Language and Cognition in Bilinguals and Multilinguals_ An Introduction

186 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS be activated simultaneously, but not under all were asked to identify the targets. To decrease circumstances. the legibility of the primes and targets, they were interspersed between two masking signals (the A related study by Van Leerdam, Bosman, and “forward” and the “backward” mask). The De Groot (2009), testing Dutch–English bi- participants were not aware that primes were linguals, provided converging evidence of parallel presented and thought they were participating phonological coding in a bilingual’s two lan- in experiments aimed at discovering how briefly guages. These authors used a new task, the presented words (the targets) are recognized. “bimodal matching task”, in which on every trial a printed word was displayed while at the Using this methodology, Brysbaert, Van Dyck, same time a speech segment, consisting of a vowel and Van de Poel (1999) asked monolingual followed by a consonant, was presented aurally. French speakers and Dutch–French bilinguals The printed word was always a word in the to identify French targets that were preceded by participants’ L2 English (e.g., mood ). The speech either Dutch or French primes. In a “homo- segment was or was not the correct pronunciation phonic” Dutch–French condition, the primes of the printed word’s body and the participants’ were Dutch words or Dutch-like nonwords task was to decide whether or not the speech that, if pronounced according to the L1 Dutch segment matched the printed word’s body by letter–sound conversion rules, were homophonic pressing either a “yes” or a “no” button. The to the French (L2) target words. Examples are the critical comparison was between two types of prime–target pair voet–voute (where voet is Dutch “no” trials. In one of these types of trials (the for “foot” and voute is French for “vault”), and “catch” trials) the speech segment was derived the pair soer–sourd (where soer is a Dutch-like from a Dutch enemy of the printed English word; nonword and sourd is French for “deaf ”). that is, from a Dutch word with the same printed Performance in this condition was compared word body as the English word but pronounced with a “graphemic control” condition, where, on differently in Dutch. For instance, the English average, the prime (again a Dutch word or a word mood might be accompanied by the spoken Dutch-like nonword) shared as many letters with body of Dutch lood, which (just as the spoken the target as the corresponding homophonic bodies of other Dutch neighbors of this word, prime did, but without sharing phonology (e.g., such as dood, rood, and brood ) is pronounced as volk–voute, where volk is Dutch for “people”, and the spoken body in English road. The second type siard–sourd, where siard is a Dutch-like nonword). of “no” trials were control trials in which the In the French–French condition the primes were English printed word had no such enemy neigh- French-like nonwords that were either homo- bors in Dutch. A huge effect of type of “no” trial phonic with the target (fain–faim), graphemically was observed, the catch trials leading to extremely similar to it (faic–faim), or unrelated controls high numbers of false positives. This finding (fint–faim). The dependent variable was the suggests that the printed English words gave rise percentage of correct target identifications. Any to parallel phonological activation in English and difference in performance to the targets in the Dutch, in other words, that phonological coding homophonic and graphemic control conditions of printed L2 words is language-nonselective (the to emerge would have to be attributed to the analogous experiment with printed L1 words was homophony relation between prime and target in not performed). the former condition and would, by implication, suggest that automatic spelling-to-sound coding In a series of studies, Brysbaert and his of the primes had occurred. The conclusion that colleagues approached the question whether this coding procedure occurred automatically language-nonselective phonological activation follows from the fact that the primes could not be occurs from a different angle, using the masked perceived consciously. The results are presented in priming technique. In these studies, briefly pre- Figure 4.7. sented visual targets were preceded by briefly presented visual primes and the participants As shown, even though the participants were

4. COMPREHENSION PROCESSES 187 unaware of the primes, in the French–French Probability of correct target condition identification performance was better word identification as a for targets preceded by homophonic primes than function of prime type for targets preceded by homographic control (homophonic vs. graphemic), primes. This “phonological priming effect” was participant group equally large in monolinguals and bilinguals, (monolinguals vs. bilinguals), indicating that L2 speakers exhibit automatic and prime language (French phonological coding to the same extent as L1 and Dutch in conditions speakers. Of particular interest, bilinguals also French–French and Dutch– showed the effect in the Dutch–French condition, French, respectively). Data and it was statistically as large as in the French– from Brysbaert et al. (1999). French condition. Duyck, Diependaele, Drieghe, and Brysbaert (2004) replicated the Dutch– including two groups that differed in their level French part of this study and extended it by of fluency in L2 French. Again all targets were in L2 French. The phonological priming effect materialized once more and turned out to be equally large for a group of nearly balanced Dutch–French bilinguals and a group of Dutch– French bilinguals less proficient in French. Finally, Van Wijnendaele and Brysbaert (2002) wondered whether the effect would also materialize if the primes were words from L2 and the targets were L1 words. Now testing both Dutch–French and French–Dutch bilinguals, they obtained the effect for both groups in an experiment with Dutch primes and French targets

188 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS (that is, with targets in L2 for the Dutch–French words but not otherwise). Similar adaptive group but in L1 for the French–Dutch group). In behavior was not possible in the masked-priming fact, the effect was equally large for both groups. experiments because it requires conscious percep- All in all, this set of studies provides strong tion of the primes. As we have seen before evidence that upon the presentation of a printed (pp. 180–181), recent models of bilingual memory word the spelling–sound correspondences of both incorporate a control system that can account for of a bilingual’s languages can be automatically such adaptive behavior. activated in parallel. Especially the fact that the primes were not perceived consciously makes this Evidence from bilingualism across conclusion hard to avoid. different alphabets and dialects The studies discussed in this section (and the All studies discussed so far tested bilinguals interlexical homophone studies discussed earlier; whose two languages share the same, Roman, e.g., Haigh & Jared, 2007) lead to the following alphabet. One could argue that the parallel three conclusions: (1) During visual recognition phonological activation of representations in of L2 words bilinguals assemble the phonological both languages that is suggested by these studies forms of these words just as native speakers do; results from this use of a shared alphabet. After (2) This process comes about automatically and all, no matter how irregular the grapheme– unconsciously; (3) Under certain circumstances phoneme relations within some of these lan- the grapheme-to-phoneme conversion rules of guages may be (as in English), languages that both languages are activated in parallel. Such share the Roman alphabetic script are all likely to parallel activation seemed a universal phenom- have at least a subset of their grapheme–phoneme enon in the visual masking experiments per- correspondences in common. For such language formed by Brysbaert and his colleagues but was pairs, applying the conversion rules of one modulated by context (the “block-order” language implies applying them for the other manipulation) and the participants’ L2 pro- language as well. The evidence for language- ficiency level in Jared and Kroll’s (2001) study. nonselective phonological activation would there- What remains is to explain why parallel activation fore be even more convincing if it would also was a less robust phenomenon in the latter study. materialize in forms of bilingualism that involve A possible reason may be the use of a different the command of languages with different scripts. methodology: the identification of masked targets following masked primes (Brysbaert and Thierry and Wu (2004) cleverly demonstrated his colleagues) versus unprimed naming (Jared the occurrence of parallel activation of languages and Kroll). The results of the Brysbaert studies written in different scripts, in so doing providing clearly showed that the cross-language primes, indirect support that this parallel activation also though not consciously perceived, accessed and involves phonological memory representations. activated the non-target language system. These In this study English monolinguals and Chinese– primes may thus have acted as a warm-up of the English bilinguals performed a semantic decision elements in the non-target language, just as a task: They were presented with pairs of English prior block of non-target language trials (the words and had to decide for each pair whether French block) in Jared and Kroll’s study or not it consisted of related words (e.g., post– appeared to have done. That Jared and Kroll mailbox: “yes”; novel–violin: “no”). Unknown to obtained evidence of parallel phonological the participants, the stimulus materials were activation in this very condition supports this manipulated on a hidden variable: The Chinese analysis. Alternatively, the different results may translations of the words in half of the semantic- be due to the participants in Jared and Kroll’s ally related and unrelated word pairs shared a study having somehow adapted themselves to the logographic character. The Chinese translations specific experimental condition (e.g., expecting of the words in the remaining half of the stimulus French target words following a block of French pairs did not share any form resemblance. The

4. COMPREHENSION PROCESSES 189 four experimental conditions emerging from the One of the pertinent studies, by Gollan, semantic relatedness and form similarity manipu- Forster, and Frost (1997) tested Hebrew–English lations are illustrated in Table 4.5. and English–Hebrew bilinguals using a priming paradigm. Hebrew and English use completely As expected, the monolingual English par- different alphabets (they contain completely dif- ticipants were not affected by the hidden form ferent sets of letters, differ in the number of similarity factor but only exhibited an effect of letters, and the Hebrew alphabet consists entirely the semantic manipulation. The bilingual par- of consonants). The primes were masked such ticipants were similarly affected by the semantic that they could not be perceived consciously and factor. In addition, they showed a different the targets were clearly visible letter strings to pattern of responses for the shared character be categorized by the participants as words or word pairs on the one hand and the non-shared nonwords (lexical decision). On the trials wherein character pairs on the other, both on behavioral both prime and target were words, they could measures (response times and error rates) and on either both be words from the same one of the ERPs recorded during task performance. This participants’ languages (both Hebrew or both finding suggests that the stimulus words’ Chinese English) or the language of prime and target dif- translations were activated during English task fered (the “between-language” condition). I will performance. Although it is not absolutely certain focus my discussion on the results of the latter that this co-activation involved the phonological condition. forms of the Chinese translations in addition to their orthographic forms (the logographic char- The primes and targets in the between- acters), the interconnectedness assumed between language condition were either translations of orthographic and phonological codes in memory one another or they were unrelated words. Some (see Figure 4.4) suggests that in this experiment of the primes in the translation stimuli shared a the Chinese co-activation implicated phono- cognate relation with their targets, whereas others logical co-activation. The fact that the partici- did not. Because Hebrew and English are written pants were totally oblivious of the hidden form in alphabetic scripts that bear no orthographic manipulation furthermore suggests that co- resemblance to one another, in this particular activation of non-target Chinese orthography and study cognate relatedness meant that prime and phonology was automatic. A couple of other target had similar phonological forms. A reliable studies provide more direct evidence of automatic translation priming effect (the difference in phonological coding in a non-target language response times for translation stimuli on the when the target and non-target languages are one hand and matched, unrelated prime–target written in different scripts. These (and some stimuli on the other hand) occurred, but only related studies) will be discussed in the remainder with primes in the stronger L1 and targets in of this section. the weaker L2 (see also p. 140). But especially TABLE 4.5 Experimental conditions in Thierry and Wu (2004) Semantically related Semantically unrelated Partly shared forms post–mailbox novel–violin you jü–you xiang xiao shuo–xiao tiqin Completely different forms bath–shower rabbit–desk xi zao–ling yu tu zi–ke zhuo The participants performed a semantic decision task in English: They decided whether pairs of English words consisted of semantically related or unrelated words. Unknown to the participants, half of the related and unrelated pairs shared a Chinese character; the other half consisted of two words with completely different visual forms.

