Translation priming asymmetry

Earlier studies of translation priming (e.g., Schwanenflugel and Rey, Reference Schwanenflugel and Rey1986; Grainger and Beauvillain, Reference Grainger and Beauvillain1988; Chen and Ng, Reference Chen and Ng1989; Jin, Reference Jin1990; Keatley et al., Reference Keatley, Spinks and de Gelder1994) used a lexical decision task (LDT), in which the prime was plainly visible, and the stimulus onset asynchrony (SOA) was at least 150 ms. The general findings revealed both translation and semantic facilitation effects. This would suggest that translation equivalents and semantically related words are somehow interconnected across languages. However, the main problem with using this type of priming paradigm is that participants would quickly realize that the target might be a translation of the prime, and hence might prepare for that eventuality by actually carrying out the necessary translation. Alternatively, there may be a post-access effect in which participants realize that the target was a translation of the prime, and this leads to a faster lexical decision response (e.g., Perea et al., Reference Perea, Dunabeitia and Carreiras2008). Therefore, the priming effect observed with a visible prime may not be an accurate indicator of the extent to which the representation of the prime automatically activates the representation of the target. For these reasons, a masked priming paradigm is preferred. In the standard masked priming paradigm (Forster and Davis, Reference Forster and Davis1984), a forward mask is followed by a briefly presented prime (40–60 ms), which is then immediately followed by the target. Under these conditions, the prime cannot be identified. In bilingual translation priming experiments the target is a word in one language and the prime is either the translation-equivalent word in the other language, or a completely unrelated word. In some experiments, the prime is followed by a backward mask so that participants have more time to process the prime, but are still unaware of the prime. Since participants are unaware of the prime, it is obvious that they would not be able to predict the target as the translation equivalent of the prime. Nor would they derive any benefit from a retrospective strategy, in which the relatedness of the target and the prime acts as a cue for the decision (Neely, Keefe and Ross, Reference Neely, Keefe and Ross1989). Thus, the claimed benefit of using this technique is that it is more sensitive to automatic processes, and less sensitive to strategic processes. A priming effect is observed when the primed target is responded to faster than the unprimed target, which is interpreted as indicating that the lexical entries in both languages are linked in some way (either at the lexical level or semantic level or both).

Another benefit of using this technique is that it provides the possibility of testing participants in one language (that of the target word) and manipulating the language of the prime without the participant's knowledge that the experiment is about their bilingualism. This bilingual paradigm is operating in a “monolingual mode” (e.g., Grosjean, Reference Grosjean1998, Reference Grosjean and Nicol2001), and thus accurately measures the automaticity of language processing.

Early work with this technique suggested that translation priming was restricted to cognate terms, i.e., words that are descended from the same form, for example, the Spanish–English translation equivalent rico–rich (de Groot and Nas, Reference de Groot and Nas1991; Sánchez-Casas, Davis and García-Albea, Reference Sánchez-Casas, Davis and García-Albea1992), but subsequent research with Hebrew–English and Chinese–English bilinguals showed that non-cognates also produced strong priming (Gollan et al., Reference Gollan, Forster and Frost1997; Jiang, Reference Jiang1999; Forster and Jiang, Reference Forster, Jiang and Nicol2001), suggesting that the difference in script might be a critical factor (Forster and Jiang, Reference Forster, Jiang and Nicol2001).

Importantly, one thing that seems clear from masked translation priming experiments using a lexical decision task is that words in L1 have a marked effect on the recognition of L2 words, but not vice versa (e.g., Keatly et al., 1994; Williams, Reference Williams1994; Gollan et al., Reference Gollan, Forster and Frost1997; Jiang, Reference Jiang1999). That is, an asymmetrical pattern of cross-language priming is observed: L1 can prime L2, but L2 cannot prime L1. This is not what one would expect if the two languages overlapped completely at the semantic level, which should produce equal amounts of priming in both directions, i.e., symmetry. This might suggest that bilinguals are unable to effectively process L2 primes within such a short time. However, this can't be true because L2–L2 repetition priming was reliably observed in both Chinese and Hebrew in lexical decision (e.g., Gollan et al., Reference Gollan, Forster and Frost1997; Jiang, Reference Jiang1999).

