Language and Cognition in Bilinguals and Multilinguals_ An Introduction

236 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS play in the way this conceptual system maps onto in bilinguals has pursued the question of whether language or, more precisely, instantiates language bilingual language production is language- (Francis, 2005, p. 252): selective or language-nonselective or perhaps both, depending on specific characteristics of the Any of the concepts a person can know linguistic and extra-linguistic context and/or ought to have the potential to be expressed the bilingual speaker. A number of the relevant in any human language. Of course, the studies used one version or other of the picture- concepts actually realized in an individual’s naming task. Of these, a subset has employed the language input or output will vary with sys- picture–word interference paradigm, where the tematic patterns across languages. Semantic picture is accompanied by a distracter word and representations may be those concepts that the influence of this distracter on picture-naming are referred to by particular words or sen- performance is determined. This paradigm tences [. . .]. Word meanings, or semantic developed from the Stroop task and was first representations of words, would be a par- used by Rosinski, Golinkoff, and Kukish (1975) ticular type of concept. to study automatic reading skills of children (see Levelt, 1999, for an historical overview of the According to this view, word meanings do not task’s roots). exist separately from a knowledge store that contains non-linguistic conceptual information In an early bilingual study using the distracter but are represented by subsets of units in this methodology Ehri and Ryan (1980) showed that store. One and the same conceptual element in English–Spanish bilinguals named pictures more this store may serve as a component part of the slowly, in both of their languages, when a word meaning of many words. As Francis puts it: “any from the non-target language was superimposed word meaning is identified with a subset or a over the picture than when a neutral distracter particular pattern of activation across the entire stimulus (a series of Xs) was superimposed. The [conceptual] system” (Francis, 1999, p. 195). This interfering effect of the superimposed word view appears to be most consistent with models occurred despite the fact that the participants of word production that do not assume different were asked to ignore the word. In a further study levels of conceptual and lexical-semantic repre- Mägiste (1984b; see also Mägiste, 1985), testing sentations but it remains unclear what it is then German–Swedish bilinguals at various levels of that activates the subsets of nodes that instantiate proficiency in their L2, obtained this same inter- word meanings in the conceptual system. ference effect and showed that its size depended on the learners’ relative proficiency in the two Having thus familiarized the reader with the languages: The stronger the non-target language, general outline of models of monolingual and the larger the interference effect. Both these bilingual speech production and the associated studies thus hinted at the occurrence of language- concepts and terminology, I can now proceed nonselective activation in bilingual word pro- with a discussion of the empirical evidence and, duction. Furthermore, Mägiste’s finding that the first, the details of the general methodology relative proficiency in the two languages influ- employed in many of the pertinent studies. ences the size of the interference effect converges with similar results in the comprehension studies PICTURE NAMING discussed in Chapter 4: The stronger the non- target language, the larger the chance it will per- Introduction meate in processing the target language. However, not having been specifically designed to do so, As mentioned, just as in the study of bilingual these studies did not provide any information on comprehension, research on language production the exact level or levels in the bilingual production system where co-activation of the non-target language occurred. Specifically, they did not

5. WORD PRODUCTION AND SPEECH ACCENTS 237 state whether it stopped at the lemma level or types of distracter words have been used (not proceeded on to the phonological level. necessarily all of them in one and the same study). Table 5.1 shows example materials used by More recent bilingual picture-naming studies, Levelt et al. (1991) in a monolingual Dutch study: using both the distracter methodology (Costa & The distracter word is identical to the picture’s Caramazza, 1999; Costa, Miozzo, & Caramazza, name (Identical), semantically related to the 1999; Hermans et al., 1998) and versions of the picture’s name (Semantic), phonologically related task in which pictures are presented without to the picture’s name (Phonological), phono- distracters (Colomé, 2001; Costa et al., 2000; logically related to a word that is semantically Gollan & Acenas, 2004; Kroll, Dijkstra, Janssen, related to the picture’s name (Phonological- & Schriefers, 2000; Rodriguez-Fornells et al., Semantic; e.g., in Table 5.1 stoep is phonologically 2005), focused on various aspects of the theo- related to stoel), or unrelated to the picture’s retical contrasts introduced earlier. For example, name (Unrelated). they examined whether both the lemma and phonological form of words from the non- In comparison to neutral or unrelated dis- response language are activated or whether or not tracters, semantic distracters (e.g. a picture of a backward spreading of activation between the desk accompanied by the distracter word chair) different types of representations occurs. The dis- have been shown to slow down the picture- tracter studies among this set were closely mod- naming response (e.g., Bajo, Puerta-Melguizo, & eled on a number of highly influential mono- Macizo, 2003; Levelt et al., 1999; Roelofs, 1992; lingual production studies that used a sophisti- Schriefers et al., 1990; Starreveld & La Heij, 1995, cated methodology trying to answer the analo- 1996). In production models that assume the gous questions regarding monolingual process- existence of lemma representations this effect is ing. In the next sections I will first describe this attributed to a competition between the lemmas methodology and I will then continue with a of picture and distracter, which are both activated review of the bilingual studies. upon presentation of the picture (see, e.g., Figure 5.3). The occurrence of this effect depends on The picture–word interference task the time relation between picture and distracter: Schriefers et al. (1990; a study in which the dis- Methodology tracters were presented aurally) obtained the effect only when the semantic distracter preceded The monolingual picture–word interference stud- the picture (i.e., with a negative SOA), but ies alluded to above investigated the processes of not when they coincided or when the picture was lemma selection and phonological encoding by presented first. manipulating both the type of distracter words presented with the picture and the time interval Interestingly, this same study showed com- between the presentation of distracter and pic- pletely different results for phonological dis- ture. Distracter and picture were presented simul- tracters. First, the effect of these distracters was taneously (a “stimulus onset asynchrony”, SOA, facilitative rather than inhibitory: they speeded of 0 ms), the distracter preceded the picture (a up the picture-naming response. Second, these negative SOA) or followed the picture (a positive facilitative effects occurred when picture and dis- SOA). Across the various studies both visual tracter coincided and when the picture was pre- and auditory distracters have been used. The sented first; that is, with SOAs that had shown main dependent variable in most of these studies null effects of semantic distracters. Conversely, no (as well as in their bilingual counterparts) such effect was observed when the phonological was picture-naming latency: the time duration distracter preceded the picture. Again, this result between the onset of picture presentation and the is opposite to the one observed with semantic onset of the participant’s naming response. Five distracters. The combined data thus suggest early semantic inhibition of semantic distracters and later phonological facilitation of phonological

238 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS TABLE 5.1 Example materials used by Levelt et al. (1991) Picture’s name Identical Semantic Phonological Phonological-Semantic Unrelated distracters distracters distracters distracters distracters bureau bureau stoel buurman stoep muts (“desk”) (“desk”) (“chair”) (“neighbor”) (“pavement”) (“cap”) cactus cactus stekel kakkerlak steno tas (“cactus”) (“cactus”) (“sting”) (“cockroach”) (“shorthand”) (“bag”) geweer geweer oorlog gewei oorzaak koets (“rifle”) (“rifle”) (“war”) (“antlers”) (“cause”) (“coach”) vinger vinger ring vink rits kwast (“finger”) (“finger”) (“ring”) (“finch”) (“zipper”) (“brush”) radio radio muziek radar museum kerk (“radio”) (“radio”) (“music”) (“radar”) (“museum”) (“church”) Example materials used by Levelt et al. (1991) in a monolingual Dutch picture–word interference study. The meanings of the Dutch words are in parentheses. distracters. The authors attributed the facilitative to depend on the time relation between picture effect of phonological distracters to boosted acti- and distracter. This is illustrated in Figure 5.6, vation of the phonological segments that need to in terms of simplified processing models that be selected in the final stage of picture naming: distinguish between lexical semantics (meaning) The phonological analysis of, say, the auditory and phonology only. distracter buurman (Table 5.1) activates a subset of the same phonemic segments that must be As shown, depending on the SOA, either the retrieved to name the target picture (this name meaning activation stage or the phonology acti- being the phonologically similar bureau). This vation stage of picture naming and distracter additional activation of the targeted phonological processing coincide. For the phonological com- elements from a second source speeds up the ponents to coincide, the distracter must be picture-naming process. presented relatively late because phonological activation is the final step in picture naming The interaction between SOA and type of dis- but the first step in distracter processing (if the tracter, semantic or phonological, can be under- distracter is presented aurally). In contrast, stood if we consider the various processing stages, because conceptual activation is an early process- and their order, in the two processes that operate ing stage in picture naming but a late step in dis- in parallel in the picture–word interference task: tracter processing, for this processing component naming the picture and processing the distracter. to coincide in picture and word processing the These two processes operate in opposite direc- distracter must be presented earlier. Note that tions, top-down and bottom-up: Following a per- this analysis assumes that semantic interference ceptual analysis of the picture, picture naming and phonological facilitation result from coincid- begins with the activation of a set of nodes in the ing conceptual and phonological processing, conceptual level of representation. From there respectively. activation floods downward to the phonological level, activating the elements in the intermediate From the fact that semantic and phonological level (or levels) en route. Distracter processing distracters are effective at different SOAs, starts off at the phonological level and works Schriefers et al. (1990) concluded, contrary to upward from there. At some point in the system interactive activation models of lexical access (see the two processes will meet and affect each other. p. 227), that lexicalization in word production is What this meeting point is and, consequently, composed of two unidirectional successive stages, what the nature of the interference will be, is likely one of meaning activation and a second of phonological activation. Levelt et al. (1991) and

5. WORD PRODUCTION AND SPEECH ACCENTS 239 The dependency of the semantic interference and phonological facilitation effects in the picture–word interference task on the stimulus onset asynchrony between word distracter and picture. Jescheniak and Schriefers (1998) built on this synonyms of this word are also phonologically study by posing the question of whether phono- activated (see Peterson & Savoy, 1998, for logical activation only occurs for the lemma that additional evidence). However, phonological emerges as the winner from the cohort of initially activation does not appear to occur for words that activated lexical candidates (the semantic cohort) are less-strongly semantically related to the target. or also for the other members of the semantic cohort. The latter is the position taken by the Bilingual studies cascaded-processing models presented earlier p. 227). (This question can be studied with the Hermans and his collaborators (1998) were the phonological-semantic distracters; see Table 5.1.) first to use the above distracter methodology to A difference between these two studies was that examine whether bilingual word production is the semantic distracters in Levelt et al.’s study language-nonselective. More precisely, they posed were (non-synonymous) members of the same the question of whether, when bilinguals are semantic category as the picture’s name (e.g., pic- asked to name pictures in their L2, the represen- ture’s name: sofa; semantic distracter: chair), tations of the corresponding translations in L1 whereas Jescheniak and Schriefers also used are also activated. If this were to turn out to be semantic distracters that were near-synonyms of the case, they furthermore wanted to know the picture’s name (e.g., picture’s name: sofa; whether this activation stops at the translation’s semantic distracter: couch). The combined results lemma representation or is transmitted further of these two studies suggest that phonological down to the phonological level so that both words activation does not occur exclusively for the in a translation pair are phonologically encoded. targeted word (the picture’s name) but that near- The above results of Jescheniak and Schriefers

240 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS (1998) and Peterson and Savoy (1998) in particu- to the word it translates. I will refer to these as lar suggested the latter might be the case because, “phonological translation” distracters. The plausibly, translation equivalents share at least fourth group of distracters consisted of words as much meaning between one another as near- that were completely unrelated to the pictures’ synonyms of one and the same language do. names (e.g., present). These unrelated distracters again provided a baseline condition from which The participants in this study were Dutch– the effects of the other distracters could be English bilinguals with a high level of proficiency assessed. in L2 English, yet clearly dominant in L1 Dutch. Each picture to be named in L2 English (e.g., a Two experiments were run and in both of them picture of a mountain) was presented with an the stimulus onset asynchrony (SOA) between aural distracter, which could be of four types. picture and distracter was varied. Distracters Two of the types were the semantic and phono- preceded the onset of the picture by 300 ms logical distracters presented above (semantic: (SOA –300) or 150 ms (SOA –150), coincided valley; phonological: mouth). A further type of with the onset of the picture (SOA 0), or followed distracters concerned words phonologically the picture 150 ms after the latter’s onset (SOA similar to the Dutch translations of the pictures’ +150). The crucial difference between the two names (e.g., bench for berg, where berg is the experiments was that in one of them the distrac- Dutch translation of mountain). Note that this ters were all English words (as in the examples type of distracter is conceptually similar to above), whereas in the second the distracters were the “phonological-semantic” distracters in the Dutch words. Table 5.2 shows some example monolingual studies (see Table 5.1). After all, a materials from both experiments. Note that the word’s translation is semantically closely similar semantic distracters in the second experiment TABLE 5.2 Example materials used by Hermans et al. (1998) Picture’s name Semantic distracters Phonological distracters Phonological Unrelated translation distracters distracters mountain valley mouth bench present (berg) chicken people power blanket leather belly reef jaw peacock beetle snare slack bag (pauw) button pig span bill belt (riem) snail (slak) pin (speld) mountain dal mouw berm kaars (berg) (“valley”) (“sleeve”) (“roadside”) (“candle”) peacock kip piste paus verf (pauw) (“chicken”) (“ring”) (“pope”) (“paint”) belt leer berk riet park (riem) (“leather”) (“birch”) (“reed”) (“park”) snail kever sneeuw slag kroon (slak) (“beetle”) (“snow”) (“blow”) (“crown”) pin knoop pint spek koek (speld) (“button”) (“pint”) (“bacon”) (“biscuit”) Example materials used by Hermans et al. (1998) in a bilingual picture–word interference study in which Dutch–English bilinguals named pictures in L2 English. Top half: English distracters; bottom half: Dutch distracters. The Dutch names of the pictures and the meanings of the Dutch distracters are presented in parentheses.

5. WORD PRODUCTION AND SPEECH ACCENTS 241 allow a test of the hypothesis that the semantic and picture coincided (SOA 0). This finding was cohort activated by the picture includes not only interpreted as evidence of co-activation of the the lemma of the Dutch translation equivalent pictures’ L1 names. of the English target (berg), but also lemmas of words semantically related to the target’s transla- The authors suggested this evidence of tion (dal, “valley”). language-nonselective word production might not have been more robust (occurring in more The results of these two experiments are pre- SOA conditions) because, due to differences in sented in Table 5.3, in terms of the difference the pronunciation of corresponding phonemes in scores between the unrelated control condition English and Dutch, the phonological overlap on the one hand and all three experimental between the English phonological translation conditions on the other hand. distracters (bench) and the Dutch names of the pictures (berg) may in fact have been rather small. The experiment with English distracters This is why they ran their second experiment, showed a facilitation effect of phonological dis- assuming that Dutch phonological translation tracters (mouth) in all four SOA conditions, distracters (e.g., berm) share more phonology suggesting (contrary to Schriefers et al., 1990) with the Dutch name of the pictures. With this that phonological coding of the picture’s name set-up, the phonological translation distracters already starts during an early stage of lexical exerted an inhibitory effect on English picture access. Furthermore, it showed an inhibitory naming in all but one of the SOA conditions, effect of semantic distracters (valley), except in condition SOA +150 being the exception. How- the condition where the distracter followed the ever, this time the effects of the phonological picture (SOA +150). As with the analogous effect distracters disappeared in three of the SOA con- in the monolingual studies, this finding was ditions. The reason may be that now the overlap attributed to the occurrence of co-activation of between the (Dutch) phonological distracters semantic competitors during lemma selection. (mouw) and the target picture’s English name Finally, phonological translation distracters (mountain) may have been too small to cause an (bench) slowed down picture naming, but only effect. Interestingly, in this experiment too the reliably so when the presentation of the distracter semantic distracters caused an effect (although it was only significant in Condition SOA –150), TABLE 5.3 providing some evidence that the semantic cohort Hermans et al.’s (1998) results activated by the picture also includes lemmas of words semantically related to the target’s transla- SOA − SOA SOA SOA tion (dal, “valley”). 300 ms −150 ms 0 ms +150 ms Hermans and his coworkers discussed these results in terms of the three types of production English distracters 19* 24* 31* 64* models introduced earlier: the discrete two-stage Phonological –44* –19* –31* 10 models, the unidirectional cascade models, and –14 –28* –5 the interactive activation models (p. 227). They Semantic –5 concluded the evidence to be equivocal on this point because the time course of the various Phonological- effects, as assessed with the different SOAs, did translation not fit the predictions of any model. Importantly though, the results—and especially the strong Dutch distracters –8 –5 –10 35* phonological translation effects in their second Phonological –17 –37* –19 –13 experiment—did indicate that at some point dur- –30* –38* –35* ing picture naming in the weaker L2 the picture’s Semantic –7 translation in L1 is also activated. In other words, they suggested that bilingual word production Phonological- translation Differences (in ms) between the mean latencies for the unrelated condition on the one hand and the phonological, semantic, and phonological-translation conditions on the other hand at each Stimulus Onset Asynchrony (SOA). Differences that are significant are marked with *. Adapted from Hermans et al. (1998).

