Introduction
Simultaneous interpreting (SI) is a type of interpreting which requires an interpreter to continuously receive information from a source language (e.g., English) while producing the translation of a previously received message in the target language (e.g., Chinese). The lag time between receiving the speaker's speech and producing the interpretation of that segment is usually less than 3 seconds (Christoffels & De Groot, Reference Christoffels, de Groot, Kroll and de Groot2005, p. 457). This means that the processing of perception, comprehension, memorisation, translation, and production has to be fulfilled within this short lag period. One explanation that has been proposed for this efficient multi-tasking of interpreters in real time is that all the concurrent processes can be conducted in parallel by sharing the limited available cognitive capacity (Gile, Reference Gile, Danks, Shreve, Fountain and McBeath1997; Christoffels & De Groot, Reference Christoffels, de Groot, Kroll and de Groot2005). According to this view, the multi-tasking performance is not impaired unless the additive capacity required by all the concurrent processes exceeds the limited capacity (Gile, Reference Gile2009).
However, this capacity-sharing assumption has been challenged. A number of studies have shown that when fulfilling lemma, phonological word-form (Ferreira & Pashler, Reference Ferreira and Pashler2002), and phoneme selection (Cook & Meyer, Reference Cook and Meyer2008) during word production, speakers are subject to the central processing bottleneck, and can conduct only one task, and specifically its central stage, at a time. Furthermore, Rohrer and Pashler (Reference Rohrer and Pashler2003) found that memory retention as well as memory recall can be impaired and postponed by performing a concurrent non-linguistic task which is also subject to the same central processing mechanism. In other words, language production can also impact on memory retention and memory recall.
If interpreters are also subject to the bottleneck during language production, then the continuous interpretation production required in SI should impair memory retention for the new incoming information and postpone the recall of the previous message. The performance of interpreters should therefore be adversely affected. Contrary to this inference, however, professional interpreters are typically able to offer high-quality interpretation covering almost all the key information delivered by the speaker with few pauses (Wang & Li, Reference Wang and Li2015). When considering the accuracy and fluency of interpretation from professionals, and existing research suggesting that certain language processes in simultaneous interpreters become automatic after extensive practice (Campbell & Wakim, Reference Campbell and Wakim2007; Riccardi, Reference Riccardi2005; Šeňková, Meylaerts, Hertog, Szmalec & Duyck, Reference Šeňková, Meylaerts, Hertog, Szmalec and Duyck2014), it appears possible that interpreters may not encounter structural limitations such as the bottleneck during word production, and therefore, do not suffer its consequent impairments. Simultaneous interpreters may develop a way of ‘avoiding’ the bottleneck, as a consequence of their constant practice in language production under conditions of time pressure. Some support for this assumption comes from findings showing that extensive practice can reduce or even eliminate the interference caused by performing a dual task (Pashler & Baylis, Reference Pashler and Baylis1991a, Reference Pashler and Baylis1991b).
Against this background, this study aims to, firstly, explore whether professional simultaneous interpreters are subject to the bottleneck during word production, and secondly, investigate whether there is a difference between professional simultaneous interpreters, proficient bilinguals and monolinguals in the duration of the word production bottleneck. The investigation is carried out by closely imitating Experiment 1 of Ferreira and Pashler (Reference Ferreira and Pashler2002). Professional simultaneous interpreters, proficient bilinguals and monolinguals performed a dual task consisting of a picture naming task in sentence context (Task 1) and a pitch tone discrimination task (Task 2). Bilinguals were expected to have a longer bottleneck duration than matched monolinguals in word production because of the need to select the intended word across their two languages (Dijkstra & Van Heuven, Reference Dijkstra and Van Heuven2002). Interpreters were not expected to be subject to the bottleneck during word production because of their simultaneous interpreting experience. The word production bottleneck stage is assumed to be reduced to a minimum after extensive practice of speeded processing in multi-tasks.
Processing stages of word production by monolinguals and bilinguals
Producing a word is more complex than is often assumed, especially for people who can speak more than one language. The models that have been developed to account for bilingual lexical access assume a similar sequence of lexical access as models of monolingual lexical access, except that two languages are activated during the lexical stage in bilinguals (Costa, Miozzo & Caramazza, Reference Costa, Miozzo and Caramazza1999; Costa & Caramazza, Reference Costa and Caramazza1999; Green, Reference Green1998). The controversy of bilingual word production models centres mainly on whether the selection of a word takes place within the target language (Costa et al., Reference Costa, Miozzo and Caramazza1999; Costa & Caramazza, Reference Costa and Caramazza1999) or between languages (Dijkstra & Van Heuven, Reference Dijkstra, Van Heuven, Grainger and Jacobs1998; Green, Reference Green1986, Reference Green1998; Grosjean, Reference Grosjean, Groot and Kroll1997). The general word production model (Ferreira & Pashler, Reference Ferreira and Pashler2002) accounting for both bilinguals and monolinguals can be illustrated as shown in Figure 1.
Figure 1. Outline of the general word production model. Information flows from top to bottom. The black lines present the word production process of monolinguals while both the black and red lines indicate the language production process of bilingual. Adapted from Ferreira, V. S., & Pashler, H. (Reference Ferreira and Pashler2002). Central bottleneck influences on the processing stages of word production. Journal of Experimental Psychology: Learning, Memory, and Cognition, 28(6), 1187–1199.
Word production starts with the conceptual (or semantic) stage, where the meaning and conceptual features of the intended word are activated. Activation then spreads to the lemma stage, where the syntactic features of the word are encoded. During the lemma stage, a ‘group’ of unintended conceptually related lemmas (or words) are activated along with the target lemma. Thus, lemma selection is necessary to select the intended lexically specific lemma. Activation then flows to the lexeme (or phonological word-form) stage. Phonological word-form nodes, which represent whole sound segments, are activated. Phonological word-form selection for the appropriate phonological word-form node is also required. The last stage before articulation is the phoneme stage. The phonemes (or sounds) that constitute the intended word receive activation, and therefore, selection is required for the right order of these phonemes. This selection is referred to as phoneme selection. After all the above procedures have been successfully completed, motor action begins and articulation commences.