190 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS important in the present context was that these phonological code and the requested Hebrew effects from L1 on L2 were larger when primes response. and targets were cognates than when they were non-cognates. The enhanced translation priming Whereas the studies by Gollan et al. (1997) effect for cognate translations, which (in addition and Tzelgov et al. (1996) demonstrated that to meaning) share phonology but no orthography, automatic phonological activation of a non-target once again suggests that under certain circum- language occurs in different-alphabet bilingual- stances bilinguals may automatically activate ism, a further bi-alphabetic study indicated that and apply spelling–sound correspondences in the one and the same letter string can give rise to non-target language (here, the language of the the simultaneous activation of the phonological prime). The reason why the activation must have rule systems of both alphabets. Lukatela, Savic´, occurred automatically is that the primes could Gligorijevic´, Ognjenovic´, and Turvey (1978) not be consciously identified. demonstrated this phenomenon to occur in Serbo-Croatian, a Slavic Indo-European lan- Employing a completely different paradigm, guage written in two different alphabets, the Tzelgov, Henik, Sneg, and Baruch (1996) pro- Roman and the Cyrillic. Some of the letters of the vided converging evidence that phonological Serbo-Croatian writing system appear in both activation in a non-target language may occur in alphabets but sound differently in each of them. different-alphabet bilingualism. The researchers The researchers presented letter strings to their employed the Stroop color-naming task (or, (monolingual, bi-“scriptal”) participants for simply, the Stroop task) and, as Gollan and lexical decision. In the critical condition a letter her colleagues, they tested Hebrew–English string was a word if pronounced according to bilinguals. In Stroop experiments the participants the letter-to-sound conversion rules of the one are presented with color words printed in dif- alphabet but a nonword if pronounced according ferent colors and are asked to name the color to the rule system of the other alphabet. If of the words and ignore their names. When the phonological coding occurs simultaneously color word stimulus and the color it is printed in according to the rule systems of both alphabets, mismatch (e.g., the word red printed in green ink), the stimuli in this condition should give rise the color naming response is substantially slower to response conflict and, thus, slow down than when they match. This difference is the responding. This is exactly what the researchers Stroop effect. Tzelgov and his colleagues observed found. a Stroop effect for stimuli that they coined “cross- script homophones”: nonsensical letter strings As mentioned earlier, automatic phonological written in the Roman alphabet that, when pro- activation has also been shown to occur in lan- nounced according to the English spelling-to- guages that use scripts that are not sound-based, sound conversion rules, sound like Hebrew color such as ideographic Chinese (e.g., Tzeng et al., words. For instance, if the English pronunciation 1977). An interesting feature of Chinese is that its rules are applied to the nonwords kahol and one written script is pronounced differently in the adom, they sound just like the Hebrew words for various dialects of Chinese so that bi-dialectal blue and red, respectively. When the color of Chinese speakers have two different phonological the printed stimuli had to be named in Hebrew, representations stored in memory for one and such cross-script homophones were responded to the same written character. This state of affairs slower when their English pronunciations mis- gives rise to the question whether word recogni- matched the color name to be produced than tion in bi-dialectic Chinese speakers involves the when they matched with the color response. simultaneous computation of two pronunciations This Stroop effect suggests that the nonsensical of one and the same character, just as the strings were automatically coded according to the Serbo-Croatian study has shown simultaneous English sound–spelling rules, causing interference phonological coding of one and the same letter in the case of a mismatch between the emerging string according to the rule system of two alpha- betic languages.

4. COMPREHENSION PROCESSES 191 Lam, Perfetti, and Bell (1991) addressed this and DS stimuli (“same” in both cases) should not question in a study that is conceptually similar differ from each other. to the Thierry and Wu (2004) study with which I started this section. They compared the per- These predictions were all borne out by the formance of bi-dialectal Chinese speakers with data and the authors concluded that a Chinese Cantonese as their native and strongest dialect character simultaneously activates the two associ- and Mandarin as their weaker dialect with ated phonological forms in bi-dialectal speakers. the performance of Mandarin mono-dialectal The similarity with the bilingual case is striking Chinese speakers on a same–different decision and warrants an explanation of the analogous task: The participants were presented with pairs bilingual and bi-dialectal effects within one and of Chinese characters and had to decide for each the same theoretical framework. Attributing all pair whether, within a given dialect, the two had of these effects to one and the same source is even the same pronunciation. The bi-dialectal speakers more compelling considering a final result that received the materials in two blocks, one with Lam and his colleagues obtained: An asymmetry Cantonese as the target dialect, the second was observed in the bi-dialectal speakers with Mandarin as the target dialect, and were such that the interfering effects of non-target instructed to focus on the target dialect Cantonese (the stronger dialect) on target exclusively. Obviously, the mono-dialectal group Mandarin were considerably larger than the inter- only performed the task in Mandarin. Four fering effects of non-target Mandarin (the weaker groups of materials were tested. dialect) on target Cantonese. This finding will remind the attentive reader that the analogous In Group 1 (DD) the two characters presented asymmetry has repeatedly been obtained in on a trial had different pronunciations in both bilingual studies. Considering the evidence Cantonese and Mandarin; in Group 2 their pro- assembled from the variety of sources so far it nunciations were the same within both dialects seems safe to conclude that a stronger language (SS); in Group 3 their pronunciations were the invades a weaker language more radically than a same in Cantonese but differed in Mandarin weaker language penetrates the stronger and that (SD); finally, in Group 4 they were different in the same holds for a pair of dialects of different Cantonese but the same in Mandarin (DS). If strengths. pronunciations are computed simultaneously in the two dialects, so the authors reasoned, To summarize, the studies discussed in this character pairs of the SD and DS types should section point out that reading by different- lead to conflict in the bi-dialectal speakers alphabet bilinguals and bi-dialectal language because the computation according to one dialect users involves the same automatic phonological will lead to a “same” response whereas com- activation across languages as observed in same- putation according to the other dialect will lead alphabet bilingualism. to a “different” response. This response conflict will have to be resolved and, therefore, lead to a Language-nonselective phonological relatively slow response, as compared to the activation in spoken language same target response (“same” or “different”) in comprehension the corresponding non-conflict condition (SS or DD). In addition, the response conflict might lead In all studies on parallel phonological activation to relatively many errors. In contrast, because in two languages discussed so far, printed stimu- such a response conflict would not exist for lus materials were used, reflecting the dominance the Mandarin mono-dialectal speakers, their of visual over auditory stimulus presentation in responses to SD and DD stimuli (in both cases research on word recognition, both monolingual “different” is the correct response) should not dif- and bilingual. Yet the primary focus of these fer from each other in terms of response times studies was on phonological activation in the and errors. Similarly, their responses to the SS bilingual’s mental lexicon upon the presentation of a word and not on the pattern of activation

192 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS established in memory units that represent of a set of orthographic word representations in orthography. The pattern of phonological acti- the mental lexicon that all share orthography vation in the lexicon was thus established with the target word (pp. 181–182), a spoken word indirectly, via automatic phonological coding of initially activates a set of phonologically similar orthographic input. In addition to the evidence word candidates in memory. In studies on spoken of language-nonselective phonological activation word recognition this set of activated representa- thus gathered, converging evidence has been tions is often called the “cohort”, Marslen- obtained from a set of studies on spoken lan- Wilson’s cohort model of auditory word recogni- guage comprehension, in which activation of tion (1987) being the eponym of the activated set. phonological memory units was established directly by a speech input. Marian and Spivey employed the eye- movement tracking paradigm in three bilingual A sizable number of these studies have used studies to see whether the activated cohort also the eye-movement tracking paradigm that was first includes phonological representations of words in used by Tanenhaus, Spivey-Knowlton, Eberhard, the non-target language (Marian & Spivey, 2003a, and Sedivy (1995) to study the process of on-line 2003b; Spivey & Marian, 1999). The participants spoken language comprehension. In experiments in these studies all had Russian as (stronger) L1 employing this methodology the participants are and English as (weaker) L2 and were residents of presented with aural instructions to manipulate an English-speaking country. They sat in front real objects on a visual display. For example, they of a board with four objects on it, among them, might be asked to “touch the starred yellow for instance, a stamp. They received spoken square” when the display holds a number of instructions such as (in the L1 Russian condition) blocks differing in marking, color, and shape, or the Russian equivalent of put the stamp below the they might be given the instruction to “put the cross: poloji marku nije krestika (where marku five of hearts that is below the eight of clubs is an inflected form of marka, which is Russian for above the three of diamonds” when the display is stamp). In a corresponding L2 English condition composed of seven miniature playing cards the instructions were in English. In a between- including two fives of hearts. The participants language competitor condition, in addition to a wear a head-band with equipment that tracks stamp, the board carried an object whose name in their eye movements while they carry out the the non-target language shares word-initial instructions. Using this technique Tanenhaus and phonology with the target word. For instance, his colleagues gathered fine-grained temporal with Russian marku as the critical stimulus word, information about the process of spoken lan- one of the distracter objects was a marker. The guage comprehension on a millisecond time scale. names of the remaining two (filler) objects on the They found, for instance, that visual context board were dissimilar from both the Russian and affects the resolution of temporary syntactic English names of the target object. This situation and lexical ambiguities. As an illustration of the is visualized in Figure 4.8. latter situation, consider the situation wherein the visual display contained both a candy and a In a control condition, none of the three non- candle and the participants were asked to “pick target objects’ names was similar to the name of up the candy and put it above the fork”. The the target object in either language. Marian and mean time to initiate an eye movement to the Spivey (2003a, 2003b) also included a within- target candy turned out to be longer in this case language competitor condition, with the target than when the display did not contain a distracter word and its competitor in one and the same object whose name shared initial phonology with language (as in Tanenhaus et al.’s, 1995, original the target (candle), thus suggesting that the visual experiment; e.g., target: marker; within-language context can influence the process of word recogni- competitor: marble, in an L2 English condition). tion. Furthermore, the example suggests that, just The dependent variable was always the pro- as a printed word initially gives rise to activation portion of trials on which eye movements were made to the competitor object in comparison to

4. COMPREHENSION PROCESSES 193 Example display presented to Russian–English bilinguals in an eye-movement tracking study. On this trial the participants receive the spoken instruction poloji marku nije krestika (put the stamp below the cross). Russian marku (“stamp”) is the target stimulus, which is similar to English marker, the between-language competitor. Adapted from Marian and Spivey (2003a) with permission from Wiley- Blackwell. a filler object. Such eye movements suggest co- both within-language and between-language activation of this object’s name during the word competitor effects: They made more eye move- recognition process. ments to within- and between-language com- petitors than to filler objects. Control groups of The three studies differed from one another monolingual English speakers showed competitor with respect to the care that was taken to prevent effects in the within-language competitor con- the participants becoming aware they partici- dition only, ruling out the possibility that the pated in a bilingual experiment. In Spivey and between-language effects in the bilinguals might Marian (1999) and Marian and Spivey (2003a) have been an artifact of poor stimulus selection. the same bilingual participants took part in both The within-language effects converged with a Russian L1 and an English L2 condition, in two Tanenhaus et al.’s (1995) results and with other separate sessions. Therefore it is plausible they evidence that during speech recognition a word suspected their bilingualism was being studied. input activates not only its own lexical representa- As especially Grosjean (e.g., 1997a, 1998; see tion but also a whole set of lexical items similar to pp. 288–291 for details) has argued, this might the presented word (e.g., Marslen-Wilson, 1987). increase the level of activation of the non-target The between-language effects demonstrated that language and, thus, the chance that this language the cohort of activated lexical representations causes interference. Therefore, to be able to included elements of both the target language and conclude that the recognition of spoken words the non-target language. This finding indicates is generally language-nonselective, evidence of that, as visual word recognition, spoken word cross-language competitor activation should recognition can be language-nonselective. ideally be collected under testing circumstances in which the participants have no idea they are par- However, the between-language effects varied ticipating in a bilingual experiment; that is, between studies. Whereas Marian and Spivey when the participants are in a monolingual mode. (2003a) and Spivey and Marian (1999) observed Marian and Spivey (2003b) installed such a con- them in both languages, in the monolingual-mode dition by testing different groups of participants, condition (Marian & Spivey, 2003b) they only drawn from the same population, in an English materialized when the task was carried out in L2 L2 and a Russian L1 experiment. English. In other words, competitor interference was obtained from the stronger L1 onto the In all three studies, the bilinguals showed