242 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS is language-nonselective. However, the evidence target language. In the next section I will present does not indicate whether only lemma selection or the data that led them to propose this alternative also phonological encoding is language- account. nonselective. Lexical activation versus lexical selection What has remained somewhat implicit in the above discussion is the common assumption that The type of production model Costa and his response selection in word production tasks colleagues take as a starting point in their studies is directly related to the amount of nuisance acti- does not assume lemma representations but dis- vation in the language system. Regarding the tinguishes between semantic nodes, lexical nodes, bilingual studies this means that activated lexical and sublexical nodes (see Figure 5.5). Each lexical nodes in the non-target language (e.g., the lemma node represents a word’s complete phonological of Dutch berg) compete with lexical nodes from form and each sublexical node represents a com- the target language during the selection process ponent part of a word’s phonology, for instance, (e.g., of mountain). In a series of experimental one phoneme. As in Caramazza’s (1997) model and theoretical articles, Costa and his colleagues (Figure 5.4) activation at the semantic level considered an alternative—that there may be flows directly to the (phonological) lexical nodes activation in the non-targeted language system (which Caramazza, 1997, called “phonological that does not compete with elements of the target lexemes”). language during the selection process but is ignored instead. One of the pertinent papers Within the theoretical context of such a pro- (Costa, Colomé, Gómez, & Sebastián-Gallés, cessing system, Costa et al. (1999) carried out a 2003) is a replication of Hermans et al. (1998) study employing the distracter methodology. The with nearly balanced, early Spanish–Catalan participants in this study were highly proficient, bilinguals. The phonological translation effect balanced, Catalan–Spanish bilinguals, who again materialized in some of the conditions named pictures in their L1 (Catalan). Contrary to (as did the effects of the other types of related Hermans et al. (1998) and Costa et al. (2003), the distracters). However, in their theoretical analysis distracters were now presented visually rather of this result Costa et al. (2003) argued this than aurally. This difference is not a trivial one effect is also compatible with the view that the because the modality of the distracter determines selection process is language-selective, meaning the exact order in which the various types of that only activated words from the target lan- memory representations will become activated. guage compete for selection. This alternative view Specifically, when an aural distracter is presented, was first introduced by Costa et al. (1999) and it will immediately activate the corresponding Costa and Caramazza (1999) and encompasses phonological sublexical nodes, and this prior to the theoretically important distinction between access of the corresponding phonological lexical language-nonselective versus language-selective node and, finally, word meaning (see also Chapter activation on the one hand and language- 4, Figure 4.9, BIMOLA, a model of auditory nonselective versus language-selective selection word recognition). In contrast, a visual distracter on the other hand. The crucial point these will activate phonology only indirectly, via acti- authors make is that during word production vated orthographic sublexical and lexical nodes lexical elements in both of the bilingual’s lexicons (see Chapter 4, Figure 4.4, SOPHIA). As a con- may be activated simultaneously, but that from sequence of this difference, the SOA range this it does not imperatively follow that the acti- suitable to test the above theoretical distinction vated elements in both languages are considered between discrete two-stage and unidirectional during the actual selection process. An alternative cascade models of speech production should dif- to language-nonselective (or “language- fer between studies that present visual distracters nonspecific”) selection is that the selection and those that present auditory distracters (see mechanism ignores the activation in the non- Starreveld, 2000, for a similar line of reasoning).

5. WORD PRODUCTION AND SPEECH ACCENTS 243 However, the most important new feature of questions. First, if the participants can switch off this study was that it added a new distracter con- the non-target language completely, no effect of dition, the so-called “identity” condition. In presenting a distracter in the non-target language this condition the distracter was the name of the should emerge. This state of affairs would receive picture in the target language or in the non-target support if picture-naming time in the identity language. For instance, a picture of a table to be condition and in an unrelated control condition named in L1 Catalan (taula) was accompanied (with, e.g., the Spanish distracter jamon, “ham”) by the visual word taula or by its L2 Spanish were the same. In contrast, a difference in picture- translation mesa. I will confine my discussion to naming latency between these two conditions the rationale of including this condition and the would suggest co-activation of the lexical node related results. in the non-target language. Second, if such a difference indeed occurs, the direction of the The authors started out questioning the com- effect would reveal the nature of the selection mon assumptions in studies on monolingual word process: The language-nonspecific (or language- production that the most highly activated lexical nonselective) selection hypothesis predicts longer node is selected from the set of activated lexical latencies (that is, inhibition) in the cross-language nodes and that the ease with which a target is identity condition than in the unrelated condi- selected depends on the level of activation of the tion, whereas the language-specific (or language- non-target competitors: The higher one or more selective) selection hypothesis predicts shorter non-target competitors are activated, the harder latencies (facilitation) in the identity condition. target selection is. In other words, ease of selec- tion in monolingual word production is thought The grounds for these predictions are illus- to be a function of the amount and fierceness of trated in Figure 5.7 (note that the level of sublexi- co-activation in the system. Costa and his associ- cal nodes is not shown). If lexical node activation ates hypothesized that this imperative relation is language-selective, a picture of a table will between activation and selection might not hold first activate the corresponding semantic nodes for the bilingual case: Activation at the semantic and from there the lexical node representing level of representation may activate lexical nodes Catalan taula but not the lexical node repre- in both languages simultaneously, but only the senting Spanish mesa. In a language-selective activated lexical nodes of the target language activation account the cross-language identity may be considered for selection (and those of the distracter mesa may activate the corresponding non-target language ignored). The authors thus semantic nodes (indirectly, via its orthographic divided the question of whether bilingual word lexical node in an orthographic lexicon, as shown production is language-selective or language- in Figure 5.7) but this activation is not trans- nonselective into two sub-questions: (1) Do or do mitted down to the (phonological) lexical node not activated semantic nodes transmit their acti- level. Alternatively, and in line with a strong vation to lexical nodes in both languages, and if version of language selectivity, the processing of they do, (2) do the activated lexical nodes of a cross-language identity distracter is blocked both languages compete for selection or does the altogether so that it also does not activate the selection process only consider the activated corresponding semantic nodes. If, instead lexical lexical nodes of the target language? In other node activation is language-nonselective, both words, if activation is language-nonselective, the picture and the cross-language identity dis- selection may still be language-selective. tracter mesa will first activate the corresponding semantic nodes; from there, the lexical nodes The identity condition, and especially its cross- representing taula and mesa will both be acti- language version (in which the picture, to be vated, and both relatively highly because the named in Catalan, was accompanied by its level of activation in the activation-sending name in Spanish; e.g., a picture of a table, taula in semantic nodes, coming from two sources, is Catalan, accompanied by mesa, its Spanish relatively high. name) provided the means to answer both these

244 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS Picture naming with visual word distracters in Catalan– Spanish bilinguals. The picture must be named in L1 Catalan. Language- nonspecific lexical selection means that the highly activated lexical nodes representing Catalan taula and Spanish mesa compete for selection. Language- specific lexical selection means that the highly activated mesa node does not compete with the taula node during selection because it is ignored by the selection mechanism. Adapted from Costa et al. (1999). If the selection process is language- identity condition than in the unrelated control nonselective, both these highly activated nodes are condition. The fact that no null effect of dis- considered for selection. Non-target mesa will act tracter type (identical vs. control) was obtained as a strong competitor in the selection process suggests that lexical activation is language- (as compared with the lexical node representing nonselective. The fact that the observed effect was an unrelated control distracter, which is only facilitative, not inhibitory, is consistent with the activated by the distracter word and not by the idea that lexical selection is language-selective: picture) and slow down the picture-naming Even a strongly activated lexical node in the non- response. In contrast, if the selection mechanism target language does not slow down processing. is language-selective, only considering activated Apparently, only the activated lexical nodes in the lexical nodes in target Catalan, it will not suffer target language are considered for selection. from the fact that Spanish mesa is also highly activated. At the same time the fact that, as com- In a follow-up study, Costa and Caramazza pared to an unrelated distracter, the Spanish dis- (1999) obtained exactly the same pattern of tracter mesa has increased the level of activation results, now testing proficient but non-balanced in the taula lexical node (via the activated seman- Spanish–English and English–Spanish bilinguals tic nodes), selection of this node should be faster naming pictures in L1 Spanish (the Spanish– in the cross-language identity condition than in English group) or L2 Spanish (the English– the unrelated control condition. Spanish group), and with the distracters printed in English. The data pattern consistent with The data agreed with a processing model that language-selective lexical selection occurred not assumes language-nonselective activation fol- only when the pictures were named in the par- lowed by language-selective selection: Picture- ticipants’ dominant language but also when naming latency was shorter in the cross-language naming was in the weaker L2. This suggests that

5. WORD PRODUCTION AND SPEECH ACCENTS 245 even lexical node activation in the stronger L1 production specify an interval in which the can be ignored during the selection process. As distracter manipulation is potentially effective. noted by the authors, it remains to be seen at what He then concludes that: “It is only when these stage during second language acquisition the two intervals overlap that a specific context effect mechanism involved in language-specific selection [= distracter effect] can be found” (Starreveld, becomes functional (assuming it is not functional 2000, p. 521). A particular null effect (e.g., of a right from the onset). But whatever the answer to phonological distracter with a negative SOA) that question, the results of this second study may therefore be due to a failure to catch this demonstrate that balanced bilingualism is not a critical point of interaction rather than indexing a prerequisite for the mechanism to be functional. theoretically important feature of the speech pro- Yet it remains possible that during earlier stages duction process (e.g., discrete serial processing of acquisition selection is language-nonselective of meaning and phonology). For similar reasons, (as inhibition caused by a cross-language identity Bock (1996), in a review of methods used in prime would demonstrate). speech production research, made a plea for a more analytic consideration of these methods. Evaluation She generalized this statement to tasks used in research domains other than speech production As the reader may have experienced, the picture– such as word recognition, where the reverse some- word interference task is a complex paradigm times happens: Here the actual process of inter- and, consequently, the interpretation of the data est, recognition, is sometimes confounded by a emerging from the relevant studies is complex as production component of the task, as is the case well and equivocal at times. As pointed out by in word naming (see Chapter 4). Just as a theory Kroll et al. (2005), a likely cause of this data of speech production based on the distracter ambiguity is the simultaneity (or temporal prox- methodology cannot do without a theory of imity) of two stimuli: the picture and the aural recognition, so a theory of recognition based on or visual distracter. As a consequence of this word naming cannot do without a theory of pro- simultaneity, two processes proceed in parallel duction, she argues. and, initially, in opposite directions: An encoding process works from the conceptualized picture So where does all this leave us now? Can we down to the actual naming response and a build our theory of bilingual speech production second, decoding, process works upward, from on evidence collected by means of a methodology a phonological or orthographical analysis of the that is surrounded with uncertainty? As the distracter word to the activation of a conceptual reader may recall, this state of affairs is not at all representation. At some point these two processes uncommon in experimental research and is meet and influence one another. Sophisticated caused by the fact that many tasks, in addition to experimental procedures are required to deter- reflecting the theoretically important cognitive mine with certainty the exact point of interaction, process of interest, pick up additional nuisance procedures that may contain elements based on processes that are hard or impossible to avoid. In inaccurate assumptions. These aspects of the task Chapter 4 (p. 170) I have presented a common complicate the interpretation of what is going on solution to this problem; namely, to look for con- and may lead to inaccurate conclusions. verging evidence from other tasks. If a particular finding emerges across a number of tasks that It is for similar reasons that Starreveld (2000) arguably all capture the core of the process of concluded that, in order to understand the origin interest, this effect is likely to reflect this process of distracter effects in picture-naming studies, a (see also Figure 4.1). In fact, converging evidence theory of how the distracter is perceived and pro- of language-nonselective activation in speech cessed is as indispensable as a theory of how the production has already been gathered. It has, picture is perceived and processed. According for instance, been obtained in studies that have to him, both such a theory of perception and of employed versions of the picture-naming task

246 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS that do not involve two simultaneous processes “non-cognates”. The authors compared the that during their initial stages work in opposite naming of pictures that lexicalize in cognate directions through the lexical system. It is to these translations in the participants’ two languages studies that I will turn to next. with the naming of pictures that lexicalize in non-cognates. Figure 5.5 illustrates the word pro- Simple picture naming duction processes assumed for Catalan–Spanish cognate and non-cognate words, respectively. The Evidence of language-nonselective phonological representations of Catalan gat and Spanish gato encoding share a large part of their constituent sublexical nodes (/g/, /a/, and /t/), whereas there is little over- The main theoretical issue addressed in the simple lap between the sublexical phonological represen- picture-naming studies to be presented here is tations of Catalan taula and Spanish mesa (/a/). If whether phonological encoding occurs for the activated lexical nodes in the non-target language translation of the picture’s name in the non- indeed send activation down to the phonological response language. (The qualification “simple” level, picture naming should be faster for cognates refers to the fact that in this version of the task than for non-cognates because in the case of cog- the picture is presented on its own, unaccompan- nates the sublexical phonological representations ied by a distracter; in other respects this task of the targets to be produced receive activation too may require rather complex processing.) Such from two sources: from the target’s lexical node, a finding would support (the bilingual version but also from the lexical node of its translation. of) word production models that assume (uni- If, however, a single lexical node (the one repre- directional or interactive) cascaded processing. senting the target word) is first selected from In contrast, results suggesting otherwise would the set of activated lexical nodes and only this support the discrete two-stage models of word node sends down its activation to the phono- production. Part of the relevant evidence was logical sublexical nodes, then picture naming provided by Costa et al. (2000). In choosing should be equally fast for pictures with cognate their stimulus materials they were guided by and non-cognate names. The results were clear- Jescheniak and Schriefers (1998) and Peterson cut: Picture naming in L2 (Spanish) by Catalan– and Savoy (1998), who had examined the occur- Spanish bilinguals was faster for cognate pictures rence of phonological encoding (henceforth also than for non-cognate pictures. More recent “phonological coding” or “phonological acti- studies have shown this same effect to occur in vation”) of non-target items in monolingual picture naming by Dutch–English bilinguals picture naming (see p. 239). These authors (Christoffels, De Groot, & Kroll, 2006; Christof- obtained evidence of phonological activation of fels, De Groot, & Waldorp, 2003; Kroll et al., near-synonyms of the target but not of words 2000) and Spanish–English bilinguals (Hoshino less-strongly related to the target. This suggested & Kroll, 2007). Interestingly, these latter authors that nearly complete semantic overlap with the have shown that the effect also emerges in target is a prerequisite for phonological encoding different-script bilingualism: A cognate effect was of a non-target to occur. Because translation observed for Japanese–English bilinguals as well, equivalents share most of their meaning between and it was as large as the effect obtained for the languages, if phonological encoding of non-target Spanish–English bilinguals. This suggests that the lexical nodes occurs at all it should at any rate additional orthographic similarity of same-script occur for the picture’s name in the non-target cognates does not contribute to the effect. language. In a further experiment Costa and his col- To demonstrate such phonological coding, leagues tested the hypothesis that the amount Costa et al. (2000) exploited the fact that trans- of activation sent down to the phonological sub- lation pairs may or may not share phonology. lexical level is proportional to the lexical node’s The former are known as “cognates”, the latter as activation level. If true, a relatively large cognate

5. WORD PRODUCTION AND SPEECH ACCENTS 247 effect should be observed when the non-response embracing the idea that cognates and non- language is the participant’s dominant language. cognates have different semantic representations The ground for this hypothesis is that under was that they found it hard to see how degree of those circumstances a cognate target’s sublexical form similarity between a pair of translations representation should receive relatively much could affect the representation of meaning. activation from the lexical node of its translation (for instance because for words in the dominant However, a reason why the meanings of language the connections between the various cognates and non-cognates might be represented layers of nodes in the system are relatively differently in bilingual memory is that L2 learners strong). The data confirmed this prediction: are likely to notice the form similarity between Larger cognate effects were obtained when cognate translations and may assume that when pictures were named in the weaker language than translations have similar forms they are likely in the stronger language. These experiments thus to have more similar meanings than when they suggest that the picture’s names in both languages have different forms. As a consequence they may are phonologically encoded. This finding rules exploit different learning strategies when learning out discrete two-stage models of bilingual word cognate and non-cognate translations and differ- production and supports cascaded models. Both ent meaning representations may ensue. Evidence the unidirectional and interactive versions of that the cognate status of translation pairs affects cascaded models are compatible with the results L2 vocabulary acquisition has been presented in (Figure 5.5 illustrates the unidirectional cascaded Chapter 3 (pp. 108–110 and 119–121). Further- models). In a later study that examined the lexical more, direct evidence that the meaning represen- bias effect (see p. 226) in bilingual speech pro- tations of cognates and non-cognate pairs do duction, Costa, Santesteban, and Ivanova (2006) indeed differ from one another does in fact exist opted for the interactive models. (e.g., Taylor, 1976; Van Hell & De Groot, 1998a). Rejecting differential meaning representation as a The present interpretation of the cognate effect possible source of the cognate effect in picture in picture naming is primarily one in terms of naming thus seems premature. In conclusion, the processing, not representation: It is attributed to a form similarity of cognate translations may have process of language-nonselective phonological become reflected in one or more representational coding that differentially affects cognates and differences between cognates and non-cognates non-cognates. In Chapter 4 (pp. 203–205) it was that may underlie the cognate effect in picture suggested that the cognate effects obtained in naming and other tasks. bilingual word recognition studies might origin- ate from representation differences between cog- Further evidence nates and non-cognates. If such differences exist, the possibility should be considered that these The possibility that representational differences also underlie the above cognate effects in picture between cognates and non-cognates are a source naming. Aware of this possibility, Costa et al. of the cognate effects forces us to conclude that the (2005) considered two options: shared morpho- data are equivocal: The cognate effects observed in logical representations for cognates but not for the simple picture-naming task might not index non-cognates (e.g., Sánchez-Casas & García- language-nonselective phonological encoding Albea, 2005) and shared semantic representations after all. However, the authors of a couple of other for cognates but not for non-cognates (e.g., Van studies that did not rely on the distinction between Hell & De Groot, 1998a). However, they regarded cognate and non-cognate translations seem to neither of these options as plausible. They won- have a very strong case indeed when arguing that dered how, if cognates share one and the same phonological encoding is language-nonselective. morphological stem, ultimately two different Particularly noteworthy about one of these studies phonological forms can be realized in speech, one (Colomé, 2001) is that the participants were kept for each language. Their main reason for not completely ignorant about the bilingual nature of

248 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS the experiment and that the non-response lan- no reason to believe their bilingualism was being guage was in no way physically present. From the tested and non-target Spanish was in no way discussion of the recognition studies presented in present in the experiment. Chapter 4 the reader may recall that a common weakness of many studies in this research area is In addition to trials that required a “yes” that the non-target language is actually physically response, there were two types of trials that each present in the form of stimulus materials from that required a “no” response: trials in which neither language or implied by the general context of the the Catalan nor the Spanish name of the picture experiment. It is not surprising that under those contained the specified phoneme, and trials in circumstances the non-target language is also which the picture’s Catalan name (e.g., taula, activated. Unequivocal evidence of language- “table”) did not contain the specified sound (e.g., nonselective processing demands complete ignor- /m/) but whose name in Spanish did contain this ance on the part of the participants that their sound (mesa). If the picture’s name is encoded bilingualism is being tested, as well as an experi- phonologically in both languages, trials of the mental set-up and stimulus set that do not trigger latter type would be relatively hard to reject. The the non-target language into activity, if complete reason is that the specified phoneme would be deactivation is an option at all (this being the activated and cause interference because the par- debated issue). Colomé’s study meets these ticipant might be inclined to respond “yes”, an requirements. In contrast, the bilingual picture– inclination that must subsequently be suppressed. word interference studies discussed before did not In agreement with this prediction, in all three do so because the non-target language was physic- experiments response times were longer for ally present in the distracters. this latter type of trial, providing evidence of language-nonselective phonological encoding. Colomé (2001) used a version of the “phoneme-monitoring task”, which was first used Also looking for evidence of phonological to study monolingual speech perception. In the activation of the picture’s name in the non-target original version of this task participants are pre- language, Rodriguez-Fornells et al. (2005) modi- sented with series of sentences or words and are fied this clever paradigm: Spanish–German asked to monitor each of these stimuli for the bilinguals and monolingual German speakers presence of a particular phoneme, specified prior were presented with pictures and asked to perform to the onset of each trial (Blank, 1980). Colomé a go/no-go task: When the picture’s name in the adapted this task such that it became a com- target language started with a vowel they had to ponent of a picture-naming task. Her participants, push a button and when it started with a conson- Catalan–Spanish bilinguals, were presented with ant they had to let it pass without responding. pictures, but instead of being asked to name each Alternatively, they had to push the button when picture aloud they had to generate its name tacitly the picture’s name in the target language started and to then monitor the internally generated with a consonant and had to withhold a response name for the presence of a particular pre- when it started with a vowel. To be able to perform specified sound. In a first experiment the partici- this task the phonological representation of the pants first saw a letter on a screen and were asked picture’s name needs to be mentally inspected. to transpose it mentally into its sound. Sub- For bilinguals, the target language switched sequently, the letter was removed from the screen between blocks of trials (which is why in this study and a picture appeared, to be named tacitly. The the bilingual participants presumably suspected participants’ task was to indicate as rapidly as their bilingualism was being studied; furthermore, possible whether or not the presented letter’s both languages were physically present). Obvi- sound occurred in the picture’s name in L1 ously, the German monolinguals only performed Catalan. In two further experiments the presenta- the task in German. Across a number of experi- tion of the picture preceded the letter. As men- ments behavioral responses (response times tioned, care was taken that the participants had and errors), electrophysiological (ERP) brain responses, and fMRI data were gathered.