If the two languages of a bilingual share a common semantic system (at least for translation equivalents), one might ask how the asymmetry occurs. One might expect that the Revised Hierarchical Model (RHM) of bilingual representations (Kroll and Stewart, Reference Kroll and Stewart1994) could explain this priming asymmetry. The RHM assumes connections between L1 and L2 at the lexical level and a direct access from the form to the meaning (i.e., conceptual level) in both languages. According to this model, access to semantic information from L2 needs to be mediated by L1 at the initial stages of L2 learning. As L2 proficiency increases, a direct connection between L2 and the semantic system is established and L1 mediation is no longer useful in processing. Thus, an asymmetry is created in this model: L1 words have stronger connections to concepts than the newly learned L2 words (also see Kroll and Tokowicz, Reference Kroll, Tokowicz and Nicol2001, Reference Kroll, Tokowicz, Kroll and DeGroot2005). As suggested by Kroll and Tokowicz (Reference Kroll, Tokowicz and Nicol2001), the lexical asymmetry may arise in part from the differential reliance of L2 on L1, and may also be a result of the different mappings from a small L2 lexicon to a large L1 lexicon for which some L2 translation equivalents are absent. However, if there are strong lexical links between L2 words and their L1 translations (as the RHM proposes), then one would expect strong L2–L1 priming, which definitely is not the case. The alternative explanation might be that the relative speed of processing in L1 and L2 is responsible for the asymmetry. If L1 processing is unaffected by L2 primes because bilinguals process the L2 primes too slowly, the translation asymmetry in lexical decision seems obvious. In fact, Jiang (Reference Jiang1999) designed an experiment to explore this possibility by inserting a blank after the presentation of L2 primes before L1 recognition to ensure sufficient time for L2 processing. The results showed no priming from L2 to L1 in lexical decision and suggested that the absence of priming was not due to slower processing of L2 primes.

Task effects in translation priming

The most puzzling aspect of this asymmetry is its dependence on task. For example, Grainger and Frenck-Mestre (Reference Grainger and Frenck-Mestre1998) tested highly skilled English–French bilinguals in both lexical decision and semantic categorization using the masked priming paradigm. Non-cognate translation equivalents of English and French were selected to serve as primes and targets. Prime words were always presented in French (L2) and target words always in English (L1). The results showed that reaction times in semantic categorization and lexical decision tasks were similar. However, there was a robust translation priming effect in the semantic categorization task, but not in the lexical decision task. Grainger and Frenck-Mestre (Reference Grainger and Frenck-Mestre1998) proposed that this task effect followed logically from the fact that semantic categorization requires access to semantic information whereas lexical decision does not. They argued that the translation priming effects observed with the highly proficient bilinguals were mediated by semantic representations shared by translation equivalents and not by excitatory connections between distinct form representations. However, this explanation of the task effect does not account for the existence of L1–L2 priming in lexical decision. If lexical decision does not require access to semantic information, then why has L1–L2 translation priming been consistently observed in this task in several earlier studies?

The Sense Model

In a recent article, Finkbeiner et al. (Reference Finkbeiner, Forster, Nicol and Nakamura2004) attempted to explain this task effect by proposing a new model of translation priming: the Sense Model, which proposes that translation priming depends on the degree of overlap in the senses associated with the prime and target, and that a semantic categorization task accentuates this overlap.

The Sense Model starts with the assumption that most words are polysemous and that the range of senses that a word has differs across languages. Translation equivalents share one sense (typically, the dominant sense), but may differ in the remaining senses. For examples, the English word black and the Chinese word 黑 are translation equivalents, sharing the core sense (COLOR) in common; however, in English black can also be used to refer to a type of humor or a calamitous day on Wall Street, while in Chinese, 黑 can refer to those who are evil-minded or something that is secret. Thus, the senses of L1 and L2 words can extend well beyond the shared semantic sense that determines translation equivalence. Since bilinguals are normally more proficient in their L1 than L2, it follows that an L2 speaker would likely know fewer senses of L2 words compared with L1 words. This would result in a representational asymmetry between L1 and L2 at the semantic level.