Each of the selection stages involved in word production (lemma, phonological word-form, and phoneme selection) has been shown to postpone another task in the dual task condition, and it has been suggested that they are subject to the central processing bottleneck for both monolingual and bilingual speakers (Cook & Meyer, Reference Cook and Meyer2008; Declerck & Kormos, Reference Declerck and Kormos2012; Ferreira & Pashler, Reference Ferreira and Pashler2002). If the lexical selection of bilinguals is across languages (Hermans, Reference Hermans2004), the intended word needs to be selected from a group of related lemmas as well as their translation equivalents in the other language. In this case, bilinguals are expected to take longer to fulfil lemma selection, and have a longer bottleneck duration in word production compared to monolinguals. The methodology used to investigate the central processing bottleneck is discussed in more detail in the following sections.
The central processing bottleneck and dual task performance
Dual task performance has been widely explored in research on divided attention, attempting to understand the limitations of human processing capacity. A typical dual-task experiment involves two tasks with independent speeded responses that are separated by a variable period of time, termed the stimulus onset asynchrony (SOA). When decreasing the SOA, performance on the first task usually remains the same, while the response to the second task is prolonged (see Fig. 2). This phenomenon is known as the psychological refractory period (PRP).
Figure 2. The chart illustrates the central processing bottleneck model. Two choice reaction time tasks are involved, indicated as Task 1 and Task 2. The white boxes represent the processing stage that can overlap with other stages while the black boxes represent the bottleneck stage that cannot be conducted with another bottleneck stage, but only in serial. The reaction time of Task 2 is postponed when decreasing the SOA but not the reaction time of Task 1.
The PRP effect has been shown to be robust across various non-linguistic behavioural tasks (Levy, Pashler & Boer, Reference Levy, Pashler and Boer2006; Pashler, Reference Pashler1984; Pashler, Reference Pashler1990) with different types of response modalities (for a review, see Pashler, Reference Pashler1994). A large number of studies converge on the finding that the PRP effect is due to the limitation of the central processing mechanism, which can only be occupied by one task at a time (Levy & Pashler, Reference Levy and Pashler2001; Pashler, Reference Pashler1984; Ruthruff, Pashler & Hazeltine, Reference Ruthruff, Pashler and Hazeltine2003; but see Tombu & Jolicœur, Reference Tombu and Jolicœur2003). That is, the particular processing stage in Task 2 (e.g., the response selection stage) cannot commence until the particular stage in Task 1 (e.g., the response selection stage) is completed at short SOA, even if its previous stage (e.g., the perceptual stage) is finished. This processing bottleneck is therefore suggested as a structural limitation, rather than a processing strategy (Pashler, Reference Pashler2016).
The assumption that the bottleneck is located at the response selection and decision-making stage was first proposed by Welford (Reference Welford1952) and has been tested and strongly supported by Pashler and Johnston (Reference Pashler and Johnston1989). The logic of distinguishing the location of the bottleneck is set out in Figure 3. When increasing the pre-bottleneck stage in Task 2, an under-additive interaction can be expected because the prolonged part should be ‘absorbed’ and reaction time to Task 2 will not be prolonged (see Fig. 3, Manipulating Perceptual Stage). When increasing the bottleneck stage in Task 1, the additive effect of Task 1 also propagates to Task 2 because the bottleneck stage of Task 2 cannot begin until the bottleneck stage in Task 1 is completed. Therefore, the reaction time to Task 1 (RT1) and Task 2 (RT2) should increase to the same extent (see Fig. 3, Manipulating Response Selection Stage). When manipulating the post-bottleneck stage, the additive effect of one task cannot impact on the other one (see Fig. 3, Manipulating Response Execution Stage). The predictions are distinctive if the bottleneck is located at different stages (e.g., perceptual stage, response selection stage or response execution stage) when manipulating the pre-bottleneck, bottleneck and post-bottleneck stages of Task 1 and Task 2.
Figure 3. The predictions of the response selection bottleneck model when manipulating each processing stage: perceptual stage where information (e.g. picture) is perceiving; response selection stage where the compatible stimulus-response binding is selected, and response execution stage where the selected response is executed. The black boxes represent the bottleneck stage which cannot be conducted with another bottleneck stage, while the white boxes indicate the stages that can overlap with other stages.
The results of Pashler and Johnston (Reference Pashler and Johnston1989) showed that the reaction time to Task 1 and Task 2 is prolonged to the same extent when increasing the perceptual stage or response selection stage in Task 1. Both RT1 and RT2 remain the same when increasing the duration of the perceptual stage in Task 2, and only RT2 is increased when the duration of the response selection stage in Task 2 is prolonged. Only RT1 or RT2 is impacted when manipulating the response execution stage of Task 1 and Task 2, respectively. These results are consistent with the predictions of the assumption that the response selection stage constitutes the bottleneck and that the response selection stages from two tasks cannot be conducted in parallel but must be carried out serially.
The dual task experiment
This study aimed to explore whether professional simultaneous interpreters are subject to the central processing bottleneck during word production, and to compare their performance with untrained proficient bilinguals and monolinguals. The experiment closely imitated Experiment 1 of Ferreira and Pashler (Reference Ferreira and Pashler2002), in order to explore the replicability of the study and to determine whether the results are only applicable to certain types of subjects. Participants were required to carry out a dual task with speeded responses: Task 1 was a picture naming task in sentence context (in English), requiring a verbal response, while Task 2 was a non-linguistic pitch tone discrimination task requiring a button-press response. The sentence context in Task 1 was presented visually, one word at a time preceding the line-drawn picture, to guarantee that any differences in performance between the three groups were not the consequence of more efficient auditory processing among interpreters as a result of their interpreting experience.