5. WORD PRODUCTION AND SPEECH ACCENTS 249 ERPs elicited at a right fronto-lateral recording site by two types of pictures. In the “coincidence” condition (solid lines) the pictures’ names in German and Spanish either both begin with a vowel or they both begin with a consonant. In the “non-coincidence” condition (dashed lines) they begin with a vowel in German and a consonant in Spanish, or vice versa. For German monolinguals no coincidence effect occurs. The bilinguals show an enhanced negativity for non-coincidence trials, between 300 and 600 ms following target onset in the go trials and from 600 ms onward in the no-go trials. From Rodriguez-Fornells et al. (2005), © 2005 Massachusetts Institute of Technology, by permission of MIT Press Journals. The pictures were selected such that on half of gested phonological activation of the non- the trials (constituting the “coincidence” condi- response language: For bilinguals, but not for tion) the names in both languages would lead to monolinguals, the percentage of correct responses the same response (either go or no-go) because was lower in the non-coincidence condition they both started with a vowel or both with a than in the coincidence condition and, overall, consonant (e.g., Esel–asno, “donkey”; Kerze–vela, bilinguals made more errors and responded more “candle”). On the other half of the trials (consti- slowly than monolinguals. This finding indicates tuting the “non-coincidence” condition) the two that the picture’s name in the non-target language languages invited different responses because the was also retrieved at least some of the time, caus- picture’s name in one language started with a ing interference on the trials where it invited a vowel whereas its name in the other language response that deviated from the targeted response. started with a consonant (e.g., Trichter–embudo, Part of the electrophysiological evidence is pre- “funnel”; Eidechse–lagarto, “lizard”). An effect sented in Figure 5.8. It shows the ERPs, recorded of the coincidence/non-coincidence manipulation at right fronto-lateral electrode sites, for monolin- would suggest that the phonology of the picture’s guals and bilinguals on go trials and no-go trials name in both languages would be activated and and in both the coincidence and non-coincidence thus provide evidence of language-nonselective condition. phonological encoding. Of course, null effects of this manipulation should be obtained for mono- As can be seen, on both go and no-go trials, linguals, because they only have one name for the ERPs in the bilingual group showed an each picture. increased negativity for non-coincidence trials as compared to coincidence trials (recall the Both the behavioral and the brain data sug- convention to plot negative signals upwards and

250 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS positive signals downward). This coincidence engaged; the prefrontal cortex responds to this effect emerged between 300 and 600 ms in the go with increased brain activation. The relevant trials and between 600 and 800 ms in the no-go point to make here is that it seems that bilinguals trials (see the original report for an explanation recruit these same two brain areas to prevent the of this latency difference). In agreement with contextually inappropriate language overtaking expectation, the monolingual German group did responding in the present go/no-go task. In other not show a coincidence effect. Again, this pattern words, it appears that language control in of results suggests interference from the picture’s bilinguals exploits more general processes and name in the non-target language, thus at the same mechanisms of cognitive control. This important time suggesting that language-nonselective pro- issue will be addressed more explicitly in cessing occurs all the way down to the level of Chapters 7 (pp. 393–401) and 8 (pp. 435–446). phonology. In conclusion, both of the above “covert The fMRI data showed a differential effect of naming” studies suggest that word production in the coincidence manipulation for monolinguals bilinguals is language-nonselective all the way and bilinguals in three brain regions: the left pre- down to the stage of phonological encoding. Of frontal cortex, the supplementary motor area the two, Colomé’s (2001) study provides the (SMA), and the anterior cingulate cortex (ACC; strongest evidence because the participants in see Chapter 8 for a localization of these brain her study were kept completely ignorant about areas). The activation in these areas was larger in the bilingual nature of the experiment and the bilinguals than in monolinguals. The left pre- non-target language was in no way implied frontal cortex has also been shown to be involved by the general context of the experiment. The when bilinguals have to switch between their two special merit of Rodriguez-Fornells et al.’s (2005) languages on command during overt picture investigation is that it provides information on naming (Hernandez, Dapretto, Mazziotta, & the temporal aspects of the interference caused by Bookheimer, 2001; Hernandez, Martinez, & activation in the non-response language. Further- Kohnert, 2000; see also pp. 440–441): Further- more, it identifies the brain areas involved more, in monolingual studies the involvement of and, in so doing, suggests that when coping with the left prefrontal cortex has been demonstrated cross-language interference bilinguals exploit in tasks that require executive functioning, such mechanisms that are also used to control other as choosing between response alternatives, switch- forms of behavior. ing between tasks, and inhibiting irrelevant stim- uli. Rodriguez-Fornells and his colleagues there- Evidence of language-nonselective fore concluded that to inhibit responses based on grammatical encoding activation in the non-target language, bilinguals recruit brain areas that are typically involved in If phonological encoding is language- executive functioning. The ACC has earlier been nonselective one might expect grammatical shown to be involved in, among others, conflict encoding also to be language-nonselective. This tasks such as the Stroop task. It has been sug- prediction follows most naturally from serial word gested that its role is to monitor performance in production models that assume the existence of conflict situations. I will present some of this evi- lemmas that (1) specify the associated words’ dence in Chapter 8, including a set of mono- syntactic information and that (2) imperatively lingual studies (e.g., Carter et al., 2000; MacDon- mediate between the conceptual and phono- ald, Cohen, Stenger, & Carter, 2000) that suggest logical levels in the production system (see a control system in which (part of) the prefrontal p. 232): It is hard to see how language- cortex and the ACC fulfill complementary goals nonselectivity could apply to a later stage in the to ensure control when conflict looms: The ACC, production process (phonological encoding) but monitoring performance, signals to the prefrontal not to an earlier stage (grammatical encoding) cortex when control must be more strongly that provides input for the later phonological

5. WORD PRODUCTION AND SPEECH ACCENTS 251 stage. It is easier to accommodate evidence of that syntactic encoding is language-nonselective. language-selective grammatical encoding on the Both behavioral measures and ERPs served as one hand and of language-nonselective phono- dependent variables. Costa and his collaborators logical encoding on the other hand in terms of (a employed the more common overt picture- bilingual version of) Caramazza’s independent naming task (without distracters), asking the network model (1997; p. 233, Figure 5.4). As we participants to respond with a noun phrase have seen, this model denies the existence of consisting of both a noun and the corresponding lemma representations and posits that a word’s definite article. They tested bilinguals whose syntactic make-up is prepared independent of two languages have similar grammatical gender and in parallel with its phonology. Yet, also in systems (Catalan–Spanish; Italian–French; all of terms of this model, there is no compelling reason these languages have only two gender values, why the activation spreading down from the feminine and masculine) as well as bilinguals semantic representations to representation units whose two languages have more dissimilar gender that specify syntax should be confined to the systems (Croatian–Italian; Croatian also has response language. A couple of studies have neutral gender). Also in this study the gram- addressed this issue, with mixed results. matical gender of the two names of a picture could be the same or different in the two lan- Rodriguez-Fornells, Lutz, and Münte (2010; guages, and monolingual control groups were see also Rodriguez-Fornells, De Diego Balaguer, included. Finally, Lemhöfer and her colleagues & Münte, 2006), Costa, Kovacic, Franck, and used the same procedure as Costa and his col- Caramazza (2003), and Lemhöfer, Spalek, and laborators, testing German–Dutch bilinguals in Schriefers (2008) examined whether the assign- L2 Dutch picture naming and including both ment of grammatical gender to a word in the pictures with names that were German–Dutch response language is influenced by the gender of cognates as well as “non-cognate” pictures; that its translation equivalent in the other language. is, pictures with totally dissimilar names in All three studies exploited the fact that grammat- German and Dutch. ical gender can match or mismatch between a bilingual’s two languages. Rodriguez-Fornells Despite their common characteristics the three and his collaborators had German–Spanish studies produced disparate results. The bilinguals bilinguals and German monolinguals perform in Rodriguez-Fornells et al.’s study responded a go/no-go task similar to the one used in more slowly and made more errors on trials Rodriguez-Fornells et al. (2005; see above). This with a gender difference between the two lan- time however the go/no-go response involved a guages than on same-gender trials. These findings decision regarding the gender of the tacitly suggest an interfering influence from the non- named word: The participants had to push a response language and, thus, language- button when the grammatical gender of the nonselective grammatical encoding. The effect picture’s name in the response language was could not be attributed to inappropriately masculine and had to let it pass without respond- matched stimulus materials in the same-gender ing when it was feminine (or vice versa). The and different-gender conditions because the pictures were selected such that on half the trials monolinguals performed equally well in both the picture’s name in German and Spanish had conditions. Analogous effects were obtained in the same gender (e.g., der Speer—el dardo, “dart”; the ERP data: Monolinguals did not show an both masculine). On the remaining half of the effect of gender coincidence but in bilinguals trials the picture’s name in the two languages had different-gender trials elicited a large negativity a different gender (e.g., die Rakete—el cohete, when compared with same-gender trials, just as “rocket”; feminine and masculine, respectively). the vowel-consonant go/no-go task of Rodriguez- An effect of this manipulation in bilinguals Fornells et al. (2005) had shown a relatively large would suggest an influence of the picture name’s negativity on non-coincidence trials. In contrast, gender in the non-response language and, hence, in none of the five experiments performed by

252 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS Costa and his colleagues was a difference noun phrases (such as the hair or the trousers) into obtained between the same-gender and different- complete sentences (e.g., the hair . . . is short; the gender conditions, not even when the bilinguals’ trousers . . . are black). two languages had very similar gender systems, and in all cases the performance of bilinguals Hatzidaki observed more subject–verb equaled that of monolinguals. Accordingly, the number-agreement errors on divergent nouns authors concluded that the two gender systems of (e.g., a participant might complete the hair with a bilingual are functionally autonomous. Appar- are short) than on convergent nouns, suggesting ently, variations in the experimental procedure that the number of a noun’s translation in the and/or the stimulus materials can have a major non-response language was also activated and influence on the data and therefore lead to com- interfered with processing. This effect consistently pletely different conclusions. Lemhöfer et al.’s materialized across a series of experiments that study revealed that the cognate status of the manipulated the relative strength of the par- stimulus materials is one variable that affects the ticipants’ two languages, and whether or not the results: As with Rodriguez-Fornells and his presented noun and the invited completion were colleagues, these authors obtained clear effects in the same language or switched between of cross-language gender compatibility versus languages. In the “non-switch” (monolingual) incompatibility, but these effects were larger for conditions the effects were larger when the test cognates than for non-cognates and in a further language was the participants’ weaker language experiment in which the picture-naming task than when it was their stronger language. This was preceded by a period of pre-training the finding converges with a similar asymmetry gender identity of the L2 test materials the effect demonstrated in research on bilingual language of gender (in)compatibility occurred for cognates comprehension (see Chapter 4) and shows that only. The authors suggested that the high per- the stronger language is relatively immune to centage of cognates (50%) in their experiments influences from the weaker language. Further- had probably boosted the activation of non- more, the effects were larger in the switch target L1 German, including this language’s conditions than in the monolingual conditions, gender information. a finding that is not so surprising given the fact that task performance in the switch condition Hatzidaki (2007) employed a very different necessarily requires that the participants keep paradigm to study the question of whether both linguistic subsystems activated. grammatical encoding is language-selective or language-nonselective; namely, the bilingual This evidence of language-nonselective gram- version of the sentence completion paradigm matical encoding was strengthened by the results introduced earlier (p. 224). She compared the per- of a second, conceptually related, technique formance of English–Greek and Greek–English which focused on pronoun–antecedent number bilinguals in a “divergent” and a “convergent” agreement. It used the same convergent and condition. In the divergent condition one of divergent nouns as used in the subject–verb the elements in a pair of Greek–English number-agreement experiments but this time translation-equivalent nouns was plural and the the participants were presented with complete other was singular (e.g., the plural scissors and sentences including the critical noun and a pre- trousers are both singular in Greek and the dicate, and were asked to complete it with a “tag” singular hair and money are both plural in question (e.g., stimulus: The scissors in the drawer Greek). In the convergent condition the nouns got stolen; possible completions: didn’t it or didn’t of a translation pair were either both singular or they). The critical dependent variable was the both plural. The dependent variable was the number of times a pronoun agreement error number of subject–verb number-agreement errors occurred in the tag question (it instead of they in that were made in these two conditions in a task the example, where scissors is the antecedent of that required the participants to complete simple the pronoun it/they). These errors occurred more frequently in the divergent-noun condition than

5. WORD PRODUCTION AND SPEECH ACCENTS 253 in the convergent-noun condition. This suggests but had to give some response to its internally that the contextually inappropriate language was generated name.) This version of the task does also activated during performance. not require the non-response language to be physically present at all (although it was in To summarize, the few studies that examined Rodriguez-Fornells et al.’s study) nor does the whether grammatical encoding in bilingual critical evidence have to be derived from a speech production is language-selective or cognate manipulation. The behavioral and brain language-nonselective do not all converge on a evidence collected in these two studies strongly consistent set of outcomes. Nevertheless they suggests that during word production the warrant the conclusion that under certain circum- response word’s translation in the non-response stances grammatical encoding is language- language is co-activated with the target all the nonselective, just as phonological encoding has way down to the level of sublexical nodes. been shown to be. The studies that focused on grammatical Summary and conclusions processing produced more mixed results: The data patterns of three of them (one actually The evidence presented in this section warrants employing a task different from picture naming the conclusion that under many circumstances and therefore somewhat improperly included in word production in bilinguals, as assessed by this section; Hatzidaki, 2007) suggested language- means of the picture-naming task, is language- nonselective grammatical encoding whereas the nonselective, irrespective of whether L1 or L2 is fourth pointed at language-selective grammatical the response language. That this evidence has encoding. As argued above (p. 251), a bilingual been assembled in studies that exploited different version of Caramazza’s (1997) independent net- versions of the task is important, because two work model (which assumes that grammatical task versions used in the studies discussed above knowledge and lexical-phonological knowledge bring along potential interpretation problems: are stored independently from one another) The evidence collected with the simple picture- could in principle account for evidence of both naming task that compared cognate and language-selective and language-nonselective non-cognate naming is equivocal because repre- grammatical processing. Therefore the present set sentational differences between cognates and of data seems to fit this model best, although it non-cognates might underlie the apparent sup- would still need to be explained what caused the port for language-nonselective phonological different results across the above three studies. encoding. The evidence collected with the pic- ture–word interference task is difficult to interpret The evidence of language-nonselective word because it is not entirely clear how the decoding production in the majority of the studies dis- process aroused by the distracter word interacts cussed above converges with the analogous with the encoding process incited by the picture. evidence assembled in the “out of context” word- In addition, the suitability of this task can be recognition studies reviewed in Chapter 4. In that questioned because the non-response language is chapter, in addition to the out of context studies, physically present in the distracter words. This a couple of experiments were discussed that likely boosts the level of activation of the non- examined the recognition of words presented in target language and, therewith, its potential to a sentence context. As you may recall, in those influence target processing. In view of these con- studies a much weaker influence of the language siderations, it may be concluded that of the task not in use was detected and in some conditions versions presented above, the one used by Colomé none whatsoever. Studies that examine bilingual (2001) and Rodriguez-Fornells et al. (2005) is best word production in sentence context are suited to address the question of (non)selectivity strikingly absent in the literature and, indeed, in of word production. (Recall that the participants all picture-naming studies reviewed above single in these studies did not name the picture overtly words were produced out of context. It therefore remains to be seen whether an influence of the

