Language and Cognition in Bilinguals and Multilinguals_ An Introduction

286 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS in the digit-switching condition? What pair of the words out loud to themselves. If so, the out- memory subsets does the participant switch put switch might be implicated. They considered between in this particular case? And if the this much less likely with auditory presentation results are not specific to language control, of the materials. exactly how do they inform the more general pro- cess of control involved? Whatever the answers Once again, language switching took an to these questions, some related later work, observable amount of time in all experiments. performed in a different theoretical context, In most cases language switching took around has also noted the similarity between language .20 to .30 seconds per switch, irrespective of switching and switching between other pairs of whether the presentation was visual or aural. tasks (e.g., Green, 1998; Hernandez et al., 2001; There was one clear exception: The fixed alterna- Meuter & Allport, 1999). This work as well as tion condition resulted in a longer switch time the theoretical implications of this observation (.43 seconds) than the random alternation will be discussed later on in this chapter and in condition (.33 seconds). If input switching were Chapter 8. under voluntary control, the opposite pattern, a smaller switch cost in the fixed-alternation condi- Contrary to Macnamara et al. (1968), tion, should have been obtained. It is for this rea- Macnamara and Kushnir (1971) studied the input son the authors interpreted their results as sup- switch, once again testing French–English port for their view that input switching occurs bilinguals. Having verified that a cost of language automatically, perhaps brushing aside too lightly switching also occurs in comprehension tasks the curious result that no null effect but a reversed (silent reading of unilingual or language-mixed effect of the fixed–random alternation manipula- texts and sentences) in two preliminary experi- tion was obtained. ments, in two further experiments the authors set out to test their hypothesis that the input switch Further studies and evaluation operates automatically. In one of these experi- ments the participants silently read blocks of That switching the language of input or output sentences in which unilingual English sentences incurs a cost has been replicated many times since alternated with language-mixed sentences, and the switching paradigm was first introduced by were asked to judge the truth or falsehood of Macnamara and Kolers. That the effect is robust each of the sentences (e.g., unilingual true: turnips is therefore beyond all doubt, but what causes it are vegetables, language-mixed true: turnips sont has been disputed ever since these early studies vegetables versus unilingual false: horses smoke were rediscovered and the questions they potatoes, language-mixed false: horses smoke addressed were picked up again. In two studies pommes de terre). The crucial manipulation here on input switching, Grainger and Beauvillain was the pattern in which the unilingual English (1987) and Soares and Grosjean (1984) were and mixed sentences alternated: fixed (English the first to challenge the interpretation of the sentence, mixed sentence, English sentence, mixed switch cost in terms of a language switch. sentence, etc.) or in random order. If the input Grainger and Beauvillain, in a French–English switch operates automatically, no effect of this study, used the visual lexical decision task intro- manipulation should occur and language switches duced in Chapter 4 (p. 157). Soares and Grosjean in the language-mixed sentences should take presented unilingual or mixed spoken sentences equally long in the fixed and random conditions. to their Portuguese–English participants. The The second experiment was a replication, except majority of the words in the mixed sentences that the participants were now presented with belonged to the participants’ one language spoken sentences. The authors reasoned that this (the “base” language) but some of them belonged modified procedure ensured that input switching to their other language (the “guest” language). was indeed being tested because in silent reading The task used by the researchers was the the possibility remains that the participants say

6. LANGUAGE CONTROL 287 “phoneme-triggered lexical decision task”: On concept of a language switch is their assumption each trial the participants were presented with a that it implies the language not in use is switched sentence preceded by a particular phoneme and off, deactivated, or—in terms of the terminology had to detect the stimulus (word or nonword) that used in Chapters 4 and 5—that bilingual word began with this phoneme. Upon locating this recognition is language-selective. But, as I pointed stimulus in the speech input, they had to indicate out above, this is not the essence of Macnamara’s as quickly as possible whether it was a word or a account of the switching cost, nor does it follow nonword. In both studies the switch cost was from it logically: The language not in use, what- measured locally. ever its state of activation, may be ignored one way or the other. Thus employing different tasks, these researchers replicated the main finding of the In later reports (Grainger, 1993; Grainger & early studies; namely, that language switching Dijkstra, 1992), Grainger provided an account of incurs a cost. As we have seen, this finding is con- the pattern of switch costs obtained in the 1987 sistent with the notion of some device that study in terms of two models of bilingual word switches between language subsets. At first sight, recognition that both assume that processing a further finding of Grainger and Beauvillain is initially language-nonselective. One of these (1987) seems problematic for this account of the is the BIA model introduced in Chapter 4 switch cost: Unlike words with a language- (pp. 177–181). Exactly how this model (and its nonspecific spelling pattern (that is, words that successor BIA+) may account for the cost of lan- are orthographically legal in both languages), guage switching will be detailed later in this chap- words with a language-specific spelling pattern ter. The second model, the bilingual activation (orthographically legal in one language only; for verification model (BAV), assumes the existence instance, English words containing the letter clus- of two lexicons, one for each language, with word ters wh, ck, or ght, or French words containing representations in both of them responding to a the letter sequences eau, oie, or oux) did not show word input. Following this initial activation, the the expected switch cost. According to the in-out activated nodes in one of the lexicons are first account of language switching, a switch to the searched for a match with the input (in a so-called other language should always incur a cost, “verification” stage). If no match is obtained, the irrespective of the orthographic characteristics of activated nodes in the other lexicon are searched the word presented on a switch trial. However, in through. The lexicon that resulted in a correct a later study Thomas and Allport (2000) attrib- match on the previous trial is the one searched uted this null effect to a missing control condi- first, and this is the source of the switch cost: If tion. With the missing condition reinstated, they the language of the current trial differs from the obtained a switch cost for words with language- language of the previous trial, the search starts specific and language-nonspecific spelling pat- out in the wrong lexicon, resulting in a relatively terns alike. The results of both studies thus slow response. appear to be consistent with the concept of a lan- guage switch, so why is it that these later authors Soares and Grosjean (1984) chose a serial rejected it? Grosjean appeared to be doing so search explanation of their data similar to the because of the inconclusiveness of the origin of one assumed in BAV but one embedded in the switching cost: “It is not because bilinguals language-mode theory to be presented in the next may, at times, process code switches more slowly section. Language-mode theory distinguishes than base-language words that researchers can between a base language and a guest language. conclude there is a language switch [. . .] involved The base language is the contextually most in the processing; the delay could be due to appropriate language and, therefore, the currently numerous other factors” (Grosjean, 1997a, most activated language. Whereas according to p. 250; see also Grosjean, 1988). The main reason BAV the language of the previous trial determines for Grainger and Beauvillain (1987) to reject the which lexicon is searched first, Soares and Grosjean assumed that the base language lexicon

288 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS is always consulted first. Only if this search ACTIVATION AND DEACTIVATION OF fails to produce a match with the input, is the THE LANGUAGE SUBSYSTEMS: THE other language, the guest language, searched. The BILINGUAL’S LANGUAGE MODES language switches that they examined in their mixed condition always concerned guest language Language-mode theory words and the delay in recognition observed for these words were seen as support for this serial Language users may find themselves in many search view. Notice, however, that a stringent test different communicative settings. To name just of this particular serial search view requires that a few: they may talk to a child, their partner, a response times associated with switches from the friend, or to some official; they may take part in guest language words back to the base language a formal or an informal conversation, comment should also be measured. Soares and Grosjean’s upon the weather or discuss the opacity of a explanation of the switch costs predicts that in country’s tax system. All these different settings these cases no cost occurs. invite the choice of a specific “register”; that is, the use of the specific vocabulary and gram- It is noteworthy that both these accounts, in matical constructions that fit the communicative terms of BAV and language-mode theory, assume context. For monolinguals the language of dis- the existence of lexical subsets in bilingual course in this multitude of communicative memory, one for each language. These appear to worlds will always be the one language they have be similar to, if not the same as, the language mastered. But for bilinguals, especially those who subsets that Macnamara and his colleagues are reasonably fluent in both of their languages, hypothesized to exist and set out to test by means language is yet another source of variability. The of the language-switching paradigm. Perhaps specific characteristics of the communicative even more noteworthy is the earlier-mentioned setting determine what language is being spoken, conclusion drawn by Macnamara et al. (1968) just as they determine the registers used by that language switching does not require a special monolinguals. In addition, the properties of the psychological skill but a skill that is also deployed communicative context appear to determine in other situations where people have to switch the incidence of code switches and the fluency between tasks. It thus seems that they never of the bilingual’s speech. François Grosjean looked upon the language switch as some mental launched the concept of the bilingual’s “language device developed through bilingualism and dedi- modes” to account for the apparently systematic cated exclusively to language control in bilinguals. way the bilingual’s speech responds to the Yet it is as if the above (and present-day) critics environment and to changes therein. More of the language-switch concept have assigned it specifically, and especially in his later work, he the more narrow meaning of such a specialized describes the various states of activation of the device. As the next sections will show, 40 years bilingual’s language system that result from the later both the subset view and the view of lan- prevailing circumstances and that, in their turn, guage switching as an instance of task switching command the language of output and affect in general feature prominently in theories on the way speech and text is recognized (e.g., bilingual language control. A further central issue Grosjean, 1988, 1989, 1997a, 1997b, 2001; Soares in contemporary theories on control is whether & Grosjean, 1984). An important underlying and to what extent the level of activation within assumption in his theory is that the bilingual’s the language subsets can be affected by sources language system is organized in two subsets, external to the language system, such as the inten- one for each language, and that these subsets tion to speak one language and not the other. can be activated or deactivated as a whole and That language-external sources have this power independently from one another. is a central tenet in language-mode theory, the theory of language control in bilinguals to which I will now turn.

6. LANGUAGE CONTROL 289 In his early publications, Grosjean (1989, 1994, bilinguals operate in a bilingual mode) may strike 1995) describes how bilinguals “find themselves at the reader as inconsistent with the evidence pre- various points along a situational continuum” sented earlier (pp. 285–287), where a language (Grosjean, 1994, p. 1657) (italics added) and how switch was always associated with a slowing down these different points induce different language of processing. A way to resolve this apparent modes (occasionally called “speech modes”). conflict is to consider the fact that Grosjean’s When talking to monolingual speakers of their theory primarily deals with language switching one or other language, they inevitably restrict under natural circumstances, whereas the themselves to speaking this one language switches in the studies discussed earlier were pro- exclusively. This monolingual behavior is enabled duced in experimental situations that were a far by the “monolingual mode” they are in. At the cry from natural language use. In natural other end of the situational continuum they are bilingual speech the speaker himself may govern communicating with bilinguals with whom they language switching, presumably exploiting this share their two languages, happily mixing the option at points where the other language two languages frequently. When this behavioral expresses the targeted concept most faithfully and pattern occurs, the bilingual is said to be in a when at the same time a switch to the other lan- “bilingual mode”. These two conditions are guage does not involve the risk of a communica- considered the end points of a continuum of tive breakdown (because the interlocutor is fluent modes, which also covers intermediate states. in this other language as well); not to switch lan- guages at that point (for instance, because it is not In cognitive terms, the bilingual’s monolingual an option since the interlocutor is a monolingual) behavior is enabled by a mental state, the mono- might actually slow down processing. Conversely, lingual mode, in which the language in use is max- in the switching experiments discussed earlier the imally activated and the other language is switches were experimenter-imposed and were deactivated as best as possible but, as stressed by arguably often enforced in unnatural places where Grosjean in various publications, with presum- there was no reason to switch in the first place. ably always some residual activation remaining. This might cause the slowing down of perform- In contrast, the bilingual mode concerns a mental ance at the switch point, possibly through a sys- state in which both languages are highly acti- tem reset that is required at that point. vated, but still to different levels. The reason the two are differentially active is that bilinguals In Grosjean’s later work (e.g., Grosjean, always “choose” one of their two languages as the 1997a, 1997b, 1998, 2001) the focus shifted from main language of the interaction. The choice of the situational continuum that induces different this language results in enhanced activation of the mental language modes to a continuum of lan- language subsystem in question. The currently guage modes itself and the correlated patterns of chosen language is called the “base” (or “matrix”, activation of the two language subsystems. In or “host”) language or Language A. The other this later work Grosjean also emphasized that language is called the “guest” language or Lan- the language-mode concept encompasses two guage B. The behavioral effect of a monolingual- factors: (1) which of the two language repertoires mode setting is the occurrence of relatively few is chosen as the base language, the language to switches to the other language, as well as rela- use, in a particular setting, and (2) what the state tively many hesitations at moments when the of activation of each of the two languages is at a selected language lacks the means to express a particular moment in time. Figure 6.1 represents particular conceptualized message easily. The these two dimensions of the concept on the verti- behavioral effect of a bilingual-mode setting is cal and horizontal axes, respectively. Three states more language mixing and more fluent speech. are depicted. In all of them Language A is the base language, activated the most. The three The correlation between degree of language states differ in the level of activation of Language mixing and speech fluency just suggested (more B, the guest language. The latter is deactivated as switches and more fluent language use when

290 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS The language-mode continuum. The level of activation of Language A (the base language) and Language B (the guest language) is represented by the degree of darkness of the squares (black = maximally activated; white = inactive). The horizontal dimension represents the state of activation of a bilingual’s two languages at a particular moment in time, the activation of Language B varying between nearly inactive and reasonably highly activated (but less so than Language A). From Grosjean (1997b). With kind permission by John Benjamins Publishing Company, Amsterdam/ Philadelphia. best as possible on the monolingual side of the Earlier in this chapter I introduced four con- continuum and is gradually activated more when trasts according to which theories on bilingual moving towards the bilingual side. In this later control can be distinguished from one another: work some suggestions are also made as to how proactive versus reactive control, global versus the language-mode concept could be extended to local control, internal (endogenous) versus account for language use in speakers of more external (exogenous) control, and whether than two languages such as trilinguals (Grosjean, the language system proper or the output from 2001) and in simultaneous interpreting (Grosjean, the language system is affected by the control 1997b; see pp. 323–324). A final point to note here process. Language-mode theory assumes that a is that the choice of the base language (or the base language is chosen prior to the onset of a dis- “intention” to speak one language and not the course—a choice that determines the activation other) and taking position on the language-mode setting of the language system. This implies that continuum presumably often do not involve bilinguals are assumed to control the use of their deliberate, conscious decisions on the part of the languages proactively. But in addition the theory interlocutors. Instead, they emerge automatically states that during ongoing discourse the activation from the current circumstances, the interlocutors’ levels of the two language subsets are not rigidly knowledge thereof, and the nature of the input fixed but fluctuate, depending on, for instance, the the interlocutors receive. Similarly, because lan- nature of the input or changed goals. So control is guage switching and how frequently it occurs is a not exclusively governed by a proactive setting of consequence of the state of the mental system, it the system, but “goes with the flow” as well. is quite plausible that language switching will Regarding the second distinction, it is obvious the more often come about automatically, rather than notion of global control is adhered to: The mode resulting from a conscious choice to switch. In setting affects the level of activation of all elem- other words, non-intentional switches plausibly ents within both language subsystems at the same outnumber intentional switches. time. With respect to the third contrast, it appears

6. LANGUAGE CONTROL 291 the theory assumes the occurrence of both Finally, in a third condition (Bilingual B) they external and internal control, because both the were informed the person to address had lived in intention to speak a particular language (internal) the United States for 7 years working for a local and the type of input the system receives (external) firm, had French and American friends, and exert an influence on the activation levels of the spoke both languages at home. two language subsystems. Finally, the control pro- cesses exert their effect on the level of activation of The second variable was the topic described the entities stored within the language system by the French stories: They either concerned proper and not on the output from the system. situations found in France or typical American activities. The stories in the latter condition Although, as we will see, language-mode the- (called “bilingual” stories) contained a number of ory can account for the situational dependency of code switches into American at places where a the number of code switches in, and the fluency code switch might naturally occur (because guest of, bilingual speech, a problematic feature of the language English but not base language French theory is that it contains some circularity in has a word that exactly expressed the concept to the reasoning on how relative activation of the be verbalized). The dependent variables were the language subsystems relates to the speech data: number of French and English syllables that were The bilingual, responding to the specifics of the uttered in the summaries the participants came communicative context, adopts a language mode up with, and the number of hesitations they con- that becomes reflected in the different activation tained. Figure 6.2 shows, for the bilingual stories, states of the two language subsystems. This the numbers for each of these three variables in pattern of mental activation then leads to a par- the three different addressee conditions. ticular speech pattern characterized by the degree of language switching that occurs and the degree The participants clearly adjusted themselves to of speech fluency. Finally, it is this speech pattern the addressee profiles provided by the experi- that then leads to the conclusion as to what mode menter. When they imagined addressing the the speaker is in; that is, what the relative level of “French” interlocutor, more French and fewer activation of the two language subsystems is. English syllables were uttered and more hesita- tions occurred than when Bilingual B was spoken Evidence to. Speaking to Bilingual A produced an inter- mediate pattern of results. These data suggest that In an experiment that copies natural communica- the instructions had led the participants to believe tion rather veraciously, Grosjean (1997a) pre- that the French addressee was not yet fluent sented a convincing demonstration of the ability enough in English not to be impeded by their code of bilinguals to adapt their speech to the specifics switching into English, that Bilingual A was prob- of the communicative setting. He presented ably fluent in English but was a purist who didn’t French–English bilinguals, fluent in both lan- appreciate language switching, and that Bilingual guages, with stories in French and instructed them B, also fluent in English, didn’t mind code switch- to summarize them in spoken French. One of the ing and presumably code switched a lot himself. variables in this study was the (imaginary) person Apparently, to maximize the communicative addressed. In one condition (French) the partici- effect, the participants had adapted their speech pants were told the person to address had just to the bilingual profile of the person spoken to. arrived in the United States, spoke only French at home, and, although he could read and write Eng- The differences between the addressees can lish, still had difficulties speaking it. In the second also explain the variation in the number of hesita- condition (Bilingual A) they were told the person tions across the conditions. As mentioned, the addressed had lived in the United States for 7 stories presented to the participants contained a years, worked for a French government agency number of English words because French was teaching French, and only spoke French at home. less suited to express the associated concepts. But apparently, bearing in mind the addressees’ pro- files, the participants in condition French, and