194 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS weaker L2, but not from weaker L2 onto stronger The studies discussed in this section so far all L1, an asymmetry that we have also encountered employed the eye-tracking paradigm, but con- when discussing the evidence for language- verging evidence has been obtained in two studies nonselective processing of written words (e.g., that used the gating technique. Grosjean (1988) Jared & Kroll, 2001; see pp. 184–185). That this employed it in a French–English study, testing asymmetry is a robust phenomenon, also when the participants in L2 English only. Schulpen, speech input is processed, is suggested by the Dijkstra, Schriefers, and Hasper (2003) tested results of a couple of further studies that closely Dutch–English bilinguals in both language con- replicated Marian and Spivey’s (2003b) study. ditions, mixing the languages during presentation Weber and Cutler (2004) tested Dutch–English of the materials (thus having the participants bilinguals using the same eye-tracking paradigm perform the task under bilingual mode circum- and also taking care the participants had no rea- stances). In both studies, gating performance son to suspect their bilingualism was being tested. to interlexical homophones and matched non- Under these circumstances, these authors also homophonic controls was compared (see p. 159). obtained cross-language competitor effects when Grosjean’s homophones and controls were verbs the participants performed the task in weaker L2 presented in an all English sentence context, (English), but not when testing was in stronger L1 whereas the stimuli used by Schulpen and col- (Dutch). Similarly, Blumenfeld and Marian leagues were all nouns presented in isolation. (2007), obtained a cross-language competitor Despite the differences between these two studies, effect when German–English bilinguals (with their results converged: In both of them success- German the stronger language of the two) per- ful identification of interlexical homophones formed the task in L2 English but not when required a longer fragment to be presented on a English–German bilinguals (with German the gating trial than successful identification of non- weaker language of the two) performed it in L1 homophonic control words. Furthermore, the English. It thus appears that speech perception in participants were less confident in their identifica- the stronger L1 may be immune to an influence of tion of homophones. Schulpen and collaborators a weaker L2 but not vice versa, and that this obtained these effects in both language con- asymmetry especially holds when the participants ditions. Grosjean’s findings indicate that a homo- have no reason to believe their bilingualism is phone’s reading in the stronger non-target being tested. Interestingly, in Blumenfeld and language was co-activated with its reading in the Marian’s study, immunity of the strong L1 only weaker target language and that this impeded held for L1 English targets that have non-cognate the target’s identification. Schulpen et al.’s results translations in weaker L2 German. L1 English point out that in their study the reverse also held: targets with cognate translations in L2 German that a homophone’s reading in the weaker non- did show a cross-language competitor effect (in target language was co-activated with its reading the condition with L2 English targets both cog- in the stronger target language and interfered nates and non-cognates manifested the effect). with recognition. At first sight this finding seems This finding suggests that the presentation of a not to be in agreement with the data obtained in cognate boosts the activation of a relatively weak the above eye-tracking studies, which had only non-target language to a sufficiently high level for shown competitor interference from stronger L1 it to interfere with target processing. onto weaker L2. A plausible reason for this deviant result is that in this study the L1 and Earlier in this chapter (p. 170) I briefly L2 stimulus words were presented mixed. This explained why it is desirable that a particular procedure may have put the participants in a theoretical issue, here whether the recognition of bilingual mode, boosting the activation of the spoken words is language-selective or language- weaker language to a sufficiently high level for it nonselective, is investigated by means of a “con- to interfere with processing the words from the verging evidence” approach in which evidence is stronger language. collected using more than one experimental task.

4. COMPREHENSION PROCESSES 195 All in all, the results of the above studies con- of non-native speech learning which assume that verge on the conclusion that spoken L2 words non-native sounds that are similar to native (when L2 is the weaker language) give rise to sounds are assimilated to the latter’s phoneme language-nonselective phonological activation in categories. This idea is advanced both in the bilingual lexical memory. The eye-tracking data perceptual assimilation model of Best and her suggest that the opposite does not hold; that is, colleagues (Best, 1994; Best et al., 1988) and in the that spoken L1 words (with L1 the stronger lan- speech learning model of Flege and collaborators guage) do not activate the phonological represen- (e.g., Flege 2002; Flege, Schirru, & MacKay, tations of similarly sounding L2 words. Finally, 2003; see also pp. 22–23 and 271–274). the gating data are not altogether conclusive due to a procedural infelicity contained in one of the Because phoneme identification is a waystation pertinent studies. Nevertheless it seems legitimate en route to word recognition in the aural modality to conclude that their results are compatible with (see Figure 4.9, to be discussed later), such those obtained by means of the eye-tracking deficiencies in the representations of non-native methodology. speech sounds and the assimilation of non-native sounds to native phoneme categories may be Spurious activation in non-native listening expected to affect the way non-native speakers recognize words. And indeed they do, as has been To conclude this section, one related further demonstrated in a variety of ways. Specifically, line of research deserves to be discussed here. It they have been shown to result in an excessive demonstrates that there is more lexical com- amount of spurious lexical activation in the non- petition in the bilingual system than in the mono- native lexicon. The experimental materials in the lingual system from a different angle than the relevant studies were always based on sound con- studies reviewed above. The studies in question trasts that are hard to distinguish by non-native (e.g., Broersma, 2002, 2005, 2006; Escudero, speakers. An example is the contrast between /æ/ 2007; Weber & Cutler, 2004) build on the (as in accident and flash) and /ε/ (as in execute observation that non-native listeners of a lan- and flesh), which are not represented separately guage often experience problems perceiving the in Dutch phonetic space. In one experiment speech sounds of this language accurately and (Broersma, 2002, 2005) native and non-native confuse non-native sounds that are not distinctive speakers of English (the non-native speakers in their native language. Probably the best-known being Dutch–English bilinguals highly fluent in example is the confusability of /r/ and /l/ by English) listened to monosyllabic English words native speakers of Chinese and Japanese with the and nonwords and had to make a lexical decision effect that they cannot tell the difference between, to each of them. All words contained one vowel for instance, English wrist and list. A plausible or one word-final consonant known to be confus- cause of these perceptual problems is that the able with one particular other vowel or conson- information stored in the associated non-native ant, respectively, by Dutch native speakers (e.g., it phoneme categories in memory is imprecise or contained /æ/ or /ε/). The nonwords were all incomplete, or perhaps that the two sounds of a “near-words”, constructed from the real words by confusable pair are not represented in separate replacing the critical vowel or consonant by its phoneme categories in memory but fused into a confusable counterpart. The results were clear: As single category. What is more, the one phoneme compared to native speakers, non-native listeners category onto which both non-native sounds are mistook relatively many of the near-words for mapped does not even have to be a non-native- words, suggesting that the near-words activated specific category but can be an extant native the representations of the corresponding actual category that represents native speech sounds words beyond their recognition threshold. similar to both elements of the confusable non- native pair. This possibility follows from models Two further studies (Escudero, 2007; Weber & Cutler, 2004), both testing highly proficient Dutch–English bilinguals in their non-native

196 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS English, employed the eye-tracking paradigm dition. The aural primes were subsequently con- discussed above. On each trial in the Weber and structed by, first, recording all of the selected Cutler study the participants heard a word that words as spoken by a male native speaker of they had to match to the corresponding picture, British English and, next, excising the first part presented together with a set of distracter of each recorded word up to and including the pictures. While performing this task, the par- vowel of the second syllable. (In addition, filler ticipants fixated their eyes longer and more often word-target materials as well as nonword-target on pictures whose names only differed from the materials were constructed, to be ignored in the targeted pictures’ names in that they contained a ensuing discussion.) During data gathering, each confusable sound (e.g., target: a picture of cattle; critical target could be preceded by an Identity competitor: a picture of a kettle) than on pictures prime, a Mismatch prime, or a Control prime. An with less confusable names (e.g., a picture of a Identity prime was the sound segment excised beetle). Native English speakers did not show from the target itself (e.g., prime: acci, from this effect of distracter type. Escudero (2007) accident; target: accident). A Mismatch prime extended these findings to the recognition of was the sound segment excised from the word previously unknown non-native words. She first the target had been paired with (e.g., prime: exe, had her participants learn the meanings of a set from execute; target: accident). A Control prime of nonsense English words such as /tεnzə/, was a sound segment excised from an unrelated /tændək/, and /tunzər/ by pairing them with word (e.g., prime: pove, from poverty; target: pictures of nonsense objects. This training phase accident). was followed by an eye-tracking phase: On each trial the participants heard one of the newly The author hypothesized that a Mismatch learned words (e.g., /tεnzə/) and their eye prime (exe) would activate the subsequent target fixations to the corresponding picture (the picture (accident) more in non-native speakers than of the object called /tεnzə/ before) and to dis- in native speakers, thereby facilitating target tracter pictures (the pictures showing the objects recognition to a relatively large extent. An called /tændək/ and /tunzər/) were registered. Identity prime was hypothesized to facilitate Again the pictures with confusable names target recognition equally in both groups. These (/tændək/) were fixated on relatively often. Both predictions were borne out by the data: As these findings suggest that non-native listeners compared to the unrelated control condition (and easily confuse spoken words that only differ in with proportion correct serving as the dependent one, confusable, sound (in word-initial position). variable), the non-native speakers showed a clear priming effect of Mismatch primes, whereas the In a final experiment (Broersma, 2005, 2006), native speakers showed no such effect. Further- fluent but non-native (Dutch) and native speakers more, the priming effect of Mismatch primes in of English performed a lexical decision task on the non-native listeners was equally large as the visually presented trisyllabic words (and non- priming effect of Identity primes in both listener words) that were all preceded by an auditory groups. These data thus suggest that non-native prime. During the construction of the materials speech input causes a relatively large amount of each critical trisyllabic word was paired with a spurious activation in the non-native lexicon: In second one with which it shared first-syllable non-native listeners the cohort of lexical repre- stress. Furthermore, up to and including the sentations activated by a speech fragment is larger vowel of the second syllable the sounds of the two than the cohort that is activated by this same words in a pair were identical except that the speech fragment in native speakers (here, for first vowel of one was /æ/ whereas in the other it instance, the speech fragment exe activated both was /ε/ (e.g., accident versus execute). For each the representations of words starting with exe word pair a phonologically and semantically and acci in non-native listeners but only those of unrelated trisyllabic word was also selected, words starting with exe in native listeners). Cutler which was to serve as target in a control con- (2005) and Broersma (2006) argued that already