254 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS non-response language will again materialize the participants had to ignore the words and in when pictures have to be named in a sentence the word-reading condition they had to ignore the context. Given the effect of context on bilingual colors in which they were printed. Performance in word recognition it would not come as a surprise these two conflict conditions was compared to if those experiments also failed to show an performance in two corresponding non-conflict influence of the non-response language under (or neutral) conditions: reading color words certain circumstances. As noted in Chapter 4 printed in black and naming the color of colored (p. 206), the most natural and parsimonious patches in the form of squares. (Ironically, in interpretation of such null effects would be that a further experiment, and with World War II the language not in use is deactivated. Neverthe- looming, colored swastikas instead of squares less, as also pointed out there, a second possibility served as non-conflict stimuli in the color-naming must be considered; namely, that elements of condition!) the non-response language are activated after all but that such activation is ignored during With this methodology Stroop introduced an selection and therefore does not hamper process- alternative to the methods used at the time to ing. That this is a feasible option is suggested study the effect of response conflict or inter- by the studies by Costa et al. (1999) and Costa ference. The standard technique had been to have and Caramazza (1999; see pp. 242–244), in participants learn pairs of stimuli of the type which strongly activated lexical nodes in the a–b (e.g., the pair of nonsense syllables pluk– non-target language did not slow down word malk) in a paired-associate learning experiment production. (see pp. 87–89) until an associative bond has been formed between the two elements of the pair such COLOR NAMING WITH DISTRACTERS: that when the a term is presented b is readily given THE STROOP TASK as the response. Subsequently, the a terms of the learned pairs are paired with new b terms Introduction (e.g., pluk–krek), a procedure that causes strong response interference when subsequently the ori- The previous section discussed evidence of ginal b terms have to be given as responses to the language-nonselective bilingual word production a terms again. The reason is that a conflict is as gathered in experiments using the picture- created between two responses, the two different b naming task. The methodology of one of the terms malk and krek. The important new aspect task versions that were described there, the of Stroop’s methodology is that the response picture–word interference task, is very similar to conflict exists in one and the same stimulus where that of the so-called Stroop task and, in fact, the two inherent aspects of the stimulus potentially picture–word interference task was introduced compete with one another, the color and the by Rosinski et al. (1975) as a transformed version name. Furthermore, the task does not require an of the original Stroop task. The Stroop task has experimental training phase to create the com- been an important piece of equipment in the petition because the stimulus has already become cognitive psychologists’ tool kit since 1935, when associated with both possible responses through the eponymous author J. Ridley Stroop published past practice (naming colors and reading words). the results of a study in which he first introduced Stroop observed that a mismatching color only it. In his original study Stroop had participants caused a small interfering effect on word reading either name the color of color words printed in a but that the interference of a conflicting word conflicting color (e.g., say “red” to the word green on color naming was substantial. He attributed printed in red) or read the words (e.g., say “green” this differential effect to a stronger associative to this word). So in the color-naming condition bond formed in the past—due to more extensive practice—between a word and its name than between a color and its name. In contemporary terminology, the differential effect for color

5. WORD PRODUCTION AND SPEECH ACCENTS 255 naming and word reading may be due to the fact Bilingual studies that word reading is a more highly automated process than color naming. The bilingual research community has embraced the Stroop methodology, transformed into its It can readily be seen that the word component bilingual counterpart, the bilingual Stroop task, within these compounded Stroop stimuli (word to see whether color naming is also hindered if and color integrated within one and the same the “carrier” of the color to be named is a con- stimulus) is comparable to the distracter in the flicting color word in the non-response language. picture–word interference experiments discussed For instance, if a Spanish–English bilingual is to above and that the printed words’ color is com- name the color of printed words in English, will parable to the picture. The resemblance with the the response be relatively slow if the Spanish picture–word interference task is even larger for a word azul (“blue”) is printed in green (and the later version of the Stroop task in which color invited response thus is English green)? A positive and word are not two different dimensions of one answer to this question would suggest that and the same stimulus but are separated, as when the non-target language cannot be switched off a white or black color word is printed on a while producing the color-naming response, thus colored background. With such versions of the providing evidence of language-nonselective task, as in the picture–word interference processing. paradigm, the temporal relation between word and color can be systematically varied and the The earliest of these studies (Dyer, 1971; Fang, time course of the interference can be tracked Tzeng, & Alva, 1981; Kiyak, 1982; Obler & (e.g., Glaser & Glaser, 1982). In accordance Albert, 1978; Preston & Lambert, 1969) com- with this analysis that the two tasks involve very pared performance in a neutral condition (with, similar processes, similar effects of distracters for instance, color patches or colored nonsensical have been obtained in both tasks: As we have strings of letters as the stimuli) with two incon- seen, semantically related distracters inhibit gruent experimental conditions, one “intralin- picture-naming performance whereas phono- gual”, the other “interlingual”. In the intralingual logically related distracters facilitate responding. incongruent condition the language of the color These same effects have been obtained in Stroop words and the invited responses was one and color-naming studies. the same (e.g., the Spanish word azul, “blue”, printed in green and requiring a color response in The Stroop methodology has been adopted in Spanish: verde). In the “interlingual” incongruent countless studies on, especially, the development condition the language of color words and of automaticity and attention control (see responses differed (azul, “blue”, printed in green, MacLeod, 1991 for a review). In this line of to be responded to with English green). Across research new conditions have been added to these studies, different language combinations the studies’ design: In addition to a conflict (or were examined: Chinese–English; Japanese– “incongruent”) condition and a neutral condi- English; Spanish–English; Turkish–English; tion, a “congruent” condition is often included, French–English; Hungarian–English; German– in which color and word match. Dependent English; and Hebrew–English. Some of these on the exact conditions included, the Stroop language combinations involved one and the effect is the difference in response time either same writing system (e.g., German–English, both between the congruent and incongruent condi- alphabetic languages), whereas others involved tions or between the neutral and incongruent two different writing systems (e.g., Chinese– conditions. The inclusion of all three conditions English, involving an idiographic and an alpha- makes it possible to decompose the Stroop effect betic language). into a facilitation component (for color words printed in the matching color) and an inhibitory A major theoretical issue tackled by the component (for color words printed in a conflict- majority of these early studies was whether in ing color). bilingual memory the two languages are stored

256 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS in a single integrated system, with connec- In contrast, with the idiographic–alphabetic tions both between words of the same language combination Chinese–English, the interlanguage and between words from different languages (the interference effect was 213 ms smaller than the “interdependent storage” or “shared storage” intralanguage effect. Fang and his colleagues view), or in two separate systems, one for each hypothesized that different writing systems language (the “independent storage” or “separate employ different processing mechanisms whereas storage” view), with no connections between similar writing systems compete for the use of the words of different languages. The occurrence of same processing mechanism. They then argued an interlingual Stroop effect was seen as support that both the size of the interference effects and for interdependent storage; the absence of such the difference in size of the interference effects in effects was thought to imply separate storage. the intra- and interlingual conditions depend on Smaller interlingual than intralingual effects were whether or not one and the same processing interpreted as a mixture between these two: mechanism is involved during task performance. Words from the bilingual’s two languages are In contrast, most other authors interpreted connected with one another, but the connections the pattern of results in terms of the degree of involved are weaker, and/or there are fewer of within- and between-language interconnectedness them, than those between words of one and the in bilingual memory: Within bilingual memory same language. These studies also looked at the the representations of color words from the two role of interlingual form similarity of the color languages are connected but the between- words. For instance, would German braun inter- language interconnectivity is not as strong as the fere more in an incongruent English color- within-language interconnectivity, and the more naming condition because of its form resem- dissimilar the languages, the smaller the between- blance with English brown than would Spanish language interconnectivity. azul, which does not share form with the English word blue? Furthermore, these studies looked at However, a couple of more recent studies have the role of language similarity as a whole (rather shown larger interlanguage than intralanguage than at the level of individual color–word transla- effects (e.g., Gerhand, Deregowski, & McAllister, tion pairs): Would the interlingual Stroop effects 1995; see Brauer, 1998, for a review). This finding be larger when the bilingual participants’ two is inconsistent with both an account in terms of languages share the same alphabet than when shared versus separate processing mechanisms they use a different alphabet? Would they be and one in terms of differences in degree of inter- larger when the participants’ languages share the connectedness between a bilingual’s two lan- same type of script (e.g., both alphabetic scripts) guages and suggests that, in addition to language than when they are written in different scripts distance, other variables may play a role. What (e.g., an alphabetic script and a logographic might these additional variables be? An interlin- script)? gual Stroop study by Mägiste (1984b; see also Obler & Albert, 1978) suggested that the relative Interlingual Stroop effects consistently proficiency in the two languages might be one occurred, and generally (but not always; see of them. She manipulated this variable in a below) they tended to be smaller than the German–Swedish study and showed that the analogous intralingual effects. Reviewing the relative size of the intra- and interlanguage literature, Fang et al. (1981) observed that the dif- Stroop effects depended on it: If the response ference between the magnitude of the interference language was the dominant language, the intra- effects in the intra- and interlanguage conditions lingual interference effect (with printed words in decreases as the similarity between the bilingual’s this same dominant language) was larger than the two languages increases. For instance, with the interlingual interference effect (with printed orthographically similar languages German and words in the weaker language). In contrast, if the English the interlingual interference effect was response language was the weaker language of only 36 ms smaller than the intralanguage effect. the two, more interlingual (from the stronger

5. WORD PRODUCTION AND SPEECH ACCENTS 257 language of the printed word to the weaker bilinguals, language similarity did play a role: response language) than intralingual interference English–German and German–English bilinguals (from weak to weak) was obtained. Notice that (similar languages) showed equally large inter- these findings converge with results from various language and intralanguage effects, but for word recognition studies reported in Chapter 4 English–Greek and Greek–English bilinguals and from a number of picture-naming studies (dissimilar languages) the intralanguage effects reported earlier in the present chapter, which had were larger than the interlanguage effects. In all shown that the stronger language interferes other words, across all studies the pattern was the with processing the weaker language more than same for low-proficiency bilinguals, the largest vice versa. interference caused by words of strong L1 irrespective of the response language. With Although the important role of relative lan- high levels of proficiency, depending on language guage strength has not been challenged since, similarity, the intralanguage effects were either three later bilingual Stroop studies qualified larger than (dissimilar languages) or equally large Mägiste’s (1984b) conclusion. Chen and Ho as (similar languages) the interlanguage effects. (1986), testing Chinese–English bilinguals at dif- ferent levels of proficiency in L2 English, repli- The pattern of results, though complex, cated Mägiste’s results when L1 Chinese, the appears to be consistent (but see Lee, Wee, Tzeng, stronger language, was the response language: At & Hung, 1992). But how should it be explained? all proficiency levels Chinese words produced Like the investigators in the earliest bilingual larger interference effects than English words Stroop studies (e.g., Dyer, 1971; Kiyak, 1982), (in other words, in all cases the intralingual Chen and Ho (1986) interpreted their results in interference was larger than the interlingual terms of interconnectedness of the two languages interference). However, when L2 English was the in bilingual memory. However, they were more response language, words of the stronger L1 specific about the types of memory connections (Chinese) only produced more interference than they had in mind and also added a processing words of the weaker L2 during the initial stages account: They explained their results in terms of acquiring L2 English. In participants with an of the word association model and concept English proficiency beyond these initial stages, mediation model of bilingual memory presented L2 words produced more interference than L1 in Chapter 3 (see pp. 129–131 and Figure 3.7). words despite the fact that L1 was still the They suggested that during the initial stages of stronger language of the two. In other words, in foreign language acquisition learners always all conditions except when beginning learners had access the meaning of L2 words indirectly, via the to respond in their weak L2, the intralingual lexical representations of their translation equiva- interference effects were larger than the inter- lents in dominant L1 (access through “word lingual effects. Tzelgov, Henik, and Leiser (1990) association”). In contrast, more proficient L2 replicated these results in an Arabic–Hebrew learners were thought to access the meaning of study. Notice that in both these studies (and L2 words directly, without L1 lexical mediation unlike in Mägiste, 1984b) the participants’ two (access through “concept mediation”). Tzelgov et languages are written in dissimilar scripts: differ- al. (1990) and Brauer (1998) followed Chen and ent alphabets (Arabic and Hebrew), and different Ho (1986) in their interpretation of the effects, writing systems (Chinese and English). except that Brauer added the hypothesis that pro- ficient bilinguals with very similar languages More recently, Brauer (1998) extended these access the meaning of L2 words both directly and results: In English–German and English–Greek indirectly (via their L1 translation), and that these bilinguals with relatively low levels of L2 pro- two routes are exploited simultaneously (he calls ficiency, words from the stronger L1 produced this the “multiple-access” strategy). more interference than words from the weaker L2, regardless of response language and language Because these studies typically do not detail similarity. In contrast, in high-proficiency the various mental processing stages in the Stroop

258 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS task and where along the route the printed word discussion). It is only natural that the way a comes in, it is often quite hard to see how exactly particular set of experimental results is presented the observed pattern of results fits the authors’ and explained varies with the exact topic under views on L2 word processing at different L2 study. The last bilingual Stroop study to be dis- proficiency levels and the associated differences in cussed here, a Hebrew–English study by Tzelgov the underlying memory representations. Of the et al. (1996; see also p. 190), had yet a different latter three studies, the one by Brauer (1998) is focus and, as a consequence of its specific focus, most explicit in this respect. He attributes smaller provided detailed information on how the dis- effects to longer routes in the underlying systems tracter stimulus is processed. so that the activation caused by the distracter word arrives in some critical memory repre- Tzelgov and his collaborators used a modified sentation too late to still cause an effect. To give version of the bilingual Stroop task in which, in one example, the smaller interference with addition to words, “cross-script homophones” weak L2 words than with strong L1 words in low- were used as distracters: nonwords written in the proficiency bilinguals was attributed to the larger letters of the non-response language but that, if number of mental processing steps in the former pronounced in this language, sound like a color case: Because a weak L2 word has to pass through name in the response language. Examples are the the corresponding L1 lexical representation Englishlike nonwords kahol and adom that, if before accessing conceptual memory (see Chapter pronounced according to the spelling-to-sound 3, Figure 3.7, the word association model), its correspondence rules of English, sound just like interfering effect is relatively small. The implicit the Hebrew words for blue and red, respectively. additional assumptions apparently are that by The main question these investigators addressed the time it gets there, in relatively many cases (and was whether a written word makes direct con- more often than with a strong L1 distracter word) tact with its lexical representation or whether, (1) the conceptual analysis of the color has instead, the word’s phonological form is first already taken place (the color has been recog- automatically generated via the application of nized), (2) the result of this analysis—activation grapheme–phoneme conversion rules, after which in a layer of conceptual nodes—has already the generated code makes contact with the word’s been transmitted to the corresponding lexical lexical representation. If the latter holds and the representation of the response language, (3) the non-response language cannot be blocked off, phonology corresponding to this targeted repre- cross-script homophones might also produce an sentation has also been accessed, and (4) the interference effect. targeted word has been articulated. In other words, the activation induced by the weak L2 Englishlike cross-script homophones like word arrives in the relevant place (conceptual kahol and adom (English being the participants’ memory) too late to still affect the response. L2, the weaker language) indeed produced a reliable Stroop effect when Hebrew was the The reason that in the above studies the pro- response language. This finding indicates that the cessing stages in the Stroop task were not participants automatically applied the English described in great detail presumably is that their (L2) spelling-to-sound conversion rules when primary goal was to become informed on performing the task in Hebrew. However, this bilingual lexical representation, the task just effect was not always as large as the correspond- serving as the tool to reach this goal. Other ing intralingual effect (when, e.g., the Hebrew studies have used the bilingual Stroop task to word for blue was printed in red and Hebrew was address other questions, such as whether or not the response language again). This replicates the orthographically different and similar languages common finding that generally (but not always; pose the same demands on one and the same see above) the intralingual effects are larger than central processor (see the study by Fang et al., the interlingual effects. 1981, reviewed above, and Lee et al., 1992, for a To summarize, bilingual Stroop studies have reliably produced both intra- and interlingual

5. WORD PRODUCTION AND SPEECH ACCENTS 259 Stroop effects and the size of these effects response language (here English) and the non- depended on the participants’ relative proficiency response language (Spanish). All these lemmas in their two languages and on the degree of (e.g., the lemmas for green, blue, verde, and azul) similarity of the languages. The effects have usu- compete for selection (according to the standard ally been explained in terms of the degree of view, but see pp. 242–244). Among the competi- within- and between-language interconnectedness tors is the lemma that corresponds to azul. This of the linguistic elements in bilingual memory. lemma has also received activation from a Only occasionally have the exact processing steps bottom-up process triggered by the word azul (cf. involved been described explicitly. the activation sequence incited by the distracter mesa in Figure 5.7). As a result, this lemma Conclusions becomes a strong competitor of the color name to be produced, green, causing the inhibitory Stroop An account of performance in any task includes effect. One can also see how verde (“green”) implicit or explicit hypotheses about representa- printed in green would implement the between- tion structures, processing operations, and the language identity condition which Costa et al. interdependence between the two. Considering (1999; see p. 244) have shown to produce facilita- this truism, it may have struck the reader that, tion in the picture–word interference paradigm. despite the procedural and conceptual similarities between the bilingual picture–word interference Looking at bilingual Stroop effects in terms of task and the bilingual Stroop task, different bilingual word production models this way is not aspects of the full processing account are high- meant to radically dismiss the above accounts of lighted in these two research lines. The bilingual these effects in terms of, for instance, differences picture–word interference studies (pp. 239–242) in strength between the representations of related focused more on processing than on representa- words within and between languages or in terms tion. In contrast, most of the bilingual Stroop of a development from indirect to direct mapping tasks stressed aspects of representation, focusing of L2 words onto meaning. Instead, the exercise especially on the type and degree of intercon- is simply meant to highlight the similarities nectedness of meaning and form representations between the Stroop task and the picture–word between the bilingual’s two language systems. interference task (as some other authors have This difference in theoretical focus involves the done; La Heij et al., 1990). An understanding of risk that the tasks’ commonalities are obscured. It their shared features will no doubt facilitate is therefore instructive to take one of the process- the interpretation of the data collected in studies ing models tested in the picture-naming studies as that exploit either task, an interpretation that, a starting point and see how Stroop effects can be as the reader may have experienced, can be quite explained in terms of this model. challenging at times. Ultimately, a complete understanding of the effects observed with both In terms of Figure 5.2, analogous to the bike tasks will have to take both the processes involved in this figure, the color of the printed word in the and the knowledge structures exploited by these Stroop task would be the conceptual content processes into consideration. to be verbalized. Let us take as an example the Spanish word azul, “blue”, printed in green and WORD TRANSLATION an English–Spanish bilingual serving as partici- pant in a bilingual Stroop experiment following Word translation as word production an instruction to name the colors in English. Just like an object concept (e.g., “bike”), a color con- In the previous section the similarities between cept (“green”) activates a set of conceptual the bilingual picture–word interference task and features that, in their turn, activate a number of the bilingual Stroop task were highlighted. There lemmas. The lemma set activated by the color green includes lemmas for color words in both the