Two factors were manipulated in Task 1: sentence constraint and word frequency. The manipulation of sentence constraint (medium and low) impacts on lemma selection, such that lemma selection for the pictures is eased by increasing the predictability of picture names in accordance with the sentence information (Levelt, Roelofs & Meyer, Reference Levelt, Roelofs and Meyer1999; Schachter, Rauscher, Christenfeld & Crone, Reference Schachter, Rauscher, Christenfeld and Crone1994). The manipulation of word frequency (high and low) appears to influence the phonological word-form selection. Phonological word-form selection has been shown to be easier when the words are more frequently used (Jescheniak & Levelt, Reference Jescheniak and Levelt1994; but see Navarrete, Basagni, Alario & Costa, Reference Navarrete, Basagni, Alario and Costa2006). Task 2 was a pitch tone discrimination task (high, medium, and low) separated from Task 1 by varying SOAs (50, 150, and 900 ms).
If there are bottleneck stages in word production, the responses to Task 2 will be prolonged as the SOA decreases since the critical stage in Task 2 has to wait until the bottleneck stage in Task 1 is finished at short SOA. If the manipulations of sentence constraint and word frequency in Task 1, which can prolong lemma and phonological word-form selections respectively during word production, propagate their effect to Task 2, it suggests that lemma and phonological word-form selections are subject to the central processing bottleneck. Furthermore, if a group has a comparatively longer bottleneck duration in word production, the commencement of the bottleneck stage in Task 2 should be delayed more than for other groups at short SOA, but not at long SOA where the responses to Task 1 should have been provided before Task 2 begins.
Method
Subjects
To determine the total number of subjects required, an a priori power analysis of repeated measure ANOVA was conducted by using GPower 3.1. At least 66 subjects were required to achieve a power of 0.95 for a medium effect of 0.25 (Cohen, Reference Cohen1969, p.348) at a significance threshold level of 0.05.
Thirty participants in each of three distinct participant groups were recruited for this experiment as follows:
Thirty professional simultaneous interpreters, with at least five years’ working experience in SI (see Table 1), were invited to participate in the study as Group I. Interpreters were recruited via email invitation from the researcher or via the networks of the participants and lecturers in SI at Macquarie University. Potential participants were identified by accessing information on interpreting and translation community websites, specifically the websites of the International Association of Conference Interpreters (AIIC), and the National Accreditation Authority for Translators and Interpreters (NAATI). Because of the difficulty of recruiting professional simultaneous interpreters, the criterion of having English as their first acquired language or dominant language was not stipulated. More than two thirds of the interpreters who participated in this study had English as their second language, and therefore were performing the picture-naming task in their second language. The other working languages of interpreters were: (a) Asian languages: Chinese (8), Japanese (3), and Korean (1); (b) European languages: Spanish (7), French (7), German (3), and Portuguese (1).
Table 1. Interpreter, bilingual and monolingual participants’ age and interpreters’ simultaneous interpreting experience (standard deviation). All the numbers are calculated in years
* “SI experience” indicates the number of years of experience specifically in doing simultaneous interpreting.
**“First language” mentioned here does not specifically represent English for interpreters and bilinguals, since most of them have English as their second acquired language.
Thirty proficient bilinguals who use English as their working language were recruited as Group II. Bilinguals in Group II included translators (who do not do interpreting work), lecturers and researchers in universities located in Sydney, and students at Macquarie University.
Thirty monolinguals who only speak English were recruited as Group III. Some monolingual participants reported very limited exposure to a second language in school. However, they indicated they almost never use this language. Participants for Group II and Group III were recruited via email, notices around campus as well as through the social networks of participants. All the participants in the study were resident in Australia.
Age range was not stipulated in this study due to the difficulty of recruiting professional interpreters in certain age ranges. The age range of the interpreter group in this study exceeded 40 years. To avoid age effects in comparing the results of the three groups, participants in Group II and Group III were selected according to the age (within ± 2 years deviation) and gender of each interpreter (see Table 1). Further language-specific selection criteria were required for Group II; participants in this group were matched to the interpreters by language profile: each bilingual participant was matched to an interpreter with the same two main languages and sharing the same dominant language. Only two interpreters and two bilinguals were simultaneous bilinguals, where the age of becoming fluent in the first and second language is within a 5 year period.
As a consequence of the recruitment difficulties due to the very small numbers of suitable participants, some exceptions needed to be made as follows. In controlling for age, one interpreter-monolingual pair and one interpreter-bilingual pair were matched within ± 5 years, and one interpreter-bilingual pair was matched within ± 10 years (with the bilingual the younger of the two); in controlling for gender, one interpreter-monolingual pair was not matched; and, in controlling for dominant language, two English–German interpreter-bilingual pairs were mixed matched.
Eleven subjects were replaced for the following reasons: Eight participants (1 interpreter, 6 bilinguals and 1 monolingual) were replaced as a consequence of having to discard more than 30% of the total number of trials in the experiment. The possible reason for the high number of discarded trials for bilinguals might be due to cultural bias in the stimuli (Jared, Poh & Paivio, Reference Jared, Poh and Paivio2013) or disadvantage for the bilinguals on lexical selection in the L2 in the dual task (Declerck & Kormos, Reference Declerck and Kormos2012). One monolingual participant was replaced because of un-speeded responses in Task 1 (mean RT exceeded 3000 ms). Two monolingual participants were replaced after reporting occasionally using a second language in short conversations.