4. COMPREHENSION PROCESSES 197 this one ambiguity of /æ/ and /ε/ in Dutch of lexical items that all share phonology with this non-native speakers of English may increase the spoken word and that, if the presented word is number of (English only) competitors during from weaker L2, the activated cohort includes the lexical selection process by tens, maybe even words from both of the bilingual’s two languages. hundreds, on average. Given the fact that more If the presented word is from stronger L1 the than one of these ambiguities may exist within a same only holds under specific circumstances particular non-native language, it becomes clear or for a specific type of stimulus materials that “the increase of lexical activation may be (cognates). very large in non-native listening” (Broersma, 2006, p. 1522). Because, as we have seen in the Earlier on I introduced SOPHIA (Figure 4.4), first part of this section, the lexical cohorts the extended version of the bilingual interactive formed during non-native listening also include activation model that includes phonological lexical representations belonging to the other lan- representations of different sublexical word units guage, the cohort sizes during L2 processing may (phonetic features, phonemes, phoneme clusters) be truly impressive. and of whole words (phonological words). Upon the presentation of visual word input, the corre- But by far the most noteworthy—because sponding orthographic representations of both counterintuitive—conclusion to draw from the languages are activated first. These then transmit present work is that, apparently, despite the fact activation not only upwards to the next level of that the non-native lexicon contains fewer lexical orthographic representations but also laterally to elements overall than the native lexicon, more the phonological nodes they are connected with lexical elements compete for selection in non- and further upwards (as well as downwards native than in native listening. (Incidentally, the again) until the lexical competition process has same might hold for non-native language pro- been resolved and the input word is recognized. duction, as the following anecdotal evidence Conversely, when auditory input is presented, the suggests: While working on this book, a zillion corresponding phonological representations of times writing errors such as “. . . this hypothesis both languages are first activated. These then was threatened [instead of strengthened]” and “. . . transmit activation laterally to the orthographic greater then [for than]” miraculously emerged on representations as well as further up to the next the screen.) level of phonological nodes. The model has not yet been tested against bilingual behavioral data. Explaining language-nonselective Nevertheless, given the components that it con- phonological activation effects tains and the way these are interconnected, it has the potential to successfully account for both The combined studies reviewed in this section so the evidence of orthographic and phonological far provide a substantial amount of evidence that, language-nonselective activation. The biggest when presented with a written word, under many challenge will be to also accurately model the cir- circumstances (but not always) bilinguals apply cumstances under which language-nonselectivity the spelling-to-sound conversion rules of both does not hold, such as has been shown often to be languages in parallel; that the phonological codes the case when the target stimuli are words from generated via one of these sets of rules contact stronger L1. similar sounding representations in the other language; and that generating such phonological A second model that may ultimately success- codes may occur automatically in the non-target fully account for parallel phonological activation language (as in the target language) even under in a bilingual’s two languages is BIMOLA, the circumstances in which the participants are led bilingual interactive model of lexical access to believe it is a monolingual study they are par- (Grosjean, 1997a; Léwy & Grosjean, 1996; see ticipating in. Furthermore, the data show that the Figure 4.9). Whereas BIA (Figure 4.4) was presentation of a spoken word activates a cohort originally developed as a computational model of written word recognition, BIMOLA was

198 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS The bilingual interactive shows many similarities between BIA (and model of lexical access SOPHIA) and BIMOLA, such as the core (BIMOLA), a model of spoken assumption of interactivity: Activated nodes at word recognition in different levels can activate (via the excitatory bilinguals. Arrowheads connections) and inhibit (via the inhibitory represent excitatory connections) nodes at neighboring levels. Fur- connections. Bullet heads thermore, within-level lateral inhibition occurs in represent inhibitory both models. But there are also crucial differences connections. From Léwy between the models. For instance, whereas in BIA and Grosjean (1996). and SOPHIA the lateral inhibition exerted by a Adapted from Grosjean word affects words from the same and the other (1997b). With kind language, BIMOLA restricts lateral inhibition to permission by John units (phonemes, words) of the other language Benjamins Publishing (between subsets). Maybe more importantly, Company, Amsterdam/ BIMOLA does not assume language nodes on the Philadelphia. grounds that “There is no empirical evidence such www.benjamins.com developed to account for spoken word recogni- tion from the outset. Whereas BIA extended an earlier monolingual model of visual word recognition (the interactive activation model of McClelland & Rumelhart, 1981), BIMOLA was modeled on TRACE, a computational model of monolingual spoken word recognition (McClelland & Elman, 1986). BIMOLA represents the auditory features, phonemes and the spoken forms of words in three layers of nodes. The feature-level nodes are shared between the two languages, whereas the nodes representing phonemes and word forms are organized in separate language subsets that are both part of a larger system (e.g., Paradis, 2004). A comparison of Figure 4.4 and Figure 4.9

4. COMPREHENSION PROCESSES 199 a node exists, nor do we know how this node is effects. What is furthermore in favor of BIMOLA created when a new language is learned” (Léwy & is that it can do without language nodes, which Grosjean, 1996). An important function of the one cannot help thinking of as being rather arti- language nodes in BIA (but not in SOPHIA, nor ficial constructs. in BIA+; Dijkstra & Van Heuven, 2002) was to suppress the activation of the non-target lan- THE PROCESSING OF COGNATES guage. BIMOLA assumes other ways to see to it that the targeted language is activated most: Bilingual studies The two word subsets receive a different amount of top-down pre-activation, based on external The studies on parallel activation in bilingual information that specifies what the base language memory that I discussed in the previous sections is (see pp. 289–290 for details). Furthermore, once all focused on the effect of various types of a word or a phoneme in a subset becomes acti- between-language word form similarity on pro- vated it sends a small positive signal to the other cessing. Pairs of interlexical homographs share words and phonemes, respectively, in the same their complete orthographic form between lan- subset. This way the subsets corresponding to the guages. Interlexical orthographic neighbors target language will be activated relatively highly. share most, but not all, of their orthographic form with their counterparts in the other lan- Given this set up, an auditory word input will guage. Interlexical homophones and homophonic give rise to activation of phonological representa- neighbors share phonology between languages. If tions in both languages, first at the language- the two languages in question use the same independent level of the phonetic features and alphabetic script, interlexical homophones and from there higher up in the system. Two of the homophonic neighbors also share orthography effects the model has successfully simulated (and are, by implication, interlexical orthographic so far suggest that it can indeed account for neighbors at the same time). The words in none language-nonselective phonological activation: of these types of cross-language word pairs share the “unit similarity effect” and the “base language any meaning between the languages. homophone effect” (Grosjean, 1988; Léwy & Grosjean, 1996). The unit similarity effect is the A further type of between-language word rela- phenomenon that a unit, for instance a phoneme, tion is the cognate relation, where a pair of L1 presented in one language and sharing properties and L2 words share both form and meaning. In with a unit in the other language, will co-activate other words, a pair of cognates constitutes a pair the latter. The base language homophone effect of translations with identical or similar orthog- is the phenomenon that words from the other raphy and/or phonology. I have already treated language (the guest language; see pp. 289–290) cognates at length in Chapter 3 (pp. 119–120), that have close homophones in the base language discussing the finding that foreign words sharing are more difficult to process than other guest a cognate relation with the corresponding L1 language words. words are easier to learn than non-cognate for- eign words. Some related findings regarding To conclude, both SOPHIA and BIMOLA cognates were also discussed there, such as the may ultimately be able to account for the present phenomenon that an unknown form that evidence of language-nonselective phonological resembles the form of an L1 word automatically activation. Because of its interconnectedness with triggers the latter. It may be obvious that the pro- the visual word recognition system and, hence, its cess underlying this phenomenon is the same as potential to explain more phenomena, SOPHIA the one underlying the above effects of interlexi- appears to be the more attractive of the two cal neighbors, homographs and homophones: In models. Nevertheless, because it was specifically all cases the effect is due to the stimulus exciting developed as a model for spoken word recogni- tion, BIMOLA seems the most natural candidate to explain language-nonselective phonological

200 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS (the representations of) a set of similar forms in nodes representing both terms in a cognate pair the mental lexicon. It will therefore not come as a and each of these nodes will transmit activation surprise that cognates have also been used to to the corresponding phonological nodes. When study word recognition in bilinguals. In this sec- the two terms in a cognate pair are phonologically tion I will review some of the pertinent studies. dissimilar, the activated phonological node Processing differences between cognates and non- corresponding to the non-target element of the cognates have also been the object of study in pair, preparing a response that mismatches the word production research that employed tasks response solicited for, will act as a nuisance com- like word association, word translation, and pic- petitor in the naming process. The consequence is ture naming. These will be reviewed in Chapter 5. a delayed naming response. We have already seen (p. 168) that the direction This analysis can also explain similar inhibi- of interlexical homograph effects (that is, whether tory effects in two further word naming studies interlexical homographs are processed faster or that were presented in previous sections: Jared slower than their non-homographic controls) and Szucs (2002) obtained longer naming varies with the specific demands of the task the latencies for interlexical homographs than for participants must perform. The word recognition non-homographic controls (p. 169), and Jared and studies that have compared the processing of Kroll (2001) obtained longer naming latencies for cognates and non-cognates suggest that this also target words that had an enemy neighbor in the holds for the “cognate effect”, the difference in non-target language (e.g. the English word bait processing time (and error percentages) for cog- having French fait as an enemy because the word nates on the one hand and non-cognates on the body –ait is pronounced differently in the two other hand. The far majority of these studies have languages see pp. 184–185) than for control shown a facilitating effect of cognates; namely, words. By definition, the phonological form of that bilinguals respond faster to cognates than to (the word body of) an enemy neighbor in the non- matched non-cognates (matched on, for instance, target language (fait) differs from that of the cor- word frequency, concreteness, and length). Across responding word in the target language (bait) and these studies a number of different tasks have also interlexical homographs will usually be pro- been used such as primed or unprimed lexical nounced differently across languages. The co- decision (Caramazza & Brones, 1979; De Groot activated homograph’s or neighbor’s phonology et al., 2002; De Groot & Nas, 1991; Gollan in the non-target language will therefore interfere et al., 1997), progressive demasking (Dijkstra et al., with processing and slow down the naming 1999), and semantic categorization (Sánchez- response. Only in cases where the homograph’s Casas et al., 1992). However, at least one study, an two pronunciations would happen to be identical English–Spanish word naming study by in the two languages, might a facilitating effect of Schwartz, Kroll, and Diaz (2007), has shown interlexical homography be expected to occur. longer latency times for cognates than for non- cognates. How can this deviant result be If this analysis of cognate inhibition in word understood? naming is correct, one might expect naming performance to improve if the competitor’s The answer to this question is likely to be phonology is more similar to the phonology of found in the specific requirement of the naming the target. Possibly even, if the two words in a task to pronounce the target word, a require- cognate pair are pronounced exactly the same, ment that does not hold for the other tasks. The naming cognates might be faster than naming previous section has provided much evidence that matched controls. The former prediction was phonological activation of visually presented borne out in Schwartz et al.’s (2007) English– words is language-nonselective. As a consequence Spanish study, in which the cognate effects were of its cross-language orthographic similarity, a examined in much more detail than suggested visually presented cognate stimulus will give rise above, the authors attempting to unravel the to strong activation of the orthographic word effects of degree of orthographic and phono-