260 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS is still a third bilingual task that has been likened was slower in the semantic distracter condition to these two tasks—word translation. The than in the unrelated control condition, again, word translation task has a long history in the just as this type of distracter inhibits picture cognitive study of bilingualism, but it has been naming. used primarily to study the question of how L2 word-forms are mapped onto meaning (e.g., De Miller and Kroll (2002) replicated these results Groot, 1992b; Kroll & Curley, 1988; Kroll & in an English–Spanish study and showed that the Stewart, 1994; Potter et al., 1984; see Chapter 3, effects also occurred when the participants had pp. 136–137 for a discussion of this work), or to translate in the reverse direction, from their simply to find out what aspects of words make L1 into their L2. Importantly, these authors also them hard or easy to translate. On the assumption included a condition in which the distracters that word translation is conceptually mediated appeared in the same language as the stimulus (this in fact being a debated issue in the form-to- input. For example, yellow could be accompanied meaning mapping studies), a quick glance at, for by the semantic distracter brown or the unrelated instance, Figure 5.7 shows how it implicitly com- distracter cap. Under these circumstances, the prises the word translation process: Ignore the orthographic and semantic distracters did not picture of a table and read mesa as the word to be noticeably affect translation performance. translated; translation then is the process of get- ting from mesa to taula. What could have caused this effect of dis- tracter language, and especially, why do not Noticing the cross-task similarities, La Heij distracters in both languages affect responding, as et al. (1990) were the first to wonder whether they do in picture naming (see p. 240)? In their word translation is similarly influenced by analysis the authors point to an important distracter words as are picture naming and the difference between translation and picture traditional Stroop tasks. They addressed this naming: The former but not the latter provides question in a study in which Dutch–English relevant information about the language of out- bilinguals translated L2 English printed words put in the form of the language information into L1 Dutch. Each stimulus word was paired present in the stimulus. If, say, silla (“chair”) is with a printed Dutch distracter word, which presented to a Spanish–English bilingual, the could be of three types: It could be orthographic- language identity of the stimulus is known as ally related, semantically related, or unrelated well. In contrast, in picture naming no such to the Dutch translation of the stimulus. For language information is incorporated in the instance, the stimulus yellow (geel, in Dutch) stimulus. The authors suggest this early language could be paired with the orthographic distracter information available in translation has the effect geest (“spirit”), the semantic distracter bruin that the participants can constrain the selection (“brown”) or the unrelated distracter muts process to the language of production. In other (“cap”). Notice that the distracters appeared words, in word translation, when the distracter in the language of output, the language the words are elements of the production language, participants translated into (also notice that they compete for selection, but when they do not orthographic distracters were at the same time belong to the production language (but to phonological distracters because they also the language of input) they do not compete for shared a phonological relation with the targeted selection. In contrast, because no language response). When the stimulus words preceded the information is encompassed in a picture to be distracters by 140 milliseconds, translation was named, distracter words in both languages faster in the orthographic distracter condition compete for selection with the targeted word in than in the unrelated distracter condition, just as picture naming. this type of distracter (but called a phonological distracter) has been shown to facilitate perform- Two recent studies (Bloem & La Heij, 2003; ance in the picture-naming task. Conversely, it Bloem, Van den Boogaard, & La Heij, 2004) have examined the effects of distracter words in word translation in greater detail, comparing them with

5. WORD PRODUCTION AND SPEECH ACCENTS 261 the effects of distracter pictures on word transla- and explain how the emerging view of word tion and manipulating the time interval (stimulus translation can account for a couple of common onset asynchrony; SOA) between the presentation findings obtained when participants perform of the distracter and the word to be translated. this task. Unlike in the studies just reviewed, The most interesting new findings were that the studies in question have not used the dis- semantic distracter pictures quickened rather tracter methodology but simple word translation than slowed down the translation process and instead. that the semantic interference caused by semantic distracter words was reversed into semantic Explaining cognate and concreteness facilitation when the distracters preceded the effects in word translation stimulus word by 400 ms instead of being pre- sented in close temporal proximity to the target. The cognate effect The authors accounted for these results in terms of a new model of word production, the con- A robust finding in simple word translation ceptual selection model (CSM), that localizes studies is that cognates are translated faster than semantic facilitation at the conceptual repres- non-cognates (Christoffels et al., 2006; De Groot, entational level and semantic interference at the 1992b; De Groot et al., 1994). Earlier on in this lexical level and that, in addition, assumes that chapter (p. 246) I have already discussed the activation in lexical nodes decays faster than acti- similar finding of Costa et al. (2000) that in vation in conceptual nodes. A crucial assumption simple picture naming (without distracters) in this model is that not all concepts that are pictures that depict cognate words are named activated by an input automatically activate their faster than pictures that depict non-cognates. lexical representations. Instead, it is assumed that Because in simple picture naming neither of the at the conceptual level one concept is selected and cognate terms of a translation pair is physically that only this concept is lexicalized. This is the present, the effect cannot be due to early per- model’s feature from which it derives its name. ceptual processing somehow favoring cognates. At the same time, the selected concept also Accordingly, Costa and his colleagues localized spreads activation to the lexical representations the effect in a later processing stage (see Figure of semantically related words so that CSM also 5.5, p. 235): As a consequence of language- assumes a “semantic cohort” of activated lexical nonselective phonological encoding, many of the representations, as do the production models sublexical phonological nodes that together discussed earlier in this chapter. represent a cognate’s name receive activation from two sources: from the target’s lexical node So far, the details of the full translation pro- and the lexical node representing its translation. cess remained implicit, as did a salient difference In contrast, the sublexical phonological nodes between word translation and picture naming that represent the name of a non-cognate receive (or Stroop color naming, for that matter); namely, activation from the target’s lexical node only. The that processing starts off at different points in boosted activation in the former set of sublexical these two tasks: After a first structural analysis of nodes facilitates name retrieval. the visual information, picture naming involves the immediate access of conceptual memory. In Given the present analysis that the final stages contrast, in translation, following a perceptual of word translation, downward from the con- analysis, the target word first accesses the ortho- ceptual (or semantic) level, are similar to those in graphic lexicon (or the phonological lexicon, if picture naming, this account can also serve to words are aurally presented for translation) and explain the cognate effects in word translation. only then conceptual memory (see Figure 5.7). But this is unlikely to be the complete story. In the next section I will describe the various Earlier on (p. 247) I have already discussed the processing stages and representational structures possibility that representational differences may involved in word translation in more detail exist between cognates and non-cognates. For

262 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS instance, cognates may share more semantic increased activation of the semantic nodes repre- nodes between languages than do non-cognates senting cognates is that the nodes at later levels in (e.g., De Groot, 1992a; Taylor, 1976; Van Hell & the translation production process, those repre- De Groot, 1998a). Furthermore, they may share senting the phonological lexical representations a morphological representation in memory and sublexical phonology, will also be activated whereas non-cognates do not (e.g., Sánchez- relatively highly and the cognates’ names will be Casas & García-Albea, 2005; see Chapter 4, relatively easy to retrieve. p. 204). Support for the former of these hypotheses comes from word association studies that have To summarize, three sources of the cognate compared within- and between-language word effect in word translation can be identified: facili- associations in bilinguals (e.g., within-language: tated name retrieval due to language-nonselective stimulus boy, response girl; between-language: phonological encoding that favors cognates stimulus boy, response fille, “girl”, in an English– over non-cognates; representational differences French bilingual). The studies in question have between cognates and non-cognates at the shown that within- and between-language associ- semantic-representational level and/or a mor- ations to cognate stimuli are more often transla- phological level (not shown); relatively strong tions of one another (as is the case in the activation of the associated semantic nodes when example) than those to non-cognates. Several a cognate is presented for translation. In trans- authors (Taylor, 1976; Van Hell & De Groot, lation production, these various sources may all 1998a) have accounted for this effect in terms of join forces in privileging the processing of cog- distributed representations in semantic memory nates over non-cognates. that are partly shared between a pair of transla- tions augmented by the hypothesis that the Given the fact that the activated semantic amount of sharing is larger for cognate transla- nodes feed activation to the phonological lexical tions than for non-cognate translations (see Van nodes of both elements in the translation pair, not Hell & De Groot, 1998a, for details). just to the lexical node that corresponds to the targeted response, what remains is to explain why, In word translation there is a further possible satisfying the instructions, the input word is source of the cognate effects. If the two languages indeed generally translated, not repeated. (Notice involved in a translation act share one and the that the accounts of picture naming advanced same alphabetic script, a cognate translation pair in this chapter give rise to the analogous question will consist of two words that share orthography of why the picture is generally named in the (in addition to phonology). If word recognition is target language.) To account for this it may be language-nonselective, as we have seen it is under assumed that some general control mechanism, many circumstances (see Chapter 4), a visually external to the language system proper, translates presented cognate word will automatically trigger the exact task requirements in a set-up of the (via activation of the orthographic sublexical language system in which the phonological lexical nodes) the orthographic lexical representations of representations of the output language are both terms in a translation pair (see pp. 199–202) more highly activated than those of the input lan- and both of these then feed activation to the guage. This idea will briefly return later in this same set of semantic nodes. The likely con- section and will be elaborated in more detail in sequence is that these semantic nodes will be Chapter 6. Alternatively, we may resort to the activated relatively highly as compared to the language-specific lexical selection mechanism situation where a non-cognate is presented for that Costa and his colleagues (1999) introduced translation (which does not activate the ortho- to account for their picture naming data graphic lexical representation of its translation (pp. 243–245). Only the activated lexical nodes of equivalent or only very weakly so). Figures 5.9a the target language are considered for selection; and 5.9b visualize this state of affairs, building activation in non-target language nodes is on and extending Figure 5.5. The effect of the ignored. Arguably, this process of selective atten- tion also demands the involvement of a control

5. WORD PRODUCTION AND SPEECH ACCENTS 263 A model of translating cognates (a) and non-cognates (b). The Catalan–Spanish participants’ task is to translate Spanish gato and mesa into Catalan gat and taula. Spanish gato strongly activates the orthographic lexical nodes of both gato and gat. As a consequence, the semantic nodes that represent the meaning of this translation pair are activated relatively strongly, as compared to the semantic nodes that represent the meaning of the non-cognate pair mesa–taula. system of the type just suggested. In Figure 5.9 activated orthographic lexical nodes can account this idea of language-specific lexical selection is for these effects. After all, if the concrete and illustrated by means of the bracket underneath abstract words presented for translation are non- the phonological lexical node of the response cognates, these two sources of the effect cease to language. exist. Out of the three sources of differential translation performance for cognates versus non- The concreteness effect cognates that were considered above, only one remains to account for the concreteness effect: a A second variable that affects word translation is difference between the semantic representations word concreteness: Concrete words are translated of concrete and abstract words in the sense that faster than abstract words (De Groot, 1992b; the former share more semantic nodes between De Groot et al., 1994; De Groot & Poot, 1997; languages than the latter Again, some evidence Van Hell & De Groot, 1998b). Because this effect comes from word association studies: Within- and holds for non-cognates and cognates alike, neither between-language word associations to concrete the convergence of activation in the relevant words are more often translations of one another set of phonological sublexical nodes from two than those to abstract words (Kolers, 1963; Van phonological lexical nodes nor differential acti- Hell & De Groot, 1998a). Another source of evi- vation of the relevant set of semantic nodes dence is the finding that bilinguals rate the words caused by the convergence of activation from two in concrete translation pairs as more semantically

264 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS similar to one another than the words in abstract nodes for woman and wife), and these will com- translation pairs (Tokowicz, Kroll, De Groot, & pete for selection, causing the delay in responding Van Hell, 2002). (the example is borrowed from Laxén & Lavaur, 2010). However, a complete account of the concrete- ness effect in word translation is likely to be more If ambiguity or synonymy in the input lan- complex than just suggested, because concrete guage indeed underlies the number of trans- and abstract words differ on more dimensions lations effect, the above interaction between than concreteness alone. To mention three: the number of translations and word concreteness in meaning components of concrete words tend to word translation is in actual fact an interaction be more correlated than those of abstract words between number of meanings and concreteness. (McRae, De Sa, & Seidenberg, 1997), the If true, a similar interaction may be expected meanings of abstract and concrete words are to occur in monolingual tasks. This prediction represented in a qualitatively different way in was confirmed by Tokowicz and Kroll (2007), conceptual memory (e.g., Crutch & Warrington, who obtained the expected interaction between 2005), and concrete words have fewer translations number of meanings and concreteness in a mono- on average in the language to be translated into lingual lexical decision task. As concluded by the than abstract words (Schönpflug, 1997; Prior, authors, the similarity of the results for the MacWhinney, & Kroll, 2007; Tokowicz & Kroll, bilingual translation task and the monolingual 2007), perhaps because concrete words tend to lexical decision task suggests that a general have a relatively small number of fixed meanings property of the language processing system is whereas abstract words have more, but context- likely to underlie both sets of data, one that has dependent, meanings (Crutch & Warrington, nothing to do with bilingualism per se. Whatever 2005). Tokowicz and Kroll, in an English–Span- this general system property may turn out to be, ish study, have shown that the speed advantage of these data point out that number of meanings of concrete words over abstract words disappears the stimuli presented for translation is a factor to when the stimuli presented for translation have take into account when interpreting concreteness only one translation in the other language. In effects in word translation. contrast, words with multiple translations showed the common advantage for concrete words. In addition to cognate status, word concrete- ness, and number of meanings, there are other A further finding obtained by Tokowicz and word characteristics that affect word translation, Kroll (2007) was that words with multiple trans- maybe most notoriously word frequency (see lations were translated more slowly than words e.g., De Groot, 1992b; De Groot et al., 1994). I with just one translation. Laxén and Lavaur will refrain from detailing all of these in terms (2010) replicated this finding in a French– of the view of word translation developed in this English study, employing the translation recogni- section. Instead, I will conclude with an analysis tion task (see p. 91). The reason a particular word of how this view can account for one further has more than one translation usually is that it is specific feature of translation performance—that ambiguous or has a close synonym in the lan- translation errors are often words semantically guage of input. As a consequence, words with related to the stimulus word, but in the language multiple translations (e.g., French femme, mean- of output (e.g., stimulus: Dutch cirkel, “circle”, ing both “woman” and “wife” in English) are response: square; stimulus: Dutch kaars, likely to activate more semantic nodes than words “candle”, response: flame; stimulus: Dutch hand- with just a single translation, especially when they doek, “towel”, response: blanket). These trans- are translated out of context (as in the present lation errors are similar to speech production type of experiments). The joint set of activated errors in unilingual settings where, instead of the semantic nodes then transmit their activation targeted word, a semantically related word is onto more than one phonological lexical node in accidentally output by the system. The similarity the output language (e.g., both the English lexical of this error type in translation and unilingual

5. WORD PRODUCTION AND SPEECH ACCENTS 265 word production suggests a common underlying and word translation is based on activation cause (thereby constituting independent evidence patterns in one and the same underlying, layered, for the main tenet of the present section: that memory system augmented by a general control word translation and monolingual and bilingual system that, one way or the other, sees to it that word production can be accounted for in terms of the current goal is met (for instance, by boosting one and the same set of processes and associated the activation in task-relevant memory nodes system architecture). Specifically, the likely source and/or suppressing the activation in task- of this type of error (see also De Groot, 1992a) is irrelevant memory nodes and/or by ignoring the fact that a set of activated semantic nodes task-irrelevant activation). activates more lexical representations than just the two focused on in Figures 5.9a and 5.9b. As COMPREHENSION AND PRODUCTION: we have seen in earlier parts of this chapter (see, MAKING A CONNECTION e.g., Figure 5.2) the lexical representations of semantically related words, in both languages, are In the preceding section a model of word trans- also activated, and one of these may accidentally lation was presented which regarded it as a form be selected. of word production in which the content to be expressed in speech is established in the layer of How then, if performance in such a varied conceptual nodes through an initial word recogni- collection of tasks exploits the same set of tion process. As with every visual illustration of memory nodes at various levels in the system, a mental process that intends to highlight the does the response required by a specific task and component structures and processes involved, not one required by an alternative task generally ignoring others, the model’s illustration involves emerge from the system? A solution to this the risk of seeding a misunderstanding: Figure problem was already suggested above, when 5.9 may be worth a thousand words but the accounting for the fact that in word translation it story it tells so pointedly may give rise to the seldom occurs that the input word is inadvertently unintentional and presumably false inference repeated as output. The suggested solution was that activated nodes representing sublexical to assume the existence of some general control orthography do not directly spread their acti- system in addition to the language system proper, vation to the corresponding sublexical phono- as some have done (e.g., Dijkstra & Van Heuven, logical nodes. Instead it suggests that semantic 2002; Green, 1998; Roelofs, 1992). Depending nodes and orthographic and phonological lexical on the task’s demands, a specific system configur- nodes mediate between the transmission of ation is chosen that suppresses or ignores task- activation from orthographic to phonological irrelevant activation and/or fosters task-relevant sublexical nodes. To draw such a conclusion activation. This control system may fail at times, would not do justice to the results of myriad perhaps due to a lapse of attention or to system studies, particularly in the domain of research on overload. At such a moment an error will occur; word reading, that suggest otherwise. Some of for instance, a semantic error in translation or a this evidence was presented in Chapter 4 switch to another task or language. Plausibly, (pp. 184–186), where I discussed a number of the language-specific selection process proposed studies which showed that orthographic sublexi- by Costa et al. (1999) is a manifestation of the cal nodes automatically activate the correspond- workings of a control system of the type outlined ing phonological sublexical nodes, suggesting dir- here. In Chapters 6 and 8 I will provide further ect connections between these two types of nodes. details on how such a system might work and on Also models of speech production assume the the brain mechanisms that appear to be involved. existence of these connections, sometimes impli- citly, at other times explicitly (e.g., Roelofs, To conclude, it seems that performance in a large number of word production tasks such as monolingual and bilingual picture naming, monolingual and bilingual Stroop color naming,

266 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS Meyer, & Levelt, 1996). Similarly, through past translated are presented in printed form, the input reading, the complete orthographic forms of word first activates the orthographic nodes and words are likely to have become connected with from, there activation spreads to both semantic their phonological counterparts. As a result, the and phonological nodes. In a word production activation of an orthographic lexical node will task such as picture naming, the semantic nodes activate its phonological counterpart directly. are activated first. Figure 5.10a presents the same types of mem- This alternative presentation illustrates the ory nodes as Figure 5.9 but in a format different imaginative power a visual illustration of a men- from the hierarchically layered one that is con- tal process can have: This new layout immediately ventionally used in production studies and with- suggests the hypothesis that it may represent, in out differentiating between sublexical and lexical an admittedly very sketchy way, language pro- nodes. But contrary to Figure 5.9 it explicitly cessing in general, encompassing both language acknowledges the fact that orthographic nodes comprehension (both reading and understanding not only feed activation to phonological nodes spoken input) and language production. This indirectly, via semantic nodes, but also directly. In hypothesis is rendered more explicit in Figure a word translation task where the words to be 5.10b, which includes bidirectional instead of (a) A language-processing system that assumes unidirectional spreading of activation from orthographic to phonological nodes, both indirectly via semantic nodes as well as directly. (b) The same system but now including bidirectional connections between the various types of nodes. It suggests that word production from a conceptual input and the recognition of spoken and written words exploit one and the same underlying processing architecture.