292 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS Number of French, English, Dutch–French–English study by Dewaele (2001). and hesitation syllables as a Both studies involved a manipulation of the function of the (imaginary) addressee and again showed that this variable person addressed. The affects the number of code switches and speech addressees in conditions fluency. Dewaele (2001) added formality of the French, Bilingual A, and setting to the set of variables that bilingual/multi- Bilingual B differ in their lingual speakers adapt to in their speech. He tested command of French. From Dutch L1 speakers who were learners of both Grosjean (1997a). French and English in a formal and an informal interview setting where French was the base lan- to a lesser extent in condition Bilingual A, never- guage of the interviews. The informal setting theless attempted to express them in French. This involved a relaxed conversation between inter- caused the relatively large numbers of hesitations viewer and participants about their hobbies, stud- in these conditions as compared to condition ies, and politics. The participants were told that Bilingual B, but presumably also the large vari- the content of their speech was more important ation in the number of French syllables across the than the form. The formal condition concerned an conditions: Retelling the English code switches in oral exam that was to assess the participants’ pro- French required rather elaborate, long-winded ficiency in French. Here it was stressed that the expressions. As a consequence, the overall evaluation of the test depended on both speech number of French syllables was larger in con- content and speech form. The formality of the dition French than in condition Bilingual A, and interview setting had a large impact on the number larger in the latter than in condition Bilingual B. of code switches. On average, in the informal set- ting 9% of the utterances were mixed, whereas in In terms of the language-mode continuum the formal setting the number of mixed utterances illustrated in Figure 6.1 these data suggest that dropped to 3%. These results are consistent with when talking to the addressees French and the language-mode hypothesis, suggesting that the Bilingual B the participants positioned them- selves towards the monolingual and bilingual end of the continuum, respectively. When talking to Bilingual A they took an intermediate position on the continuum. The second manipulation (topic) provided additional support for this adaptability of bilingual speakers by showing that the bilingual stories produced about 10 times as many English syllables as the monolingual stories. Similar results were obtained in a Turkish– German study by Treffers-Daller (1998) and a

6. LANGUAGE CONTROL 293 participants in the formal interview setting oper- the view that the different speech patterns across ated more towards the monolingual end of the settings are mediated by differential activation of continuum than those in the informal setting. the two language subsystems would lose ground. Further on (pp. 307–308) I will present a theory of Evaluation: An alternative account? bilingual control (e.g., Green, 1986, 1998) that encompasses a control mechanism that supervises On the face of it all three studies discussed above processing and acts upon the imminent output of provide convincing support for language-mode the bilingual language system, and which thus theory. Yet a remark by Dewaele (2001) triggers appears to bear more than a superficial resem- an alternative interpretation of the observed blance to the monitor introduced here. fluctuations in the number of code switches and hesitations, one that may turn out to provide a In fairness to language-mode theory, this data real challenge for the theory. Dewaele noted that ambiguity—the fact that one and the same data a bilingual in a bilingual mode—where, according set can be explained in more than one way—is by to the theory, both language subsystems are no means a unique feature of the work discussed highly activated—might nevertheless produce here but is omnipresent in much of the related few mixed utterances because a conscious output- work and beyond. Indeed, the studies on lan- monitoring process may prevent imminent lan- guage switching reviewed earlier in this chapter guage switches from actually emerging in the also produced data patterns consistent with output. As we have seen in Chapter 5 (p. 224), a more than one theoretical interpretation, and the similar monitor (and plausibly the very same sections to come will add yet further explanations monitor) is also assumed to operate on internal of the switch data. This data equivocality partly speech during monolingual speech production reflects the limitations of the tasks and pro- thus preventing speech errors from popping out cedures available to the researcher. Regarding of the system (e.g., Levelt, 1989). But given the language-mode theory, the strongest evidence to fact that the incidence of language switches is a support it would arguably involve first finding major marker of the bilingual’s current language a means to locate the language subsets in the mode, how can it be justified that this time this brain, and next finding out whether and how the important marker is ignored and overruled; that activation in these brain areas would respond to this time the few switches that occur are not taken manipulations of the type presented above. to indicate what they seem to indicate—that the Unfortunately, even though over the past 20 years person is operating in a monolingual mode? the tool kit of the neurocognition researcher has Whatever the answer to this question, introducing become enriched with impressive new techniques a monitor in a model of the bilingual speaker may to spot and measure brain activation, these tech- be problematic for the theory as it is, because, niques may turn out to be unsuitable for the pres- instead of the relative activation levels of the base ent purpose. The reason is that evidence is and guest languages, flexible behavior of the accumulating that the bilingual’s two languages monitor may underlie the different numbers of are stored in exactly the same brain areas (e.g., code switches across conditions: Base and target Fabbro, 1999; Hernandez et al., 2000; Klein, language may, in theory, always be equally Milner, Zatorre, Meyer, & Evans, 1995; Paradis, activated, but the monitor may operate in a strict 1990, 1997, 2000) so that identifying the acti- or a more lenient way, depending on the charac- vation levels associated with the different lan- teristics of addressee and setting (see also De guages may not be possible. Evidence suggesting Groot, 2002). If this account were true, it would otherwise—namely, that the left hemisphere is rip the heart out of language-mode theory in its the home of the native language whereas later present form. The claim that bilinguals respond languages are subserved, at least partly, by the to the specific characteristics of the communica- right hemisphere (e.g., Kim, Relkin, Lee, & tive setting would not be challenged. However, Hirsch, 1997)—can be accommodated with this view by assuming that the right hemisphere is

294 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS primarily involved in coping strategies when a subsystems and focuses more on what the con- language is still underdeveloped. In other words, sequence of language choice is for the activation activation patterns in the brain that at first sight level of individual lexical items rather than on the suggest distinct cortical localization of native subsets as wholes. Grosjean’s focus, instead, is on and later languages may in fact result from the the way contextual factors affect the activation participants’ unbalanced bilingualism and the levels of the linguistic subsystems. need to use compensatory strategies while using the weaker language (e.g., Paradis, 1997; see also A central assumption in Paradis’s theory is Chapter 8, pp. 427–435). that a memory item (e.g., a stored word) becomes available for further processing when a sufficient To summarize, various sources of evidence number of positive neural impulses have reached suggest that a bilingual adapts to the specific its neural substrate. The amount of impulses that characteristics of the current communicative con- is required to reach this point constitutes the text such as its formality and the person being item’s “activation threshold” (Paradis, 2004). It is talked to. However it remains to be seen whether assumed that activation thresholds vary with the this adaptability concerns fluctuations in the frequency and recency of the item’s use and, con- degree of activation of the bilingual’s two lan- sequently, differ between items: Frequently used guage subsets or fluctuations in the attentiveness items have lower thresholds than infrequently of a mental monitor that watches over the output used items, and recently used items have lower of the language system. thresholds than items used longer ago. In other words, frequent items and recently used items A NEUROLINGUISTIC THEORY OF BILINGUAL require fewer neural impulses to reach their acti- CONTROL vation thresholds than infrequent items and items used further back in time. This proposal is func- As we have seen, the central idea of language- tionally equivalent to a view on word recognition mode theory is that bilinguals exert control by that we have already encountered in Chapter 4 balancing the activation levels of the two lan- pp. 177–181; e.g., Dijkstra & Van Heuven, 1998). guage subsystems in such a way that they meet the There, fixed activation thresholds but different goal of the current communicative setting. The baseline levels of activation for frequent and activation levels of the two language subsystems infrequent words were assumed as well as a tem- can be set proactively and globally, simul- porary increase in a word’s baseline activation taneously affecting all elements within both level after it has just been recognized. An item is language subsets. Despite a difference in termin- selected if its activation exceeds the activation of ology, these ideas are very similar to those all possible competitors. This is ensured not only encompassed by a comprehensive neurolinguistic by activating the target item with neural impulses theory of bilingualism (and language processing but also by concurrently raising the activation in general) that Michel Paradis has developed thresholds of competing items, thus inhibiting over many years (e.g., 1994, 1997, 2001, 2004). As them. A further assumption is that language pro- language-mode theory, Paradis’s theory assumes duction requires lower threshold settings than language control in bilinguals to result from dif- language comprehension. The reason is that in ferential activation of the bilingual’s two lan- production the neural impulses to activate the guage subsystems that is triggered by the inten- relevant memory items must be generated within tion to speak one language and not the other. But the system, whereas in comprehension they are in the presentation of their ideas, these scholars excited by an external source; namely, the input employ somewhat different emphases: Paradis is that impinges on the eyes or ears. Generating the more explicit about the neural mechanisms neural impulses within the system is thought to be involved in the differential setting of the a mentally more demanding process than having them aroused by external stimulation. Like Grosjean, Paradis assumes that the

6. LANGUAGE CONTROL 295 bilingual language system is organized in two and proactively, and this process is set into subsets, networks of strongly interconnected motion the moment one of the languages is elements. Additional assumptions are that both selected for use. When a bilingual intends to use are part of a larger neurofunctional language one of her languages, neural impulses are fed to system which is independent from a further the targeted language subset and, at the same system, a neurofunctional cognitive system. The time, the activation thresholds of all elements in two language subsets and the cognitive system the other subset are automatically raised, thus combined are known as the “three-store” model. preventing interference from this language in Each language subsystem stores the grammar of production (although it may still be possible to that language as well as the language’s vocabulary comprehend input from the non-selected lan- and each individual lexical item is assumed to guage, as a consequence of the relative ease contain the word’s lexical meaning, its syntactic to excite the neural substrate from external constraints, its morphological characteristics, and stimulation; see above). Paradis likens this process its orthographic and phonologic forms. The con- to the activation and inhibition of different sets ceptual system stores concepts in a language- of muscles when one raises one’s arm, and claims independent way as a set of conceptual features the processes involved are a general property of onto which the words’ lexical meaning map. all cerebral systems and subsystems (Paradis, These views on the content of lexical items and 2004, p. 211). The intention to use one language the existence of a separate language-independent and not the other—which triggers these processes cognitive system are similar to those advanced in of differential activation and threshold raising— Levelt’s (1989) model of the monolingual is part of the conceptualized message in the speaker, in which a word’s lexical meaning and non-linguistic cognitive system. Paradis’s model syntactical specifications together constitute its shares this assumption with many other views on lemma and its morphological and phonological bilingual speech production (De Bot & Schreuder, properties constitute the word’s lexeme (see 1993; La Heij, 2005; Poulisse & Bongaerts, 1994). pp. 224–226 for a more detailed account). A special attraction of the activation threshold Paradis rejects the idea that language identity hypothesis (and of its sibling, language-mode of a word is explicitly represented in the bilingual theory) is that it, if complemented with a control system by means of language tags (e.g., Green, mechanism, can explain a number of symptoms/ 1998; Poulisse, 1997; Costa et al., 1999) or lan- recovery patterns of bilingual aphasia (Paradis, guage nodes (e.g., Dijkstra & Van Heuven, 1998, 1997, 2001, 2004; see also Green, 1986, 2002): 2002), considering such entities totally redun- blending, selective recovery, sequential recovery, dant. According to him, a word’s position in one and differential recovery. Blending (the inability of the linguistic subsets and the unique mapping to avoid switching between the languages; within this subset of the word’s form to its lexical e.g., Fabbro, Skrap & Aglioti, 2000; Meuter, meaning, as well as the unique mapping of the Humphreys, & Rumiati, 2002; Perecman, 1984) latter to a subset of the conceptual features in the may be due to a control mechanism failing to conceptual system, enable successful recognition keep the activation thresholds of the elements in and production, irrespective of the language the non-selected language high enough for these identity of the word. These mappings are the elements not to be selected inadvertently. product of past language experience. The fact Selective recovery (where there are lasting aphasic that bilinguals can tell to what language each symptoms in one language) may result from per- of the words in their vocabulary belongs is—so manent inhibition (too-high threshold settings) of Paradis claims—a form of metalinguistic know- one of the languages. Sequential recovery (where ledge that is not represented in the language one language only begins to reappear after the system proper. other one has been restored) suggests temporary inhibition of one of the languages. Finally, The language subsets can be activated and differential recovery (where both languages are inhibited independently of one another, globally

296 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS recovered in parallel but at different speeds) sug- into French but translation from French into gests greater inhibition of the one than the other Arabic failed. An interpretation in terms of a language during the recovery process. In other damaged language system is implausible, because words, all four types of aphasia may result from how then could each language be available at least the corrupted workings of some control mechan- some of the time? Green (1986) explained this ism responsible for regulating the setting of acti- puzzling linguistic behavior in terms of a control vation thresholds, rather than to the content of mechanism that fails to generate sufficient one or both language subsystems being damaged. resources to activate relevant parts of the lan- This idea that control failure rather than dam- guage system the moment they are needed and to aged linguistic knowledge may at times cause the simultaneously inhibit other parts. bilingual aphasia is strengthened by the fact that the speech errors produced by non-aphasic people In conclusion, the hypothesis that permanent under stressful circumstances are similar to the or ephemeral failures of a language-independent errors produced in language pathology (Dornic, control system underlie various forms of lin- 1978): The stressful circumstances may deplete guistic failure is a productive one that can the available mental resources leaving too few of account for pathological and non-pathological them to secure the impeccable operation of the linguistic behavior alike. control system. Similarly, non-aphasic bilinguals operating in a bilingual mode may at times exhibit THE EMERGENCE OF LANGUAGE SUBSETS the excessive language-mixing characteristic of the above “blending” type of bilingual aphasia It should be clear by now that many theorists (Grosjean, 1985). adhere to the notion of language subsets to account for the fact that bilinguals are rather As argued by Green (1986), a fifth type of successful in keeping their languages separate. bilingual aphasia first reported by Paradis, The subset/subsystem notion is a core concept in Goldblum, and Abidi (1982), “alternate antagon- the views on bilingual control presented in this ism” combined with “paradoxical translation”, chapter so far. In a number of places I have briefly can also be accounted for in terms of a failing hinted at how these subsets may come into being: control system operating on a possibly intact by co-occurrence in past linguistic experience of language system. The Arabic–French aphasic the elements of one and the same language such patients in question manifested a rather bizarre as the contiguity between different words in the form of linguistic behavior in which the avail- sentences we encounter, between word forms and ability of the two languages seemed to take turns. their meanings, between lexical meanings and the These patients alternated between relatively conceptual information in a language-neutral unobstructed spontaneous production of one of conceptual system onto which the lexical mean- their languages accompanied by an inability ings map (if such difference between lexical and to spontaneously produce the other, and the conceptual meaning exists at all; see pp. 234–236). exact reverse pattern at some later point in time, The consequence of co-occurrence is intercon- say the next day or later the same day, while nectedness of the memory units (“nodes”) that comprehension was always good in both lan- represent the co-occurring entities. The more fre- guages. More curiously even, the language avail- quently they occur together, the stronger the con- able for spontaneous production could not be nections. And because, of course, the language translated into, whereas the language unavailable units of one and the same language are for spontaneous production could: When, say, encountered in contiguity more frequently than French was available for spontaneous production, language units that belong to different languages translation from French into Arabic was possible (which may co-occur at code switch points and but translation from Arabic into French was not. during translation), the within-language memory Conversely, when Arabic was available for spon- taneous production, Arabic could be translated

6. LANGUAGE CONTROL 297 connections get stronger than the between- spread resonance in the language subsystem. The language connections and the two languages corollary of being stuck within one language is become insulated from one another, encapsulated. a cost when external circumstances enforce a switch into another language, as happens when an The underlying process that creates these experimenter imposes one. It should be obvious bonds between memory nodes is their co- by now that entrenchment does not imply the activation, a process that Hernandez, Li, and non-selected language subsystem is totally MacWhinney (2005) call resonance: According deactivated, but that whatever activation is going to a learning principle called Hebbian learning, on over there does not cause any serious trouble co-activated memory nodes get linked up because it is outweighed by the much larger with one another (what “fires together, wires activation of the selected language, or because together”, as the maxim goes; see Murre, 2005, activation in the non-selected subsystem is for an overview of this and other learning prin- ignored. ciples that may drive first and second language acquisition, as it drives learning in other Hernandez and his associates (2005) explain domains), and the more often nodes fire together, why early simultaneous bilinguals in particular the stronger the bond between them becomes. are likely to develop a strong firewall between the When a link between nodes has first been estab- two language subsets in this way, without denying lished by co-occurrence, further strengthening that late bilinguals may also develop a state of occurs when an input corresponding to one of functional language independence not only for these nodes is presented to the system: As a con- their L1 but also their L2. A hint of evidence sequence of spreading activation along the links supporting this suggestion was provided by that connect the activated node with other nodes, Linck, Kroll, and Sunderman (2009) in a study the latter are also activated, all resonating with that at the same time indicates that immersion the activation-sending node and all getting more learning promotes the emergence of L2 language strongly connected with the former and each independence. These researchers compared two other as a consequence of this. The ultimate result groups of American-English learners of Spanish of frequent co-activation of the nodes represent- on their performance on a language fluency ing the elements of one language is widespread task. The groups had had between four and resonance within a language system once it has six semesters of training in Spanish and were been selected for use. As a result of this state the matched as closely as possible on self-rated language user is “caught” within the language, proficiency in Spanish and a number of other a state that Hernandez and his colleagues variables. The crucial difference between them call “entrenchment”. Because in this state the was that one group had acquired Spanish in a elements (e.g., words, phonemes, meanings) of Spanish immersion environment and was tested the language in use are much more activated than in this environment, whereas the other group had those of the non-selected language, the former acquired it in a classroom setting and was tested are much more available and, therefore, other- in the L1 English environment. language intrusions only occur infrequently. In the fluency task the participants were On p. 284 I cited Macnamara et al. (1968), who presented with a series of category names (e.g., likened the bilingual to a musician who, after clothing, furniture) and were asked to generate as observing the notation for a key at the beginning many exemplars as possible to each category of a piece of music, performs the piece in the name within 30 seconds (e.g., dress, skirt, trousers, appropriate key effortlessly and automatically, sock . . .). Both an L1 English and an L2 Spanish completely forgetting about the key the piece is condition were included. The finding of special meant to be played in. It appears the analogous interest here was that in the L2 Spanish condition case of the bilingual who has no trouble staying in the immersed learners generated more exemplar one language once having accessed it is enabled names within this same unit time than the class- by this process of entrenchment through wide- room learners did, whereas the opposite result