According to the Sense Model, translation priming depends not only on the overlap in the semantic senses activated by the prime and the target, but on the ratio of primed to unprimed senses associated with the target. The representational asymmetry indicates that most, if not all, of the L2 senses are associated with L1, but not vice versa. In order to produce priming, it is necessary to activate a sufficient proportion of the target senses. Priming from L1 to L2 occurs because the L1 prime can activate a high proportion of the L2 target senses. However, priming from L2 to L1 is weaker because the L2 prime might activate only the dominant sense of the L1 target, and hence the ratio of primed to unprimed senses associated with the L1 target will be rather low, compared to that in the L1–L2 direction. This differential activation is assumed to affect the degree of priming in a lexical decision task. However, in a semantic categorization task, the category is assumed to act as a kind of filter which limits activation to just the category-relevant features of the target, thereby increasing the ratio of the primed senses to the unprimed senses in the case of L2–L1 priming (see Figure 1). Of course, a similar effect would not be expected in the L1–L2 direction because all of the L2 senses would be activated by the L1 prime in any case. However, in lexical decision, there is no category information, and therefore there is no filtering effect that could increase the ratio of primed to unprimed senses in the L2–L1 direction, therefore no priming is observed. We will refer to this explanation as the Category Restriction Hypothesis.

Figure 1. A schematic representation of the activation of the L1 target in a lexical decision task and a semantic categorization task. The dominant sense that determines translation equivalency is shared between L1 and L2, depicted by the dark grey circle. The non-dominant senses (peripheral senses, not shared between L1 and L2) are depicted by the white circles. Translation priming is assumed to depend on the proportion of target senses that are activated by the prime. In lexical decision, an L2 prime will only activate the dominant sense, leading to a small proportion of primed L1 target senses. But in semantic categorization, when the peripheral senses are suppressed by the category, the proportion of primed senses will be greater, leading to stronger priming.

The evidence supporting the Sense Model comes from Japanese–English bilinguals’ performance in lexical decision and semantic categorization (Finkbeiner et al., Reference Finkbeiner, Forster, Nicol and Nakamura2004). In this study, Japanese–English bilinguals were presented with items that were blocked according to semantic category and asked to decide whether the displayed item belonged to the indicated category or not. Each category consisted of the same number of exemplar and non-exemplar L1 targets. All the targets were preceded either by a masked translation prime or a control prime in L2. During the debriefing, all the participants denied any awareness of the English primes. The findings of this experiment revealed a significant L2–L1 priming effect of 19 ms. However, the same items used in a lexical decision task showed no priming. Thus, L2 primes appeared to activate their L1 translation counterparts in a semantic categorization task, but not in lexical decision. Furthermore, Finkbeiner et al. found that the same Japanese–English bilinguals showed L2–L2 repetition priming in lexical decision and concluded that L2 primes were processed effectively in lexical decision.

To date, the Sense Model is the only theory that is able to provide an account for the priming asymmetry and its dependence on task. However, there are four important aspects associated with the Sense Model that have never been empirically investigated. First, there is the assumption that the category in a semantic categorization task functions as a filter restricting the activation of the target semantics, unlike lexical decision; namely, the Category Restriction Hypothesis. Second, there is the possibility that a congruence effect might be contributing to priming in the semantic categorization task. That is, when subjects try to make a category decision on the targets, the same decision process could also have been carried out on the prime. If the prime elicits a different response to the target, a response conflict is generated. No such effect occurs on translation trials where both prime and target would elicit the same response (by definition). In fact, priming caused by implicit decision processes on the prime has been observed in the monolingual literature (Damian, Reference Damian2001; Davis, Kim and Sánchez-Casas, Reference Davis, Kim, Sánchez-Casas, Kinoshita and Lupker2003; Forster, Mohan and Hector, Reference Forster, Mohan, Hector, Kinoshita and Lupker2003; Forster, Reference Forster2004; Quinn and Kinoshita, Reference Quinn and Kinoshita2008). Third, previous studies (e.g., Gollan et al., Reference Gollan, Forster and Frost1997; Jiang, Reference Jiang1999; Finkbeiner et al., Reference Finkbeiner, Forster, Nicol and Nakamura2004) have inserted a backward mask between the prime and target in order to allow more processing time for L2 primes. If the Sense Model is correct, this should not be necessary for L2–L1 priming if a semantic categorization task is used. Finally, it is of interest to find out whether just any task that requires access to word meaning can also produce L2–L1 priming, or whether some more specific requirement is involved.