5. WORD PRODUCTION AND SPEECH ACCENTS 267 unidirectional connections between the different In fact, the relation between comprehension types of nodes in the system. In other words, and production and the underlying architecture Figure 5.10b expresses the hypothesis that of the language system have been studied word production, visual word recognition, and thoroughly since the second half of the last cen- auditory word recognition exploit one and the tury. For instance, the motor theory of speech same underlying processing system instead of perception (e.g., Liberman & Mattingly, 1985) there being separate systems dedicated to each assumes that we perceive phonetic elements in one of these processes. It is based on visual terms of the motor commands in the brain that illustrations of a model of monolingual reading drive the muscles used when we produce these and spelling (e.g., Van Orden, Pennington, & sounds in speech. A further example concerns the Stone, 1990; Bosman & Van Orden, 1997), the possibility of grammatical encoding and gram- phonological coherence model, and also bears a matical decoding being “entangled” processes striking resemblance to SOPHIA (see Figure 4.4), rather than operating independently from one a model of bilingual word recognition. another, as they are most often thought to do (Kempen, Olsthoorn, & Sprenger, 2009). Allport, Although attractive because of its parsimony, MacKay, Prinz, and Scheerer (1987) dedicated a Figure 5.10b does not enjoy the status of a truism complete volume to the relation between com- but summarizes a multifaceted hypothesis of prehension and production. In one of the contri- which each component must be validated butions to this volume, Monsell (1987) considered empirically. Whether or not comprehension and six possible models for the relationship between production share representational structure speech comprehension and speech production, needs to be settled empirically for each of the and reviewed the evidence in support of each of representation domains that we have dis- them. All of these models shared the assumption tinguished: word meaning, word form (both that the conceptual representations are common orthographic and phonological), and sublexical to speech comprehension and production. They (again, both orthographic and phonological). differed from one another in terms of the degree Shared representation may hold for some of interconnectedness between, and the sharing domains but not for others. Furthermore, if the of, the other types of representations that he dis- data point out that for a particular domain tinguished: phonological word nodes, sublexical separate representations for comprehension and phonological nodes, acoustic representations and production exist, these representations may or articulatory representations. may not be connected. In view of the topic of the present chapter, speech production, learning to To be able to choose between the alternative know the exact relationship between comprehen- models, Monsell considered evidence from three sion and production is particularly important sources of experimental data: cross-modal because speakers are known to monitor their own repetition priming studies, dual-task studies, speech, employing their comprehension apparatus and neuropsychological studies. In the relevant in doing so (see Figure 5.1). If comprehension cross-modal studies the participants first had to and production share one and the same set of perform some “priming” task that involved the underlying representations, this means that this production of words (e.g., speak the word out one set is exploited simultaneously (or maybe loud) and, subsequently, some task that involved in rapid alternation) during the production and the auditory comprehension of the words pro- monitoring of speech. Similarly, revealing the duced earlier (e.g., perform an auditory lexical relation between comprehension and production decision). In the latter task, called the “probe” is likely to inform our understanding of simul- task, also words occurred that had not been taneous interpreting, which involves the simul- produced before. The question of interest was taneous use of these two processes most of the whether probe words that had been produced time (or, again, their use in rapid alternation; see before would be responded to faster than those Chapter 6, pp. 316–319). not produced before. Such priming effects

268 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS would suggest that phonological word nodes are sion and speech production, but with connections shared between comprehension and production. in both directions between the phonological Alternatively they could be accounted for in sublexical input nodes and the phonological terms of models that assume separate phono- sublexical output nodes. In addition, he assumed logical word nodes for production and com- that the acoustic and articulatory representations prehension that are, however, connected in two are connected in both directions. These conclu- directions via the sublexical nodes. In contrast, sions tally with those drawn by speech production the absence of a priming effect of earlier pro- researchers (Levelt et al., 1991; Roelofs et al., duction on later comprehension would suggest 1996). However they seem to be at variance with the existence of separate phonological word the way speech accents in bilinguals are nodes for comprehension and production and, explained. These accents, and especially the at the same time, the non-existence of the above accent in bilinguals’ L2 speech, are the theme of indirect connection between them via the sub- the final part of this chapter. To anticipate, lexical nodes. Priming from earlier production the relevant work suggests that phonological to later comprehension of the same words sublexical representations are shared between occurred under many circumstances, ruling out speech comprehension and speech production. the latter type of model (see Monsell, 1987, for details). SPEECH ACCENTS IN NON-NATIVE SPEAKERS The dual-task studies in question (e.g., Shallice, McLeod, & Lewis, 1985) compared a Introduction dual-task condition where participants have to simultaneously perform a speech comprehension So far this chapter has dealt with the mental task (e.g., detect words belonging to a prespeci- processing that takes place in the mind of (mono- fied semantic class in a series of aurally presented linguals and) bilinguals while they produce words, words) and a speech production task (e.g., read ignoring the physical characteristics of the speech aloud visually presented words) with a control signal that emerges from this mental work. In this condition in which participants only performed section I will remedy this neglect by focusing one task. If performance in both tasks exploits on one particular characteristic of this signal; a single set of phonological word representations, namely the fact that it often accented. This is true a performance decrement in the dual-task con- of the speech of the majority of non-native dition as compared to the control condition speakers and even bilinguals with a high level should occur. Contrary to this prediction (and of fluency in their L2 are more often than not contrary to the above results obtained in cross- recognized as non-native speakers just from the modal priming studies), the decrement in per- way their speech sounds. However, the degree of formance in the dual-task condition as compared “accentedness” varies greatly between individual to the single-task condition was negligibly small, L2 speakers and in a small minority of them not a suggesting that separate phonological word nodes hint of an accent can be detected. These facts for speech comprehension and speech production raise a number of questions, such as what causes exist. Neuropsychological evidence supports the accent and what aspects of speech are affected this conclusion: Both cases of impaired auditory (e.g., particular phonemes or suprasegmental word comprehension with preserved word pro- prosodic features). To what extent does the duction and cases of impaired word production accuracy in pronouncing L2 speech sounds but preserved word comprehension have been depend on the accuracy with which these sounds reported. Weighing the combined evidence from can be perceived? Does an accent affect the these three lines of research, Monsell (1987) opted for a model that contains both separate phonological word nodes and separate phono- logical sublexical nodes for speech comprehen-

5. WORD PRODUCTION AND SPEECH ACCENTS 269 comprehensibility of an L2 utterance and the Lenneberg (1967) were the first to posit this view, time taken by a perceiver to process it and, if so, which has come to be known as the “critical to what degree? Does familiarity of a native period hypothesis” (CPH). They attributed the listener with the non-native speaker’s L1 affect loss of the facility to learn a language after this the perceived accentedness and/or the compre- period—which they thought to end around hensibility of this speaker’s L2? In this section puberty—to a loss of neural plasticity. If lan- I will present a selection of studies that have guage acquisition coincides with the critical tackled one or more of these questions. I will first period a nativelike pronunciation (as well as review a number of studies that were primarily nativelike performance in other subdomains of concerned with identifying causes of the non- language; see Chapter 2) will be attained. If lan- native accent and will conclude with a discussion guage acquisition only coincides with part of this of studies that focused on the relation between critical period or starts after its offset, complete accent and comprehensibility. L2 mastery can no longer be attained and L2 speech will remain accented. Whereas Lenneberg Causes hypothesized that the acquisition of language in general is constrained by the critical period, A factor that has consistently been shown to others have proposed that especially the capacity affect the degree of accentedness of L2 speech is to acquire a nativelike pronunciation is con- the age at which the speaker began to learn this strained by it (e.g., Scovel, 1988). A large part of language: The younger the learner was at the Chapter 2 dealt with the hypothesis that L2 time, the less accented L2 speech is. Global accent grammar learning is determined by a critical differences between young and older L2 learners period. The present section discusses the critical can easily be detected by having native speakers period hypothesis in relation to L2 speech perform a general accentedness rating of the accents. L2 speech (e.g., Flege, Munro, & MacKay, 1995a; Flege et al., 1999). Experiments in which Birdsong (2005) evaluated a couple of pre- L2 speakers and native speakers are asked to dictions that can be derived from the critical produce individual vowels and consonants pro- period hypothesis. One of them is that no L2 vide the data for a more fine-grained analysis speaker who started learning this language after of the accent differences between native and the closure of the critical period can attain native- non-native speech (e.g., Flege, Munro, & like performance. The existence of just a single MacKay, 1995b). A similar age of acquisition late learner whose L2 speech is indistinguishable effect occurs in perception: Early L2 learners out- from native speech would constitute irrefutable perform late learners in the identification of L2 counterevidence. Such evidence has, however, speech sounds (e.g., Broersma, 2005; MacKay, been obtained repeatedly, just as late L2 learners Meador, & Flege, 2001; Weber & Cutler, 2004; see have been identified with nativelike L2 gram- also Chapter 4, pp. 195–197, and see Chapter 2, matical ability (Chapter 2). For instance, Bongaerts pp. 47–77 for age of acquisition effects on L2 (1999) asked native English speakers to rate grammar). Whereas the reliability of these effects English speech samples spoken by advanced L1 is beyond doubt, there is an ongoing debate on Dutch learners of English on a 5-point scale ran- what causes them. ging from 1 (very strong accent, definitely non- native) to 5 (no foreign accent at all, definitely The explanation that has inspired a substantial native). All learners had started their English amount of work in this research area, and that, training relatively late in life, around the age of due to its appeal, seems to have assumed the 12. Nevertheless, the judges mistook the speech status of uncontestable truth among the general of about half of these advanced learners (five out public, is that there is some critical period for of eleven) as being spoken by native speakers of the acquisition of language (be it the L1 or any English. This result suggests that starting to learn further language) early in life. Penfield (1963) and a second language relatively late in life is not an

270 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS insuperable impediment for acquiring an authen- acquisition on the one hand and degree of L2 tic pronunciation. accent on the other hand should be related non- linearly. Specifically, a discontinuity in the age Dutch and English are both Germanic (of acquisition)–accentedness function should languages and have similar sound systems. The be observed at an onset age of about 15 years. support for the possibility of successful late Various studies have shown that such discon- speech learning would be even more convincing tinuity does not occur but that the relation if it could also be shown for learners of an L2 between onset of learning and degree of accent is typologically more distant to the L1. In a second linear across the whole age spectrum. study Bongaerts (1999) addressed this issue by replicating the Dutch–English study with Figure 5.11 presents the results of a cross- advanced Dutch L1 learners of French (French sectional study by Flege et al. (1999; see also being a Romance language). This time fewer of Birdsong, 2005, and Flege, 1999). These authors the advanced learners were judged to be native- had a group of American-English native speakers like, but still there were three (out of nine) who rate a set of English sentences spoken by 240 passed the “native” test. Abu-Rabia and Kehat native Koreans (and by 24 native English (2004) have since obtained a similar result for speakers) for overall degree of foreign accent. The late learners of Hebrew who differed in L1 Korean participants had all immigrated to the background. The findings of all three studies (and United States between the ages of 1 and 23 and see also Bongaerts, Van Summeren, Planken, & had lived there for at least 8, and on average 15, Schils, 1997) thus indicate that an early start is years at the time the rating was done. As shown, not a mandatory requirement for acquiring an not even a hint of a discontinuity can be observed unaccented pronunciation. As hypothesized by in the age–accent function, neither at an arrival Bongaerts, it is a combination of factors—inten- age of 15 years, nor at any other arrival age (note sive training in perceiving and producing the that onset of L2 learning was equated with the L2 speech sounds, massive L2 input, and a high age of arrival in the L2-speaking country; this is motivation to sound nativelike—that leads to common practice in this research area). Instead, unaccented L2 speech. degree of L2 accent was linearly related to arrival age: the later the Koreans had arrived in the If at some point early in life, at the offset of the United States, the stronger their L2 accents. This critical period around puberty, a special ability same pattern had also been obtained in an for speech learning is lost, the degree of L2 speech earlier study, in which the accentedness of native accent must be linearly related to the age of Italians’ L2 English was rated (Flege et al., acquisition for L2 speakers first exposed to the 1995a). These data thus suggest that there is no L2 before the offset point (recall that “age of age-bounded period of time in which humans are acquisition” refers to the age the learner first especially well equipped to learn (the phonology starts learning the L2): The longer before the of a) language. In addition, they do show that the closure point of the critical period the learner age at which L2 learning starts is an important began learning the L2 and, thus, the longer the predictor of degree of accent. learner has taken advantage of the special lan- guage learning ability, the less accented his or her Age of acquisition is not the only variable to L2 speech should be. In contrast, the degree of affect the degree of L2 accent. In a series of accent in L2 speakers who started learning the studies Flege and his colleagues have identified L2 beyond the offset point should not be amount of continued L1 use after immigration to systematically related to age of acquisition (see the L2-speaking country as a further relevant Chapter 2 for a thorough discussion of these variable (e.g., Flege, Frieda, & Nozawa, 1997; predictions). In other words, the CPH predicts Piske, MacKay, & Flege, 2001). In one of the per- that, across the whole age of acquisition spectrum tinent studies (Flege et al., 1997) the researchers covering both the years before and after the had native English speakers rate the accentedness assumed offset point of the critical period, age of of English sentences spoken by both native

5. WORD PRODUCTION AND SPEECH ACCENTS 271 Degree of accent in L2 English in Korean immigrants in the United States as a function of age of arrival. The accent ratings were produced by native speakers of English. The rated accents for a control group of native English speakers are shown for comparison. Reprinted from Flege et al. (1999), Copyright 1999, with permission from Elsevier. English speakers and native Italian speakers who critical period hypothesis because it suggests that had immigrated to an English-speaking area in an early onset of L2 acquisition is by no means a Canada at a young age (5.7 years on average). The guarantee that nativelike L2 speech is attained. raters’ task was to judge the language back- ground, English or Italian, of each speaker. To account for the effect of amount of L1 use Ratings were done on a 4-point scale: definitely on L2 speech accent, Flege and his collaborators Italian; probably Italian; probably English; developed a new model of L2 speech learning, definitely English. The native Italians consisted the “speech learning model” (e.g., Flege et al., of two groups that were matched on age of arrival 1997). The model assumes a common system but that differed according to self-reported use of that holds both the phonetic elements (also Italian in their daily lives (36% in the “high-use” “phonetic categories”, “phonemic categories”, or group versus 3% in the “low-use” group). “phoneme categories”) of the L1 and the L2. (These phonetic elements can be likened to the The data showed a clear effect of amount of phonological sublexical nodes in the speech- L1 use: The Italians who still spoke Italian rela- production models presented earlier in this tively often were perceived to have significantly chapter.) The L1 and L2 elements in this common stronger accents than the Italians who still spoke system influence one another, in both directions, Italian only occasionally. A further finding of during production. The strength and directional- interest was that, despite their early arrival in ity of this cross-language influence is thought to Canada, both L1 Italian groups had detectible vary as a function of a number of factors; among foreign accents. This finding challenges the others, relative language dominance and the

272 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS present (related) variable, amount of L1 and L2 Flege and his co-workers assume this assimila- use. In this view, the stronger accent of the L2 tion process to take place more often in late speech of the high-L1-use participants is caused bilinguals than in early bilinguals. The reason by the L1 phonetic elements influencing the is that late bilinguals have already established targeted L2 sounds more in high-L1-use partici- relatively strong L1 phoneme categories in pants than in low-L1-use participants. It is memory, which, as a result, constitute relatively plausible that the interaction takes the form of strong attractors of new L2 sounds. In addition co-activation of L1 phonetic elements during L2 to number of years of prior L1 use, amount of production. If true, these findings once again current L1 use affects the strength of the L1 show that during bilingual speech production phoneme categories and, therefore, their attrac- both language systems are activated in parallel, tion potential. These merged representations thus and this all the way down to the level where contribute importantly to the accentedness of L2 the speech output takes its phonetic shape. speech of, especially, late learners and L2 speakers Importantly, the present view on the cause of who still use their L1 relatively often. Interest- speech accents also predicts a speech accent when ingly, because the merged representations contain bilinguals produce L1 speech. Evidence in characteristics of both the L1 and the L2 sounds support of this prediction will be presented in that they represent (the original L1 phonetic Chapter 7 (pp. 363–365). category having shifted away somewhat from its original place in phonetic space), not only L2 In addition to the occurrence of such co- speech but also L1 speech should become activation of the language not in use, Flege and accented as a consequence of phonetic category his colleagues hypothesized a second type of assimilation (see pp. 363–365 for evidence). In interaction between the L1 and L2 phonetic Chapters 2 and 4 we have encountered a model of elements (Flege et al., 2003; see also Flege, 2002), L2 speech perception that assumes a similar pro- one that leads to phonetic representations in cess of assimilation, the perceptual assimilation bilingual memory that differ from the corre- model (Best, 1994; Best et al., 1988). A difference sponding representations in the phonetic memory from the current model is that Best’s model does system of native speakers. The authors suggested not appear to assume this position shift of the that during the development of the L2 phonetic attracting category. This difference can perhaps system, the degree of similarity between the L1 be traced back to the different phenomena the phonetic categories stored in memory and the two models primarily address, Best trying to new L2 categories to learn determines whether or account for perceptual discrimination problems not separate new categories for the L2 sounds are that non-native listeners often experience and established in memory: The greater the perceived Flege attempting to explain the ubiquitous speech dissimilarity between an L2 sound and the most accents in non-native speakers. similar L1 category, the larger the chance that a new category will be formed for the L2 sound. If In addition to phonetic category assimilation, the new L2 sound is very close to an extant L1 Flege and his colleagues assumed a process of category, and is therefore perceived as an instance phonetic category dissimilation. This process of this category (even though slight differences becomes operative when an L2 sound to be between the L2 and L1 sounds may be detected), acquired differs relatively much from the closest the former is “attracted” by the latter and the two L1 sound. Under these circumstances a separate merge into a single phonetic representation in L2 phonetic representation is established in memory. This process is called phonetic category memory. But in order to obtain maximal distinc- assimilation. During this process of merging, the tiveness between the new L2 sound and the closest attracting L1 category is assumed not to remain L1 sound, the two adopt slightly exaggerated fixed in phonemic space but to shift somewhat to values on the phonetic characteristics that dis- a position that reflects the acoustic characteristics tinguish them. For instance, English /e’/ is charac- of the L2 sound that it absorbs. terized by more tongue movement than Italian /e/.