Selection criteria
All ninety participants were asked to fill out two questionnaires in addition to the main study. The questionnaires were the Edinburgh Handedness Inventory (MQ) to measure handedness, and the Language Experience and Proficiency Questionnaire (LEAP-Q; Marian, Blumenfeld & Kaushanskaya, Reference Marian, Blumenfeld and Kaushanskaya2007) to measure language background. Participants who were right-handed, and had no history of speech, language or hearing deficits, or any other neurological deficits, were invited to participate in this study. No significant difference was obtained between interpreters and bilinguals in self-evaluated language proficiency in both languages in T-tests (p > .05; see means in Table 1).
Apparatus and stimuli
The stimuli were delivered and responses collected using Presentation® software (developed by Neurobehavioral Systems) running on a Windows 10 system on a MacBook Air 13.3" personal computer. Auditory stimuli were presented and verbal responses were collected using a microphone headset connected to the computer via a USB port. Manual responses were collected using a Microsoft 600 Wired keyboard, connected to the computer via a USB port. Verbal responses were recorded and saved for accuracy checking.
The sentence constraint and picture materials in Task 1 were based on Experiment 2 in Griffin and Bock (Reference Griffin and Bock1998). American English materials were used for this study instead of Australian English materials to avoid a potential advantage for monolingual participants who were born and grew up in Australia, in contrast to bilingual and interpreter participants who did not. Most of the pictures were taken from Snodgrass and Vanderwart (Reference Snodgrass and Vanderwart1980), and the remainder from Griffin and Bock (Reference Griffin and Bock1998) and Ferreira and Pashler (Reference Ferreira and Pashler2002). Pictures were black-and-white line-drawings, and agreement on the names for the pictures was tested with a large population (Snodgrass & Vanderwart, Reference Snodgrass and Vanderwart1980). The names of high (e.g., DRESS) and low frequency (e.g., DRUM) pictures were carefully matched in terms of naming agreement, number of syllables, number of phonemes, and initial phoneme of the names (for details, see Griffin & Bock, Reference Griffin and Bock1998, p. 319). The mean frequency of the high and low frequency target names, respectively, was 110 and 15 occurrences per million words, as measured in the CELEX database (Baayen, Piepenbrock & Van Rijn, Reference Baayen, Piepenbrock and van Rijn1993).
Medium-constraint and low-constraint English sentences were used in this study, avoiding the possible artificially shorter bottleneck duration for interpreters in high-constraint sentences because of their skill in anticipation compared to the other two groups. The probability of the sentence constraint was based on first response (the name was provided as the first response) and overall response (the name was provided within the three required responses without order required). The medium constraint sentence (e.g., the sentence “The princess wore a. . .” followed by a picture of DRESS) probability for high frequency and low frequency pictures, according to Griffin and Bock (Reference Griffin and Bock1998), is .04 vs .42 and .68 vs .06 as first response and overall response, respectively. The low constraint sentence (e.g., the sentence “Peter saw a drawing of a . . .” followed by a picture of DRUM) does not constrain for a particular word, and can match all the pictures.
The pitch tones in Task 2 were 180 Hz, 500 Hz, or 1200 Hz. The subject viewed the display from a distance of approximately 65 cm. The display was presented in white on a black background and was viewed under normal room illumination. All participants were tested individually in a comfortable and quiet environment.
Design
The main experiment was divided into 2 blocks of 60 trials each to explore the possible blocking impact on the bottleneck stage in word production. The order of blocks was counterbalanced within each group. Three independent variables were manipulated in the experiment: sentence constraint (medium and low), word frequency (high and low), and SOA (50, 150, and 900 ms). These variables were counterbalanced across items and manipulated within each group. Each picture was presented once in each block, with a total of 60 pictures presented twice in the whole experiment. Cloze constraint was manipulated within items and counterbalanced across blocks (Block 1 and 2). None of the items that were presented followed the same sentence constraint across blocks. SOA was also manipulated within items and counterbalanced across each group. The picture items in each experimental condition were rotated across each group. Each subject saw 5 picture stimuli in each experimental condition, and all 30 participants in each group saw each picture in each experimental condition 10 times. The sequence of trials in each block was randomised independently, and in each trial the pitch tone was randomly selected.
Procedure
Instructions to subjects were given on the computer screen before each block. If subjects had questions about the instructions, an explanation was provided by the researcher. The instructions (with the exception of the instructions for the first practice block) emphasised that the subject should provide responses to both Task 1 and Task 2 as quickly as possible while maintaining accuracy. The instructions also informed participants that verbal responses would be recorded for later error detection.
The experiment began with 2 practice blocks, followed by 2 main blocks. Each practice block had 30 trials and each main block had 60 trials. Practice Block 1 was a single tone discrimination task. Each of the 3 pitch tones was presented 10 times in random order. Practice Block 2 was a dual task in the same paradigm as the main blocks, including a picture-naming task and a tone-discrimination task. None of the testing materials presented in the practice blocks were included in the main blocks. After every 15 or 20 trials in Practice Block 2 and the main blocks, respectively, a pseudo-random sentence-repeat-section appeared, asking subjects to repeat or provide the main idea of the last sentence that had been presented on the screen. This measurement was included to ensure the subjects actually read the sentences. After repeating the sentences, subjects were told to press the “spacebar” to continue when they were ready.
The sequence of each trial is presented in Figure 4. The screen was blanked for 500 ms, before the presentation of a fixation point (a plus sign) at the centre of the screen. The fixation point disappeared after 1000 ms, followed by a blanked screen as the foreperiod (500 ms). After the foreperiod had elapsed, the constraint sentence was presented. The sentence was displayed one word at a time in 12-point Times New Roman at the centre of the screen. Each word lasted for 285 ms and was presented in a rapid serial visual presentation (RSVP) paradigm. After the offset of the final word of the sentence, the picture stimulus, which required subjects to name it in English as quickly as possible, was immediately presented and remained on the screen until both responses had been detected. The picture stimulus was separated by the pre-determined SOA from the tone discrimination task. After 50, 150, or 900 ms of picture stimulus onset, a randomly selected pitch tone was presented and lasted for 285 ms. The pitch tone was varied among the trials and there was an equal number of each pitch tone presented within each block. A response to the pitch tone had to be provided by pressing the Red, Yellow or Green colour key for the low, medium and high pitch tone, respectively, with the right hand.