To address the above-mentioned issues, the current paper will provide experimental evidence for the Sense Model's explanation of the task effect in masked translation priming. In Experiment 1, we consider whether L2–L1 priming is restricted to exemplars of the category, as required by the Sense Model. Second, in Experiment 2 we consider whether the effect is a genuine priming effect, or is a response congruence effect. Experiment 3 investigates whether L2–L1 priming can still survive in the conventional masked priming procedure without the 150 ms backward mask that has been used in previous studies. Finally, we take up the question of whether the increase in priming produced by the categorization task is simply a consequence of the fact that in order to respond, the semantic properties of the target words must be accessed, whereas this is not necessarily the case in a lexical decision task.

In the current study, we focus on late but proficient second-language learners to further examine these assumptions and possibilities in the semantic categorization task. The bilingual population recruited in the study has similar language experiences and proficiency as those in earlier masked priming studiesFootnote ² (e.g., Gollan et al., Reference Gollan, Forster and Frost1997; Jiang, Reference Jiang1999; Finkbeiner et al., Reference Finkbeiner, Forster, Nicol and Nakamura2004).

Experiment 1: The Category Restriction Hypothesis

The Category Restriction Hypothesis implies that L2–L1 priming should be observed for exemplars, but not for non-exemplars, because the category would not restrict the semantic senses of L1 targets that are non-exemplars. If translation priming is observed for non-exemplars as well as for exemplars, then some other explanation for the task effect must be found. This issue was not specifically addressed in Finkbeiner et al. (Reference Finkbeiner, Forster, Nicol and Nakamura2004).

Therefore, the purpose of the first experiment is: (1) to confirm the results predicted by the Sense Model, testing a different group of bilinguals (Chinese–English); and (2) to confirm that L2 to L1 priming is only observed for exemplars.

Results and discussion

In analyzing the results of this experiment and all subsequent experiments, data from trials on which an error occurred were discarded and outliers were treated by setting them equal to cutoffs established at two standard deviations above or below the mean for each participant. Subjects who made errors on more than 25% of the trials would have been excluded from the analysis, but none made more than 25% errors. As shown in Table 1, mean response time for exemplars was 591 ms in the translation prime condition and 611 ms in the control condition, yielding a translation priming effect of 20 ms. A two-way ANOVA by both participant and item analysis showed that this effect for the exemplars was significant (F ₁(1,16) = 11.57, p = .004; F ₂(1,156) = 6.08, p = .015). However, the translation priming for the non-exemplars was in the wrong direction (−2 ms), and was not significant (all Fs < 1). There were no significant effects for errors.

Table 1. Mean semantic categorization times (RT) (in milliseconds) and error rates (ER) (in percentages) for both exemplars and non-exemplars from L2 to L1 (Experiment 1).

The findings of Experiment 1 confirm that a masked translation priming effect in the L2–L1 direction can be obtained in a semantic categorization task, as shown originally by Grainger and Frenck-Mestre (Reference Grainger and Frenck-Mestre1998) and Finkbeiner et al. (Reference Finkbeiner, Forster, Nicol and Nakamura2004). In addition, no priming was observed for non-exemplars, as predicted by the Category Restriction Hypothesis, which assumes that the category acts as a context that filters out the irrelevant senses in L2–L1 processing so as to increase the ratio of primed to unprimed senses. Such an effect would not occur for non-exemplars since the category is not relevant to non-exemplars. Therefore, the proportion of primed to unprimed senses in the target word remains too low to support priming, and hence there is no priming, as in a lexical decision task where there is no context effect.