5. WORD PRODUCTION AND SPEECH ACCENTS 273 The Italian learner of English, noticing this dif- that the capacity for speech learning remains ference, might establish a new category in mem- intact over a lifetime, an assumption that is ory for English /e’/ that exaggerates the English– supported by the fact that many cases have been Italian contrast with the effect that English /e’/ documented of late L2 learners without detect- spoken by this native Italian is produced with able L2 speech accents (e.g., Bongaerts, 1999). more tongue movement than this /e’/ uttered by However, it requires effort as well as an optimal native speakers of English. Just as category L2 learning environment to be able to counteract assimilation, category dissimilation is thus a the effects of the above two types of interaction in source of accented speech. Flege et al. (2003) bilingual phonetic memory: extensive training in obtained experimental support for the occurrence perceiving and producing the L2 (and the of both category assimilation and dissimilation, motivation to sound nativelike that goes with it) showing that early Italian–English bilinguals who and massive L2 input. In terms of Flege’s speech use L2 infrequently exaggerated tongue move- learning model, the effect of this extensive L2 ment in the production of English /e’/ (suggesting exposure and drill should be that in the shared dissimilation), whereas late Italian–English vowel space the L2 vowels ultimately occupy the bilinguals demonstrated less tongue movement same positions as they do in native speakers’ than L1 native speakers would (suggesting phonetic memory. In other words, the effects of assimilation). the processes of assimilation and dissimilation should be undone or the processes themselves Figure 5.12 illustrates these processes of must be prevented from occurring. The speech assimilation and dissimilation (adapted from learning model assumes a close relation between Flege, 2005). Figure 5.12a depicts the position of the production and perception of speech sounds: each L1 vowel in a (hypothetical) phonetic vowel The accents in L2 production arise from phon- space at the onset of L2 acquisition. The vowel emic categories, either shared by L1 and L2 or space is delineated by the vowels’ first two unique to the L2, which came into being as a formants (the concentrations of acoustic energy result of perceptual processes. Finally, perhaps in particular frequency ranges in the speech the most interesting aspect of the model is that it signal). In Figure 5.12b the L2 vowel system to be predicts speech accents not only in bilinguals’ L2 learned is superimposed on the L1 vowel system. but also in their L1. More evidence to support Finally, Figure 5.12c illustrates the processes of this prediction will be presented in Chapter 7 assimilation and dissimilation that occur during (pp. 363–365). the process of L2 speech learning. As can be seen, both processes result in the vowels of both L1 and The comprehensibility of accented L2 taking slightly different positions in the vowel L2 speech space as compared to those taken by the vowels in monolingual speakers of the two languages. The Given the fact that speech accents abound in the result of this is accented speech in both languages. L2 speech of many bilinguals and that traces of an accent may even be detectable in the L2 To summarize, Flege and his colleagues speech of highly fluent bilinguals, a question that suggested that the effects of age of L2 acquisition presents itself is whether and to what extent an and amount of current L1 use on L2 speech accent hampers listeners’ comprehension, causing accents and L1 speech accents result from two comprehension failures, delays in comprehension, forms of interaction among the L1 and L2 phon- or perhaps both. Munro and Derwing addressed emic categories in a phonemic space shared by this question in a series of studies in which both. In other words, there is no longer a reason English was always the bilinguals’ L2 (Derwing & to attribute the age of acquisition effects to the Munro, 1997; Munro & Derwing, 1994, 1995, existence of a circumscribed period in life in 1999). In these studies degree of accent was which humans are especially well equipped determined by having native speakers of English to acquire speech (and language in general). Accordingly, Flege and his collaborators assume

274 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS (a) An L2 learner’s L1 vowel system at the onset of L2 acquisition. (b) The targeted L2 vowel system added onto the L1 vowel system. (c) The processes of assimilation and dissimilation that occur during L2 speech learning. Adapted from Flege (2005).

5. WORD PRODUCTION AND SPEECH ACCENTS 275 rate the speech of L2 English speakers on a The most noteworthy result of these studies is 9-point scale. The scale varied between 1 (no that the accent ratings were generally consider- foreign accent) to 9 (very strong foreign accent). ably harsher than the comprehensibility ratings: Degree of comprehension was measured in two Although accent did affect comprehensibility, in ways, by means of “comprehensibility” ratings many cases material that was judged to be highly and “intelligibility” scores of the L2 speech. The accented received a low score on the compre- first of these concerned listeners’ judgments on hensibility scale (indicating that the accent did a 9-point scale varying between 1 (extremely easy not cause a comprehension problem). The dis- to understand) and 9 (extremely difficult to crepancy between the accent ratings and the understand). The second was a more objective intelligibility scores was even larger: Despite the measure which had been introduced earlier by fact that relatively few of the speakers’ utterances Gass and Varonis (1984). It was based on the were rated to be without an accent, in the far listeners’ transcriptions of the L2 speakers’ majority of cases the listeners’ transcription of speech following an instruction to the listeners the speakers’ utterances closely matched the to do so as carefully and completely as possible. targeted utterances, suggesting that most words The intelligibility scores were derived from the were perfectly well understood. This pattern held deviations between the transcripts thus obtained for both proficient (Munro & Derwing, 1999) and and the intended utterances. intermediate (Derwing & Munro, 1997) L2 speakers. The study comparing the L1 English The L2 English speakers in these studies were listeners’ performance on the true–false sentence from different language backgrounds: Mandarin, judgment task (Munro & Derwing, 1995) showed Cantonese, Japanese, Polish, and Spanish. Their that the L1 English speakers’ utterances were task was either to describe the events depicted in slightly more often verified correctly (98% a series of cartoons presented to them (Derwing correct) than the L2 speakers’ utterances (93%). & Munro, 1997; Munro & Derwing, 1999) or to Furthermore, the former were verified somewhat read aloud a set of sentences that were either true faster than the latter. (e.g., Elephants are big animals) or false (e.g., Most people wear hats on their feet; Munro & These findings suggest that a foreign accent Derwing, 1995). Regarding the latter test, in does incur a cost on the part of the listeners, both addition to judging the degree of accent and in terms of comprehensibility and in terms of comprehensibility, the listeners judged each processing time. These costs may be the source of sentence to be either true or false, and the time to the irritation that an L2 accent has been reported perform this judgment was measured. These to cause in native speakers of the language data were compared to the listeners’ responses to (Munro & Derwing, 1995). However, the data the same set of sentences but spoken by native also show that the adverse effects of accent on English speakers. In a fourth study (Munro & comprehensibility are generally quite small, and Derwing, 1994) the researchers had L2 speakers that many L2 speakers’ utterances, despite being first perform the cartoon description task and, accented, are perfectly intelligible and do not subsequently, had them read aloud the trans- hamper communication noticeably. Irritation criptions of their own utterances. This was done therefore seems a rather disproportional reaction to find out whether there are differences in to the minor inconvenience an accent causes for accentedness between extemporaneous speech the listener. and read materials. No such differences were observed and I will therefore ignore this study in A final noticeable finding is that listeners per- the ensuing discussion. A final question the formed well above chance in identifying the L2 researchers addressed was whether listeners can speakers’ L1 background in a forced choice select the specific language background of an L2 task. Not surprisingly, the languages that are speaker when given several options to choose typologically relatively close to one another from. were mistaken for each other more often than typologically more distant languages: The two

276 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS Asian languages Cantonese and Japanese and could it be that the listeners in these experiments the two Indo-European languages Spanish and could often successfully detect an L2 speaker’s Polish were more often confused with one L1 language background? (2) A bilingual’s L1 another than the other language pairs that can be phonetic system leaves its mark on L2 speech. formed from this set of four languages (Derwing This latter conclusion, of course, repeats the main & Munro, 1997). These findings point towards the conclusion drawn above, that L2 speech accents following two conclusions: (1) Even though it are caused by an interaction between the L1 and may appear that we, hearing a foreign language, L2 phonetic systems, resulting in L2 phonemic are completely ignorant of the language, we must categories that deviate from the corresponding at least have picked up some knowledge about its categories in monolingual speakers. phonology through past experience. How else SUMMARY • Models of speech production can be classified in three categories: discrete two-stage models, unidirectional cascade models, and interactive activation models. Discrete two-stage models assume that activation can only flow in a forward direction from the lemma level to the lexeme level and that activation is not transmitted from the former to the latter until after one lemma has been selected from the initially activated set of lemmas, the “semantic cohort”. Unidirectional cascade models also assume that activation only flows forward. In addition they assume that the moment a lemma becomes activated it immediately starts sending activation down to the lexeme level. As a result, all elements in the semantic cohort are temporarily phonologically encoded. Finally, like the unidirectional cascade models, the interactive activation models assume that all initially activated lemmas immediately pass on their activation to the corresponding lexemes. But unlike the unidirectional cascade models they assume that activation may also flow back from activated lexemes to the corresponding lemmas. • Any bilingual speech production model should deal with the fact that languages differ in the way they lexicalize the conceptual information in the preverbal message: A particular subset of the preverbal message’s conceptual components may be expressed in a single word in one language but may require a whole phrase to be expressed in another language. One solution that has been proposed is to add a so-called “verbalizer” to the model, a processing component that receives input from the preverbal message and carves it up in such a way that it matches the semantic information in the targeted lemmas. • A second issue to explain by any model of bilingual speech production is how the selection of the currently intended language comes about. One way it can do this is by including a piece of information that specifies this intention (a “language cue”) in the preverbal message and to include in each lemma a piece of semantic information that specifies to what language it belongs (a “language tag”). • Monolingual and bilingual models of speech production often zoom in on the lexicalization component of the full process; that is, on lexical access. These models usually include three levels of representations, representing preverbal concepts, lemmas, and the words’ phonological forms (“lexemes”), respectively. Lemma and lexeme representations can be represented in a single memory node (that is, they can be of the “localist” type) or the information they contain can be spread out over multiple memory nodes (“distributed” or “componential” representations). Multiple lemma activation in models that assume localist lemma representations can be accounted for in terms of spreading activation in a lexical-semantic network.

5. WORD PRODUCTION AND SPEECH ACCENTS 277 • Recent speech production models no longer assume that lemma representations contain both semantic and syntactic information, but syntactic information only. According to most models of this type, activation in the phonological level is mediated by the syntactic information in the lemma. Caramazza’s independent network model dismisses lemmas altogether. • Monolingual picture–word interference experiments have shown that the aural presentation of semantic distracters slows down picture naming but only when the distracter coincides with the picture or precedes it. In addition these studies have shown that aural phonological distracters speed up picture naming but only when the picture is presented first. These semantic interference and phonological facilitation effects are attributed to the convergence of, respectively, semantic and phonological activation in the two types of processes involved in the picture–word interference task: top-down picture naming and bottom-up word processing. • Bilingual studies using the picture–word interference task have suggested that at some point during picture naming in the weaker L2, the picture’s translation in L1 is also activated. This finding indicates that word production in bilinguals is language-nonselective. It is however not clear whether only lemma selection or both lemma selection and phonological encoding are language-nonselective. • According to a common view of language-nonselective lexical access in bilingual speech production, strongly activated lexical representations in the non-response language should hinder production. A couple of bilingual picture–word interference experiments proved this assumption wrong and led to a model of bilingual lexical access in which a stage of language-nonselective activation is followed by a stage of language-specific lexical selection. • In picture naming by bilinguals a cognate facilitation effect is observed. This effect suggests that phonological encoding is language-nonselective, thus supporting cascaded models of bilingual speech production. The effect may however also result from representational differences between cognates and non-cognates. • Conclusive evidence in support of language-nonselective lexical access in bilinguals requires that the participants in an experiment do not suspect that their bilingualism is being tested and that the stimulus materials are exclusively unilingual. Experiments that employed the bilingual version of the picture–word interference task do not fulfill these requirements. • Whereas the picture-naming studies legitimate the conclusion that both lexeme activation and phonological encoding are language-nonselective, it is too early to tell whether grammatical processing is also language-nonselective. • Bilingual Stroop studies have reliably produced both intra- and interlingual Stroop effects, and the size of these effects depended on the participants’ relative proficiency in their two languages and on the degree of similarity of the languages. The effects are usually explained in terms of the degree of within- and between-language interconnectedness of the linguistic elements in bilingual memory. • Careful task analysis reveals that word translation, Stroop color naming, and picture naming share many processing components. • The cognate effect in word translation may be caused by (1) facilitated name retrieval due to language-nonselective phonological encoding that favors cognates over non-cognates, (2) representational differences between cognates and non-cognates, (3) relatively strong activation of the associated semantic nodes when a cognate is presented for translation. • Performance in a large number of word production tasks such as monolingual and bilingual picture naming, monolingual and bilingual Stroop color naming, and word translation is based on activation patterns in one and the same underlying, layered, memory system augmented by a general control system that sees to it that the current goal is met (that is, that the requested task and not another one is performed).

278 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS • Word production, visual word recognition, and auditory word recognition may exploit one and the same underlying processing system with components that are either fully shared or highly interconnected between production and recognition and between the aural and visual modality. • Onset age of L2 acquisition (“age of acquisition”) and degree of L2 accentedness are linearly related: The younger the L2 speaker was when first starting to learn the L2, the more nativelike his or her L2 sounds. Despite popular wisdom, for various reasons this age of acquisition effect cannot be explained in terms of the existence of a critical period for language learning early in life. A further variable that determines degree of L2 accentedness is the extent to which the L2 speaker still speaks the L1. • L2 speech accents can partly be explained in terms of the merger of L2 phonetic categories with extant L1 phonetic categories in a phonetic memory system shared by the L1 and L2: L2 sounds that closely resemble an L1 phonetic category in this system assimilate with this L1 category. A second cause of the L2 accent is a process of dissimilation, which applies to L2 sounds that differ strongly from all L1 sounds: They take a position in phonetic memory which exaggerates the actual physical distance between the new L2 sound and its closest L1 sound. • The stronger the L1 phonetic categories in memory, the stronger their assimilative and dissimilative effects during L2 speech learning and, consequently, the stronger the L2 accent. The strength of the L1 phonetic categories depends on amount of previous L1 use. Therefore, the older the L2 speakers were when they first started learning the L2 and the more often they still use the L1, the stronger the L2 accent. • Accented speech does affect its comprehensibility but this adverse effect is small, even in cases where the accent is a strong one. • Listeners can detect the L1 background of L2 speakers. This is independent evidence of the influence of L1 phonetic categories on the L2 categories.