Figure 4. The sequence of each trial in the dual task practice and main blocks.
After two responses had been detected, feedback was provided in accordance with the responses. Feedback was displayed for 1500 ms on the screen. If a verbal response was given before the correct manual response, the message “Correct!” was given. If an incorrect manual response was offered after a verbal response, the feedback “Incorrect!” was presented. If a manual response was given before the verbal response, or two verbal responses, or two manual responses were provided, then a warning message “You have the response in the wrong order!” or “Please only respond verbally once!”, or “Please only press button once!” was presented, respectively. The “Correct!” feedback was not provided in the main blocks. After the trial was completed, the next trial was not presented until the intertrial interval (1.3 s) had elapsed.
At the end of each (main) block, feedback on the speed of verbal and manual responses and the accuracy of manual responses for the preceding block was provided. In addition, information on the number of block(s) that had been finished was provided. Subjects were encouraged to rest during this period and continued by pressing the spacebar to move on to the next block when they felt ready.
Analysis and results
Prior to analysis, reaction times faster than 200 ms or slower than 2500 ms in Task 1, or under 200 ms or exceeding 3000 ms in Task 2, were discarded as deviations. Trials with undetected verbal responses, or with verbal responses involving stammering, noise before the response, or reversed response order were discarded (a total of 2.59 % of trials). Trials in which synonyms of the intended picture names were provided were counted as incorrect verbal responses and were discarded from the data analysis.
In this section, the results of the main analyses are reported. First, the differences between the three groups in error rates are analysed, using one-way analysis of variance (ANOVA), with post-hoc comparisons between groups. If statistically significant differences between groups are found across blocks, then there is a possibility that the RTs for Task 1 and Task 2 might not reflect the true performance of three groups. The main analysis focuses on RTs for Task 1 and Task 2. Four-way 2×2×2×3 analyses of variance (ANOVAs) with within-subjects variables of Block (block 1 vs block 2), Constraint (medium vs low constraint sentence), Frequency (high vs low frequency word), and SOA (50, 150 or 900 ms), and between-subjects factors of Group (interpreter group vs bilingual group vs monolingual group) were used to analyse the results for Task 1 and Task 2, unless noted otherwise in the sections below. Post-hoc comparisons between groups were performed where relevant. The significance level of all reported effects was set at .05, with a Bonferroni adjustment of significance levels for post hoc pairwise comparisons.
If lemma and phonological word-form selection are subject to the central processing bottleneck, then the manipulation of sentence constraint and word frequency in Task 1 will be propagated to Task 2. In this case, the expected significant constraint and frequency effect should be found in Task 2. If interpreters are not subject to the central processing bottleneck during word production, a non-significant SOA effect should be evident for interpreters in Task 2. Furthermore, a significant interaction between GROUP and SOA would provide evidence that there is a difference in the bottleneck duration in word production for the three groups.
Error rates
One-way ANOVAs were conducted to analyse the differences between the three groups in error rates (see Table 2) in each block for Task 1 and Task 2. Only one error rate reached significance among the three groups. the number of errors in naming the pictures in Block 2 was significantly different between the three groups, F (2, 87) = 7.517, p = .001, MSe = 0.04. Post-hoc tests showed that the interpreters made significantly more errors than the monolinguals, and marginally significantly more errors than the bilinguals. However, no significant difference in error rates was obtained between the bilingual and monolingual groups.
Table 2. Mean error rates for three groups in Task 1 and Task 2.
Note: “ICR” represents incorrect responses rate.
Reaction times
Task 1
Reaction times (RT) at SOA 900 ms in Task 1 were discarded from the data analysis, because a significant SOA effect was obtained, F (2, 174) = 63.086, p < .001. MSe = 161,441.784, suggesting that the RT was prolonged with increasing SOA due to subjects adopting the grouping strategy (the strategy of holding back the execution of the first response until the response selection of the second task had been finished, with the two responses then produced in rapid succession or at the same time; see Fig. 5). Thus, four-way 2×2×2×2 ANOVA was used to analyse the results for Task 1. The ANOVA results of the three groups for Task 1 are presented in Table 3.
Figure 5. The performance in the picture naming task (Task 1) with three SOA.
Table 3. Analyses of variance results (Reaction time) for Task 1: Picture naming task.
Note: * p < .05; ** p < .01; *** p < .001.
All significant results for Task 1 are reported in this Table.
The group difference was significant, showing that bilinguals (1142 ms) were slower than monolinguals (979 ms) and interpreters (1061 ms). However, post hoc comparisons showed that only the difference between bilinguals and monolinguals reached significance (see Fig. 6). The picture-naming RT was significantly faster 1) in Block 2 than in Block 1 (178 ms faster); 2) in medium-constraint sentences than in low-constraint sentences (24 ms faster); 3) when word frequency was high than when it was low (74 ms difference); and 4) at SOA 50 ms than at SOA 150 ms (17 ms difference), as shown by the significant main effects of block, constraint, frequency, and SOA, respectively. The word frequency effect was significantly reduced after a block of practice. The significant interaction of block and word frequency showed that high-frequency words were named 99 ms faster than low-frequency words in Block 1, but this difference was reduced to 49 ms in Block 2. High-frequency words and low-frequency words were 153 ms and 203 ms faster in Block 2 than when in Block 1, respectively. Consistent with previous studies showing that the word frequency effect diminishes as the predictability of sentence constraint increases (Griffin & Bock, Reference Griffin and Bock1998), the interaction of sentence constraint and word frequency was significant. A simple effect analysis showed that the word frequency effect was significant in both the medium- and low-constraint conditions (54 ms and 94 ms frequency difference, respectively). However, the constraint effect was only significant when the word frequency was low (44 ms constraint difference) but not when it was high (5 ms difference). The interaction of block and constraint was significant. This interaction revealed that the constraint effect was no longer significant in Block 2 (11 ms constraint difference), but only in Block 1 (38 ms difference).