298 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS emerged when L1 English was invited. The higher how they come into being, in the form of two scores for the immersion participants performing connectionist learning models that have both in L2 suggest that they had developed, likely shown that language subsets develop from through resonance, stronger links between L2 bilingual language input: the bilingual simple words than the classroom participants. Linck and recurrent network (BSRN; French, 1998; her colleagues suggested the smaller scores French & Jacquet, 2004) and SOMBIP, the self- when performing in their L1 were due to them organizing model of bilingual processing (Li & inhibiting their L1 in the immersion context. Farkas, 2002). In Chapter 4 two other connection- An alternative explanation is that there was no ist models were presented, one that accounts for active inhibition of the L1 but that it was simply visual word recognition in bilinguals (BIA) and a less available as a result of its less-frequent cur- second that models the recognition of spoken rent use and the consequence this has for the words in bilinguals (BIMOLA). Both of these average L1 word’s most recent use (that is, it models are “localist” connectionist models in might be a recency effect). In fact, this account which the represented entities such as words and receives support from the further finding that, phonemes are each represented by a single pro- returned to their L1 environment a couple of cessing node in the network (see p. 132 for details) months later, the immersed group’s fluency in and neither model was designed to explain how L2 Spanish remained superior to that of the learning comes about. Instead, they represent a classroom group while their L1 English fluency fixed state of a hypothetical bilingual at one was equally as good as the classroom learners’ L1 particular level of L2 proficiency. In modeling fluency. this state, all the representations (“nodes”) in the system, their baseline activation levels and More anecdotally, while I spent a week in activation thresholds, and the inhibitory and France recently I was appalled to discover that the excitatory connections between them, must be country’s national language with which I was set by hand. Modeling two different levels of pro- once reasonably familiar was completely unavail- ficiency would require performing this handiwork able to me, while at the same time I was aware I twice, once for each level of proficiency. had not lost it (see pp. 347–351). What appeared to be happening was that I was kept hostage In contrast to these stationary models, con- by my rather weak Italian, a language I had nectionist learning models simulate the emergence been immersed in not long before for a couple of the representations and their subsequent of months while working on this book. Whenever evolution with increased language experience. a foreign language escaped from my mouth, Both BSRN and SOMBIP learn by computing it was always Italian not French, and even statistical regularities in input sentences. BSRN attempts at mediation by my native Dutch could received as input simple artificial French and not set me free. It was as if my Italian immersion English noun-verb-noun sequences that were experience had created the beginning of a func- generated by a “language generator”. SOMBIP tional subset that, once caught in it, frustrated was presented with realistic child-directed my attempts to access French. That it was speech produced by an English-speaking father Italian—not my native Dutch or any other of the and a Cantonese-speaking mother who followed languages I have some knowledge of—that I was the one-parent/one-language principle when entrenched in may be a manifestation of both addressing their child. the foreign language effect and the typological distance effect to be discussed in Chapter 7 BSRN’s architecture is based on Elman’s (pp. 343–346). What the experience also illus- (1990) simple recurrent (SRN) model, which con- trates is the importance of maintaining a cerns a “distributed” connectionist network; that language for it to remain available for use. is, a network in which the represented entities such as words and phonemes are each stored not But there is also more formal evidence to in a single processing node (as in “localist” support the existence of language subsets and models) but distributed across many different

6. LANGUAGE CONTROL 299 processing nodes (see Chapter 3, p. 133 and (a) A simple recurrent Figures 3.8 and 3.11b). Figure 6.3a provides an network (SRN) model illustration of the structure of the network. consisting of three layers of nodes and a set of context The system consists of three layers of nodes: a nodes. The latter enable the layer of input nodes (24 in all), a layer of hidden model to retain a previous nodes (32 in all), and a layer of output nodes internal state of the model. (24 in all). In addition there is a set of context (b) The clustering by nodes (in SRN called a “buffer”), the inclusion of language and grammatical which allows these models to retain a previous class that emerged after the internal state of the model. Memorizing a prior model was fed with a large state is a prerequisite for regularities in the input number of French and presented to the model to be detected and, thus, English three-word sentences that were constructed from one French and one English mini-lexicon. Each mini-lexicon consisted of four subject nouns, four verbs, and four object nouns. Reprinted from French and Jacquet (2004), Copyright 2004, with permission from Elsevier. a prerequisite for learning simple grammar (see e.g., Murre, 2005, for details). Retention of the prior material is effectuated by copying the state of activation in the hidden nodes into the buffer after each learning cycle. BSRN processes a con- tinuous stream of sentences one word at the time and its task is to predict every next word. The input sentences were all constructed from 24 words: 12 French words and their 12 English translations. Each mini-lexicon consisted of four subject nouns (English: boy, girl, man, woman; French: garçon, fille, homme, femme), four verbs

300 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS (English: lifts, takes, sees, pushes; French: separation to be successful. A further point to souleve, prend, voit, pousse), and four object stress here is that language separation does not nouns (English: toy, ball, book, pen; French: jouet, seem to require the use of two separate under- ballon, livre, stylo). French (1998) stressed the fact lying neural substrates, because in the model that these words carried no semantic information functional separation emerged from patterns of and that he could have chosen other, arbitrary activation over one and the same set of nodes. symbols for the purposes of his simulation (in other words, the ensuing structure of the lexicon Li and Farkas’ (2002) SOMBIP exploits a dif- to be discussed next did not require semantic ferent architecture and different learning prin- input to be built). The language generator ciples. It combines the use of a type of learning combined elements from this 24-word vocabulary model called “unsupervised”—in which learning into legal sentences (e.g., boy lifts toy man sees pen occurs without an analogue of a supervisor or man takes book femme souleve stylo fille prend teacher who provides information on the targeted stylo femme pousse ballon woman takes book . . .). behavior—with the use of a word co-occurrence Language switching was only allowed at sentence detector (WCD) and also exploits the Hebbian boundaries and switching probability was fixed at learning principle mentioned earlier. The specific 0.001; that is, switching occurred once per 1000 unsupervised learning model used is Kohonen’s sentences. There were no special markers to (1982) self-organization model (SOM) that self- indicate a switch of languages nor were sentence organizes an input, extracting a representation boundaries marked. from it and projecting the extracted input onto a two-dimensional topological map. SOMBIP A total of 300,000 words (100,000 sentences) contains two such maps, one storing the lexical were fed into the model this way, and at various phonological forms of words and the other points during training the state of activation in storing word meanings. They are shown in the hidden-layer nodes was inspected for each of Figure 6.4a. Each map contains a collection of the words. French (1998) found that these hidden- nodes. At the beginning of learning, an input, say layer representations were distributed over many a lexical phonological form, randomly activates a nodes and highly overlapped between languages. set of nodes on the corresponding map according But a cluster analysis showed that they were also to how similar the input pattern happens to be, clustered into clear sets that reflected, on the one by chance, to the weight vectors of the nodes. The hand, the grammatical categories subject, verb, weight vectors of the activated nodes and their and object, and, on the other hand, the two neighbors on the map are then adjusted so that languages French and English (see Figure 6.3b). they become more similar to the input. Upon The latter finding demonstrated that a separation the next presentation of this or a similar input in language subsets had emerged despite the fact these nodes will, as a consequence, respond more that the input had not been explicitly marked for strongly to it. Upon further presentations of language. These language clusters and grammat- the input the collection of nodes that respond ical clusters were already stable after the presenta- will gradually become smaller up until the tion of 60,000 words (20,000 sentences). This moment the input gives rise to a very high level pattern of results was replicated in a further study of activation in just a few neighboring units. This wherein the lexicons contained 768 words each set is the ultimate representation of the input (instead of 12) and the number of nodes in the stimulus. layers of input, hidden, and output nodes was increased (to 48, 100, and 48, respectively). After The phonological forms that were fed into the the presentation of 30,000 sentences the lan- lexical phonological map were the lexical phono- guages separated out again. These simulations logical forms of the 400 most frequent words in with BSRN thus support the view, propagated the Chinese–English child corpus mentioned by Paradis and Grosjean, that no language tags above, consisting of 184 Chinese words and 216 or language nodes are required for language English words, together covering 56% of all word tokens in the corpus. The corresponding meanings

6. LANGUAGE CONTROL 301 (a) Two self-organizing maps, storing the lexical phonological forms of words (SOM 1) and word meanings (SOM 2), respectively. The stored representations are extracted from phonological word forms and word meanings presented during training. (b) The organization in SOM 2 after training. Chinese and English representations are clustered in different areas on the map. Similarly, words of different grammatical classes are clustered in different areas. Adapted from Li and Farkas (2002), Copyright Elsevier 2002. that were fed into the semantic map were not Figure 6.4b shows the organization in SOM 2, readily available but had to be derived first by the semantic map, after the network had been running the whole corpus through the WCD trained on the 400-word vocabulary. As can be one by one. For each word, the WCD computed seen, the Chinese and English representations on with which other words it co-occurred and how the map were not randomly intermingled but are often, only considering the 400 target words and clustered in different areas on the map. In other ignoring all other words. The result of this com- words, despite the fact that the material that was putation for each word was regarded as the fed into the WCD and the self-organizing maps word’s meaning (notice that this amounts to a did not contain any explicit information about definition of what constitutes a word’s meaning the language every single word belonged to, two that may not be shared by everyone). Finally, language subsets had emerged (but see Thomas & when WCD had done its job of extracting word Van Heuven, 2005, who argue that explicit meanings for the 400-item vocabulary, the process language membership information was provided of self-organization on the phonological and to the system after all). Furthermore, the words semantic maps could start. Hebbian associative within the separate lexicons were clustered in learning (“what fires together, wires together”) grammatical categories, and (not shown) within was ensured by presenting each phonological a grammatical category words with similar form and its associated meaning to the corres- meanings also tended to be grouped together. ponding maps simultaneously. These results all suggest that categories may

302 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS emerge from statistical learning (see Chapter 2 bilinguals can switch their languages on and off, for evidence from infant studies). A final note and that an “in” state implies the elements in the to make here is that Li and Farkas adhere subsystem are activated whereas an “out” state to the view that lexical word meaning exists implies they are deactivated. However, as already separately from more general conceptual know- mentioned earlier, this inference does not follow ledge (see Chapter 5, p. 234) and that the clusters imperatively from the in-out account of language that emerged on the semantic map concerned switching: Elements in the subsystem that is cur- lexical-semantic meaning. This offers a way to rently not accessed can nevertheless be activated reconcile the clear language separation that but this activation may be ignored by some emerged in this map with the rather general view control system. An in-out account can thus that in bilinguals meaning representations are be reconciled with the substantial evidence of largely (but not exclusively) shared between the language-nonselective activation reviewed in two languages. Chapters 4 and 5. To summarize, language subsets emerge from The second view, Grosjean’s language-mode the co-occurrence of linguistic elements and the theory (pp. 288–291), explicitly acknowledges that ensuing co-activation of their memory represen- there is always some minimal level of activation in tations, the co-activation creating a bond between the non-selected subsystem but that bilinguals these representations. The more frequent the can nevertheless control the output because the co-occurrence of particular linguistic elements level of activation differs between the two sub- and, hence, the more frequent the co-activation systems: The more highly activated subsystem, of their memory representations, the stronger the selected one, dominates the output. This the connections between the latter become. difference in degree of activation of the sub- Connectionist models of bilingual language systems emerges from both top-down (internal, acquisition that exploit this learning mechanism endogenous) and bottom-up (external, exogen- have shown language subsets to emerge even ous) sources; in other words, from the specific though the words and sentences that were fed intention of the speaker (e.g., to produce uni- into these models were not explicitly marked for lingual output in the selected language) and the language. This finding supports the view that nature of the input (unilingual or mixed). the bilingual language system can do without Although a different terminology is used, these language tags or language nodes or any other views are very similar to those assumed by piece of information that explicitly states the Paradis in his neurolinguistic theory of bilingual language identity of a linguistic entity. control (pp. 294–296). In that theory language choice results in differential activation-threshold LANGUAGE CONTROL IN A LANGUAGE- settings within the two language subsets, a pro- NONSELECTIVE ACTIVATION FRAMEWORK cess that is functionally equivalent to activating one subsystem while at the same time deactivating Introduction or inhibiting the other. So far, three views on bilingual language control Whereas the most central idea in Grosjean’s have been presented and in all of them the notion and Paradis’s conceptions of language control is of language subsets played a crucial role. The that language choice induces, proactively and early language-switching work of Macnamara globally, differential levels of activation in the and colleagues (pp. 283–286) assumed that bi- two linguistic subsets, the two views on control linguals go in and out of their language sub- that I will present here both stress that bilingual systems upon each language switch. Later language processing is basically language- authors have taken this account to imply that nonselective. This means that the representations in both language systems may be highly activated, at least during the initial stages of processing. One of these views accounts for language control

6. LANGUAGE CONTROL 303 in visual word recognition, the other for control As a result of this interconnectedness between in word production. The subset notion is not a word nodes and language nodes the latter operate central concept in either of these two views as “language filters” that modulate activity of although, as we shall see, in one of them it the word nodes in the word recognition system, appears to slip in after all, but in disguise. It con- as follows: An activated word node sends on cerns the bilingual interactive activation model activation to the corresponding language node, introduced earlier in another context (Dijkstra boosting the latter’s activation. In turn, this & Van Heuven, 1998; Grainger, 1993; Grainger & language node will then send inhibitory feedback Dijkstra, 1992). To account for language control, to all the word nodes of the other language, the model assumes the existence of two language deactivating the latter. Therefore, if the language nodes, one for each language. The second of the of the word presented on the current trial is two views to be presented here assumes that lan- different from the language of the previous guage control in word production can be secured word, the level of activation in the word node if the conceptualized message contains a lan- representing the current word is relatively low and guage cue. As a result, the words that are output by this slows down its recognition. This accounts the system will generally be ones belonging to the for the cost of language switching. Conversely, if targeted language. the language of previous and current word is the same one, the word node representing the Language control in bilingual visual currently presented word is on average more word recognition highly activated than any competing word nodes in the non-target language and this secures In Chapter 4 (pp. 177–179) I described the details language control. of BIA’s architecture and the way the model processes a visual word that is presented to it (see In addition to this bottom-up, stimulus-driven also Figure 4.4). I will not repeat that exercise language node activation, Dijkstra and Van here. It suffices to say that a word stimulus Heuven (1998) assumed that the activation of the activates, via two levels of feature and letter language nodes can also be influenced top-down nodes, orthographic word nodes that represent by non-linguistic contextual sources of informa- words of both languages. How then, with all this tion from outside the word recognition system language-nonselective activation going on in such as, in an experimental setting, the specific the word recognition system, is the input word instructions given to the participants or, in a ultimately recognized correctly and, specifically, natural context, specific knowledge about the how does the model account for the processing interlocutors. This top-down modulation of cost caused by experimenter-imposed language language node excitation then again affects the switches? It is here that the language nodes are activation of nodes within the word recognition brought into play. These two nodes, one for system. In this respect BIA resembles the views on each language, are connected by means of an control encompassed in Grosjean’s language- excitatory link to each of the orthographic word mode theory and in Paradis’s neurolinguistic nodes of the corresponding language. Further- theory, where language choice resulted in a more, each of them is connected by means of an differential setting of the activation levels of the inhibitory link to all orthographic word nodes representations in the selected and non-selected of the other language (see Figure 4.4). It was language. What is new is the assumption that this mentioned above that BIA encompasses a lan- differential setting is mediated by language nodes, guage subset organization in disguise. It is this constructs that do not exist in Grosjean’s language-specific linkage between orthographic language-mode theory or in Paradis’ activation word nodes on the one hand and language nodes threshold hypothesis. The implication of allowing on the other hand that provided the ground for language node activation and, as a consequence, that claim. activation of nodes within the word recognition system, to be influenced by top-down processes

304 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS is that word recognition according to BIA is not a prerequisite for performance in the semantic rigorously language-nonselective after all. categorization task. For this reason, the absence of a cost of language switching cannot easily be Whereas the above account attributes the cost reconciled (if at all) with a view that assumes the of language switching to a process of inhibition lexicon is the locus of the switch costs: Delayed within the word recognition system, more recently lexical access on the switch trial should have several authors have put forward the view that resulted in delayed meaning assignment and, the switching costs arise from processes outside therefore, in delayed semantic categorization. the mental lexicon (Green, 1998; Thomas & Other evidence is that the magnitude of the costs Allport, 2000; Von Studnitz & Green, 1997, varies with the specific requirements of the task 2002a, 2002b). To perform a particular task, a (e.g., Von Studnitz & Green, 1997) and that test control structure (also called a “task schema”) circumstances can be created in which language has to be set up that links outputs from a system, switching incurs a benefit rather than a cost here the word recognition system, with the (Thomas & Allport, 2000; Von Studnitz & Green, response required by the task. These task schemas 2002b). may either be retrieved from long-term memory or set up anew the moment they are needed. Two Dijkstra and Van Heuven (2002) developed a such schemas are the so-called “language task successor of BIA, BIA+, which acknowledges schemas”, each of which is operative when that language switch costs have a lexicon-external material in the language in question is processed. locus. This version of the model still contains the In this scenario the switch cost does not result language nodes, but the inhibitory connections from inhibition of the targeted representation from the language nodes to the word nodes of the within the lexicon, but from the inhibition of the other language have been removed. As a con- language task schema that is currently called for. sequence, language nodes no longer suppress In other words, the switch cost reflects the time it activation in word nodes of the other language takes for the currently appropriate task schema to and, therefore, word recognition is now strictly be recovered from inhibition (Thomas & Allport, language-nonselective initially. The only remain- 2000). ing function of the language nodes is that they serve as language tags. To account for the lan- The adherents of this alternative view come guage switch costs and a number of other effects, up with various sources of evidence, direct and such as the fact that the direction of interlexical indirect, to suggest that switch costs indeed arise homograph effects varies with the specific from inhibition of a specific task schema and demands of the task (see Chapter 4, p. 168) and not from inhibition of a specific node within the the effects of non-linguistic contextual infor- word recognition system. One is the finding that mation and participant strategies, Dijkstra and switch costs do not reliably occur in all situations Van Heuven added a task/decision system onto where one would predict them to emerge if the the original BIA word identification system. As a within-lexicon account were true. For instance, result, BIA+ has come to strongly resemble Caramazza and Brones (1979) did not obtain a Green’s (1986, 1998) model of “inhibitory con- switch cost in a semantic categorization task trol”, the model of bilingual control that I will in which Spanish–English participants were turn to further on. presented with word pairs each consisting of a category name and the name of an exemplar of Language control in bilingual word this or another category. Their task was to decide production as quickly as possible for each pair whether or not the instance belonged to the category. Categoriza- La Heij (2005) presented a view on language tion took equally long when the words in a pair control in bilingual word production that builds belonged to the same language or to different on earlier work by De Bot and Schreuder (1993) languages. Lexical access is a prerequisite for and De Bot (1992; see p. 228) and on extensions of meaning assignment, and meaning assignment is