4. COMPREHENSION PROCESSES 201 logical resemblance of cognates. The study In the high-constraint condition of that study (see included (in addition to non-cognates) four types Table 4.4) cognates and non-cognate controls of cognates, namely cognates with: identical were named equally fast, suggesting language- orthography and nearly identical phonology; selective processing. In contrast, in the low- identical orthography but more dissimilar constraint condition cognates were named faster phonology; relatively dissimilar orthography and than their non-cognate controls, a finding that nearly identical phonology; and both relatively suggests language-nonselective processing. The dissimilar orthography and phonology. Note joint homograph and cognate data suggest that that these differences between the four cognate the modulating effect of context on homograph groups in terms of orthographic and phonologic processing is more pervasive than on cognate similarity concerned differences in the degree of processing: With naming time as the dependent between-language resemblance. Because Spanish variable, the homograph effect disappeared in and English are both Roman-alphabetic lan- both highly and weakly constraining contexts, but guages, cognate pairs that share orthography but the cognate effect only disappeared in the highly at the same time have completely dissimilar constraining contexts but remained in the weakly phonology are unlikely to exist. A clear effect of constraining condition. Recently, Van Hell and degree of phonological similarity occurred: De Groot (2008) replicated Schwartz and Kroll’s Naming orthographically identical cognates with cognate results in an experiment in which Dutch– nearly identical phonological forms was faster English bilinguals performed a lexical decision and much more accurate, both in L1 and L2, than task instead of word naming. As in Schwartz and naming orthographically identical cognates with Kroll’s investigation, the stimulus materials were more dissimilar phonology. presented in L2 English. Cognates processed in isolation and in a low-constraint sentence context After discussing interlexical homograph effects were responded to faster than matched non- earlier (p. 168) and suggesting why the direction of cognates. In a high-constraint context condition these effects might differ between studies, I con- the cognate effect disappeared. cluded with the statement that the very occur- rence of such an effect, irrespective of its direction, In a further Dutch–English study, Duyck, Van constitutes evidence of language-nonselective Assche, Drieghe, and Hartsuiker (2007) similarly word recognition. Similarly, the present facilitat- obtained a cognate effect when cognates and ing and inhibitory cognate effects have all been non-cognate controls were presented in isolation interpreted to reflect language-nonselective or in a low-constraint sentence context and the bilingual word recognition or, in other words, co- participants had to provide lexical decision activation of elements in the nontarget language. responses to the (L2 English) target stimuli (no Later on (pp. 203–205) an alternative account of high-constraint sentence condition was included). cognate effects will be presented. But first some These authors split up their cognate materials words are in order on the effect of sentence con- into a group of identical cognates (e.g., lip–lip) text on cognate effects and, in a separate section, and a group of non-identical cognates (e.g., ship– on cognate processing in trilinguals. schip) and obtained a cognate advantage for both groups, but it was relatively large for the identical Earlier on we have seen that interlexical- cognates. A further Dutch–English experiment homograph effects can disappear if the homo- suggested there may be a limit to the cognate graphs and their controls are presented in a effect in low-constraint sentence contexts and, sentence context (Elston-Güttler et al., 2005; thus, to language-nonselective activation. This Schwartz & Kroll, 2006), suggesting that sentence experiment, with again L2 words as test stimuli, context prevents the activation of the homo- employed the eye-tracking methodology. This graph’s representation in the contextually technique allows relatively natural reading on the inappropriate language subsystem. Schwartz and part of the participants and does not require any Kroll also looked at the effect of sentence context secondary task such as lexical decision that on the processing of cognates and non-cognates.

202 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS possibly interferes with the normal reading pro- effects in low-constraint sentence contexts as cess. An analysis of first fixation duration, gaze evidence of language-nonselective word pro- duration, and regression path duration (see cessing. However, the differential results for p. 161) showed a robust cognate effect for identi- interlexical homographs on the one hand and cal cognates but none for non-identical cognates. cognates on the other hand may also be due Apparently, degree of between-language cognate to possible representational differences between similarity is a factor to be taken into account. interlexical homographs and cognates. These This conclusion is strengthened by the results will be discussed further on, following a discus- of a subsequent study (Van Assche, Duyck, sion of a few trilingual studies to which I will Hartsuiker, & Diependaele, 2009) in which the now turn. researchers again studied the processing of cognates and non-cognates in low-constraint Trilingual studies sentence contexts employing the eye-tracking technique, but this time the test was done in In one sense, the modifier “bilingual” concerns an (stronger) L1 Dutch. Even under these circum- imprecise qualification of the studies that we have stances a cognate effect materialized, suggesting considered so far in this chapter. The suggestion also that the more dominant language is not is that the participants in these studies had a immune to an influence from the other language command over exactly two languages, no more under all circumstances. The data furthermore and no less. However, many of them may have showed that first fixation duration, gaze duration, possessed at least some knowledge of more lan- and regression path duration all decreased with guages than just the two under study and actually an increasing cross-language overlap of the two use it in communication. Although it is in agree- terms in the cognate pairs, thus once again point- ment with common practice, with hindsight this ing at a role for degree of between-language cog- labeling convention is somewhat unfortunate, nate similarity. However, in a near replication of especially now that a study by Lemhöfer, this study (Van Assche, 2009), now including both Dijkstra, and Michel (2004) has shown that a high-constraint and a low-constraint sentence multiple non-target languages may have additive context condition, none of the reading measures effects on processing the target. These researchers showed a cognate effect, also not in the low- tested Dutch–English–German trilinguals in a constraint condition. Contrary to Van Assche et German (L3, the weakest language) lexical deci- al. (2009), this finding suggests that the dominant sion study. Two groups of cognate stimuli were native language is processed in a language- presented. One group consisted of German selective way. words that shared orthography with both Dutch and English (“triple” cognates); a second group To summarize, although the evidence is some- consisted of German words that only shared what mixed, the general picture that emerges orthography with their Dutch translation equiva- from the sentence context studies with interlexical lents (“double” cognates). The double cognates homographs or cognates as the critical stimuli were responded to faster than a group of matched is that both highly and weakly constraining German non-cognate control words. In their turn, sentence context prevents co-activation of an the triple cognates were processed faster than interlexical homograph’s representation in the the double cognates. Apparently, during L3 word non-target language but that only highly con- recognition both L1 and L2 were co-activated straining contextual information prevents co- with L3 and each of these two sources of co- activation of a cognate’s representation in the activation contributed separately to the cognate non-target language. Generally, a weakly con- effect. In terms of the authors’ vivid metaphor: straining sentence context does not nullify the Upon the presentation of a cognate word, the cognate effect, although it does seem to do three languages together provide one “echo” (as so under special circumstances. It is tempting the authors called the phenomenon) which speeds to regard the relatively persistent cognate

4. COMPREHENSION PROCESSES 203 up the response. It thus appears that all languages of these. The implicit assumption was that known join forces in generating a response, just as cognate and non-cognate pairs are represented in all languages known have earlier been shown bilingual memory in similar types of represen- to be exploited in acquiring foreign vocabulary tational structures, with each element of both (pp. 98–101). a cognate translation pair and a non-cognate translation pair being represented in a separate A second trilingual lexical decision study, by set of memory nodes (together representing Van Hell and Dijkstra (2002), makes the—by now orthography, phonology, and meaning). However, familiar—point that level of proficiency in a several authors have suggested an alternative to non-target language determines whether or not this account of the cognate effects in terms of co- and to what extent the latter exerts an effect on activation of representational units in the non- processing the target language. The participants target language. This alternative is that cognate were Dutch–English–French trilinguals with L1 effects result from qualitatively different repre- Dutch as their strongest language, English as sentations for cognate and non-cognate pairs. their strongest foreign language, and French as a An important reason to consider such an account weaker foreign language. Across two experiments, is that in one important respect cognates differ the participants’ proficiency in L1 Dutch and L2 from interlexical homographs and interlexical English was kept constant, but L3 French pro- neighbors: They share both form and meaning ficiency differed between experiments. A further between a bilingual’s two languages. Two versions noteworthy aspect of this study was that the of this alternative view have in fact been authors took great care to let the participants proposed. believe they were taking part in a monolingual experiment. Dutch words that have an English From the different patterns of cross-language cognate translation were consistently responded semantic priming effects occurring for cognates to faster than matched Dutch words with a and non-cognates, De Groot and Nas (1991) non-cognate English translation. In contrast, the suggested that cognates and non-cognates might analogous cognate effect of L3 French only be represented differently at the conceptual materialized when the participants’ fluency in (meaning) level of representation. Van Hell and French was beyond some minimal level. The De Groot (1998a) drew a similar conclusion from combined results show that also a weaker lan- an experiment that compared bilingual word guage may influence the processing of a dominant association behavior within a language and L1, but only if some minimal level of proficiency between languages. To account for the results of in it has been attained. These effects occurred these studies, two possibilities were considered, despite the fact the participants were not aware one in terms of localist representation models their trilingualism was being tested and thus have and a second in terms of distributed representa- no reason whatsoever to consciously activate the tion models. According to the localist view, a non-target languages. word meaning is represented in a single node in memory. According to the distributed view the Explaining cognate effects: An meaning of a word is spread out over a number of alternative account nodes, each of them representing one aspect of the word’s meaning (cf. Figures 3.6 and 3.8, and Up until this point, just like the effects of inter- Figures 3.11a and 3.11b; see pp. 132–133 and lexical homographs and interlexical neighbors, 139–140 for further details). In terms of the local- cognate effects were attributed to co-activation ist view, a pair of cognates may share a single of memory nodes representing elements of the conceptual representation in bilingual memory non-target language during target processing. whereas the two members of a pair of non- Depending on the exact nature of the materials, cognates might each have a separate conceptual the co-activated nodes may represent orthog- representation. In terms of the distributed view, raphy, phonology, or meaning, or a combination both cognates and non-cognates share a number