Experiment 2: Controlling for response congruence

According to Dehaene, Naccache, Le Clec'H, Koechlin, Mueller et al. (Reference Dehaene, Naccache, Le Clec'H, Koechlin, Mueller and Dehaene-Lambertz1998), responses to the target stimulus in a semantic categorization experiment can be influenced by implicit categorization responses to the prime, even though the prime is masked. In a number categorization task, Dehaene et al. required subjects to decide whether a target number was bigger or smaller than 5. Prior to the target, a masked number was presented. When the masked prime and target both fell on the same side of 5 (congruent trials), responses to the target were faster than when they fell on opposite sides (incongruent trials). In addition, both electrical and hemodynamic measures of brain activity indicated that subjects were in fact covertly classifying the masked primes. In a masked translation priming experiment, such an effect could easily be misinterpreted as translation priming since, on related trials, both the prime and the target would elicit the same categorization response, but on control trials they would typically elicit a different response, since unrelated control primes would usually be non-exemplars of the category. If the target was an exemplar, the translation prime would also be an exemplar (by definition), and hence this would be a congruent situation. However, if the control prime was a non-exemplar, then this would be an incongruent situation, which would lead to slower responses. However, no such effect would occur when the target was a non-exemplar, because the translation prime and the control prime would both be non-exemplars. If the prime was in fact an exemplar, then a response congruence effect would be observed for exemplar targets, as shown by Forster et al. (Reference Forster2004, Experiment 3).

The response congruence effect implies that the primes are semantically interpreted. This conclusion has been challenged by Damian (Reference Damian2001), who pointed out that with a limited number of stimuli (the digits 1–4 and 6–9), the masked primes would have been presented many times as targets on previous trials. This pairing of a visible target with the appropriate motor response could form a low-level S–R association, so that when that target number was subsequently presented as a masked prime, an implicit response tendency would be generated, which could be responsible for the congruence effect. Damian showed that when a larger range of stimuli was used, so that no target was ever repeated as a masked prime, no congruence effect was observed. If Damian's argument is correct, then there is no problem in designing suitable items for a translation priming experiment, since the primes have never been presented before as targets. But if Dehaene et al. are correct, and primes are categorized on the basis of their semantic properties, then the items in a translation priming experiment have to be designed so that the prime has the same status with respect to the category as the target.

Experiment 2 was therefore designed to determine whether the robust priming effect from L2 to L1 in semantic categorization was partially attributable to a congruence effect. To control for a possible congruence effect, another condition was added to the design of Experiment 1, in which the control prime was a category member. Therefore, three prime conditions were created: translation (the prime is the translation of the target), congruent (the prime is an exemplar of the category) and unrelated (the prime is unrelated to the target, but not an exemplar). If there is a genuine translation effect, then the translation primes should produce faster responses than the congruent primes, and if there is no difference between the congruent and control conditions, then one can infer that there is no congruence effect. One might argue that the congruent condition is really a semantically related condition, and that any difference between this condition and the unrelated control condition might reflect semantic priming, not a congruence effect. However, semantic priming effects with masked primes are very weak with short prime durations (50 ms and below, e.g., Rastle, Davis, Marslen-Wilson and Tyler, Reference Rastle, Davis, Marslen-Wilson and Tyler2000), even when there is a very close relationship between the prime and the target. The translation prime was the best candidate for the target in terms of semantic overlap, but the congruent/exemplar prime (e.g., CAMEL) was not closely related to the target (e.g., DOG) semantically even though it belonged to the same category (e.g., ANIMAL).

Results and discussion

As shown in Table 2, for exemplars, the mean response time in the translation prime condition was 619 ms, compared with a mean of 633 ms in the congruent control condition, and a mean of 630 ms in the incongruent control condition. The effect of prime type was significant (F ₁(2,42) = 3.44, p = 0.042; F ₂(2,174) = 5.11, p = 0.007), and this effect was obviously due entirely to the translation condition. There were no significant differences in the error rates (both Fs < 1). This result indicates that there is no congruence effect (–3 ms), and hence the translation priming observed here, and in Experiment 1, cannot be attributed to a congruence effect. Finally, planned comparisons for exemplars showed that translation priming (compared with the congruent unrelated condition) was significant in both subject and item analyses (F ₁(1,21) = 5.05, p = 0.036; F ₂(1,87) = 6.32, p = 0.014).

Table 2. Mean semantic categorization times (RT) (in milliseconds) and error rates (ER) (in percentages) for both exemplars and non-exemplars from L2 to L1 (Experiment 2).