6. LANGUAGE CONTROL 305 that work by Poulisse (1997) and Poulisse and the amount of conceptual information specified Bongaerts (1994). In developing his ideas he took in the preverbal message on the one hand and as starting point the common assumption that the the number of lexical elements that compete for word corresponding to the most highly activated selection on the other hand: The more elaborate lexical representation will ultimately be selected the information in the preverbal message, the and output by the system. In his specific solution fewer lexical items provide a satisfactory match of the control problem La Heij drew an analogy with this information and, consequently, the between, on the one hand, the state of mind larger the chance the intended item is output by monolinguals are in when they have more than the system. The reason is that, first, lexical items one lexical option available to them to label a are activated proportionally to the amount of particular conceptual content and, on the other overlap between the information contained by the hand, the bilinguals’ common state of mind of preverbal message and the information specified having two words (one in each language) for one in the lexical items, and, second, the most highly and the same concept. In the monolingual case, activated element is output by the system. In for instance, synonyms may exist that express the monolingual case, if the preverbal message one and the same concept, a concept may be contains the specification that, for instance, expressed with its category name (e.g., dog) or slang should be used, it is likely that the lexical with the name of a specific exemplar of the item containing this specification will be more category (terrier, or poodle), or a concept may be highly activated than a more formal “equivalent”. expressed in both a formal and a slang term. Consequently, it will be the slang word that emerges in the output. La Heij characterizes this If more than one lexical item matches the process as “complex access, simple selection”: If content of the preverbal message that is output by the process of conceptualization is so elaborate the conceptualizer (the component of the speech (complex) that its output contains all the essential production system that builds the conceptual information the speaker wants to express, the structure to be expressed in speech; e.g., Levelt, intended word falls naturally out of the process 1989; see Chapter 5, p. 224 for details), how then is in the far majority of cases. As he puts it, the intended lexical element selected from this set under those circumstances “[. . .] there is no con- of possible candidates and output by the system? vergence problem to be solved” (La Heij, 2005, In the monolingual literature this multiple p. 294). mapping of conceptual content onto lexical items and the problem it creates for language produc- Notice that this view presupposes that perfect tion is known as the “convergence problem” and synonyms do not exist. In principle, two different has led to much debate as to how the monolingual words mean different things, if only because they speaker solves it (e.g., Levelt, 1989; Roelofs, are used in different contexts (this being meaning- 1992). One of the solutions has been to assume ful). A slang word is not the meaning equivalent some sort of lexicon-external checking mechan- of its more formal counterpart, synonyms like ism, a monitor that before the selected item is thin and slim mean different things, and, more being output determines whether it is in fact the obviously, so also do dog and poodle. Levelt intended item. If it is not, the selected item is (1989) also adhered to this view that synonymy withheld, another one is selected from the set within a language does not exist, instead of alternatives, and the checking process starts embracing the “principle of contrast” (Clark, anew. La Heij characterizes this process as “sim- 1987, 1988), which states that all forms in a ple access, complex selection”. The “simple language contrast in meaning. This principle is access” part of this qualification refers to a considered to be an important driving force in “sufficing” conceptualizer, one that is sloppier language acquisition (when you encounter a new than could be. form, assume its referent is different from the referents of forms you already know and try to La Heij proposed a selection process that figure out what it is). presupposes a straightforward relation between

306 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS The analogous solution for the bilingual case it assumes control operates locally, on specific is to assume that, just as the intention to use lexical items, and not globally, by activating and/ formal language or slang, or a category name or a or deactivating a whole language subsystem. sub-ordinate term, the intention to speak in one Finally, a view of control in terms of a language language and not the other is part of the infor- cue as part of the conceptualized content to be mation contained in the preverbal message. In expressed suggests how fluent language mixing other words, no theoretical distinction has to be can come about in situations where such is made between, for instance, the information allowed or even desirable: The speaker can refrain about which register to use and the information from choosing a language, with the effect that the about which language to use, a view that has been preverbal message does not contain a language expressed by many other researchers in the field cue. A slightly different version of this account (De Bot & Schreuder, 1993; Paradis, 1997, 2001). has already been presented on p. 229. There it was Following Poulisse and Bongaerts (1994) and suggested that the language cue can be assigned Poulisse (1997), La Heij proposed that the lan- different weights in different situations. guage information in the preverbal message takes the form of a “language cue”. It is furthermore LANGUAGE CONTROL THROUGH REACTIVE assumed that each lexical item contains a “lan- SUPPRESSION guage tag” that specifies its language membership. The language cue and the rest of the conceptual Introduction information in the preverbal message cooperate in activating lexical items, a process that will When glancing over the figures that illustrate generally result in the highest activation for the performance in language comprehension and targeted item and, as a consequence, in its selec- production tasks in Chapters 4 and 5 it may be tion. Take as an example the English–French noticed that they all sketch various components bilingual who intends to express the concept of the language system and how these connect to “boy” in French. Among the conceptual informa- one another, but none of them includes a control tion in the preverbal message is the language cue mechanism that guides and regulates perform- +French. If lexical items indeed contain language ance in specific language tasks. Yet the language tags, French garçon, containing the French system is always used, both in natural situations language tag, will become more highly activated and in the laboratory, with a specific purpose, than English boy, which contains the English and depending on the specific goal it will be used tag. As a consequence, garçon will be selected and differently each time. I have already implicitly output. introduced this issue in Chapter 5, where I showed how one and the same part of the under- When comparing this model of language con- lying language system may be implicated in trol in word production with the other proposals naming pictures in L1 or in L2, in translating discussed so far, a distinguishing feature is that it words from L1 to L2 or vice versa, and plausibly does not require the proactive control assumed in in a number of other word generation tasks as Grosjean’s language-mode theory and Paradis’ well. The implication is that an account of lin- activation threshold hypothesis. Elements in guistic performance is never complete when it the non-target language do not have to be ignores a control system that enables the language deactivated, however slightly. The language cue system to be used as requested (in the laboratory) sees to it that a word in the intended language or intended (in a natural context). Perhaps the is output by enabling the activation in the most appealing evidence to support this claim is corresponding lexical entry to exceed that of its found in the bilingual aphasia literature, where equivalent in the non-target language. The model some forms of pathological language behavior also does not require the reactive suppression of the non-target language, to be discussed below. A further noteworthy feature of the model is that

6. LANGUAGE CONTROL 307 can only be understood by assuming an intact The inhibitory control model language system that is operated on by a failing control system (see p. 295). One researcher who A central construct in Green’s (1998) inhibitory has consistently stressed the importance of control model of bilingual language control is a efficient and unobstructed control operations in supervisory attentional system (SAS), which he fluent language use and of sufficient resources adopted from Norman and Shallice’s (1986) to “fuel” these processes is David Green (Green, theory on the control of behavior. In this theory 1986, 1993, 1998, 2005; Green & Price, 2001; SAS is a resource-limited control structure that is Price, Green, & Von Studnitz, 1999; Von Studnitz involved in the planning, regulation, and verifica- & Green, 1997, 2002a, 2002b). He and his col- tion of non-routine voluntary actions and that is leagues have played a major role in introducing assumed to reside in the prefrontal regions of these insights in research on bilingualism. the brain. SAS is a subsystem of a larger control system in which simple, well-learned actions are The insight that control plays an important carried out automatically by ready-made memory role in language behavior has also paved the way structures that specify action sequences, called to the awareness that bilingual language control “schemas”. The SAS supervises the routine and the systems involved may resemble, or even running of the schemas and intervenes when be the same as, the control processes and control necessary (e.g., to prevent action slips). It does so structures required in other situations where by altering the activation levels of the running multiple responses compete for selection. Accord- schemas. Unlike Norman and Shallice, Green ingly, a number of authors have embedded their applies the term schema not only to the above- views on bilingual control within a more general mentioned ready-made structures in long-term theory of the control of action. Some of them memory that have resulted from past experiences envision bilingual language control as a process and that underlie automatic, routine behavior, whereby the inadvertent selection of a competitor but also to “mental devices or networks that in the contextually inappropriate language is pre- individuals may construct or adapt on the spot in vented by a process of “reactive inhibition” or order to achieve a specific task” (Green, 1998, “reactive suppression” of the competitor. It is p. 69). In his view a task schema thus appears to especially this view that is advanced in Green’s be the equivalent of an understanding of the task work to be presented here, but the idea is also instructions by the participants; that is, their central in the influential work of Ellen Bialystok mental representation of the instructions, com- to be presented in the next chapter, on the cogni- plemented with a setting of the system that tive consequences of bilingualism (pp. 393–396). enables them to perform the task as requested. Here, I will first present Green’s inhibitory con- Task schemas for tasks that have been previously trol model, including a discussion of a study by performed (e.g., reading words aloud) can be Meuter and Allport (1999) that had a major retrieved from memory as such or adapted, if impact in this field of study for two reasons: In necessary, and schemas for tasks that are novel this study the language switching paradigm that (e.g., lexical decision) have to be constructed on Macnamara and his colleagues first used to study the spot. In Green’s model, which is illustrated bilingual language control (pp. 281–283) was in Figure 6.5, this process of schema retrieval, revitalized and the paradigm was modified in adaptation, and construction is commanded by such a way that it encompassed what has come to the SAS. be regarded the litmus test of the phenomenon of reactive suppression. I will then proceed by In the model, the “conceptualizer” is a non- discussing a number of studies that included linguistic system that builds conceptual represen- Meuter and Allport’s critical test of suppression tations from information in long-term memory. and produced results suggesting that control is In doing so it is driven by a goal to achieve not secured through reactive suppression in all some effect through the use of language. The cases. intention to produce a word in a specific language

308 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS The inhibitory control model of bilingual language control. The bilingual lexico-semantic system stores the bilingual’s word knowledge. The SAS (supervisory attentional system) is a general control system that regulates language behavior by constructing, or retrieving from memory, so-called “language task schemas” that see to it that the bilingual language user performs the targeted task. From Green (1998) with permission from Cambridge University Press. is part of the conceptual representations (cf. the When the goal is to perform a particular task, language cue in, e.g., La Heij, 2005). The the SAS installs the relevant language task “bilingual lexico-semantic system” is part of the schema. This schema then controls behavior actual language system and stores the bilingual’s (i.e., regulates the output of the lexico-semantic word knowledge; for instance, knowledge regard- system) in two ways: by altering the activation ing the words’ forms and meanings. Especially levels of the lemmas within the lexico-semantic central in Green’s model is a word’s lemma, system and by inhibiting pending outputs from which specifies its syntactic properties such as its that system (Green, 1998, p. 69). When, for gender and is assumed to be used in both pro- instance, the bilingual is asked to name pictures in duction and comprehension tasks. Each lemma L2, the relevant schema (presumably one that contains a language tag specifying the word’s exists as such in memory because the task is a language membership. During, for instance, word rather natural one) is activated and enables task production, conceptual information from the performance by increasing the activation level of conceptualizer activates lemmas in the lexico- lemmas with an L2 language tag in the lexico- semantic system proportional to the degree of semantic system, decreasing the activation of L1- information shared between the conceptual repre- tagged lemmas in the system, and by reactively sentations and the lemmas in the lexico-semantic inhibiting imminent L1 outputs from the system. system. So far Green’s model is very similar to the If, instead, the task is picture naming in L1, the views of, among others, La Heij (2005) presented opposite pattern of control is established by the above. But two of its characteristics set the model SAS. When the task is translating L1 words into apart from the other theories on language control L2, the SAS installs a schema that specifies L1 presented so far. The first concerns the feature as the input language, inhibits the schema for already introduced above—the (explicit) inclusion naming the input in L1, and activates a schema of a control mechanism (SAS) that regulates lan- for producing L2 as the output. The degree of guage behavior by retrieving or setting up task reactive inhibition is proportional to the level of schemas that specify in what way the language activation of the non-target language’s lexical system must be used to perform the requested items: the more active a lemma in the non- task. Its second unique feature is a process of response language, the more it will be inhibited. reactive, local suppression of contextually Therefore suppression will generally be stronger inappropriate lexical elements. in circumstances where L2, often the weaker lan-

6. LANGUAGE CONTROL 309 guage, is the language of output (and stronger into the stronger language L1 than for a switch L1 has to be suppressed) than when L1 is to be into the weaker language L2. output. (This is so because the memory represen- tations of the linguistic elements belonging to Meuter and Allport (1999) derived this pre- the weaker language will generally have lower diction from similar asymmetries that have activation levels than those belonging to the been observed when participants have to switch stronger language.) between an easy, behaviorally relatively dominant task (e.g., word reading by well-practiced readers) From this description of the workings of the and a more difficult, behaviorally weaker task model it may be inferred that Green assumes con- (e.g., color naming). The cost that accompanies a trol can be exerted, on the hand, proactively and switch between tasks is typically larger when the globally, by adapting the activation levels of participant has to switch from the weaker (color all lemmas in both language subsets before naming) to the more dominant (easier, more task performance starts and, on the other hand, practiced) task (word reading) than with a switch reactively and locally, by suppressing the acti- in the reverse direction, from dominant to weak. vation in the lemmas of words from the non- This counterintuitive effect is explained in terms response language that still threaten to slip of differences in the required level of suppression through. In theory, the model could do without of the non-target task, in combination with a proactive, global control altogether: Even if the phenomenon called “task set inertia”: To enable ultimately most highly activated element in the task performance, the non-target task must be bilingual’s lexicon were to correspond to a word actively suppressed and the target task must in the currently inappropriate language, selection be activated. The stronger the non-target task, of this word could be prevented through late the more it has to be suppressed in order not to suppression. interfere with target task performance. Task set inertia is the phenomenon that the task set of the As mentioned, Meuter and Allport (1999) previous trial carries over into the current trial. In provided a result that has been regarded as the other words, with a switch from color naming to litmus test of reactive suppression because the word reading, the relatively strong suppression of result in question indeed seems to suggest that the dominant word-reading task on the previous reactive suppression is proportional to the level trial lingers on to the subsequent trial, causing a of activation of the to-be-suppressed lemmas. relatively large delay in producing the word- Meuter and Allport used an adapted version of reading response now required (Allport, Styles, & the numeral-naming task of Macnamara et al. Hsieh, 1994). Meuter and Allport (1999) hypoth- (1968; see p. 285) in which unbalanced bilingual esized that the language-switching task imple- participants named the Arabic numerals 1 to 9 ments this general paradigm of switching in either their L1 or L2. The numerals were between a well-practiced and a less-practiced task, presented in lists of various lengths in which and, therefore, predicted a similar asymmetry in switch and non-switch trials occurred in an the data. The data are presented in Figure 6.6. unpredictable order within one and the same list. The language to use on a particular trial was As shown, language switching into both lan- indicated by the color of the numeral’s back- guages incurred a cost, a finding that replicated ground, blue or yellow. A crucial aspect of this a similar result in Macnamara et al.’s (1968) study was that, unlike in Macnamara et al.’s seminal study and that has since been replicated study, response times were registered for every by many other researchers (e.g., Costa & single trial, switch and non-switch. This way the Santesteban, 2004b; Hernandez et al., 2000; authors could test their “relative strength hypoth- Jackson, Swainson, Cunnington, & Jackson, esis”, which states that the size of the switching 2001; Kroll et al., 2000). But presently of more costs depends on the relative strength of the interest, the predicted language-switching asym- bilingual’s two languages: The switching cost was metry occurred as well: The cost of switching was (paradoxically) expected to be larger for a switch larger for switches into the dominant L1 (143 ms)

310 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS Mean numeral naming time (in ms) as a function of response language and trial type. On non-switch trials naming in L1 is faster than in L2. On switch trials naming is faster in L2. Data from Meuter and Allport (1999). than for switches into the weaker L2 (85 ms), with collected 30 years later—that there is nothing the effect that on a switch trial a response in special about language switching but that it con- stronger L1 took longer than in weaker L2 (which cerns a manifestation of task-switching behavior is why the effect is sometimes called “para- in general. To repeat the relevant quote: “lan- doxical”). The fact that the response pattern was guage switching [. . .] seems to require no psycho- similar to the one obtained with other task pairs logical skill peculiar to bilingualism, but rather consisting of a dominant and a weaker task sup- a skill which is equally applicable in a larger ports the authors’ theoretical starting point that number of operations in which persons are language switching is just a special case of this asked to switch modes of response rapidly” more general methodology and that one and the (Macnamara et al., 1968, pp. 213–214). What is same interpretation, suppression combined with more, Macnamara and his colleagues already task set inertia, may account for this specific data provided direct evidence to support their claim pattern in all these cases. by showing that the language-switching task produced exactly the same switching costs as a As mentioned earlier (p. 281), Macnamara condition where participants switched between et al. (1968) did not measure the cost of language two non-language tasks. It is for these reasons switching locally, at the very point where it that I characterized Macnamara’s work as having occurred, but by subtracting the overall response a modern flavor to it. time for complete blocks of non-switch trials from the overall response time for complete Data that qualify the inhibitory control blocks of switch trials. For that reason they could model never have detected the asymmetry in the switch costs that is critical to be able to assign it the If an asymmetry in the required level of suppres- present interpretation: that it results from sion due to an uneven command of the two unequal suppression of a stronger and a weaker languages indeed underlies the asymmetrical task (or “task schema”, in Green’s terminology). switch costs, then the smaller the proficiency Nevertheless, as we have seen, one of the main difference between the two languages, the smaller conclusions these early authors drew from their the asymmetry in the size of the switch costs data is remarkably similar to Meuter and should be. In balanced bilinguals, where both Allport’s account of the language switch data