204 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS of conceptual nodes, but cognates share more of sisting of a single morpheme only. Sánchez-Casas them than non-cognates do. and García-Albea (2005) attempted to rectify this neglect by hypothesizing various possible loci for The Spanish–English and Catalan–Spanish an extra, morphological, representational level in work of Sánchez-Casas, García-Albea and their some of the existing models. They proposed to colleagues has provided maybe the most com- insert the extra morphological layer in between pelling support for the idea that cognates and the levels of word nodes and language nodes in non-cognates are represented in qualitatively BIA (Figure 4.4, left), or in between the levels of different representations in bilingual memory orthographic/phonological words and semantics and that this difference may underlie the above in SOPHIA (Figure 4.4, right). These proposals cognate effects (e.g., García-Albea, Sánchez- still await computational modeling. Casas, & Igoa, 1998; Sánchez-Casas et al., 1992; Sánchez-Casas & García-Albea, 2005). These That ultimately special attention will have to researchers set up a detailed research program in be paid to morphology in models of bilingual which they tried to pin down the separate contri- word recognition is dictated by the results of two butions of form and meaning similarity between empirical studies that looked at the influence of primes and targets to the masked priming effects morphological complexity of words in lexical that are reliably obtained when a target stimulus decision by monolinguals and bilinguals: is preceded by its, masked, cognate translation Lehtonen and Laine (2003) found that Finnish– (e.g., masked prime: rich; target: rico). This Swedish bilinguals decomposed many more mor- work suggested that neither form nor meaning phologically complex Finnish words into their similarity on its own suffices for the priming effect constituents during recognition than Finnish to occur. Interestingly, it also demonstrated that monolinguals did. Lehtonen, Niska, Wande, the effects obtained for cognate translations Niemi, and Laine (2006) observed a similar ten- strongly resemble those obtained in masked dency when comparing the performance of priming studies in which morphologically related Swedish monolinguals and Swedish–Finnish words from one and the same language are pre- bilinguals on a Swedish lexical decision task, but sented as prime and target. These combined here the difference between bilinguals and mono- findings led the authors to conclude that mor- linguals was relatively small. Furthermore, the phology determines the way words are stored in performance of the monolinguals differed both monolingual and bilingual memory. More between the two languages: Finnish monolinguals specifically, they concluded, in line with other decomposed more words than Swedish monolin- authors before them (Cristoffanini, Kirsner, & guals did. Milech, 1986; Kirsner, Lalor, & Hird, 1993), that cognate translations share a morphological These differences between the two studies are representation in bilingual memory whereas plausibly due to a difference between these two each member of a pair of non-cognate transla- languages in morphological richness: In Finnish a tions has a separate morphological representation noun can have up to 2000 different inflectional in bilingual memory. forms, whereas Swedish is a morphologically rela- tively limited language. Degree of morphological Unlike in models of monolingual word recog- richness has consequences for the way words nition (see e.g., Taft, 1994; Taft & Forster, 1975; are stored in the mental lexicon and processed: Taft, Hambly, & Kinoshita, 1986), in models If a word has many different inflected forms (such of bilingual word recognition morphology is as is the case for Finnish) the chance that each typically the Cinderella among the various types single one of them has developed a separate of linguistic domains that are distinguished. In representation in the mental lexicon is consider- the models discussed in the previous sections so ably smaller than if a word has relatively few dif- far, morphology is blatantly absent, possibly ferent inflected forms. Without such a “full-form” because the computational modeling of bilingual representation in memory for a specific word, rec- word recognition has been based on words con- ognition will involve a decomposition process in

4. COMPREHENSION PROCESSES 205 which the word is first analyzed into its constitu- such separate representations may exist but may ents, which can then be recognized. Full-form not be the (sole) source of the cognate effect. As representations develop over word recognition we shall see in Chapter 5, cognate effects in speech practice. Because the, fewer, morphological inflec- production tasks are equivocal for similar tions per Swedish word have each been reasons. encountered far more often than each of the much larger number of Finnish inflected forms, Summary and conclusions Swedish monolinguals will have relatively many full forms stored in the mental lexicon; hence the The above discussion of cognate effects in difference in frequency of decomposition between bilingual word recognition concludes this review Swedish and Finnish monolinguals. The differ- of studies addressing the question of whether ence between the Finnish monolinguals and bilingual lexical access is language-nonselective Finnish–Swedish bilinguals performing in or language-selective. And what, from a bird’s Finnish can be understood by realizing that the eye view, can we conclude on the basis of the monolinguals, with a large amount of practice in assembled evidence? As we have seen, in the just one language, Finnish, have had more majority of studies the experimental materials opportunity to develop full-form representations consisted of lists of critical words intermingled of morphologically complex words than the Finn- with control words. The joint evidence collected ish–Swedish bilinguals. The analogous, although in these out-of-context studies clearly points smaller, difference between Swedish monolin- to an influence of the non-target language on guals and Swedish–Finnish bilinguals performing target word processing, suggesting co-activation in Swedish can be explained in the same way. The of elements of the non-target language and, reason it is relatively small plausibly is that Swed- thus, language-nonselective processing. The evi- ish, because it is morphologically limited, has also dence also suggests that the effect of the non- given the bilinguals the opportunity to develop target language is modulated by the relative many full-form representations. strength of the bilingual’s two languages: The influence of the non-target language was espe- But let us return to the cognate effects in cially strong when it was the stronger language relation to representational structure that started of the two. When the non-target language was this discussion: Whatever the exact locus of the weaker language, smaller effects or even qualitatively different cognate and non-cognate null effects on target processing materialized. representations in bilingual memory—a yet to be Boosting the activation of the non-target lan- included morphological layer in the system, a guage, for instance, by presenting a block of semantic level, or any other level yet unspeci- words from this language before a critical block fied—the important message to bear in mind is of test materials, increased its influence on that representational differences between cognates processing the target language. Under those and non-cognates may underlie the apparent circumstances even the stronger language turned manifestations of language-nonselective pro- out not to be immune against an influence from cessing. The notion of language-nonselective a weaker non-target language. All in all it appears processing implies the co-activation of one or legitimate to conclude that the vast majority of more separate language-specific memory nodes the out-of-context studies suggest language- belonging to the non-target language while pro- nonselective word processing. In fact, even the cessing a target language stimulus. The crucial apparent counterevidence, the observed null message to be communicated here is that, effects of a relatively weak non-target language, plausibly, when a cognate is processed, no such does not challenge this conclusion because these co-activation is involved because the assumed null effects do not necessarily indicate that in two, separate, language-specific representations these cases no elements in the non-target lan- of the two terms in the implicated cognate trans- guage were activated at all. An alternative lation pair in fact do not exist. Alternatively,

206 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS account is that there was co-activation in this SENTENCE PROCESSING language subsystem but that it was simply too weak to affect processing the stronger target Introduction language. The studies discussed so far in this chapter all In contrast to the out-of-context studies, in focused on word recognition in bilinguals. The the few studies that embedded the critical few studies in which sentence materials instead of materials in context and in which the language of isolated words were presented to the participants context and critical materials was the same one, also had this focus on word recognition rather no influence of the non-target language occurred than on the way whole sentences are processed: except in some of the cognate studies that, as we Their goal was to find out how contextual infor- have seen, must be regarded with caution. This mation affects word recognition in bilinguals. In result is especially important because these con- this section I will present a set of studies that were text studies resemble natural language processing designed to provide information on the way more closely than the out-of-context studies do. bilinguals process sentences, and especially on Up until this point in the discussion an influence how they parse them. Parsing is the technical of the non-target language on target language name for uncovering the grammatical structure of processing was assumed to result from co- a sentence, identifying its constituent parts as sub- activation in the non-target language subsystem ject, verb, object, and so on. As mentioned before, and, therefore, to index language-nonselective word recognition and parsing together enable the language processing. Analogously, the most nat- language user to come up with the literal meaning ural interpretation of a non-influence of the non- of a sentence. target language in the context studies is that the latter is not activated during target processing As was the case for the above studies on and, thus, that this null effect indexes true bilingual lexical access, the experimental methods language-selective processing. Yet, as we have just used in the bilingual parsing studies have been seen, the interpretation of null effects may be adopted from the research on monolingual equivocal and we must therefore keep our minds language comprehension. In this particular open for the possibility that this null effect also case, the bilingual research community took does not index language-selective processing. An on the common practice of studying parsing alternative interpretation may be that, just like by looking at the way syntactically ambiguous words presented out of context, words presented or syntactically and/or semantically anomalous in context activate representations in both of the sentences are processed. Across these studies bilingual’s language subsystems but that the sentence processing has been measured with a activation in the non-target language’s sub- varied set of techniques, such as eye-movement system is somehow ignored by an attention sys- recording during self-paced reading (Frenck- tem that supervises task performance. The effect Mestre, 2002, 2005b; Frenck-Mestre & Pynte, is that target processing is not hampered by this 1997), the moving-window technique (Dussias, co-activation. This suggestion may seem com- 2006; Dussias & Cramer, 2006; Dussias & pletely ad hoc at this point, but in Chapter 5 I Cramer Scaltz, 2008), and ERP recording (e.g., will present the results of a couple of studies on Hahne & Friederici, 2001; see pp. 160–162 for a bilingual word production that can only be description of these techniques). In addition to understood if we accept the conclusion that this these on-line measures, off-line techniques have is in fact what happens in at least some been used, such as having participants indicate bilinguals some of the time. Furthermore, it is which one of the nouns in simple noun-verb- consistent with the recent addition of a control noun sentences fulfills the role of subject system to models of word recognition that I (McDonald, 1987; McDonald & Heilenman, introduced earlier (pp. 180–181) and to which I will return in later chapters.

4. COMPREHENSION PROCESSES 207 1992) or having them make grammaticality judg- A second type of syntactically ambiguous ments on legitimate and illegitimate sentences construction with which the syntax-first principle (Mack, 1986). can be illustrated is when the main verb permits both a direct-object complement and a sentence In reviews of the literature, Frenck-Mestre complement, as does the main verb admitted in (2005a) and Kroll and Dussias (2004) presented the following pair of sentences, borrowed from the various parsing models that have been Dussias (2006) and Dussias and Cramer (2006, proposed and the evidence for each of them. 2008): Serial syntax-first models (e.g., Frazier & Rayner, 1982) assume that the parser initially ignores 1a. The ticket agent admitted the mistake syntax-external sources of information such might not have been caught (sentence as contextual-semantic information, lexical complement) information, or information on the relative frequency of occurrence of particular sentence 1b. The ticket agent admitted the mistake constructions. These models assume that the when he got caught (direct object). parser initially automatically applies a universal parsing strategy that assigns the encountered sen- The information what types of constructions are tence the simplest structure possible on the basis enabled by a particular verb is known as the of syntactical information only (where degree of verb’s subcategorization frame and is assumed complexity of the emerging structure is deter- to be stored in the verb’s lexical representation. mined by the number of “nodes” required to Importantly, according to the syntax-first describe the sentence—the fewer, the simpler the account, Sentence 1b should be easier to process structure). Only later, if and when this strategy than Sentence 1a because its resolution involves encounters a deadlock, other sources of infor- the simpler syntactic structure. mation may be implicated in a reanalysis of the sentence. Earlier on in this chapter (p. 164) lexical ambi- guity effects were discussed in terms of Fodor’s To illustrate, consider the sentence-initial (1983) modularity of mind theory. The present fragment The student graded. . . . According to syntax-first account of syntactic ambiguity reso- syntax-first models, when graded is first lution agrees with this theory because it assumes encountered it is interpreted as the main verb of that initial syntactic analysis is impenetrable by the sentence and, accordingly, the parser would other sources of information than syntactical expect to encounter a direct object next, because information. In modularity of mind theory, this this analysis produces the simplest structure. The feature of “information encapsulation” is consequence of this privileged analysis is that regarded the main characteristic of a modular the reader should experience a problem when the process. sentence continues with . . . by the professor received an A, in which graded turns out not to be This modular syntax-first view is challenged the main verb but the verb of a reduced relative by monolingual studies suggesting that parsing clause (example taken from Frenck-Mestre, is permeable by other sources of information, 2005a). As compared to unambiguous control such as semantic information provided by the sentences, such temporarily ambiguous reduced sentence fragment preceding the ambiguity or by relative-clause constructions have shown to cause the subcategorization information stored in a problems to readers (e.g., Ferreira & Clifton, main verb’s lexical representation. To illustrate, 1986), as indicated by, for instance, relatively consider the following three sentences, one of long reading times or eye fixations in the dis- which (2a) was already introduced above. ambiguating area (by the professor). The process- ing delay is likely to be due to a repair process that 2a. The student graded by the professor needs to take place when the initial analysis turns received an A. out to be incorrect. 2b. The paper graded by the professor received an A.