6 Language Control INTRODUCTION AND PREVIEW guaranteeing language-pure comprehension and production in such “monolingual” circumstances In the previous chapters we have seen that when (where a single language is selected for use), the bilinguals are conversing with their interlocutors mechanism involved must also enable language in one of their languages the mental system that control in translation situations, where there is stores the other language is not completely at rest. input in one language (usually called the source In general a picture emerged of a bilingual lin- language) while output in the other language guistic system that is noisier than the language (the target language) must be produced. Here system of monolingual language users because, switching off one linguistic subsystem (if such during both language comprehension and lan- were possible at all) is not a feasible option guage production, linguistic elements of both because the translation act requires the parallel linguistic subsystems are activated. Yet, when involvement of both languages. Alternatively, witnessing bilinguals partaking in conversations rapid mental switching between the two linguistic that are intended to be unilingual, it is obvious subsystems must take place. In this chapter I will that they are quite successful at coping with this present a number of views on language control in extra fierce mental rivalry. Generally, if a bilinguals; that is, on how they manage to adapt bilingual has selected one language for current their language use to the specific requirements of use, his or her speech contains few intrusions of the current communicative context, selecting one the non-selected language, and misunderstand- of their two languages or using both languages in ings arising from mistaking an input for a word in translation. the other language are rare. Apparently, some efficient mental control mechanism is at work A prerequisite for language control is that that prevents such infelicities from hampering the language membership of individual lexical items conversation flow, a mechanism that enables (and of other types of linguistic knowledge struc- the bilingual to mentally control the language tures) is built into the bilingual’s language system. of input and output. In addition to generally The system must somehow specify which ones of the stored units belong to the one linguistic subsystem and which ones belong to the other. A 279

280 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS number of proposals as to how this is done have languages are affected by the control process at been advanced. According to one of them, the work. For instance, some control process may elements of each language are embedded in a simultaneously activate all units of the targeted separate network of strongly interconnected (selected) language and at the same time elements within bilingual memory, called a “sub- deactivate or inhibit all elements in the non- set”. The connections between the units within a targeted (non-selected) language. Contrary to subset are formed by their co-occurrence in the global control, local control exerts its effect on environment and possibly by mental contiguity; specific elements within the language system. that is, by their co-occurrence in thought. In other The second dimension of control concerns the words, the representations of the linguistic units direction in which it operates, proactively or belonging to one particular language become reactively. Proactive control (also called “early” strongly interconnected in bilingual memory control) is involved when, for instance, the inten- because they co-occur in linguistic expressions tion to speak one language and not the other and are encountered as such by the language user. results in a preparatory setting of the system that A unit’s embedding within one of these networks increases the availability of the elements of the determines its language membership. The subset selected language through pre-activation, while concept may be applied to a complete language perhaps at the same time decreasing the avail- system (e.g., Paradis, 1981, 2004) or to sub- ability of the non-selected language, through components of the system such as a component deactivation or inhibition/suppression. This pre- that stores phonemes or one that stores the paratory setting of the system subsequently phonological forms of complete words (e.g., influences performance on the language task to be Grosjean, 1997a). A second proposal regarding executed. In contrast, reactive control (or “late” the structural implementation of language control) means that the control process operates membership posits the existence of two so-called on the imminent outputs of the language system, “language nodes” within the memory system of a preventing elements of the non-response lan- bilingual, one for each language (e.g., Dijkstra & guage from emerging in speech (in production) or Van Heuven, 1998, 2002). The form representa- preventing an input element from being assigned tions of all L1 words are connected to the L1 a meaning in the non-target language (in com- language node. The form representations of all prehension). The third dimension concerns the L2 words are connected to the L2 language node. locus of control; that is, whether the control Language membership is determined by these process exerts its effect on the language system connections between word form representations proper or on the output of the language on the one hand and language nodes on the other. system. Finally, the fourth dimension concerns A third proposal (e.g., Green, 1998; Poulisse, its source, internal (“endogenous”) or external 1997) is that the lemma representation of each (“exogenous”): The control process may be set in word contains a “language tag”, a piece of infor- motion by a stimulus presented to the system mation that specifies that this particular lemma (exogenously), or it may result from the intention belongs to L1 or L2. These three proposals are to speak one language and not the other or to not mutually exclusive, but more than one of perform some other language task such as trans- them may be assumed within one and the same lation. Different forms of control may cooperate view on bilingual control. to secure smooth processing so, for instance, con- trol may at the same time be exerted globally The various views on bilingual language by some proactive control process and locally by control that have been proposed can be dis- some reactive control process. tinguished from one another along four inter- related theoretical dimensions. The first of these Following the usual Methods and Tasks sec- concerns the scope of control; that is, whether tion, this chapter presents a number of different control is exerted globally or locally. Global views on bilingual language control in unilingual control means that all elements of one or both settings. Although not always using the above

6. LANGUAGE CONTROL 281 terminology, each of these views incorporates a the words or phrases within a sentence or a text specific set of assumptions regarding the above were in the other language. Similarly, in the switch four dimensions of control. To avoid the risk of blocks of the production studies participants had misrepresenting the original authors’ views I will to switch between the languages on request. What generally refrain from molding them in the pres- language to use on a particular trial was signaled ent terminological framework. Nevertheless, to by a cue; for instance, a circle presented prior to a get a good grasp of the differences between the picture to name cued the use of one language and various theoretical stances, it helps to bear these a triangle cued the use of the other language. four dimensions of control in mind while getting acquainted with these views. Following the sec- The blocked design is still being used tions dealing with control in unilingual settings, occasionally, but in the majority of the more the final section of this chapter is dedicated recent studies the switch and non-switch con- entirely to language control in translation and, ditions were embedded within one and the same especially, simultaneous interpreting. series of stimuli presented to the participants. Within comprehension studies of this type (e.g., METHODS AND TASKS Grainger & Beauvillain, 1987) two successive words from the same language constitute a Probably the most popular techniques for study- non-switch trial and two successive words from ing language control in bilinguals are the picture– different languages constitute a switch trial. word interference paradigm, the bilingual Stroop Similarly in production studies (e.g., Meuter & task, and the “language-switching paradigm”. Allport, 1999), two successive trials in a series The former two have already been dealt with in may be cued for a response in the same language Chapter 5. In the current chapter I will primarily (non-switch) or for responses in different lan- focus on the language-switching paradigm. In guages (switch). In the earliest comprehension studies using this paradigm the bilingual partici- studies (Kolers, 1966; Macnamara, 1967) the pants have to switch between their two languages, participants were presented with complete (uni- mostly at odd moments but sometimes in a lingual and code-switched) sentences. Instead, the regular pattern, the researcher trying to find out experimental materials in the later studies were what happens in their minds during both non- generally smaller stimuli presented “out of con- switch and switch trials. The technique has been text”. In the comprehension studies these were employed in both comprehension and production often words and nonwords presented one by one, studies and in a number of different forms, the each requiring a lexical decision response. In the production studies often encompassing simple production studies they were usually sequences of picture naming as the task to be performed digits or pictures to be named one after the other. by the participants. In the earliest language- switching studies (Kolers, 1966; Macnamara, The critical dependent variable in all these 1967; Macnamara, Krauthammer, & Bolgar, studies was the time it took to switch between 1968; Macnamara & Kushnir, 1971) a “blocked” languages, the so-called “switch cost”. In the design was used: Non-switch blocks in the com- seminal studies by Kolers and Macnamara and prehension studies showed materials in one of the his colleagues the cost per switch was measured participant’s two languages only and non-switch globally by, first, calculating the difference in blocks in production studies invited speech in overall response time for a mixed block on the just one language. In contrast, in the switch one hand and a unilingual block on the other blocks of the comprehension studies mixed lan- hand and, second, dividing this difference by the guage materials were presented: With one of the number of language switches in the mixed block. languages often serving as base language, some of In the more recent studies (where the switch and non-switch conditions were encompassed within one and the same series of trials) the switch costs were measured locally, at the very point where the switches occurred. An advantage of measuring

282 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS switch costs locally is that it enables a comparison likely at least partly driven by the particularities of the costs incurred by a switch to L1 on the one of the communication setting—determine when hand and to L2 on the other hand. As we will to switch languages, and in naturalistic bilingual see, a number of studies have shown that these speech comprehension their interlocutors deter- switching costs are not equally large, and the mine when to switch the input language (where observed directional difference in their size is one “determines” should not be taken to mean that of the cornerstones of one of the theories of language switching always involves a conscious bilingual control to be reviewed in this chapter choice on the part of the speaker). The loci of (pp. 307–313). experimenter-imposed and naturally occurring switches may not coincide, and an experimenter- A problematic aspect of measuring switch imposed switch may involve a linguistic unit that costs is that they are confounded with “mixing under more natural circumstances might never costs”: Studies on task switching—say between involve a language switch. A switch imposed by color naming or word reading in the classical the experimenter may impede processing—as color Stroop task—have shown that it not only the switch costs observed under those conditions takes longer to respond on a switch trial than on a show they do—whereas a switch that occurs non-switch trial (the switch cost) but that naturally may not hamper the conversational flow response times are also generally longer, even on or may even enhance conversational fluency. non-switch trials when responding alternates between different tasks, than when the partici- For these reasons, some researchers (and espe- pants only perform one task. This effect is called cially François Grosjean; see pp. 288–291) have the “mixing cost”. A plausible reason is that in studied language switching in experimental set- the mixing condition the participants must keep tings that had the participants themselves deter- two tasks ready in working memory (Monsell, mine when to switch languages. In these studies 2003). Especially because a substantial amount of the frequency of occurrence of such participant- evidence suggests that language switching is a motivated switches was manipulated by varying specific form of task switching, exploiting the the information the participants received regard- same cognitive machinery (see pp. 393–401 and ing their (imaginary) addressees or by varying 435–446), the occurrence of mixing costs is the context in which a conversation took place; something to keep in mind when designing for instance, formal versus informal. More language-switching studies. A further experi- detailed information on these studies, including mental procedure, the “alternating-runs para- their rationale, will be presented in due course, digm”, has been used to avoid a possible con- but a final note to make here is that naturally found between switching and mixing costs. In this occurring switches may be of two types, inten- paradigm the tasks alternate in a constant and tional and non-intentional, the latter being predictable pattern every N trials (Monsell, 2003). inadvertent lapses into the other language. It has been suggested that non-intentional switches In a number of respects the language- occur especially frequently at points where the switching studies characterized above do not con- two languages “meet” because of the use of a stitute a veridical imitation of language switching similar linguistic structure; for instance a gram- in natural discourse. One obvious reason is the matical structure shared by the two languages or fact that most of these studies have examined a word with a cognate translation in the other language switching out of context; that is, on language (e.g., Broersma & De Bot, 2006; Clyne, sequences of words that would never join 1967). together in naturalistic linguistic structures. Another reason is that the moment a language Simultaneous interpreting, the main theme of switch occurred, and the type of linguistic unit the final section of this chapter, has primarily involved, were both determined by the researcher. been tackled experimentally by identifying one Instead, in naturalistic bilingual speech produc- or more components of the full task and then tion the bilingual speakers themselves—though comparing the performance of professional

6. LANGUAGE CONTROL 283 interpreters on the one hand and ordinary Goldman-Eisler, 1972). These forced delays bilinguals on the other hand on the component increase processing load and lengthen the EVS. tasks (e.g., word retrieval, working memory, or input comprehension). Alternatively, the per- In recent research on language control in formance of ordinary bilinguals and professional bilinguals the important new insight has emerged interpreters on one or more component tasks that this form of control is a manifestation of is correlated with these same individuals’ per- cognitive control in general, and that bilingual formance on the full task. Although other language control and other forms of cognitive methods to study simultaneous interpreting are control such as inhibiting contextually inappro- available (see Christoffels & De Groot, 2005, for a priate actions or ignoring goal-irrelevant acti- review), my focus will be on this componential vation in the cognitive system are effected by the approach. The underlying assumption is that same cognitive equipment. What is more, the mental control while performing this extremely evidence suggests that bilinguals are often better complex task can be maximally secured when as than monolinguals on tasks that require these many of its component parts as possible run other forms of mental control. This evidence automatically. The interpreter’s limited conscious will be discussed in Chapter 7 (pp. 393–396) and mental resources can then all be allocated to information on the brain structures involved will the task’s components that defy automation, be provided in Chapter 8 (pp. 435–446). including a monitoring process that oversees per- formance and intervenes whenever a performance LANGUAGE CONTROL AS SWITCHING failure looms. BETWEEN LANGUAGE SUBSETS Common dependent variables in studies on Introduction simultaneous interpreting are reaction time measurements that reflect the speed with which The theoretical starting point of the earliest knowledge is retrieved from memory, error scores, language-switching studies (Kolers, 1966; quality-of-performance measures, measures that Macnamara, 1967; Macnamara et al., 1968; reflect the participants’ working memory Macnamara & Kushnir, 1971) was the idea that capacity, and the so-called ear–voice span (EVS). bilinguals form stronger associations among the The EVS is the time lag between the moment a words within each of their languages, than source-language fragment is spoken by the between words across their languages, and that speaker and the moment the equivalent of this this is why they can keep their languages fragment emerges in the interpreter’s aural functionally separate or, in terms of Macnamara rendition of the source language. It is measured (1967), how they secure “linguistic inde- as the number of words or seconds between input pendence”. These authors thus adhere to the and corresponding output and varies as a func- subset notion introduced earlier, and it is this tion of local interpreting difficulty and specific notion that they set out to test by means of the relationships that hold between the source and language-switching methodology. Bilinguals in target language. For instance, sentences in the the switch conditions were assumed to go in and source language may typically have a subject- out of their language subsets upon each language object-verb structure (as in German and change in linguistic input or output, selecting Japanese) whereas sentences in the target lan- one language subset to process the current lin- guage may generally have a subject-verb-object guistic input or to produce the current output. structure (as in English). Given this state of Retreating from one language subset and access- affairs the interpreter will often be forced to inter- ing the other was assumed to require extra effort rupt his or her rendition of the source language beyond the effort involved when no mental until the speaker has output the verb, in the switch of language subset is required. From this, meantime keeping the corresponding object phrase in working memory (e.g., Crystal, 1987;

284 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS the authors predicted slower performance on Other authors (e.g., Grainger, 1993; Grainger switch trials than on non-switch trials. Two & Beauvillain, 1987) have subsequently inter- separate mental switching devices were assumed: preted Macnamara et al.’s view of language an input switch and an output switch. The input switching as going in and out of language sub- and output switch were assumed to govern the sets to imply that the language currently accessed selection of the appropriate language subset in is “switched on” (activated) and that the lan- speech comprehension and speech production, guage not in use is “switched off” (deactivated). respectively. This implication is not stated explicitly in the original reports of Macnamara and his col- Whereas going out of one language and enter- leagues, although these authors do mention that ing the other was thought to require effort and other researchers before them seem to have pro- thus extra processing time, subsequently staying posed the switching device works by somehow in a language once switched to, was thought to switching one language off and the other on: be automatic and effort-free. Macnamara et al. “Penfield and Roberts (1959) described this (1968, p. 213) used the following vivid analogy to device as a ‘curiously effective automatic switch’ illustrate this point: which they seem to visualize as having the effect of when one language is on, the other is necessar- Indeed, the whole study of bilingualism ily off” (Macnamara et al., 1968). There is, gives one the distinct impression that the however, no a priori reason to assume that the bilingual’s linguistic performance is similar “in-out” account of language switching implies to that of the musician who observes the that the elements of the language currently not notation for key at the beginning of a piece accessed are “off”, deactivated. Indeed, the latter of music and then forgets about it though assumption would be hard to reconcile with the in his playing he performs the actions plethora of studies, discussed in Chapters 4 and appropriate to the key. Similarly, the 5, demonstrating that the non-selected language bilingual once started in one language can is often co-activated with the language currently forget about which language he is speaking in use. and yet obey the rules of that language. Another possibility, fully consistent with the As we shall see, some of the more recent in-out view of Macnamara and his colleagues, is related work appears to take a different stance, that the elements of the language not in use can assuming that some control mechanism is be activated but that this activation is ignored incessantly monitoring the imminent output of and may even go unnoticed. Such a state of the language system to see whether it might con- affairs would be consistent with one of the tain elements of the contextually inappropriate views on bilingual control that has emerged in language and, if so, suppressing them. A further the recent literature (e.g., Costa et al., 1999; view expressed and put to a test in this early Costa & Santesteban, 2004b; see pp. 242–245 and work was that the hypothesized input switch 310–312). On that account, the device that operates automatically whereas the hypothesized Macnamara and his colleagues metaphorically output switch was thought to be subject to called a “language switch” might well be some voluntary control (Macnamara & Kushnir, mental control mechanism that sees to it that 1971; Macnamara et al., 1968). This assumed dif- whatever goes on in the non-targeted language ference between input switch and output switch subsystem is neglected. Viewed from this per- was motivated by the fact that bilinguals can spective, the studies of Macnamara and his deliberately choose to speak one language rather co-workers appear surprisingly modern, both than the other but have no control over the lan- with respect to the methodology they used guage of input. Recently, in a different theoretical and the results they obtained. It is to a more framework, Costa and Santesteban (2004a) have detailed description of these studies that I will made this very same point. now turn.

6. LANGUAGE CONTROL 285 Evidence switch operates automatically. As mentioned, the authors also looked at the effect of switching Macnamara (1967) and Macnamara et al. (1968) between a pair of tasks performed in one and intended to examine output switching but the same language. In this part of the study the because of the nature of the stimulus materials participants either named all digits within a used in the first of these studies (triads of words) block (in one and the same language), added it was uncertain whether the switching mechan- 1 to all digits presented within a block (e.g., if ism at work had indeed been the hypothesized 2 appeared, the participant was to say 3; all output switch, not the input switch, or both of responses again in the same language), or they them. The 1968 study prevented this possible switched, within one and the same block, confound by presenting language-neutral stimuli between naming the presented digit or adding to the participants; namely, digits instead of 1 onto it (again in one and the same language). words. For this reason I will confine myself to a This “digit-switching” condition was included discussion of this second study. It is also the to see whether there is something special about most comprehensive study of the two, including language switching or, in other words, to see a manipulation that allows for a test of the whether such a thing as a language switch exists authors’ hypothesis that output switching is at all. subject to voluntary control. Furthermore it included a comparison between language The data showed a longer average response switching and switching between two tasks that time per digit-naming trial in the language-mixed both involved the use of one and the same lan- blocks than in the unilingual blocks, thus con- guage. It is especially this latter comparison that firming the prediction that switching languages gives Macnamara’s work its modern flavor, takes time and, according to the authors, support- because presumably the most topical question ing the subset organization of bilingual memory. addressed in current research on bilingual control When the switch patterns were regular, average is whether or not the mechanism involved in lan- response time per trial was .21 seconds longer in guage switching differs at all from the control language-mixed blocks. With irregular switch mechanism involved in switching between other patterns, the difference between unilingual and pairs of tasks that do not involve a language mixed conditions was .39 seconds per switch trial. switch. This difference between the regular and irregular language-switching conditions confirmed the Macnamara et al. (1968) used a digit-naming authors’ hypothesis that output switching is task that in modified form has been adopted in under deliberate control. Furthermore, the digit- more recent work on language switching (e.g., switching condition produced costs that were Finkbeiner, Almeida, Janssen, & Caramazza, surprisingly similar to the costs obtained with 2006; Meuter & Allport, 1999). French–English language switching: respectively .22 seconds and participants named blocks of digits in their one .39 seconds per switch in the regular and irregular language, in their other language, or switched conditions. The authors concluded that language between languages within a block of trials. In switching “seems to require no psychological skill some of the mixed-language blocks the switches peculiar to bilingualism, but rather a skill which is occurred on a regular basis and could therefore be equally applicable in a larger number of oper- anticipated. In other mixed blocks the switch ations in which persons are asked to switch patterns were irregular. If output switching is modes of response rapidly” (Macnamara et al., under voluntary control, switching costs should 1968, pp. 213–214). The authors refrain from be relatively small in the former condition, so the elaborating on the wider theoretical implications authors reasoned, because a regular switching of this result. If going out of one language subset pattern makes it possible for the participants and accessing the other is the source of the to predict a language switch. A null effect of this language-switching cost, as they assume, what manipulation would suggest that the output equivalent explanation could account for the data