-frequency unfamiliar words was the major problem; whereas in backward translation, concept activation became the major problem for low-frequency words.
In general, the dynamic and adaptable nature makes oral word translation a challenging yet fascinating field to explore. Firstly, the dynamics and adaptability of word translation indeed hinge significantly on the characteristics of the words to be translated. Each characteristic as aforementioned—word concreteness, frequency, familiarity, cognate status and translation ambiguity—plays a crucial role in shaping the translation process. In real-life translation or interpretation scenarios, these characteristics often interact, creating a multi-dimensional challenge to capture its intricacy and complexity. Secondly, it is important to emphasize the hierarchical structure of the linguistic system in translation. A single word encompasses multiple layers of information, including lexical, orthographic, phonetic and semantic aspects. Consequently, future research employing methods such as ERP (Event-Related Potentials) could yield crucial insights into how the brain processes language during translation at different linguistic levels, revealing whether and in what manner these processes occur asynchronously and interplay dynamically. Last but not least, bilinguals’ L2 proficiency may also be a critical factor which impacts translation performance. The intricacy of word characteristics and the complexity of the sub-processes involved in word translation increase significantly when considering individual variations in L2 proficiency. The specific level of skill and familiarity with L2 that each bilingual possesses can significantly influence how he or she is fine-tuned to the nuances of translation.
There are a couple of limitations in the present study. Firstly, we familiarized participants with the stimuli before the experiment, which might lead to the potential presence of learning effects and individual differences as confounding variables. Secondly, to prioritize the semantic relatedness factor, there was a lack of precise control over word concreteness, frequency and length across the three semantic relatedness conditions. Although the by-item ANCOVA in this study incorporates word frequency and concreteness as covariates to control for their potential influence on reaction times and error rates, future research should implement control measures at the item level for more robust results.
5. Summary
This study has investigated how the oral translation of a word situated in the central vision is influenced by its adjacent lexical items. The results demonstrate that forward translation is faster than backward translation, and that both directions of translation are subject to within-language (but not cross-language) semantic context effects. These results are at odds with the RHM predictions (Kroll & Stewart, Reference Kroll and Stewart1994). We have argued that the variable of frequency of to-be-translated words may provide an explanation for this discrepancy. The data we have presented demonstrate that both forward and backward translation are conceptually mediated, suggesting quantitatively rather than qualitatively distinct routes for two directions of translation. As such, our data seem to accommodate the two-process account of word translation that word retrieval and concept activation are at interplay, yielding context-sensitive translation performance. This theoretical account has an important implication for revealing the dynamic and adaptable process of oral word translation. Indeed, an integrated theoretical model that fully captures the dynamic and adaptable nature of word translation remains undeveloped. Such a model would need to address the complex, fluid and flexible processes involved in translating words. This includes considering the multifaceted interactions among word characteristics, the asynchronous processing of sub-linguistic levels such as semantics, orthography, phonetics and lexicon, as well as accounting for individual variations in L2 proficiency. By incorporating these elements, the model would offer a more holistic understanding of how different factors influence the translation process and contribute to the variability in translation outcomes. Developing this model would significantly advance our understanding of translation dynamics and enhance our insights into the cognitive processes that underpin translation.
Supplementary material
The supplementary material for this article can be found at http://doi.org/10.1017/langcog.2024.61.
Data availability statement
All data and scripts are available at https://osf.io/w3cnm/
Competing interest
The authors declare none.
Open access