6. LANGUAGE CONTROL 311 languages are equally strong, the asymmetry in between their equally strong L1 and L2. The the switch costs should even disappear altogether. researchers concluded (see also Costa, 2005) that Meuter and Allport (1999) performed a number language selection in participants highly pro- of post-hoc analyses on their data that suggest ficient in at least two languages might not come this to be the case. Costa and Santesteban (2004b) about through a process of reactive inhibition subsequently tested this prediction in a series of of the non-response language. Instead they five experiments, using a picture-naming task suggested that in the process of developing a high instead of a numeral-naming task. In a first proficiency in two languages, bilinguals may experiment, using this new task they replicated develop a control mechanism that somehow only Meuter and Allport’s finding, testing unbalanced takes lexical elements of the response language Spanish–Catalan and Korean–Spanish bilinguals. into consideration in the selection process, For both groups an asymmetrical switch cost ignoring activated lexical elements in the non- was obtained, with switching into the dominant response language. language producing the largest switch cost. More interestingly, in three experiments in which A further study that delved yet deeper into this balanced Spanish–Catalan bilinguals highly issue (Costa et al., 2006) extended and qualified proficient in both languages were tested, the these conclusions. One of the questions the asymmetry did indeed disappear. authors posed was whether such a language- specific selection mechanism perhaps only So far the results are consistent with the pre- develops in participants of the type tested by dictions of the inhibitory control model, suggest- Costa and Santesteban (2004b); namely, early ing that in both balanced and unbalanced bilinguals speaking two similar languages. The bilinguals language control requires reactive data of this new study suggested that neither age suppression of the non-target language; in un- of acquisition of the L2 nor language similarity balanced bilinguals the degree of suppression is plays a critical role: Symmetrical switching costs especially large when weaker L2 is the response also showed up when bilinguals who were highly language and stronger L1 the language to be proficient in two dissimilar languages (Spanish suppressed, whereas in balanced bilinguals the and Basque) switched between these languages. degree of reactive suppression is always the same. Furthermore, symmetrical switching costs turned But Costa and Santesteban (2004b) considered up in a group of Spanish–English bilinguals an alternative interpretation of the absence of an who had grown up in a monolingual Spanish asymmetry in the switch costs shown by balanced environment and had started to acquire English bilinguals. What if they exploit a qualitatively later in life but were quite fluent in it at the time of different sort of selection mechanism, one that testing. Yet a further experiment showed that for does not require reactive inhibition of the non- the symmetrical switching cost to occur it was no response language at all, as had been suggested by prerequisite that a strong L1 was one of the test earlier studies of this research group that had languages: Whereas Costa and Santesteban exploited the picture-naming interference task (2004b) had shown a symmetrical switching cost (see Chapter 5, pp. 242–245)? To answer this between strong L1 and weak L3, the present study question, they looked at the switching behavior showed the same result when participants highly of early balanced, proficient Spanish–Catalan proficient in both L1 Spanish and L2 Catalan bilinguals who were learners of yet another lan- switched between their strong L2 Catalan and guage (L3) in a switching task involving L1 and weak L3 English. L3. The reactive suppression account predicts that under these circumstances the asymmetrical So far it appeared that bilinguals and tri- switch cost should re-emerge. However, contrary linguals who are highly proficient in at least two to this prediction the switch costs were equally languages, irrespective of the age at which they large when switching into the L1 and the L3, as had started to acquire their L2 and the similarity was the case when these bilinguals switched of their L1 and L2, have developed a language- specific selection mechanism that does not involve

312 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS inhibitory reactive control and that they can menter imposed. As mentioned before, in natural apply generally when switching between any two language use bilinguals themselves determine of the languages they know to at least some which language to use, and language switching extent. Two further experiments, however, proved does not always slow down processing but may this latter conclusion to be wrong and showed even increase speech fluency (pp. 288–289). In an that there are limits to the exploitation of this English–Spanish study Gollan and Ferreirra non-inhibitory selection mechanism: The (2009) obtained evidence to suggest that the asymmetrical switch cost (with, again, a larger asymmetrical switch costs typically observed in cost when switching into the stronger language) unbalanced bilinguals result from the fact that in reappeared when Spanish–Catalan–English– the pertinent studies they were not free to choose French quadrilinguals, highly proficient in L1 the response language. In addition to an English- Spanish and L2 Catalan but with a much weaker only and a Spanish-only condition in which all command of L3 English and a very feeble com- pictures had to be named in the same language, a mand of L4 French, were tested in English voluntary switching condition was included in and French. This finding suggests that for the which the participants were told they could either language-specific selection mechanism to be name a picture in English or Spanish (“just say operative at least a high command of one of the whichever name comes to mind most quickly”; test languages is required. A final experiment p. 643). The participants were divided into two suggested that this is a necessary but not a groups, one consisting of balanced bilinguals and sufficient requirement, because an asymmetrical a second consisting of unbalanced bilinguals switch cost was also obtained when a further dominant in English. One finding of interest was group of participants switched between strong L1 that most of the participants in the voluntary and a small vocabulary they had acquired just switching condition indeed switched voluntarily before the actual picture-naming study. between the languages even though switching generally involved a cost in time. This finding To account for the complex pattern of data testifies to the robustness of the switch costs. But gathered in their studies, Costa and his colleagues of special interest in the present context was the tentatively hypothesized that a prerequisite for finding that the switch costs observed in this con- the language-specific selection mechanism to be dition were symmetrical in the balanced and functional is that words are solidly established in unbalanced bilinguals alike. In other words, when language-specific lexicons. If this requirement is language switching is voluntary the response not met, all participants, even those highly pro- pattern observed for unbalanced bilinguals ficient in two non-test languages, have to resort resembles that of balanced bilinguals, and neither to the inhibitory control strategy that—so the pattern suggests that control is secured by means asymmetrical switch costs suggest—is the only of a reactive inhibition. strategy available to unbalanced bilinguals (but see below). What remains to be explained under The studies by Albert Costa and his group and this account is why switch costs occur under all by Gollan and Ferreira (2009) narrowed down the circumstances, even when selection does not circumstances under which bilingual control exploit the reactive suppression mechanism but a might be exerted by means of a reactive inhibi- language-specific selection mechanism and the tory control mechanism. Finkbeiner, Gollan, and costs are symmetrical. The most plausible Caramazza (2006) went a step further, producing explanation of these costs is that they index the a set of data which, according to them, indicates resetting that is needed (i.e., the installment of a that language control might never come about new language task schema) when a new stimulus– by means of reactive suppression of the non- response association is required. response language, and that evidence suggesting otherwise may be due to a specific feature of the All studies discussed in this section so far stimuli typically used in these studies: They are examined the effects of language switching under usually “bivalent”, meaning that each of them circumstances where the switches were experi-

6. LANGUAGE CONTROL 313 can take two values, its name in L1 and its name persisting inhibition of language A in ABA in L2. The experiment in question involved series sequences, in which the switch from A to B had of numeral-naming and picture-naming trials. involved reactive inhibition of A. In other words, The pictures all had to be named in L1 (they were and contrary to Finkbeiner et al.’s (2006) con- “univalent”), whereas the numerals had to be clusion, the data from these studies support the named either in L1 or in L2 (they were “biva- inhibitory control model. lent”). The participants were all unbalanced bilinguals with English as their L1 and an L2 that Summary and conclusions varied between them; that is, they were the type of participants that in the experiments by Costa and The occurrence of asymmetrical switch costs in his associates always seemed to use the inhibitory the majority of conditions testing unbalanced control mechanism. The numeral-naming data bilinguals suggests that an account of language once again showed a pattern consistent with a control in terms of reactive suppression of elem- reactive inhibition account of language control: ents of the non-response language as assumed by Responses were slower on language switch trials the inhibitory control model at least applies to than on non-switch trials and the switch cost bilinguals of this type when they are forced to was relatively large when switching was from switch between the languages. Using a different L2 into L1. But contrary to prediction, picture paradigm, the studies by Philipp and her col- naming (always in L1) was not slower when the leagues (Philipp et al., 2007; Philipp & Koch, preceding numeral was named in L2 than when 2009) provided converging support for the it was named in L1. Accordingly the authors model. But the joint data also indicate that rejected an account of language control in terms balanced bilinguals might secure language con- of reactive inhibition of the non-response trol by means of the language-specific selection language. mechanism proposed by Costa and his colleagues. It remains to be seen whether and under what Finally, two studies by Philipp and collabor- circumstances the third control process discussed ators (Philipp, Gade, & Koch, 2007; Philipp & in the preceding sections—differential proactive Koch, 2009) contributed to the present debate by activation of elements in the two language subsets introducing a new language-switching paradigm resulting from the intention to use one language that enables a direct measurement of the reactive and not the other—contributes to language inhibition assumed by the inhibitory control control. model. In these studies German–English–French trilinguals named visual stimuli (numerals and Whereas the evidence in support of the inhibi- colors) in all three languages, a cue indicating tory control model’s assumption of reactive which language to use in naming the current suppression is somewhat mixed, the model’s stimulus. Two types of stimulus triads were core assumption that language control is effected included, so-called “n-2 repetition sequences” by a control system external to the language and “n-2 non-repetition sequences”. Presented system proper has by now become widely with the last stimulus of an n-2 repetition accepted. The language system can be used in a sequence, the participants had to switch back to multitude of different ways and the number of the language used to name the first stimulus of different options is especially large in the case the triad (ABA; e.g., naming in German followed of bilingual and multilingual language users. by naming in English, followed by naming in Exhibiting the linguistic behavior that is desired German). In contrast, n-2 non-repetition under specific circumstances (e.g., speaking in sequences required responses in three different L1, naming an object in L2, counting in L3, trans- languages (CBA; e.g., French, English, German). lating from L1 to L2) requires the involvement Responding to the third stimulus took longer in of a control system which exploits the language sequences of the ABA type than in CBA system in accordance with the language user’s sequences, a finding the authors attributed to current goals, monitors performance, and keeps

314 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS it on track. Another important assumption of and slows down the response on a switch trial. the inhibitory control model is that the control For similar reasons, a switch between pairs of system involved is not dedicated exclusively tasks involving one and the same language will to language processing but is implicated in the delay responding. Once the system has been reset control of behavior in general. The fact that to accommodate the current task requirements, asymmetrical switch costs have also been relatively little control may be required to observed in experiments where the participants maintain the existing state of control. This idea switched between pairs of tasks involving one and was already expressed in an earlier quote taken the same language (e.g., between color naming from Macnamara et al. (1968), to be repeated in and word naming) supports this hypothesis, as abbreviated form here: “. . . the bilingual once does the fact that, just as in studies on language started in one language can forget about which switching, in these task-switching studies the language he is speaking and yet obey the rules of asymmetrical switch cost does not always occur that language” (see p. 284 for the full quote). (Monsell, 2003). To conclude, it appears that language switch- Further evidence that a more general control ing is in fact a type of task switching that is taken system takes care of language control comes care of by a general control system that regulates from studies showing that bilinguals outperform both language processing and other forms of monolinguals on a number of non-linguistic tasks cognitive processing. Another way of putting this that demand a high level of cognitive control. is to claim that there is no such thing as a mental The hypothesis that motivated these studies was language switch. that bilinguals have become experts in cognitive control in general because their bilingualism CONTROL IN SIMULTANEOUS INTERPRETING requires that they incessantly control their lan- guage output. This line of work will be presented A characterization of translation and in Chapter 7 (pp. 393–396). A final source of simultaneous interpreting evidence comes from brain research showing that the brain regions activated during tasks that One specific form of language switching has been require the control of language are the same as neglected in this chapter so far—one that is those activated in tasks that require control in practiced by most bilinguals some of the time in other cognitive domains. In both cases specific naturalistic communicative settings. It concerns regions in the frontal lobe appear to be implicated “translation”, a term that refers to any act of (see Chapter 8, pp. 435–446). rephrasing a message in one language, the “source” language, in another language, the Particularly important in the context of the “target” language, while maintaining semantic present chapter (which primarily accounts for equivalence. Word translation, where the source language-switching effects) is that assuming a language input is a single word, has already been language-external control system that enables dealt with in earlier chapters (pp. 136–137 and the language system to be used according to the 261–265). In this section I will primarily deal with current goals provides a simple explanation for the translation of larger linguistic segments, the ubiquitous language-switching costs in both focusing on issues of control. Translating is prac- comprehension (pp. 286–288; and see Jackson, ticed in different ways (see Table 6.1). They differ Swainson, Mullin, Cunnington, & Jackson, 2004, from one another with respect to the modalities for a more recent study on receptive language of input and output (writing, speech, sign lan- switching) and production (the present section): guage, and combinations of these) and the timing Language switches in both input and output of the input and output relative to one another. necessitate a reset of the control system so that For instance, in consecutive interpreting the the task to be performed on the previous trial can be interrupted and the currently requested one installed. This resetting of the system takes time

6. LANGUAGE CONTROL 315 TABLE 6.1 Forms of translation Translation (used in a broad sense): All tasks where a message expressed in SL is transposed in TL. Translation (used in a narrow sense): Converting written SL into written TL. This form is also called text-to-text translation. Simultaneous interpreting: Converting spoken SL into spoken TL. In doing so the interpreter listens and speaks at the same time most of the time, trying to keep the “ear–voice span” (the lag between an input segment and the corresponding output) as short as possible. Consecutive interpreting: As in simultaneous interpreting SL and TL are both spoken, but listening and speaking occur in succession: The interpreter only starts to speak after the speaker has finished speaking. Semi-consecutive interpreting: As consecutive interpreting, but listening and speaking alternate: The speaker interrupts his or her speech once in a while, at which points the interpreter takes over to translate the prior spoken fragment. Sight translation: There is no speaker delivering a speech but a written SL text, which the interpreter translates orally into TL. Simultaneous interpreting with text: A speaker reads his or her text aloud from paper while the interpreter reads a copy of this text and produces an oral rendition simultaneously. Spoken-language translation: Converting one spoken language into another spoken language (simultaneously, consecutively, or semi-consecutively). Sign-language translation: Converting spoken SL input into a sign language (simultaneously, consecutively, or semi- consecutively). Machine translation: Using a computer to translate between natural languages. Typically interactive systems are used, in which human beings pre-edit or post-edit the text that is input into or output by the computer. Transliteration: The conversion of written names of people, places, and institutions from one script into another script, as when Russian SПУТНИК becomes SPUTNIK in English.* Transcription: The conversion of the sound of names of people, places, and institutions between languages written in different scripts, a process that often results in different spellings for one and the same name (Tchaikovsky, Tsjaikowskij, Csajkovszkij).* The term “translation” is used to refer to any act of converting a message in one language (the “source language”; SL) into another language (the “target language”; TL), while maintaining semantic equivalence. Based on Crystal (1987), De Groot (1997), and Gile (1997). * Examples taken from Crystal (1987). speaker and interpreter take turns, never speaking structure, e.g., Japanese) and SVO languages at the same time, whereas in simultaneous inter- (e.g., English). When presented with an SOV preting the speaker’s speech and the interpreter’s source language sentence to be translated into an rendition of it overlap approximately 70% of the SVO target language sentence the interpreter time (Chernov, 1994). cannot start his or her rendition of the object structure before the speaker has output, and the Each type of translation task is characterized interpreter has translated, the verb structure. In by a unique set of processing and storage the meantime, the interpreter must memorize the requirements and, therefore, each requires its own input’s object structure. The amount of attention unique set of skills to be performed optimally. to be devoted to comprehension, production, For instance, a simultaneous interpreter must and memorizing fluctuates with the continuously constantly divide attention between at least changing demands posed by the input (due to three task components: comprehending input, variations in, e.g., its rate, clarity of diction, producing output, and memorizing parts of the grammatical complexity, information density, analyzed input that must await expression in the and information complexity), the output (e.g., output because of sequencing differences between whether or not it contains tongue-twisters; the source language and the target language. To whether or not the concepts to express are lexical- illustrate the memory component, consider the ized in the target language), and the amount of difference between SOV languages (in which information to be temporarily retained (see sentences typically have a subject-object-verb

316 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS Christoffels & De Groot, 2005, for a review). This secutive interpreting were related to different necessitates the constant regulation of the exact sets of skills: The tests that reflected an ability to portion of attention that must be allocated to reconstruct information from memory and the each of the ongoing activities and this orchestra- examination ratings on consecutive interpreting tion, the “coordination component”, is regarded loaded highly on one factor; tests that all required as a fourth attention-demanding component of speed and accuracy of perceptual and productive simultaneous interpreting (Gile, 1995, 1997). linguistic processes and the examination ratings Finally, there is an ongoing debate about whether on simultaneous loaded highly on a second a complete model of simultaneous interpreting factor. should also encompass a separate translation component (e.g., Anderson, 1994; see the “trans- The (im)possibility of simultaneity coding” view of interpreting presented in the next section). So, in all, simultaneous interpreting The characteristic of simultaneous interpreting appears to challenge working memory excessively. that is likely to contribute most to its complexity In contrast, the mental bottleneck in consecutive is its eponymous component: the simultaneity of interpreting is the relatively long time interval a number of processing components, including between input and output, which creates the comprehension and production. The simultaneity problem that parts of the analyzed input may of these two task components per se is not a have dissipated from memory the moment the unique characteristic of simultaneous interpret- interpreter must take the floor. To compensate for ing because in monolingual speech production such memory loss the consecutive interpreter speakers are known to monitor their own speech. often jots down notes while the speaker is In doing so they exploit their comprehension delivering his or her speech. These can later help system (see Chapter 5, p. 224 and Figure 5.1). to reconstruct any lost information. What is, however, unique to simultaneous inter- preting in this respect is that interpreters have to To substantiate the present claim that different comprehend somebody else’s speech—speech that versions of the translation task are enabled by they have not conceptualized themselves. There different sets of skills, let us consider a study is evidence to suggest that at the same time they by Gerver, Longley, Long, and Lambert (1984). still monitor their own speech to some degree These authors provided a detailed analysis of (e.g., Gerver, 1976; Isham, 2000), although they consecutive and simultaneous interpreting, fail to detect errors more often than regular focusing on what it takes to be able to rapidly speakers do (Moser, 1978). This relative in- grasp and convey the essence of the speaker’s sensitivity to errors may be due to a conscious input. Based on this analysis they developed a strategy of interpreters to pay as little attention battery of tests, 12 in all, that were administered as possible to their own output in order to be to students selected for a graduate course on able to concentrate more on the incoming consecutive and simultaneous interpreting. The message (Fabbro & Darò, 1995; Moser-Mercer, performance of the participants on all these tests Frauenfelder, Casado, & Künzli, 2000). was correlated with the ratings they received in their final interpreting examinations. Two Christoffels and De Groot (2004) provided observations suggested that each of these tests support for the present suggestion that the assessed an ability that is relevant for skillful simultaneity requirement of the task contributes interpreting: First, all participants who passed the importantly to the task’s complexity. They examinations scored higher on all tests than all studied the effect of simultaneity of comprehen- participants who failed the examinations and, sion and production in an extremely simplified second, all test scores were positively correlated version of the task: The participants, reasonably with the examination ratings. But of special inter- fluent Dutch–English bilinguals without any est here was the finding that, as shown by a factor professional interpreting experience, were asked analysis, simultaneous interpreting and con- to interpret simple short sentences in two con-