As in Experiment 1, there was no translation priming effect for non-exemplars (–4 ms, both Fs < 1). There were no significant differences in errors. Thus, the findings revealed reliable L2–L1 translation priming for exemplars but not for non-exemplars when the variable of congruence was controlled, confirming the Category Restriction Hypothesis proposed by the Sense Model.Footnote ³

The absence of any cross-language congruence effect is interesting, because such effects have been reported in several within-language experiments (e.g., Damian, Reference Damian2001; Forster et al., Reference Forster, Mohan, Hector, Kinoshita and Lupker2003; Forster, Reference Forster2004; Quinn and Kinoshita, Reference Quinn and Kinoshita2008). One possible explanation might be that the task set established by the instructions was to categorize words in a particular script, namely Chinese. Since the prime was in an alphabetic script, no attempt would have been made to categorize it, and hence no congruence effect occurred. Nevertheless, activation of the semantic properties of the prime could still have taken place. That is, the script difference between L1 and L2 can eliminate the implicit decision process (categorization) on the primes but the semantic properties of the primes are nevertheless activated.

Another explanation could be related to the relative language proficiency of L1 and L2. If participants’ L1 is more dominant, it is reasonable to think that lexical access to L1 is faster than L2. Thus, it is possible that the time allowed for processing primes was sufficient for semantic activation to benefit L1 target recognition, but not sufficient for categorization. Therefore, translation priming was observed, but no congruence effect. To test this hypothesis, one could investigate whether a congruence effect is obtained in the reverse direction (L1–L2) in semantic categorization. If the congruence effect is still absent in the L1–L2 direction, then it would appear that the script might be critically involved. This explanation would be supported if congruence effects were observed in an experiment where targets were a mixture of L1 and L2 words, so that the task list was to categorize either type of words.

One other methodological issue that requires further discussion is the inclusion of an interpolated mask between the prime and target in Experiments 1 and 2. This procedure was adopted in previous translation priming experiments (e.g., Gollan et al., Reference Gollan, Forster and Frost1997; Grainger and Frenck-Mestre, Reference Grainger and Frenck-Mestre1998; Jiang, Reference Jiang1999; Finkbeiner et al., Reference Finkbeiner, Forster, Nicol and Nakamura2004). This issue was first raised by Gollan et al. (Reference Gollan, Forster and Frost1997), who argued that the absence of L2–L1 priming could be due to the relative speeds of processing of the two languages, the idea being that the L1 target word might generate a response before the L2 prime processing finished. In order to guard against this possibility, an interpolated mask was introduced, even though Jiang (Reference Jiang1999) failed to find L2–L1 priming in lexical decision with such a mask. Clearly, extra time for L2 processing is not relevant to priming in lexical decision. However, it is unclear whether the interpolated mask is required in semantic categorization. Experiment 3 was designed to investigate this issue.

Experiment 3: The role of the backward mask in priming

In earlier studies, it was thought that maybe the reason no L2–L1 priming was observed in lexical decision was because there was insufficient time for the L2 prime to be processed. However, inserting a backward mask between the prime and the target had no effect (Jiang, Reference Jiang1999). Subsequent work with a semantic categorization task revealed L2–L1 priming, but the practice of including a backward mask was continued, making it unclear whether this procedure was really necessary. From the point of view of the Sense Model, it is the category that is responsible for priming, and therefore the backward mask should not be necessary.

Results and discussion

As shown in Table 3, mean response times were 513 ms in the translation prime condition, which was significantly faster than the unrelated condition (525 ms) by both subject and item analyses (F ₁(1,21) = 17.27, p <.01; F ₂(1,87) = 6.25, p = 0.01). When compared to the congruent condition (518 ms), the translation effect was significant in the subject analysis, but not in the item analysis (F ₁(1,21) = 4.47, p = 0.047; F ₂(1,87) = 1.01, p = 0.317). A separate analysis of the congruent and unrelated conditions showed no congruence effect from L2 to L1 (F ₁(1,21) = 2.85, p = 0.11; F ₂(1,87) = 1.39, p = 0.24). In addition, the mean error rate was 3.3% and did not differ significantly between conditions for exemplars. The non-exemplars did not show any advantage for translation primes (562 ms) compared with the exemplar controls (564 ms), nor unrelated primes (561 ms) (F ₁(2,42) = 0.21, p = 0.81, F ₂(2,174) = 0.46, p = 0.63). The mean error rate was 4.2% and did not differ significantly between conditions for non-exemplars.