Figure 6. The performance of the three groups in the picture naming task (Task 1).
The interaction of group and block was significant. There was a significant difference between the bilinguals and monolinguals in Block 1 (1243 ms and 1037 ms, respectively) but not in Block 2 (1041 ms and 920 ms, respectively); but no significant differences were found between interpreters and the other two groups in any blocks (the RT of interpreters in Block 1 and Block 2 was 1169 ms and 953 ms, respectively). The significant interaction of group and word frequency showed bilinguals were significantly slower than monolinguals under the low word frequency condition (191 ms) and marginally significantly slower under the high word frequency condition (135 ms). The three-way interaction between group, sentence constraint and SOA was significant. Simple effect analysis showed that the SOA effect was only significant in the medium-constraint condition in the bilingual group. The RT at SOA 150 ms was 55 ms faster than the RT at SOA 50 ms. The marginally significant four-way interaction between group, block, word frequency, and SOA demonstrated that the SOA effect reached significance in the bilingual group in Block 2 in the low word frequency condition. The RT at SOA 150 ms was 58 ms faster than the RT at SOA 50 ms. Furthermore, the frequency effects were all significant except under the condition of SOA 150 ms in Block 2 (17 ms difference) in monolinguals.
Task 2
Table 4 presents the results of the analysis for Task 2. Consistent with the typical PRP effect, the significant main effect of SOA showed that RT decreased as SOA increased (425 ms difference between the shortest and the longest SOA; see Fig. 7). RT2 was postponed at short SOAs because its response selection could not begin until the bottleneck stage in Task 1 had been completed. The significant main effects of sentence constraint and word frequency and any interactions with either or both these factors suggest that the manipulations in Task 1 had been propagated to RT2.
Table 4. Analyses of variance results (Reaction time) for Task 2: Tone discrimination task.
Note: * p < .05; ** p < .01; *** p < .001.
All significant results for Task 2 are reported in this Table.
Figure 7. The performance of the three groups in the tone discrimination task (Task 2).
The main effect of group was significant. Post-hoc results showed that the difference between monolinguals (1330 ms) and bilinguals (1618 ms), and between interpreters (1380 ms) and bilinguals was significant, reflecting a slower RT for bilinguals. However, the three groups seem to have a similar bottleneck stage during language production, since the interaction between group and SOA was not significant (see Fig. 8). To further investigate whether there is a difference between the three groups in the duration of the language production bottleneck, a three-way 2×2×2 repeated ANOVA with the factors of Block, Constraint and Frequency as within-subjects variables, and Group as between-subjects variable was conducted, comparing the RT difference between the longest and shortest SOA. No significant difference was found between groups (F (2, 87) = 2.664; p = .075). Post-hoc comparison tests showed no differences between any two-way combination of groups, suggesting that Task 2 was postponed by Task 1 to a similar extent at the shortest SOA for all three groups. These results suggest that interpreters, like bilinguals and monolinguals, cannot begin the second task during word production. We can therefore conclude that SI experience does not help to ease the impact of the bottleneck during language production. Furthermore, speaking more than one language seems to have a limited impact on the duration of the word production bottleneck.
Figure 8. The performance of the three groups when reducing SOAs in Task 2.
The significant interaction between group and block showed that the bilinguals were much slower than monolinguals in Block 1 (336 ms slower) than in Block 2 (240 ms slower), while the bilinguals were significantly slower than the interpreters in both blocks (232 ms and 243 ms slower in Block 1 and 2, respectively). Furthermore, the smallest block effect was found in the monolingual group (160 ms difference). The bottleneck stage in word production was reduced when presenting the word for the second time even in different conditions, as shown by the marginally significant interaction between block and SOA. The SOA effect was significant in both blocks, and the block effect was significant in all the SOAs. The difference between the effect for the short SOA and long SOA, however, was slightly smaller in Block 2 (the difference between SOA 50 and 900 was 399 ms) than in Block 1 (the difference was 451 ms). This result supports the assumption that selection stages in language production might be reduced or eliminated after practice, such as during simultaneous interpreting practice.
Discussion
Previous work in simultaneous interpreting has suggested that professional interpreters can do more than two tasks concurrently during simultaneous interpreting by sharing their limited capacity resources (Gile, Reference Gile2009). This claim counters the evidence from other studies that suggest that two tasks cannot be conducted at the same time, but must be carried out sequentially (Pashler, Reference Pashler1994). In the present study, we aimed to explore whether professional simultaneous interpreters can produce a word without interference from another unrelated task. A dual task was conducted to explore whether professional simultaneous interpreters might not encounter the central processing bottleneck during single word production due to their simultaneous interpreting experience. Furthermore, we sought to determine whether their performance in the word production bottleneck differed from highly proficient bilinguals and monolinguals.