6. LANGUAGE CONTROL 317 ditions. In the simultaneous condition they were least some subcomponents of comprehension instructed to start translating on-line; that is, as and production simultaneously. What we do not soon as possible after the onset of the input know, though, is exactly which ones this concerns. sentence; in a delayed (off-line, consecutive) Both speech comprehension and speech pro- condition they were asked to start delivering duction involve automatic subcomponents their translation after completion of the input (e.g., perceptual analysis and articulation) as well sentence. Even in this low-demand substitute as attention-demanding subcomponents (e.g., for the real task, performance deteriorated as a conceptualization of input and output). By result of the simultaneity requirement: The definition, automatic processes—if executed by quality of the output was worse, fewer words were different processing mechanisms—can be per- output, and less of the input material was recalled formed simultaneously effortlessly. Similarly, in the simultaneous condition than in the con- automatic processes are often thought not to secutive condition. That the adverse effects of interfere with ongoing attention-consuming pro- simultaneity remain after task training was cesses. A processing bottleneck typically only demonstrated by Darò and Fabbro (1994) and occurs when more than one of a task’s subcom- Lambert (1989). These authors compared the ponents draws on our limited attention resources memory performance of advanced interpreting and when, at the same time, together they require students following listening and simultaneous more resources than actually available. The interpreting. Despite the fact that simple question of present interest is therefore whether materials were used (short stories built from interpreters can truly perform a number of the simple words and syntax), presented at a modest attention-demanding subcomponents of com- input rate, the results showed poorer recall in prehension and production simultaneously or the interpreting condition than in the listening whether, instead, at any moment in time they condition. Similarly, Gerver (1974) had interpret- manage to somehow interweave one of the ing students answer comprehension questions attention-demanding comprehension com- following, among other things, just listening to ponents with one or more of the automatic pro- stories or after simultaneously interpreting them. duction components (or, vice versa, one or more In the interpreting condition fewer questions were of the automatic comprehension components answered correctly than in the listening condition. with one of the attention-demanding production In all four of these studies the participants inter- components). preted from a spoken language to a second spoken language. Interestingly, in a study in This question concerning the possibility of which the target language was signed instead “divided” attention has been a main area of of spoken, recall following interpreting (by pro- research since the middle of the twentieth century. fessional sign-language interpreters) equaled The typical studies addressing this question recall following listening (Isham, 2000; Isham & are dual-task or multiple-task studies. In these Lane, 1993). This suggests that the relatively poor studies, as in simultaneous interpreting, par- performance following spoken-language inter- ticipants have to perform two or more tasks preting is somehow caused by the requirement of concurrently and performance under dual- or concurrent listening and speaking. multiple-task conditions is compared with per- formance in a single-task control condition. Two An important note regarding the task’s early influential studies of this type (Hirst, Spelke, simultaneity is in order here; namely, that it is Reaves, Caharack, & Neisser, 1980; Spelke, Hirst, not at all clear in what sense it is actually & Neisser, 1976) tackled the question of whether simultaneous. We do know that in simultaneous attention can be divided over several attention- interpreting the speaker and interpreter produce demanding processing operations simultaneously speech simultaneously most of the time, so in in an experimental set-up that, intuitively, seems this sense the task is obviously simultaneous. The to be equally as demanding as, or even more implication is that the interpreter executes at demanding than, simultaneous interpreting. A

318 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS small number of participants were trained in texts enable anticipation (Adamowicz, 1989). reading stories while concurrently writing to Consistent with the present suggestion that dictation sentences unrelated to the stories. (This anticipation facilitates interpreting, Gerver and dual-task setting is plausibly more demanding his colleagues (1984) have shown that the inter- than simultaneous interpreting because it involves preting performance of student interpreters is the processing of two unrelated contents; in correlated with their performance on a “cloze simultaneous interpreting it is one and the same test”, a test that required them to restore content that must be understood and rephrased materials that had been deleted from connected in the other language.) After massive practice in discourse. One of the skills this test indexes is which the task demands were gradually increased, the ability to engage in anticipatory processes: to it appeared that one or two participants managed predict what is yet to come on the basis of an to perform both tasks with comprehension while understanding of the prior discourse. their reading speed was as fast as when they read without concurrent writing. Because both reading Kempen et al. (2009) provided indirect, mono- for comprehension and writing sentences to dicta- lingual evidence to suggest that whatever goes on tion require attention resources, the authors con- in parallel in simultaneous interpreting it is not cluded that simultaneity of more than one the grammatical decoding and encoding com- attention-demanding activity appears possible, at ponents of source language comprehension least for some individuals after massive training. and target language production. They had their Yet the possibility could not be definitely ruled participants perform either one of two tasks: out that, in fact, these participants had alternated reading with correction, and reading and para- rapidly between the reading and writing-to- phrasing. In both cases visually presented dictation tasks and that this was what they had sentences appeared on the screen phrase by learned to do through extensive practice. phrase. In the reading with correction task sentences with a direct-speech construction had Similarly, in simultaneous interpreting the to be read aloud and, if they contained an error, interpreter may rapidly alternate mentally corrected on line. Part of these sentences con- between the attention-demanding components tained a reflexive pronoun that agreed with the of the task. Alternation from comprehension to corresponding subject (The headmaster com- production would be a safe option at points plained: “I have seen a nasty cartoon of myself in where elements in the source text can be predicted the hallway”) whereas others contained a mis- on the basis of prior context, background matching reflexive pronoun (*The headmaster knowledge, and information redundancy. Studies complained: “I have seen a nasty cartoon of him- on monolingual discourse processing have self in the hallway”). (The experiment was run shown that language users do indeed exploit in Dutch with Dutch participants; the example such sources of knowledge to anticipate up- sentences thus concern translations of the coming words. For instance, Van Berkum, Brown, materials actually presented.) The response time Zwitserlood, Kooijman, and Hagoort (2005) and between the reflexive in the input and in the out- Otten, Nieuwland, and Van Berkum (2007) found put served as the dependent variable. In the read- that words inconsistent with a specific prediction ing and paraphrasing task, the above two types of based on prior discourse evoke an ERP response sentences had to be paraphrased on line into their different from the ERP elicited by prediction- indirect-speech counterparts and the participants consistent words, and this prior to the moment had to correct any error they might come across. the critical word appears in the input. The ability The invited output sentence in reply to both of to anticipate specific words in simultaneous inter- the above two example input sentences thus was: preting enables the interpreter to ignore parts of The headmaster complained that he had seen a the input so that more attention can be devoted nasty cartoon of himself in the hallway. Even to other task components such as production or though these two tasks are monolingual ones, memorizing. Especially prepared, well-structured their resemblance with simultaneous interpreting

6. LANGUAGE CONTROL 319 will be obvious: Also in simultaneous interpreting task appears to be cognitively less demanding input sentences have to be transposed into output than simultaneous interpreting. Furthermore, the sentences on line. In particular, the paraphrasing materials concerned separate sentences instead task bears a close resemblance with simultaneous of longer text, thus arguably posing relatively interpreting. In both cases an input message few demands on memory as compared with sim- has to be understood and an output message has ultaneous interpreting. If simultaneous decoding to be produced with the same meaning but a dif- and encoding were possible at all, one would have ferent wording, the only clear difference between expected to obtain support for it in this relatively the two tasks being that only one of them involves easy task. a language switch from input to output. Because of this similarity, in the translation literature Following the above characterization of within-language paraphrasing has been referred simultaneous interpreting, the larger part of the to as “unilingual translation” or “intralingual remainder of this chapter will be dedicated to translation” (e.g., Anderson, 1994), and the task two questions, each dealing with a different is often used as exercise during the training of aspect of control in simultaneous interpreting. translators and interpreters (see e.g., Moser- The first concerns language control: Given the Mercer, 1994). fact that simultaneous interpretation always involves the use of two languages, how can it be In the reading with correction task, when the secured that the source language intrudes into input sentence contained a mismatching reflexive the target language output as little as possible? (himself) the response time was relatively long The second, to be called “processing control”, (myself: 460 ms; himself: 502 ms). The delay at the deals with control in a more general sense. As mismatching point is plausibly due to the decoder we have seen, simultaneous interpreting is an being taken by surprise when it encounters a extremely complex skill, involving several mismatch between the decoded structure so far attention-demanding components as well as their and the input encountered at the mismatch point. coordination and rapid sequencing. How then do In contrast, in the reading and paraphrasing task, (some) people come to master this skill and the the response time to the correct myself input sen- obviously extremely complex control operations tences was longer than to the incorrect himself it involves? Both aspects of control are plausibly input sentences (myself: 507; himself: 461). This effectuated by a language-external control finding suggests that the encoded structure has system of the type sketched earlier (p. 314). In completely overwritten the input structure, as later sections I will address each of these two if only one of them can be available at any one questions in turn. But first I will explain the moment in time. This, in turn, suggests that notions of conceptually mediated translation decoding an input structure and encoding an versus transcoding, as well as the phenomenon of output structure cannot proceed in parallel but natural translation. Without an account of these occur in rapid alternation instead. The longer concepts, a general characterization of trans- response time in the reading and paraphrasing lation processes, and especially simultaneous task when the input contains myself as the interpreting, would be wanting. reflexive form can be attributed to the mismatch with encoded himself, the encoder being taken by Conceptually mediated translation, surprise when it encounters myself instead. transcoding, and natural translation To conclude, even though these results were The term “conceptually mediated” (or “vertical”) obtained in a monolingual study, there is no translation refers to a form of translating that is obvious reason why the conclusion that gram- hypothesized to exploit the same comprehension matical decoding and encoding cannot occur and production apparatus as used in monolingual simultaneously could not be generalized to language tasks. According to this view, the task simultaneous interpreting as well. Lacking a language switch, the reading and paraphrasing

320 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS involves, first, the processing of a speech input A phenomenon called “natural translation” upward through the source language system, suggests that conceptually mediated translation from a peripheral analysis of the input up to the is a general translation strategy that is already point where the input is assigned a non-verbal used by young children. The term refers to the conceptual representation and, next, a downward translating that is done in everyday circumstances process that sets off with this conceptualized by bilinguals without any special training in representation and concludes with the articula- translating. As shown by several authors (e.g., tion of speech or with the writing down of text Harris & Sherwood, 1978; Malakoff, 1992), any (in written translation). In theory then, any indi- individual with a certain degree of competence in vidual who can communicate both receptively two languages can communicate between the two and productively in two languages separately can and, thus, act as a translator for monolingual translate between them, even when knowledge speakers of each language. This form of trans- regarding translation-equivalent structures lation behavior can already be observed in very across these languages is completely lacking and young bilingual children, starting soon after the no specialized translation strategies have been onset of their first linguistic utterances (Harris, developed. 1980). This suggests that translation ability is concomitant with bilingualism, an emergent In contrast to conceptually mediated trans- property of bilingualism that exploits the lation, the second hypothesized translation language-processing machinery and knowledge procedure, “transcoding” (or “horizontal” trans- structures that are already in place and used in lation), assumes that source language input monolingual settings. Furthermore, the fact structures, be they words, common phrases, or that even young children manifest this behavior idiomatic expressions, are directly replaced by the shows that already at a young age bilinguals have corresponding target language structures. Earlier functionally separated their languages, because in this chapter (p. 297) I described the process only when the knowledge of what language by which the memory representations of entities individual linguistic items belong to is in place, that co-occur in the environment become con- is translation possible. (This, incidentally, also nected through a process of joint activation suggests that a once-popular view of code (“what fires together, wires together”). This switching in young bilingual children as resulting learning mechanism is also likely to underlie from linguistic confusion must be wrong.) Finally, transcoding: During each translation act, the phenomenon indicates the existence of meta- translation-equivalent structures in memory will linguistic awareness (defined as “the ability to be activated in close temporal proximity, even reflect on and manipulate the structural features when translation comes about via the vertical of language independent of meaning”; Bruck & route, and as a consequence they will be linked up Genesee, 1995, p. 307) in very young bilinguals. with one another. The more often the same two This conclusion follows from the fact that trans- translation-equivalent structures are co-activated lation is only possible by stepping back “from the through translation, the stronger the link between comprehension and production of an utterance in them will become and the more easily available order to consider the linguistic form and structure one term of the pair will become given the other underlying the meaning of the utterance” term. The consequence is that especially experi- (Malakoff & Hakuta, 1991, p. 147). enced translators and interpreters will possess a large stock of memory structures that enable This evidence that bilinguals without appre- transcoding. Paired-associate learning (see ciable translation experience dispose of a natural Chapter 3) is an additional means to form and translation ability obviously does not imply that strengthen such links between translation- they translate in the same manner as professional equivalent memory structures. The present two translators. In addition to evident quantitative hypothesized translation strategies are visualized differences between “natural” translators and in Figure 6.7. professionals in translation speed and accuracy,

6. LANGUAGE CONTROL 321 Two translation strategies (based on Paradis, 1994). The grey arrows depict conceptually mediated translation: The input is fully analyzed by the source language system in a bottom-up fashion. Next, the emerging conceptual representation is processed top-down by the target language system. The white arrows depict the transcoding process in which source language structures are directly transformed into their equivalents in the target language. Adapted from Christoffels and De Groot (2005) by permission of Oxford University Press, Inc. qualitative differences are likely to exist between meaning and, therefore, translate into more than them because processionals may use special one word in another language, one can easily see strategies and skills that they developed on the how the automatic triggering of one particular job. One possible difference concerns the degree target language word by a source language word of transcoding exhibited by lay translators and (by activation spreading along the direct link professionals: Because professionals possess a between the corresponding memory structures) relatively large stock of well-established transla- may lead the translator up the garden path. tion structures in memory (see above), they Because of the relatively strong memory con- plausibly rely on transcoding more than non- nections between a source language word and its professionals do (Paradis, 1994). Yet this is a frequent translation(s), this danger is especially source of dispute in the literature. Whereas large when the currently targeted meaning is a Paradis regards the frequent use of transcoding a relatively infrequent one. But it is also easy to signature of professionalism and as beneficial see how the swift availability of translation- for translation and interpreting performance, equivalent structures along transcoding connec- according to other scholars (e.g., Seleskovitch, tions in memory can help to reduce processing 1976) it indicates a lack of expertise and hinders load, freeing mental resources for non-automatic performance. task components. As long as some resources are spared for a monitoring process that oversees As is often the case when opposing views are performance and intercepts incorrect transla- held, there may be a kernel of truth in both of tions delivered by a transcoding structure, the them. Because many words have more than one

322 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS beneficial effects of transcoding presumably out- performance the moment mental load during weigh the risks involved. For such monitoring to interpreting exceeds some critical level and con- take place no special mechanism needs to be pos- ceptually mediated interpreting fails. tulated because, as mentioned earlier, interpreters are known to monitor their speech (e.g., Gerver, Language control in simultaneous 1976; Isham, 2000). interpreting Whatever the attitude of the research com- Differential activation of language subsets and munity towards transcoding, there is evidence to processing components suggest that it does in fact occur, also in trans- lation and interpreting by professionals. Isham The views on language control in simultaneous (1994) and De Bot (2000) observed that the form interpreting are built on those developed to of recently encountered grammatical structures account for language control in monolingual in the source language determines the exact tasks and again assign crucial roles to language choice of grammatical target language structures, subsets, a language cue in the conceptualized and many studies (e.g., De Groot, 1992b; message, and a language-tag component in the Christoffels et al., 2003) found cognate effects in lexical lemma. To account for the fact that translation; that is, differential performance simultaneous interpreters must keep both lan- for cognates and non-cognates. These effects guages simultaneously active and yet produce demonstrate an influence of the form aspects of output in only one of them, the target language, the source language input on translation per- Paradis (1994, 2004) extended his views on formance. Yet in conceptually mediated trans- differential activation and inhibition of the lation the downward production process sets off two language subsets in monolingual tasks from a non-verbal conceptual message which, (pp. 294–296). Recall that Paradis assumes that a by definition, does not contain traces of the memory item becomes available for further pro- source language’s input forms. This suggests cessing when a sufficient number of excitatory that the above effects originate from another neural impulses have reached its neural substrate process than conceptually mediated translation; and that the amount of impulses required to and transcoding—enabled by ready-made trans- reach this point was called the item’s “activation coding structures in memory—is a plausible threshold”. He furthermore assumes that candidate (but see pp. 261–263 for alternative thresholds are not fixed but can be adapted to accounts). Furthermore, a study by Fabbro, the language user’s current goals, for instance, Gran, and Gran (1991) suggested that both to use one language or the other. Regarding professional and student interpreters exploit a simultaneous interpreting he suggested that the mixture of transcoding and conceptually thresholds of all the elements in both subsets are mediated interpreting, but with the professionals clearly set well below a level that would frustrate relying more on the latter than the students. source language comprehension, target language When the interpreter is particularly stressed or production, or both, but that at the same time tired, the use of form-based interpreting the thresholds of the elements belonging to the strategies increases, presumably because it source language are set higher than those of the requires less mental resources than conceptually target language elements. mediated interpreting (Darò & Fabbro, 1994). In other words, for professional interpreters trans- One might wonder whether a relatively high coding may be a convenient backup strategy threshold setting of the source language elements to resort to during circumstances of excessive would involve the risk that their comprehension is mental load. It thus appears that Seleskovitch impeded. But as we have seen, Paradis’s views (1976) was right in positing that the excessive include the idea that relatively high threshold reliance on transcoding reflects a lack of expert- settings still enable comprehension because of ise. But having this strategy in reserve benefits the relative ease with which input from an