Table 3. Mean semantic categorization times (RT) (in milliseconds) and error rates (ER) (in percentages) for both exemplars and non-exemplars from L2 to L1 without the backward mask (Experiment 3).

These results confirm the findings of Experiment 2 for both exemplars and non-exemplars, further demonstrating that the L2–L1 translation effect is not simply a response congruence effect.Footnote ⁴ Equally importantly, the results provide evidence that the backward mask is not a critical factor in L2–L1 priming. Again, the Sense Model is supported. Also, it should be noted that it seems likely that processing of the prime continues even after it has been replaced by the target, otherwise we would have to conclude that the meaning of the prime can be extracted within 50 ms, whereas evidence from eye-tracking and ERP experiments suggests that 150 ms is a more likely estimate (Sereno and Rayner, Reference Sereno and Rayner2003). Thus the relevant variable is probably the prime duration (which should be long enough to initiate lexical access of the prime), not the interval of time between the prime and the target.

One further aspect of the results is worth attention. That is, without the backward mask, the mean reaction times across three conditions for both exemplars and non-exemplars decreased by 108 ms for exemplars and 112 ms for non-exemplars, respectively, compared with the results for Experiment 2. Given that Experiments 2 and 3 were run with a similar group of participants and the same stimuli, it seems that the effect of the backward mask is to slow down the response to the targets. How this interference might have affected the processing of L2 primes is unclear.

Experiment 4: Categorizing with an ad hoc category

An obvious difference between semantic categorization and lexical decision is that semantic access is essential for semantic categorization, but not for lexical decision. It has been argued that this is the reason why the semantic task is more sensitive to translation priming (Grainger and Frenck-Mestre, Reference Grainger and Frenck-Mestre1998; de Groot, Reference de Groot and Cook2002). This argument was also adopted in long-term priming studies. For example, Zeelenberg and Pecher (Reference Zeelenberg and Pecher2003) concluded that long-term cross-language repetition priming depended on tasks that required access to conceptual knowledge (e.g., semantic categorization). Obviously, it is important to pursue the same question in masked translation priming. If a conceptual task is critical for producing L2–L1 priming, then any categorization task that requires conceptual level processing should increase the effectiveness of masked L2 primes. The following experiment was designed to test this possibility.

One way to achieve this is to ask participants to categorize words according to the physical properties of the objects they refer to, e.g., “Is it bigger than a brick?”. Such a category is completely ad hoc in the sense that it is an arbitrary classification that would never have been learned by any speaker of the language. In order to answer this question for each word, participants need to first access the lexical entry to determine the referent and then recover the conceptual properties associated with that entry. Only then can they make a decision about its relative size. If this type of ad hoc categorization can produce L2–L1 priming, the implication is that explicit conceptual level processing is all that is required to produce L2–L1 priming. If there is no observed priming, then one can conclude that the type of category is the key to L2–L1 priming. This is compatible with the Sense Model, since the ad hoc “bigger than a brick” category is unlikely to have any filtering effect.

Results and discussion

The mean classification times and error rates for English and Chinese targets are shown in Table 4. There was a significant effect of 34 ms for L1–L2 priming (F ₁(1,14) = 5.65, p = 0.035; F ₂(1, 48) = 10.02, p = 0.003). However, there was no priming in the reverse L2–L1 direction (–3 ms). There were no significant effects for errors.Footnote ⁶

Table 4. Mean categorization times (RT) (in milliseconds) and error rates (ER) (in percentages) for all items (English targets and Chinese targets) in Experiment 4.

What is crucial in this experiment is failing to find L2–L1 priming in a semantic categorization task that uses an ad hoc category. It therefore seems clear that the category plays a more critical role in translation priming than merely forcing participants to take meaning into account. The fact that a significant L1–L2 priming effect was found also shows that the failure of L2–L1 priming cannot be attributed to excessive noise in the data produced by the nature of the task. It should be noted that this result is still consistent with the Sense Model, since it is assumed that most or all of the L2 senses would be activated by the L1 translation prime.