The findings from this study show that the reaction time to a second non-related tone discrimination task was prolonged with increasing overlapping of the two tasks. Furthermore, the manipulation of sentence constraint and word frequency, which have been shown to impact on lemma selection and phonological word-form selection (Griffin & Bock, Reference Griffin and Bock1998) in Task 1 propagated their effect to Task 2. The reaction time was faster when the name of the picture was a high-frequency word than when it was a low-frequency word, and when the sentence was medium constraint than when it was low constraint. In other words, the selection of lemma and phonological word-form in Task 1 is similar to the response selection process and also engages the central processing mechanism, therefore affecting the second unrelated task in similar ways as for Task 1. These results suggest that monolinguals, bilinguals and even professional simultaneous interpreters are subject to the central processing bottleneck during word production, postponing another concurrent task during lemma and phonological word-form selection (see also Declerck & Kormos, Reference Declerck and Kormos2012; Ferreira & Pashler, Reference Ferreira and Pashler2002). Inconsistent with the hypothesis, professional simultaneous interpreters were unable to avoid the central processing bottleneck during word production even after extensive practice in SI. Consequently, interpreters might suffer the impact of the bottleneck on another concurrent task, such as memory recall and memory retention (Rohrer & Pashler, Reference Rohrer and Pashler2003). This study provides evidence arguing against the assumption that all the processing tasks can be conducted concurrently when limited capacity is efficiently allocated (Miller, Ulrich & Rolke, Reference Miller, Ulrich and Rolke2009; Tombu & Jolicœur, Reference Tombu and Jolicœur2003, Reference Tombu and Jolicœur2005). Even interpreters, who are known as time-sharing experts, cannot conduct all tasks simultaneously.
Contrary to the hypothesis, interpreting experience does not appear to have any effect on the duration of the bottleneck stage during the word production of interpreters, even in comparison with the bilingual group. Moreover, bilingual experience seems to have limited impact on the word production bottleneck. If, as suggested by Green (Reference Green1998), bilinguals fulfil lemma selection across languages, then it is reasonable to assume that more time is required to select the intended word across languages than within one language. Even though a somewhat larger bottleneck duration was found in bilinguals, the difference between bilinguals and monolinguals in the duration of the word production bottleneck did not reach statistical significance. However, as shown in Figure 6, bilinguals took more than 900 ms to produce a word, exceeding the longest SOA in this study. That is, the RT2 difference between SOA 50 ms and SOA 900 ms might not fully reflect the bottleneck stage in word production for bilinguals, because the bottleneck stage in Task 1 might not be completed at SOA 900 ms. Therefore, the results obtained in this study may not truly reflect the difference between interpreters and bilinguals, and bilinguals and monolinguals in the word production bottleneck.
Bilinguals in this study showed disadvantages in linguistic tasks, consistent with the work of Ivanova and Costa (Reference Ivanova and Costa2008), which demonstrated that bilinguals took a longer time to produce a word than monolinguals even in their dominant and first language. These results are consistent with the assumption that bilinguals’ two languages are both activated, and the non-target language causes interference during language production (Costa et al., Reference Costa, Miozzo and Caramazza1999; Green, Reference Green1998). Surprisingly, however, no significant difference was found between interpreters and monolinguals or interpreters and bilinguals in picture naming latency. This leads us to consider whether there may be a significant difference within the interpreter group based on observations made during the experiment that Asian-language and European-language interpreters appeared to demonstrate differences in performances. This question is explored further in the additional analysis presented below.
Participants named pictures significantly faster in Block 2 than in Block 1. However, the degrees of “improvement” varied between interpreters, bilinguals, and monolinguals. The difference between interpreters and monolinguals, and bilinguals and monolinguals in naming latencies were significantly reduced in Block 2 where the stimuli had already been named once, even in different conditions (e.g., a stimulus showed in the medium constraint condition only once across two blocks). The possible explanation for this result might be found, as Gollan and colleagues (Gollan, Montoya, Fennema-Notestine & Morris, Reference Gollan, Montoya, Fennema-Notestine and Morris2005) suggest, in the frequency effect. They suggest that bilinguals should be slower in picture naming because of the weak link between word processing stages as the result of comparatively less frequent use of each language in comparison to monolinguals. Moreover, bilingual disadvantages in linguistic tasks can be attenuated with increased repetition of picture naming. In line with the argument of Gollan and colleagues (Reference Gollan, Montoya, Fennema-Notestine and Morris2005), the picture naming latencies of interpreters and bilinguals in this study were reduced much more than for monolinguals after a block practice.
Additional analysis: Abnormal response latency within the interpreter group
In this study, the results suggesting that there is no significant difference between interpreters and bilinguals, and interpreters and monolinguals in Task 1 are somewhat surprising, since interpreters were 81 ms faster than bilinguals and 82 ms slower than monolinguals (across blocks). Moreover, larger standard deviations of RTs were observed for interpreters in each condition. In view of these results and the observations of differences in performance between Asian-language and European-language interpreters during the experiments mentioned above, further investigation of a potential language difference within the interpreter group was conducted. The possibility of differences caused by the number of years practicing SI was ruled out by the non-significant difference in SI experience between European-language and Asian-language interpreters (see Table 5), t = 1.277, df = 22.699, p = .214, and the SI experience (length of SI) was non-significantly related to picture naming latencies between the two interpreter groups when it was included as a covariate, F (1, 27) = .205, p = .654.
Table 5. The number of years of age, and simultaneous interpreting experience and mean error rates in Task 1 and Task 2 for language sub-groups.
Note: “Lang” represents the word “Language” while “ICR” represents “Incorrect Responses rate”.
* The English-speaking monolinguals were divided into two groups. The Asian-language matched group included the participants who were carefully matched to Asian interpreters while the European-language matched group included the monolingual counterparts of the European-language interpreters.
Therefore, each of the three groups was divided into two sub-groups: an Asian-language sub-group (n = 12) and a European-language sub-group (n = 18). For the interpreter group and the carefully matched bilingual group, the allocation was done based on the other language(s) they speak besides English. For the monolingual group, subjects were divided into two sub-groups according to the interpreters they were matched with, to avoid possible age and gender effects.