6. LANGUAGE CONTROL 323 external source brings about the neural impulses ferential threshold setting of source and target that activate the targeted memory items. So even language in the first place. if the thresholds are set at a relatively high level, sufficient neural impulses will be incited by the Grosjean (1997b, 2001) proposed a different input for the comprehension thresholds to be solution that circumvents this problem. It is reached. Instead, for production to be possible at illustrated in Figure 6.8. Just like Paradis he split all, lower thresholds are required because the up the language system into two linguistic sub- arousal of the neural impulses to activate the systems, viewing these as knowledge stores. Onto targeted memory items, now coming from each subset he added two processing mechanisms, internal sources, is more demanding. Therefore, an input mechanism (or “component”) and an the number of impulses exciting the relevant output mechanism. According to Grosjean, as memory items will be relatively small and produc- the language subsystems proper (the knowledge tion is only possible if the associated thresholds stores), these processing mechanisms can be are relatively low. The implication of this state of activated to different degrees. He assumed that in the language system for simultaneous interpreting simultaneous interpreting the source and target is clear: Even if during simultaneous interpreting language subsystems are both highly activated the thresholds of the elements belonging to the and to equal degrees (the interpreter is in a source language subsystem are raised towards “bilingual mode”; see pp. 288–291). In addition, to relatively high levels, comprehension is still allow for accurate comprehension of the source possible. At the same time, these relatively high language, this language’s input component is thresholds prevent output in the source language. highly activated. Instead, to prevent source In fact, the only difference between this account language elements from emerging in the output, of language control in simultaneous interpreting the output component of the source language is and the account of language control in mono- completely deactivated. In addition, both the lingual production tasks presented earlier is that input and output components of the target lan- the latter allow for an even larger difference in the guage are activated: The output component of threshold settings for the two languages, with, the target language must obviously be activated specifically, the thresholds of memory elements in highly to enable production in this language, the non-response language raised to maximally and this language’s input component must be high levels. activated so that interpreters can monitor their output. Because of the relative ease of Still, there may be situations in which Paradis’s comprehension as compared to production, a (1994) solution to set the thresholds of the relatively low level of activation of the target source language elements at a relatively high language’s input component (which is the level involves the risk of suboptimal comprehen- equivalent of a high threshold setting in sion and subsequent breakdown. This situation Paradis’s model) would suffice. However, because may arise when the current comprehension this component is a different one from the requirements are taxing due to, for instance, poor analogous source language input component and quality of the speaker’s diction. Adapting to is exclusively dedicated to comprehension any- these circumstances by lowering the thresholds of way, there is no reason why it should be activated the source language elements might solve this less than the corresponding output component. problem but would create a new one because the lower settings might render source language Grosjean’s view on simultaneous interpreting elements available for production, thus leading to thus explicitly distinguished between two know- language-mixed output. Because of the com- ledge stores (the language subsets) on the one prehension problem this would cause for the hand and four processing mechanisms on the (monolingual) listeners, this situation must be other hand. The exact content of the knowledge avoided under all circumstances. In fact, the stores is not explicated. Presumably lexical know- risk of mixed output was what motivated the dif- ledge such as morphological, grammatical, and form information regarding words is implied.

324 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS A view on language control in input lexicon must be activated. Finally, to pre- simultaneous interpreting. vent words from the source language emerging in The language system the output, the source language’s output lexicon contains two language- must be deactivated as much as possible. specific knowledge stores (the black squares) and two So whereas Grosjean proposed a solution in language-specific input- terms of two equally highly activated language processing and output- subsets augmented by a total of four differentially processing mechanisms activated input and output processing mechan- each (the circles). During isms, each dedicated to one language only, the simultaneous interpreting present proposal distinguishes between four both knowledge stores are differentially activated knowledge stores and highly activated, as are the remains implicit about the processing mechan- two input-processing isms that operate on these knowledge stores. mechanisms and the output- Comparing the present view with Paradis’s processing mechanism of the account of language control in simultaneous target language. The output- interpreting, the only essential difference is the processing mechanism of the splitting up of Paradis’s source and target lan- source language is guage subsystems in a part that contains elements deactivated. Adapted from dedicated to comprehension and one with Grosjean (2001). Copyright © elements exclusively used for production. This 2001. Reproduced with seemingly minor change may be crucial, however, permission of Blackwell because it does not run the risk of comprehension Publishing Ltd. failure as a consequence of too low an activation level (or, in Paradis’s terminology, too high a But form aspects of lexical entries are plausibly threshold setting) of the source language’s represented separately for comprehension and production, an assumption that in the literature emerges in the proposal of separate input and output lexicons. Extending this possibility to the bilingual case implies the existence of (at least) four lexical knowledge components: an “input lexicon” that consists of two subsets, one for each language, and an “output lexicon” again containing two subsets. Christoffels and De Groot (2005) suggested an account of language control in simultaneous interpreting in terms of such a four-component view of the lexicon, incorporating the common notion of differential activation of language subsets. This set-up is illustrated in Figure 6.9. We hypothesized that these four subsystems of the bilingual language system can be activated independently from one another. Because in simultaneous interpreting the source language must be understood and the target language produced, the input lexicon of the former and the output lexicon of the latter must be highly activated. In addition, to allow for monitoring of the target language’s output, the target language’s

6. LANGUAGE CONTROL 325 A view on language control in simultaneous interpreting. Conceptually mediated translation takes the route via the conceptual representation level. Lexical transcoding takes the shortcut via the lexical-semantic representations. Separate input and output lexicons are assumed for each language subsystem. The source language’s input lexicon and the target language’s output lexicon are highly activated to enable source language comprehension and target language production. The target language’s input lexicon is also activated to some extent so that the output can be monitored. The source language’s output lexicon is completely deactivated. From Christoffels and De Groot (2005) by permission of Oxford University Press, Inc. subsystem in Paradis’s account. In the present delicate maneuvers required to differentially set model, the activation of the source language’s the activation levels of all of them, proactively input lexicon can safely be set at a maximum level and globally. Especially in the latter respect a without running the risk that the source language further proposal, based on La Heij’s (2005) view intrudes in the output. The reason is that the of language control in monolingual word pro- elements in the input lexicon are specialized for duction presented earlier (pp. 304–306; see also De comprehension and are alien to the production Bot, 2000), is more parsimonious. As described process. there, this author advanced a “complex access, simple selection” view of lexical selection that A language cue in the preverbal message consists of two steps. First, an elaborate (“complex”) conceptualization process delivers a The common element in all three views of lan- detailed output (the “preverbal message”) that guage control in simultaneous interpreting contains all the information a speaker wants to presented so far is the assumption of differential communicate, including his or her goal to use activation of complete language subsets, sub- slang or to express the conceptualized content in, components thereof, or language-processing say, the L2. The conceptual information specify- mechanisms. The plausibility of these views ing the targeted language is called the “language ultimately depends on evidence suggesting the cue”. Second, a “simple” selection process sub- existence of each of the hypothesized separate sequently takes place in which the lexical entries components of the language system and of a that are maximally activated by the preverbal control system that can perform the obviously message are selected for production, generally one

326 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS per lexical concept in the preverbal message. keeps in mind his goal to produce output in the If a bilingual’s intention is to speak in L2, the target language, he will satisfy his audience, at inclusion of the L2 language cue in the preverbal least in this respect. message guarantees that generally L2 speech will emerge from the system because the lexical entries Processing control in simultaneous for L2 words will on the whole be more highly interpreting activated than those representing L1 words. Introduction On the assumption that simultaneous inter- preting comes about via the conceptually So far the focus of the discussion has been on mediated (“vertical”) route presented earlier, this language control in simultaneous interpreting; view on language control in monolingual settings that is, on how interpreters manage to produce does not seem to require any special adjustment output in the target language despite the fact that to be able to account for language control in both language subsystems are activated. But there simultaneous interpreting. The deverbalized is obviously more to control in simultaneous conceptual representation that is the end product interpreting than merely seeing to it that the out- of the linguistic analysis of the source language put language differs from the input language. input is plausibly the equivalent of the preverbal Earlier in this chapter I characterized interpreting message that La Heij and many others hypoth- as an extremely complex task that taxes working esize to be the output of the first (conceptualiza- memory inordinately because it encompasses tion) step of speech production (see Chapter 5). several components that all consume attention This conceptual representation must sub- and must be executed under speed stress and sequently be expressed into target language interwoven within the same unit time. A consider- speech. As long as the interpreter adds the inten- ation of the task’s requirements will easily con- tion to produce speech in the target language vince anyone that the task is excruciating, and can to this conceptual structure in the form of the only succeed under the confident guidance of an appropriate language cue, the output will unrelentingly alert control system with a large generally be in the targeted language because the capacity and that can rely on a large database of words in this language will be activated relatively swiftly available knowledge. How then is inter- highly by the conceptual structure. In other preting possible at all, at least for some dedicated words, just as the language cue guarantees that and perhaps gifted individuals? To what extent the targeted language will be output in unilingual does it exploit, on the one hand, the knowledge conversational settings, it also sees to it that structures and linguistic skills that come with a output in the target language emerges during high level of proficiency in the two languages interpreting. Furthermore, just as the monitoring concerned and, on the other hand, specialized system intercepts speech errors on their way out knowledge, skills, and strategies that are not con- in common monolingual speech production, in comitant with fluent bilingualism? And if the simultaneous interpreting it will intercept the odd latter contribute to successful performance, do word in the source language on its way out. The they reflect an innate aptitude or do they result only clear difference between the two situations from task-relevant training (are interpreters born is that the conceptual content that sets off pro- or bred)? Knowing what knowledge and skills duction is the end product of decoding the come with fluent bilingualism and which ones source language input in interpreting, but comes require specialized training (or are innate) is not from within the cognitive system in common only of theoretical interest but has practical monolingual language production. Unlike the implications as well: It informs both the devel- above views, this account of language control in opment of interpreter training programs and the simultaneous interpreting does not require the assessment procedures that admit students to the proactive setting of the activation level of various program. mental subsystems. As long as the interpreter

6. LANGUAGE CONTROL 327 In the next sections I will present the results known as the “immediacy strategy”). Further- of a number of studies that have addressed one more, as readers do, they used syntactic and or more of the above questions. As we will semantic cues in the input to chunk the reading see, rather than probing the participants’ per- materials into meaning units consisting of several formance on the full simultaneous interpreting lexical elements. These findings suggest that at task, many of the relevant studies tackle these least in some respects expert and non-expert questions by focusing on one or more of the translators use similar comprehension processes hypothesized task components, such as working during translation and that the comprehension memory, comprehension, or a sub-skill of these. processes used in translation build on those used The underlying assumption of this componential in common unilingual reading situations. As we approach—common in the study of complex shall see later, the latter conclusion does not imply tasks—is that performance on the full task is that comprehension processes during translation related to the degree of mastery of its com- are identical in all respects to those employed in ponents. This approach does not imply the monolingual language processing. assumption that full-task performance is a direct function of the performance on the component A study by Dillinger (1994) similarly indicated tasks. Instead, there is the understanding that that expert and non-expert simultaneous inter- the requirement to execute all of the constituent preters use similar comprehension strategies. In parts in concert and under severe time pressure this study interpreters and fluent bilinguals may modulate the way each of them is executed in without any interpreting experience were asked isolation, independent from the others. The com- to interpret English materials into French. ponents of the interpreting task that I will focus Dillinger’s unit of analysis was the “proposition”, on are comprehension of the source language, which in linguistic theory refers to a basic mean- word retrieval, and working memory. ing unit consisting of something or someone being talked about (called an “argument”) and Comprehension processes an assertion made about this argument (the “predicate”). With the number of processed text One set of studies examined comprehension propositions serving as the dependent variable, processes in interpreting and other tasks by com- the two groups responded the same way to a paring the performance of expert interpreters/ number of syntactic and semantic text manipula- translators and non-experts. McDonald and tions such as the density of syntactic clauses, Carpenter (1981) had experts and amateur whether a clause concerned a main or a sub- German–English translators perform a task, ordinate clause, and the density of propositions. “simultaneous translation”, which involved a The differences between the two groups that did mixture of simultaneous interpreting and written emerge were primarily quantitative: Unsurpris- text-to-text translation. The participants were ingly, the accuracy of interpreting was better for presented with written text fragments that they the experienced interpreters. had to translate orally as quickly as possible while reading them. Their translation behavior Other quantitative differences between the was inferred from the eye movements that were comprehension processes in professional inter- registered while they performed the task (see preters and non-interpreters were identified by p. 161 for details). The data showed that the pat- Bajo, Padilla, and Padilla (1999) in a study where terns of eye movements for both participant the participants did not perform any full transla- groups, the experts and amateurs alike, were tion task but simply read text for comprehension. similar to those that characterize normal reading One group of non-interpreters consisted of processes: The participants tended to assign students in a translation and interpreting pro- meaning and reference to each word as soon as gram who had finished their training in text possible, as readers typically do (this strategy is translation but had not yet been trained in simultaneous interpreting. A second group of non-professionals consisted of highly proficient

328 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS balanced bilinguals who had never participated in translate the input materials in translation influ- such a training program. The participants read ences the comprehension processes by directly through texts presented on the screen word-by- comparing the participants’ performance on a word by means of one version of the self-paced simple comprehension task and a translation reading technique (see p. 160 for details): Each task. Gerver (1974) had interpreting students time the participant pressed a button the next either listen to spoken stories or simultaneously word appeared on the screen. In experiments interpret them, after which they had to answer using this technique the time duration between comprehension questions. In the interpreting two successive button presses is considered the condition fewer questions were answered cor- reading time for the word presented last. The rectly than in the listening condition. This finding reading time per word turned out to be signifi- suggests that the requirement to simultaneously cantly shorter for the professional interpreters perform other mental operations in addition to than for the control group of proficient balanced comprehending the aural input has an adverse bilinguals, whereas those for the group of transla- effect on comprehension. In a second study, tion students were in between those of the other Macizo and Bajo (2004) had fluent Spanish– two groups. The depth of text comprehension did English bilinguals who were experienced trans- not differ between the groups. In a further test the lators/interpreters either read and repeat a set of participants were presented with exemplar– “object relative” sentences or read and translate category word pairs (e.g., finch–bird or finch– them, in two separate blocks each preceded by clothing) and were asked to decide for each of block-appropriate instructions. In object relative them whether the first word referred to an exem- sentences the noun that refers to the agent of the plar of the category named by the second word. main verb must be assigned the role of patient in The response times in this task, which presumably an embedded relative clause (e.g., The judge who reflect the speed with which meaning is assigned the reporter interviewed dismissed the charge at the to words, were shorter for the professional inter- end of the hearing, in which judge is the agent of preters than for the other two groups. These joint the main verb dismissed but the patient of inter- findings thus suggest that interpreters excel in viewed). Sentences of this type are known to word recognition, one of the sub-skills of text require more processing resources than their comprehension. As stressed by the investigators, “subject relative” counterparts (The judge who the finding that the professional interpreters out- interviewed the reporter dismissed the charge at the performed the balanced bilinguals is of special end of the hearing, in which judge is the agent of interest, because it suggests that the faster word both the main and relative verb). The reading recognition and meaning assignment of the inter- materials appeared word-by-word in the center preters did not merely result from a high level of of the screen and the presentation rate was proficiency in both languages but from specific self-paced, each next word appearing upon a key training. This hypothesis was confirmed in a press by the participant. The time between two retest taken about 10 months later. During the successive key presses was registered as the intervening months the student group but not the reading time per word. Repeating or translating non-student control group had practiced all sorts started after the final word of the sentence had of exercises aimed at improving rapid word rec- been read and a message (either the word repeat ognition and semantic access. The student group or the word translate) appeared on the screen (but demonstrated shorter reading and categorization note that the participants always knew in advance time on the retest, thus demonstrating the efficacy whether the sentence to be presented was to be of skill-relevant training. In contrast, the non- repeated or translated). student group had not become any faster on either task. The investigators argued that, if comprehen- sion processes are the same in normal reading and A further set of studies examined whether the in reading for translation, reading time per word requirement to simultaneously comprehend and should be the same in the read-then-repeat and

6. LANGUAGE CONTROL 329 the read-then-translate conditions. Contrary to et al. (2008) ran a further study in which they this prediction, reading times were longer when manipulated the syntactic congruency between the participants read for subsequent translation the source language sentences and the corre- than when they read for subsequent repetition. sponding target language sentences and the fre- This finding suggests that comprehension quency of some of the target language words processes vary depending on the specific goal of (while the frequency of the corresponding source reading. The reading time difference between language words did not vary). In the read-then- the two conditions was especially large in the translate condition, but not in the read-then- sentence area where computation demands were repeat condition, the reading times were affected highest (around interviewed in the above sentence, by both these manipulations, a finding that again which is the place where readers have to figure indicates co-activation of structures in the target out the thematic role of reporter). Level of com- language’s linguistic system in the former but prehension was the same in the two conditions. not the latter condition. Once again the authors The researchers attributed the reading time attributed these effects to transcoding operations difference between the two conditions to during comprehension in the read-then-translate additional processing in the reading phase of conditions. In other words, the results of all three the read-then-translate condition. They hypoth- studies suggest that the requirement to translate esized that this extra processing concerned the modulates the comprehension processes as transcoding operation discussed earlier. executed in unilingual reading. Converging evidence was obtained in two To summarize, the joint results of the above further studies by the same research group studies suggest that in some ways the comprehen- (Macizo & Bajo, 2006; Ruiz, Paredes, Macizo, & sion processes of professional translators and Bajo, 2008) in which professional Spanish– amateurs involved in a translation task resemble English translators/interpreters were presented the comprehension strategies that readers norm- with two blocks of L1 Spanish sentences. As in ally use in common unilingual reading. Com- Macizo and Bajo (2004) they were instructed to prehension processes in translation thus seem read then repeat one block of sentences and to to build on those used in normal reading. In read then translate (in L2 English) those of the addition, the above results revealed a number of second block. Macizo and Bajo (2006) manipu- quantitative differences between the compre- lated the cognate status of the words in sentence- hension processes of professional and amateur initial and sentence-final position. The results translators, including faster word recognition showed that when the participants were asked and faster meaning assignment in professionals. to read and translate the sentences the reading Finally, they showed that reading for translation times for cognates were faster than those for in professional translators modifies normal non-cognate control words, especially in sentence- unilingual comprehension processes. Whether final position. However, when the participants this also holds for amateur translation is were instructed to read and repeat the sentences undetermined because, to my knowledge, the this cognate effect did not materialize. The relevant experiment has not been performed yet. authors concluded that when professional trans- lators/interpreters process linguistic materials Word retrieval and working memory with the purpose of translating them, lexical elements of the target language are also activated In one of the studies reviewed above (Bajo et al., while target language activation does not occur 1999) it was found that professional translators/ when they merely process the input for com- interpreters recognize and assign meaning to prehension. It may be remembered that cognate words faster than proficient bilinguals in- effects do not constitute unequivocal evidence of experienced in translating and interpreting. A parallel activation in the target language (see couple of other studies have shown that inter- pp. 203–205 and 247). Aware of this fact, Ruiz preters possess other special abilities as well,