208 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS 2c. The student graded the professor with exactly the same structure (see Dussias & an A. Cramer, 2006, for a discussion of the pertinent monolingual studies). Sentences 2a and 2b have an identical structure and, except for one lexical element, are identical 3. The CIA director confirmed the rumor to one another in all respects. Because of their could mean a security leak identical structure, syntax-first models predict that they are equally hard to process. Further- The important point to make is that this effect more, according to syntax-first models, both suggests that lexical information associated with should be more difficult than Sentence 2c, where the main verb is implicated in syntactic analysis, graded is the sentence’s main verb. But contrary a conclusion that defies the syntax-first view. to these predictions the evidence shows that The reason why different verbs favor different Sentence 2a is more difficult than Sentence 2b structural analyses presumably is that verbs differ (Frenck-Mestre, 2005a). The source of this dif- from one another in terms of the structures in ference must lie in the use of the lexical item which they are preferentially used: In naturally student in Sentence 2a versus paper in Sentence occurring English, admit is more often followed 2b, since this is the only difference between by a sentence complement than by a direct object the two sentences. The suggestion thus is that and the opposite holds for confirm. In other semantic lexical information is implicated in the words, the bias information in the verb’s lexical parsing process. More precisely, the [-animacy] representation is based on linguistic experience. feature of the lexical element paper seems to have reduced the likelihood that readers treat Still further sources of extra-syntactical this noun as the subject of graded. This renders a information, to be ignored here, have been shown main-clause interpretation (favored by syntax- to affect sentence parsing (see Frenck-Mestre, first models but only appropriate in Sentence 2c) 2005a, for a review). These data from mono- less likely in Sentence 2b than in Sentence 2a. lingual studies all suggest that sentence parsing Apparently, the semantic information provided is an interactive, non-modular, process in which by a lexical element that precedes the syntactic syntactic and other sources of information ambiguity can help resolve the ambiguity operate in concert to come up with the proper (Frenck-Mestre, 2005a). analysis of the sentence. Adherents of the syntax- first models have parried this conclusion by Other evidence to suggest that lexical informa- pointing out that these non-syntactic sources of tion interacts with syntactic information during information might not come into play until after sentence parsing comes from studies that have an unsuccessful first parse, during a second, used the ambiguity exemplified in Sentences 1a repair, stage. Evidence that under certain circum- and 1b above, where the main verb allows both stances syntactically simpler constructions are a direct-object complement and a sentence more difficult to process than syntactically more complement. The specific type of lexical infor- complex constructions would perhaps more mation to be exploited in this case concerns a seriously challenge syntax-first models. Such a bias provided by the main verb towards one or result has in fact been obtained in a monolingual the other type of complement. The preferred study in which direct-object constructions subcategorization frame of the verb admitted is were more difficult to process than sentence- the sentence complement, honored in Sentence complement structures despite the fact that 1a but violated in Sentence 2a. In contrast, the former concern the syntactically simpler other verbs, such as the verb confirm, prefer a structures (see Dussias & Cramer, 2006, for a direct-object complement to a sentence com- discussion). Subsequent bilingual studies have plement. This presumably is the reason why produced similar results. It is to a discussion of Sentence 3 is harder to process than Sentence some of the bilingual studies on sentence parsing 1a, despite the fact that the two sentences have that I will now turn.

4. COMPREHENSION PROCESSES 209 The resolution of syntactically Mean reading times for syntactically ambiguous verbs ambiguous sentences (in ms) as a function of verb bias and sentence construction type. For DO-biased verbs the direct-object A small number of bilingual studies on sentence complement is the preferred construction. For parsing have contributed to solving the above SC-biased verbs the preferred construction contains a theoretical controversy between syntax-first and sentence complement. Reprinted from Dussias and interactive-parsing models by trying to discover Cramer (2006) and Dussias and Cramer Scaltz (2008), how bilinguals resolve syntactic ambiguities. Copyright (2008), with permission from Elsevier. As dependent variables they have generally used behavioral measures obtained on-line by a direct-object construction, latencies were means of the moving-window technique or eye- shorter when the sentence indeed continued with movement registration. In addition to informing a direct object than when it proceeded with a the above controversy, the purpose of these stud- sentence complement, and the opposite pattern ies was to find out whether L2 speakers and native held for verbs that prefer a sentence complement. speakers use similar parsing strategies, what the This interaction was statistically significant. effects of L2 proficiency and the relationship These data suggest that a syntactically simpler between L1 and L2 on L2 parsing might be, structure (the direct-object construction) is not and whether preferred parsing strategies in a necessarily easier to parse than a more complex bilingual’s one language might modify the structure (the sentence-complement structure), a preferred strategies associated with the other finding that challenges the syntax-first models. language. A second group of bilingual studies The data also show that, as native speakers, L2 employed a popular off-line technique to deter- speakers can exploit lexical verb-bias knowledge mine how L2 speakers assign function (e.g., sub- during parsing. As we shall see next, this result ject, object) to the various noun phrases in a does not imply that it is exclusively L2 lexical sentence and, especially, whether the function- knowledge that they exploit while processing L2 assignment procedures they execute in their L1 sentences. transfer to L2. The focus in the current section will be on the former class of studies, but I will Although the above interaction between verb also present some representative results from the bias and sentence construction type was statistic- second class. ally significant, subsequent analyses showed that the difference between the two conditions with In a Spanish–English study, Dussias and sentence-complement biased verbs (the two Cramer (2006; Dussias, 2006) employed the self- right-hand bars in Figure 4.10) was not. Dussias paced moving-window technique (see p. 160). The and Cramer (2006) wondered why these verbs ambiguous constructions they studied were those exemplified in Sentences 1a, 1b, and 3 above, in which the main verb allows both a direct-object complement and a sentence complement but con- tains a bias towards one of these (cf. confirmed versus admitted). The time between the moment the disambiguating information (e.g., might not in Sentence 1a and could mean in Sentence 3) appeared on the screen and the participant pressed the space bar for the next window served as dependent variable. The test materials were all in the participants’ L2 English. The critical data of this experiment are presented in Figure 4.10. As shown, verb bias interacted with sentence construction type: When the main verb preferred

210 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS had not shown a clearer effect of sentence con- In this sentence either the head of the complex struction type and hypothesized the reason might noun phrase (el hijo; “the son”) or the second be an influence of the participants’ L1, Spanish, a noun phrase in this phrase (la actriz; “the language in which these verb biases are much less actress”) can be the subject of the relative clause. salient than in English. For instance, the Spanish Spanish favors the former (“N1 attachment”, translation equivalent of an English verb with a “high-attachment”, or “non-local attachment”) clear sentence complement preference may have analysis (Cuetos & Mitchell, 1988), whereas an equally strong bias towards a sentence English favors the latter (“N2 attachment”, “low- complement and a direct-object complement. If attachment”, or “local attachment”) analysis Spanish verb-bias knowledge was in fact co- (Frazier, 1987). In other words, Spanish readers activated with English verb-bias knowledge of Sentence 4 would typically regard el hijo as the during task performance, co-activated Spanish subject of the relative clause whereas English will once in a while have primed the direct-object readers of the English translation of this sentence construction instead of the targeted sentence would usually take the actress to be the subject of complement construction, thus reducing the the relative clause. An earlier study examining difference between the two construction-type how Spanish–English bilinguals and English– conditions. A subsequent study with improved Spanish bilinguals parse this type of structure in materials (Dussias, 2006; see also Dussias & both languages had shown that both participant Cramer Scaltz, 2008) supported this inter- groups generally favored low attachment over pretation and, in so doing, provided indirect high attachment, also when they processed evidence that when bilinguals parse L2 sentences, Spanish sentences (Dussias, 2003). All testing in L1 verb knowledge is co-activated and influences that study was done in a predominantly English- processing. speaking environment in the United States. The critical finding in the present study—which only As we have seen in the earlier parts of this tested L1 Spanish sentences—was that the par- chapter, not only does L1 influence L2 word ticipants immersed in L1 Spanish behaved like recognition under many circumstances, but the Spanish control participants, showing shorter occasionally the opposite—an influence of L2 on total fixation times for sentences in which the L1 word recognition—has been shown to occur as correct solution involved high attachment. In well. Dussias (2006) showed that an influence of contrast, the participants immersed in L2 English L2 on L1 parsing can also occur (and see Dussias showed shorter total fixation times for sentences & Sagarra, 2007, for a further demonstration of whose correct solution involved low attachment, this phenomenon). Two groups of Spanish– thus behaving like English monolinguals. This English bilinguals were included, one group that second result demonstrates an influence of L2 on was immersed in an English (L2) environment L1 parsing. In addition, the fact that the English and a second that was immersed in a Spanish (L1) and Spanish immersion environments produced environment. The participants read L1 Spanish different results suggests that extensive current sentences while their eye movements were being experience with one language in particular recorded. The performance of these two groups affects parsing solutions. It is therefore plausible of bilinguals was compared to that of Spanish that, similarly, the dominant low-attachment monolinguals. The critical sentences contained preference in the 2003 study had been caused by a relative clause ambiguity for which the favored the fact that all (L1 Spanish) participants were solution differs between Spanish and English. tested in an (L2) English-speaking environment. The construction in question is illustrated in the In other words, contextual linguistic and cultural following example: information appears to influence parsing. It may be obvious that this constitutes an additional 4. Alguien disparó contra el hijo de la actriz challenge for the syntax-first models of parsing. que estaba en el balcón (Someone shot the son of the actress who was on the balcony). In two further studies, Frenck-Mestre (2002,