Four-way 2×2×2×2 repeated analyses of variance (ANOVAs) was conducted with the same within-subjects variables as in the above analyses with the three groups and two sub-groups as between-subjects variables. The performance of groups was significantly different, F (2, 84) = 5.554, p = .005, reflecting significantly faster RT in monolinguals (973 ms) than bilinguals (1150 ms). The difference between the language sub-groups was significant, F (1, 84) = 4.983, p =.028, indicating that the Asian-language sub-group (1120 ms) was slower than the European-language sub-group (1021 ms) in picture naming. The interaction of groups and language sub-groups was significant, F (2, 84) = 4.945, p = .009. Simple effects analysis showed that the Asian-language sub-group was significantly slower than the European-language sub-group in the interpreter group (277 ms slower), but not in the bilingual (80 ms slower) and matched monolingual groups (62 ms faster). Within the Asian-language sub-group, monolinguals who matched Asian-language interpreters (942 ms) were significantly faster than interpreters (1227 ms) and bilinguals (1190 ms), while the difference between the bilinguals and interpreters was not significant. Within the European-language sub-group, the only marginally significant difference was between the performance of interpreters and their bilingual counterparts, showing that the interpreters were faster in picture naming than the bilinguals (159 ms faster). These findings appear to confirm the observation that European-language interpreters were faster than Asian-language interpreters in word production made during the course of the experimental data collection and processing.
One might argue that the relatively smaller typological distance between the European languages in this study and English (in comparison to the Asian languages and English) might account for the difference between Asian-language and European-language interpreters. However, no significant difference was observed between Asian-language bilinguals and European-language bilinguals in this study. Furthermore, a similar naming latency for interpreters and bilinguals was not only obtained in this study (1227 ms and 1190 ms for Asian-language interpreters and their carefully matched bilinguals, respectively), but also in the study of Christoffels and colleagues (Christoffels, De Groot & Kroll, Reference Christoffels, De Groot and Kroll2006) even when two languages are closely related (982 ms and 959 ms for Dutch–English interpreters and their matched bilinguals, respectively). These results argue against the assumption that similarity between two languages might be the reason for the significant difference within the interpreter group in this study.
It is possible that the difference within the interpreter group might be an artefact of the inclusion of cognate words. In this case, European-language interpreters should be significantly slower than their matched monolinguals after discarding all the trials that include cognate words. Further analysis was conducted between European-language interpreters and their matched monolinguals to test this possibility. Any English word sharing the same pronunciation with any of the other languages involved in this study was counted as a cognate word. The same trials including cognate words were discarded in both interpreter and monolingual groups in consideration of the picture-items effect. For French interpreters and their matched monolinguals, the trials that included the pictures of Bomb, Baby, Plant, Badge, and Bowl were removed. For German interpreters and their monolingual counterparts, the response latencies of the pictures Hand, Bomb, Arm, Glass, Baby, Star, Ball, Ring, Box, House, and Ax were excluded from analysis. No significant difference between the two groups was obtained, F (1, 34) = .916, p = .345, suggesting that the professional European-language interpreters are as fast as monolinguals in lexical access. This result is consistent with the results including cognate words, suggesting that the cognate words are not the reason for the similar performance of European-language interpreters and monolinguals in this study. Table 6 presents the mean RT after discarding the cognate words.
Table 6. Reaction time (standard deviation) for European-language interpreters and monolinguals.
Note: Reaction times are in milliseconds. HF represents high-frequency picture names, while LF represents low-frequency picture names.
It may be that the performance of the European language interpreters in this study is a one-off exception, unlikely to be replicated in future studies. However, this interpretation does not fully explain the obvious differences in results in tone discrimination errors between Asian language and European language interpreters (see Table 5). One potential explanation for the striking difference within the interpreter group might be the difference in the amount of SI experience. SI is a highly specialised professional practice. Some interpreters may have relatively limited experience in a year measured in hours or days. It may be possible that the European-language interpreters in this study have comparatively more frequent experience of SI than the Asian-language interpreters. However, further studies are required to test this hypothesis.
Conclusion
The results of this study show that simultaneous interpreters, widely regarded as time-sharing experts, are also subject to the central processing bottleneck during lemma and phonological word form selection, just as bilinguals and monolinguals are. Experience in simultaneous interpreting does not help interpreters to ‘avoid’ the bottleneck which could potentially impact on their SI performance. This finding argues against the assumption that all tasks can be conducted in parallel and concurrently by sharing the limited capacity (Gile, Reference Gile2009; Mcleod, Reference McLeod1977). Furthermore, this study indicates that interpreters might encounter many more difficulties during SI than they are aware of, such as postponing the stored information retrieval during language production, and suggests that further exploration is necessary. Several of the findings of this study necessitate further research. Firstly, it is possible that the amount of exposure to SI may affect the performance of interpreters in different ways. Future studies using interpreters as subjects need to more clearly identify the frequency of their exposure to SI. Secondly, it remains unclear whether the lexical selection of bilinguals (including interpreters) takes place between languages or within a language. It is worth exploring whether the similarities between bilinguals and monolinguals in the duration of the word production bottleneck might be the result of language-specific selection in bilinguals (Costa et al., Reference Costa, Miozzo and Caramazza1999) or a potential bilingual advantage in response selection. Thirdly, there is the question of whether ‘ideal’ bilinguals who can perform similarly to monolinguals in lexical access do, in fact, exist. Further investigation is necessary to determine how it could be the case that the professional simultaneous interpreters in this study, for most of whom the testing language was not their dominant language, could be as fast as monolinguals in word production, especially given the robust results of many existing studies showing that proficient bilinguals are slower than monolinguals in lexical access even when testing in their dominant language (Ivanova & Costa, Reference Ivanova and Costa2008). This result provides grounds for further testing the language production performance of interpreters with explicit types of interpreting experience, and consequently, improving existing models of bilingual lexical access with this special bilingual group.