330 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS while at the same time suggesting that some of written sentences and asked to read them with the sub-skills required for interpreting can be understanding and to concurrently try to remem- developed by other means than interpreting ber the final word of each sentence. Following experience. Christoffels et al. (2006) compared each set the participant is asked to reproduce the the performance of two groups of Dutch–English final word of all sentences presented in the set. bilinguals without any professional interpreting The presentation procedure stops the moment the experience and a group of professional inter- participant fails to recall all sentence-final words preters on a number of simple tasks that were all of the immediately preceding set. The dependent assumed to reflect a basic cognitive sub-skill variable is the participant’s reading span score; of simultaneous interpreting. The two groups of that is, the largest sentence-set size for which the inexperienced interpreters differed from one participant exhibited perfect recall. Alternatively, another in L2 English proficiency (and amount the dependent variable is the total number of of L2 language training): One consisted of sentence-final words of all previous sentence sets unbalanced bilinguals, all university under- successfully recalled up until the moment of graduates, with stronger L1 Dutch and weaker failure to reproduce all sentence-final words (but still relatively strong) L2 English. The second of the just-presented set. To make sure that the concerned balanced bilinguals with Dutch as participants do indeed read with understanding their L1 who, just like the interpreters, had a uni- and do not just try to memorize the sentence-final versity degree in English, and were all teachers words, at random moments the participants of English and, as such, practiced English pro- must answer questions regarding the previous fessionally on a daily basis, again just as the sentence’s meaning. Our second test concerned interpreters did. the “speaking span test” (Daneman & Green, 1986): Participants are presented with sets of One set of tests assessed the participants’ words of increasing size and asked to construct fluency in retrieving words from memory: picture sentences that contain all the words of a set. This naming in L1 and L2 and word translation in test is thought to assess the ability to produce both directions. The importance of rapid and sentences and remember words simultaneously. automatic word retrieval for interpreting is evi- The speaking span score is the largest number of dent because any conscious effort required to words for which participants successfully create a retrieve the appropriate target language word for sentence (or the total number of words success- a concept to be expressed reduces the mental fully embedded in sentences up until the moment resources available for other non-automated of failure to construct a sentence around the processing operations. Indeed, Carroll (1978) word-set just presented). Notice that the demands assumed “verbal fluency” (as the swift availability of these two memory span tests mimic the of linguistic knowledge, and especially vocabu- requirement in interpreting to concurrently com- lary knowledge, is often called) to be one of prehend and memorize linguistic material on the the most important requirements for a successful one hand, and to concurrently produce and career as a simultaneous interpreter. In agreement memorize linguistic material on the other hand. with this view, word retrieval fluency has been shown to be positively correlated with per- Performance on the reading span test is known formance on a simultaneous interpreting task in to correlate positively with reading comprehen- bilinguals without any prior task training sion (Daneman & Carpenter, 1980). The likely (Christoffels et al., 2003). reason is that many processes that must be carried out to get a good grasp of the message communi- Two further tests that we administered are cated by a text—such as finding the antecedents commonly thought to reflect both the storage and of anaphors, resolving syntactic ambiguities, and processing functions of working memory. One making inferences—require that earlier parts of of these was the oft-used “reading span test” the text are kept in working memory. In other (Daneman & Carpenter, 1980): The participants words, good comprehension requires the con- are presented with increasingly large sets of

6. LANGUAGE CONTROL 331 current storage and processing of information, working memory capacity, is not a side-effect of and it is exactly this ability that the reading span proficient bilingualism but appears to emerge test assesses. The demands on working memory from extensive on-the-job interpreting practice. during interpreting are much larger than during reading comprehension: Not only does good However, a second possible cause of the comprehension of the source language necessitate differential span scores of interpreters on the one the concurrent processing and storage of the hand and non-interpreters on the other hand input, thus already taxing working memory, but should be considered: It might be that the high the sequencing differences between the source scores of interpreters reflect a large innate work- and target language, and the temporary retention ing memory capacity, and that the interpreting of parts of the input this requires, also demands profession appeals to individuals with this special additional capacity. Furthermore, a part of the ability (that is, they may self-select for the job). working memory resources must be allocated to However, studies by Bajo et al. (1999) and Padilla, the production of an earlier input fragment and Bajo, Cañas, and Padilla (1995) suggested that it to coordinating the various task components. is not good working memory genes but, indeed, Given these demands of the interpreting task, extensive task practice that underlies the large performance on the reading span test may be working memory capacity of interpreters. Like expected to be relatively good for interpreters. Christoffels et al. (2006), these researchers had Similarly, interpreters may be expected to excel on professional interpreters and different groups of the speaking span test because, as mentioned, control participants perform a number of tasks, it reflects the ability, required in interpreting, to including the reading span test. Among the con- concurrently maintain information in memory and trols was a group of students in a translation and construct output that expresses this information. interpreting training program who had finished their training program in text-to-text translation The upper and lower parts of Figure 6.10 but had not yet been trained in simultaneous summarize the main results of the working interpreting. A further control group consisted memory and word retrieval tasks, respectively. A of participants who had finished their under- quick glance at these four graphs immediately graduate studies in other fields and, aiming at reveals that on the word retrieval tasks (picture other careers, had not had any training in transla- naming and word translation) the interpreters tion and interpreting. The results of both studies and teachers performed at similar levels, and that showed larger reading spans in the interpreters both groups outperformed the student group than in the other groups of participants, who did (except on the task that indexes a skill that all not differ between them. The finding that the groups may be expected to be equally experienced translation and interpreting students did not in: picture naming in L1). In contrast, on the perform any better than the other control group working memory tasks the interpreters clearly points at interpreting practice as the source of the outperformed both other groups, which did not superior working memory of interpreters. differ from one another statistically. These joint findings strongly support two conclusions: First, In the foregoing, a number of skills have been one of the component skills of professional identified that characterize the professional inter- interpreting, swift word retrieval in L2, appears preter and that are all likely to be prerequisites to emerge from extensive practice in the L2, for performing the task at a professional level: suggesting that it is concomitant with proficient rapid word retrieval given a concept to express, a bilingualism. (To what extent teacher training large working memory capacity, rapid word contributed to the high word retrieval fluency of recognition and meaning assignment during the teachers is not clear. It would require a control reading. In addition to these special abilities of group of balanced bilinguals without any teacher interpreters, a further one has been identified: training to find out.) Second, a further pre- Interpreters not only have a relatively large requisite for professional interpreting, a large working memory capacity, they also appear to exploit working memory differently from

332 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS Mean number of recalled words on the reading span and speaking span tasks per group, and mean reaction times (in ms) on the picture-naming and word translation tasks per group. Adapted from Christoffels et al. (2006) Copyright 2006, with permission from Elsevier. non-interpreters. The critical evidence comes suppression; that is, by getting the participants to from studies that exploited the articulatory sup- perform a secondary task that engages their pression technique introduced earlier in the con- articulation apparatus so that it is no longer text of the working memory model of Allan available for the rehearsal required by the primary Baddeley and his colleagues (e.g., Baddeley, 1986; task. This secondary nuisance task often involves Gathercole & Baddeley, 1993; see Chapter 3, the concurrent and repeated uttering by the pp. 116–119). In addition to a “central executive”, participants of some nonsense sound (e.g., the this model contains a “phonological loop”, a sub- nonsense syllable bla). The consequence of system specialized for processing and temporarily occupying the loop this way is deteriorated per- maintaining verbal material. To be able to main- formance on the primary task as compared with tain verbal material, the phonological loop is performance in a condition where the articulation equipped with a rehearsal system. When a task apparatus is not kept busy by a secondary task must be performed that requires the involvement (e.g., Papagno et al., 1991). of the phonological loop, the proper working of the loop can be frustrated by articulatory As we have seen above, comprehension requires the temporary retention of the input

6. LANGUAGE CONTROL 333 materials, and to fulfill this requirement the concluded that the interpreters’ ability to phonological loop’s rehearsal system must be simultaneously comprehend and produce speech deployed. In simultaneous interpreting, however, is somehow related to their ability to use word the requirement to produce vocal output in the knowledge efficiently. target language while making sense of the input equals an articulatory suppression condition. The above articulatory suppression experi- How then is comprehension possible at all? Bajo ments can tell us something about how et al. (1999) and Padilla et al. (1995) hypothesized simultaneous interpreting is done because, as in that the answer may be that interpreters use simultaneous interpreting, vocal output is pro- their working memory differently from non- duced while at the same time an input has to interpreters. They set out to test this hypothesis be processed. Yet in a number of respects the by studying word learning with and without articulatory suppression condition in these articulatory suppression. Lists of words were experiments deviates from veridical simultaneous visually presented to professional interpreters, interpreting. First, the source language input in student interpreters, and non-interpreter controls true-to-life interpreting differs greatly from the who were asked to read and memorize the words. analogous test materials in the above word- In the articulatory suppression condition, but not learning tasks and plausibly poses different in the control condition, the participants were processing demands. Second, in common asked to repeat the syllable bla during learning. simultaneous interpreting the input material is Following the presentation of each word list, spoken, not written, as in the above studies. recall was tested. The results showed that the (Alternatively, the input material is signed, in the professional interpreters were not adversely case of interpreting from a sign language to a affected by the concurrent articulation require- spoken language; see below.) In these respects, ment: They recalled equally as many words in the two further studies using articulatory suppression suppression condition as in the non-suppression mimicked simultaneous interpreting more closely condition. In contrast, the other groups of by presenting the input materials aurally. In one participants showed the typical adverse effect of them (Isham, 2000), monolingual under- of concurrent articulation and to an equal degree. graduate students listened to pre-recorded short These findings thus confirm the researchers’ stories that were interrupted by a tone from time hypothesis that interpreters use their working to time. Upon hearing this tone they had to write memory differently from non-interpreters, down as much as they could recall from the story including interpreting students with relatively so far. One condition involved just listening to the little experience on the task. Specifically they story. In a second condition the participants had suggest that the working memory of professional to repeat the word “double” while listening to the interpreters does not (always) need the support stories. To guarantee that the participants would of the phonological loop to operate effectively. process the meaning of the stories, after listening What remains is to explain why and under what to each of them true/false questions were posed circumstances this is and, specifically, why regarding the stories’ content. The results showed interpreters were not affected by concurrent both poorer recall and poorer comprehension in articulation in the experiments described above. the concurrent articulation condition than in the Additional research by Padilla, Bajo, and Macizo just listening condition. These results converge (2005) ruled out several options, including the with the above evidence of an adverse effect of possibility that interpreters are generally good at concurrent articulation with visual word input combining tasks even with tasks they have not in the non-interpreter group (Bajo et al., 1999; practiced in combination before. The results Padilla et al., 1995). suggested that only when the words to learn were highly familiar to them did concurrent articula- Because Isham (2000) did not test interpreters tion not impede performance. The authors it is impossible to tell whether they would be similarly affected by concurrent articulation or whether, as in the above studies with visual input,

334 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS they would again show immunity to articulatory interpreters. A couple of studies reviewed by suppression. The results of a second study, by Isham (2000) strengthen this conclusion and Darò and Fabbro (1994), indicate that prior suggest it can legitimately be generalized to pro- interpreting experience does not counteract an fessional interpreters. In one of these studies it adverse effect of concurrent articulation when the was found that professional interpreters also input is aural instead of visual. These authors had perform worse on a verbatim recall test after advanced Italian–English student interpreters interpreting short stories than after just listening first learn and then recall series of spoken digits to them (Isham, 1994). However, when the target in three conditions. In a listening condition the language was signed instead of spoken, recall per- digit series to be learned were presented bi- formance was equally good after listening and naurally over headphones and the participants interpreting (Isham & Lane, 1993). These findings were asked to reproduce as many of them as thus suggest that it is indeed the simultaneity of possible after each series. In an articulatory sup- listening and speaking that affects recall and that pression condition the participants were asked to even professional interpreters are affected by it. continuously utter two nonsense syllables during Furthermore, the equivalence of recall following learning. Finally, in a simultaneous interpreting listening and sign-language interpreting indicates condition the participants had to translate the that, unlike in advanced student interpreters (Darò digits in the other language during learning. The & Fabbro, 1994), the requirement to translate interpreting condition led to poorer recall than has no separate detrimental effect on the memory the other two conditions and the suppression performance of professional interpreters. condition showed poorer recall than the just listening condition. The results of this experiment Summary and conclusions thus suggest that the requirement to listen and speak at the same time has a negative effect on In the preceding sections I have presented a set memory even in participants trained in inter- of findings that typify professional interpreters. preting. They furthermore point out that the As compared with other bilinguals, they are additional requirement to translate has a separate relatively fast at executing a number of tasks that adverse effect on memory. tap basic sub-skills of the full interpreting task: retrieving the name for a concept to express, The results of a second experiment in Darò recognizing words, and assigning meaning to and Fabbro’s (1994) study, which included a them. They furthermore possess a relatively veridical simultaneous interpreting condition, large working memory capacity and exploit this confirmed these conclusions. In this experiment capacity in a way different from non-interpreters. the same participants as tested in the above digit This latter characteristic of professional inter- recall experiment listened to one story in one preters was revealed in experiments looking at condition and had to simultaneously interpret the adverse effects of articulatory suppression on another story in the second condition. Right after recall in a task that involved the learning of series listening to or interpreting the story, a recall test of visually presented words: Whereas both was administered in which they had to reproduce non-interpreters and interpreting students were the story verbatim. The interpreting condition led hampered by the concurrent articulation task, to lower recall scores than the listening condition. professional interpreters were not. This suggests Because of a missing articulatory suppression that professional interpreters can do without condition that does not at the same time involve working memory’s phonological loop when they translation, it is impossible to tell to what extent process written language material. However, this concurrent articulation and translation contri- null effect of articulatory suppression does not buted to the decreased recall separately. Neverthe- warrant the conclusion that they can also do less, these results indicate that the requirement without the phonological loop when processing of concurrent listening and speaking frustrates spoken source language input, because related the memory performance of advanced student

6. LANGUAGE CONTROL 335 experiments with aural input do show detrimental interpreting. Although it is unclear whether this effects of concurrent articulation on memory. ability can ever become fully automated, this skill can also be improved with practice. Before Earlier in this chapter the question was posed practicing it on the full task, in interpreting- as to how simultaneous interpreting is possible training programs it is often first exercised with at all. At least part of the answer to this question the so-called “shadowing” task (repeating a must lie in the above special skills that profes- spoken input verbatim) and the paraphrasing sional interpreters have developed. A processing task, which both share with simultaneous inter- characteristic that is known to be correlated with preting the requirement to concurrently listen processing speed is “automaticity”. By definition, and speak. Nevertheless they are considered an automated mental process does not require to be easier, because neither paraphrasing nor attention and can be performed in parallel to shadowing requires the involvement of two lan- both other automatic processes and processes guages and, in addition, shadowing does not that require attention. To the extent that fast pro- involve rephrasing the input. Practicing para- cessing indeed signals that the process in question phrasing also increases the fluency in rephrasing has become automated, the above finding that an input in a different way while remaining faith- professional interpreters are relatively fast at all ful to the input’s content. This is yet another skill sorts of sub-skills thus suggests they have auto- that simultaneous interpreters must acquire mated these sub-skills to a higher level than (e.g., Gerver et al., 1984; Moser-Mercer, 1994). It other language users have. As a consequence, is also trained by having the students generate they have more mental resources to spend on the synonyms for words. components of the full interpreting task that defy automation, such as resolving textual ambiguities A further ability to develop concerns the task’s or working out the meaning of conceptually coordination requirement, which Gile (1995, difficult materials. In other words, it is plausible 1997) identified as a separate component skill that the relatively large working memory capacity of simultaneous interpreting (see p. 316). As of professional interpreters follows directly mentioned, this component takes care of regulat- from their fast and, presumably, automatic word ing the exact portion of attention that at each recognition and word retrieval skills. moment in time must be assigned to each of the various ongoing activities. Research on the The studies discussed above all point towards acquisition of complex skills in general (in which the conclusion that professional interpreters are the present coordination component is usually not endowed with the above special abilities by referred to as “control of attention” or “attention nature. Instead they indicate that these abilities management”) has shown that training this result from the training of task-relevant linguistic component skill separately from training the full sub-skills in translation and interpreting pro- complex skill improves performance on the grams and from extensive on the job practice. In full task (Gopher, 1992; Gopher, Weil, & Siegel, addition, one study has shown that at least one of 1989). Interestingly, this work has also shown the required sub-skills, fluent word retrieval, may that coordination skills acquired in a particular not require any specific training but comes with a context can transfer between complex tasks; for high level of proficiency in both languages. instance, from gaming to air traffic control. Plausibly then, to come to master the coordin- The special abilities of interpreters that I ation skill involved in interpreting also does not focused on in the preceding sections—that is, require it to be practiced with the full, awesome fast word-retrieval and word-recognition skills task. and a large working-memory capacity—by no means constitute the complete set of skills that A final ability to develop (if the interpreter-to- interpreters need to develop to successfully be is not naturally endowed with it) is immunity practice their profession. Earlier I presented to speed stress. It is required for interpreting the requirement to concurrently listen and speak because of the speaker-based nature of simul- as a major stumbling block in simultaneous