4. COMPREHENSION PROCESSES 211 2005b) tested the same type of ambiguous struc- solutions for a given structure, L2 parsing strat- tures as illustrated in Sentence 4 (this time in egies may change with increasing L2 proficiency: L2 French) but looking at the effect of level of L2 Initially the learner analyzes the L2 structure in proficiency on L2 parsing and using the eye- the way he or she analyzes similar structures in his movement recording methodology. Across these L1. At a more advanced level the learner analyzes studies L2 proficiency was manipulated by com- it in the same way as native speakers do. paring the performance of three groups of late learners of French: beginning Spanish–French A corollary of this conclusion is that data bilinguals, beginning English–French bilinguals, that do not clearly point towards one preferred and proficient English–French bilinguals. A con- parsing strategy—a “native-language” or a trol group of French native speakers was also “second-language” strategy—may be due to the included and first-pass gaze duration on the dis- learner being in a transitional stage in between ambiguating region (see p. 161) served as the L1 preference and L2 preference. This is how dependent variable. As mentioned above, English Frenck-Mestre (2005b) explained the fact that in favors low attachment. French, however, favors a similar study with German–English and Greek– high attachment and we have seen this also English bilinguals (Felser, Roberts, Gross, & holds for Spanish. So the three groups of learners Marinis, 2003) no clear preference for either differed on two dimensions: L2 proficiency and strategy was observed. Alternatively, a parsing whether or not their L1 and L2 share the same pattern that does not clearly point towards the use attachment preference. of one preferred strategy may reflect the use of a special strategy by L2 learners. For instance, Consistent with the preferred solution in while processing a structural ambiguity, learners French, the native speakers’ first-pass gaze dura- may rely more on lexical information and less on tion in the disambiguating area was significantly structurally based parsing strategies than native shorter for the sentences in which high attach- speakers do, as Papadopoulou and Clahsen ment provided the correct solution than in (2003) proposed on the basis of the results of a sentences where low attachment was correct. The study with learners of L2 Greek. group of beginning Spanish–French bilinguals manifested this same preference. Unfortunately, Converging evidence that degree of L2 pro- the cause of this finding is indeterminate: These ficiency determines whether an L1-preferred participants may show this pattern because they or an L2-preferred parsing strategy is applied exploited the same strategy as native French to L2 sentences, or some mixture of both, has speakers do and independent of their L1, but been obtained in off-line experiments studying they may also have used the preferred L1 parsing the way bilinguals assign functions or “thematic strategy, a strategy that happens to be the same roles” to the various noun phrases in L2 sen- as the one favored for L2 French. The pattern tences. In a typical experiment of this type the obtained for the beginning English–French participants read or listen to simple, transitive bilinguals is more informative and, in fact, sentences—both grammatical and ungrammat- suggests that less-advanced L2 learners apply the ical ones—containing two concrete nouns and an analysis preferred in their L1 when reading these action verb and are asked to choose the noun in ambiguous constructions in L2: Unlike the native each sentence that performs the action specified French speakers (and the beginning Spanish– by the verb. The sentences vary on a number of French learners) they showed shorter first-pass features (called “cues”) that language users are gaze durations for low-attachment constructions. known to exploit in figuring out such “form– In contrast, the proficient English–French bi- function” relations in natural language com- linguals were faster at reading high-attachment prehension. Across languages, common such cues constructions, thus showing the same preference are word order, subject–verb agreement in person as native French speakers. Clearly then, if a and number, case, and whether or not a noun learner’s L1 and L2 favor different parsing refers to a living entity (“animacy”). Individual languages differ from one another in the cues they

212 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS rely on most in assigning function, as has been Chinese–English study). The results thus mimic demonstrated in a bevy of studies performed those obtained by Frenck-Mestre (2002, 2005b) in in many languages (see Table 7.3, p. 366, for the on-line sentence parsing studies presented examples of the various types of sentences used in above. these studies). This work provided the corner- stone of a well-known model of sentence com- Summary and conclusions prehension called the competition model (e.g., MacWhinney, 1997; MacWhinney, Bates, & The joint results of the studies discussed above Kliegl, 1984; “competition” refers to the competi- suggest that with increasing L2 practice the pars- tion between conflicting cues in the resolution ing strategies of L2 learners come to resemble process). Native speakers of English, for instance, those of native speakers and that, just as native make especially strong use of word order. For speakers, L2 learners exploit extra-syntactic example, given a noun-verb-noun, NVN, sources of information during L2 parsing. Fur- sequence, the first noun is assigned the role of thermore, the present results warrant the con- actor in the far majority of cases. Instead, for clusion that during early stages of L2 learning, native speakers of Japanese, animacy is an espe- parsing strategies the learner prefers to use in his cially strong cue: If one out of the two nouns in or her L1 transfer to the parsing of L2 sentences. an NVN sequence refers to a living being and But perhaps most remarkable is the finding that the other to an inanimate being, the former is the opposite situation—an L2 parsing strategy assigned the subject role; in English, word order applied to L1 sentences—may occur as well may overrule this, semantically plausible, solu- (Dussias, 2003, 2006; Dussias & Sagarra, 2007). tion. For native French speakers noun–verb This finding reminds of the effects of L2 on L1 agreement is the strongest cue. word recognition reported in the earlier sections of this chapter (see pp. 365–367 for more evi- The above methodology was subsequently dence). As we have seen, the effects of L1 on L2 adopted in studies on bilingual sentence com- word recognition and vice versa are usually prehension (e.g., MacWhinney, 1987; Vaid & attributed to co-activation of the memory repre- Chengappa, 1988). A major question addressed sentations of words from the non-target language in the bilingual studies is what cues—those dom- during target language processing. Similarly, the inant in L1, those dominant in L2, or maybe a present manifestations of transfer from L1 to L2 mixture of both—L2 learners primarily rely on parsing strategies and vice versa suggest when processing L2 sentences (see Cook, Iarossi, language-nonselective activation of parsing Stellakis, & Tokumaru, 2003, for a study that procedures. tried to determine what cues bilinguals use when processing their L1). McDonald and Heilenman This evidence of language-nonselective gram- (1992) posed this question in a cross-sectional matical processing in bilinguals suggests a qualifi- study of late L1 English learners of French at cation of the conclusion drawn above that with four different levels of French proficiency. The increasing L2 proficiency L2 speakers’ parsing researchers found that when performing the strategies come to resemble the parsing strategies agent-assignment task to French (L2) sentences, of native speakers. Although this may be true, the learners’ dependence on word order and L2 parsing by proficient L2 speakers may never noun–verb agreement gradually moved towards become truly indistinguishable, under all cir- the pattern typical of French native speakers: cumstances, from native monolingual speakers’ Word order became less important and noun– parsing. If L1 parsing can be shown not to be verb agreement became more important. So immune from an L2 influence under all circum- apparently, the way bilinguals process sentences stances it is rather unlikely that the reverse in their L2 gradually changes from an L1 strategy situation, L2 immunity, would always hold for towards one that resembles L2 processing more proficient L2 users: The co-activation of the (see Su, 2001, for converging evidence from a generally stronger L1 parsing procedures will

4. COMPREHENSION PROCESSES 213 affect the procedures applied to L2 sentences. brain a particular cognitive process takes place.) Clahsen and Felser (2006a, 2006b) have indeed This “localization” question will be addressed in revealed subtle differences between sentence pars- Chapter 8 and, accordingly, it is there that I will ing procedures in highly proficient L2 speakers discuss the fMRI studies. and native speakers (although they propose a different explanation for these differences than Ardal, Donald, Meuter, Muldrew, and Luce the one advanced here; namely one in terms of (1990) provided early evidence of delayed seman- the depth of the syntactic representations that tic analysis in sentence processing by bilinguals learners and L1 speakers compute during and showed that the delay holds for both of a sentence processing). bilingual’s languages. These researchers recorded event-related potentials while early and late A final remark is in order here; namely that to English–French and French–English bilinguals be able to detect a reciprocal influence of a read sentences in their L1 and L2. A group of bilingual’s L1 and L2 parsing strategies at all, the matched English monolinguals was also included, two languages’ preferred parsing solutions for a obviously only to be tested in English. The critical particular sentence construction must differ (as sentences contained a semantic anomaly, always was indeed the case in the above studies). The in sentence-final position (e.g., I generally like reason is that if the two languages share the same menthol bottles). The primary dependent variable parsing preference for this type of construction was the N400 elicited by the anomalous words in one can never tell whether the L2 speaker used a the critical sentences (and by sentence-final words native language or a second language parsing in non-anomalous control sentences). The results strategy. In retrospect, the point seems too showed that the N400 elicited by anomalous obvious to be worth musing on, but it can easily words was delayed in bilinguals as compared with be overlooked in designing experiments aimed at monolinguals, and that it was delayed in the informing the present dispute. bilinguals’ L2 as compared with their L1. In other words, monolinguals seem to be faster at detect- The processing of semantic and ing semantic anomalies than bilinguals, and syntactic anomalies bilinguals detect anomalies faster in their L1 than in their L2. Age of L2 acquisition did not A further set of studies has examined the modulate these effects. The authors attributed way bilinguals process semantically and/or syn- these effects to more highly automated language tactically anomalous sentences using the ERP processing in monolinguals than in bilinguals and methodology (see pp. 162–163). A couple of these to more highly automated L1 than L2 processing will be discussed here (see Kotz, 2009, for a more in bilinguals. A further finding of interest was complete review). They have shown that several that in bilinguals the amplitude of the N400 to aspects of sentence processing may be delayed in anomalous words was somewhat smaller in L2 bilinguals as compared with monolinguals. They than in L1, especially in the least-fluent bilinguals. furthermore suggest that nativelike semantic A subgroup of monolinguals with a relatively low analysis of L2 sentences is in place before native- L1 fluency showed this same reduction of the like syntactic analysis, although also with respect N400 in their L1 as compared to more-fluent L1 to semantic analysis differences may remain users. Apparently, so the authors concluded, between native speakers and even relatively fluent the amplitude of the N400 somehow reflects level L2 speakers. A few other studies have looked at of language fluency. syntactic processing in bilinguals employing the fMRI methodology. The primary question Weber-Fox and Neville (1996) obtained partly addressed in these studies was whether the two similar results in a study that tested proficient languages of a bilingual are processed by the Chinese–English bilinguals varying in the age at same or different brain areas. (Recall that fMRI is which they first started to acquire English. The especially well suited to find out where in the participants performed acceptability judgments on English sentences that were either correct or

214 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS semantically or syntactically anomalous. All 5b syntactically correct Der Vulkan wurde age groups showed an N400 effect (that is, a semantically incorrect gegessen difference in the amplitude of the N400 for (semantic violation (The volcano was semantically anomalous sentences on the one only) eaten) hand and semantically correct sentences on the Das Eis wurde im other hand). Furthermore, for the participants 5c syntactically incorrect gegessen who had acquired L2 English after 11 years of semantically correct (The ice cream age, but not for those with younger acquisition (syntactic violation was in-the eaten) ages, the effect was delayed as compared with an only) Das Türschloss English monolingual control group. These results wurde im gegessen thus suggest that at least for relatively late learn- 5d syntactically incorrect (The door lock ers semantic processing in L2 is relatively slow. It semantically incorrect was in-the eaten) remains unclear why an effect of acquisition age (semantic and materialized whereas Ardal and his colleagues syntactic violation) (1990) had shown similar results for early and late learners. Importantly though, in both studies an As these sentences illustrate, both semantic and N400 occurred for bilinguals as well as mono- syntactic congruency were manipulated. The linguals. This finding legitimates the conclusion sentences could be incongruent (anomalous) on that semantic processing is qualitatively the same one or both of these dimensions or they were in bilinguals as in monolinguals. congruent on both. The ERP responses were measured time-locked to the final word of each In their analysis of the ERP signal, Weber-Fox sentence, which was always a participle. In their and Neville (1996) also looked for the occurrence analysis of the ERP signal the authors focused of ELAN and P600, the two components that are on the difference between the ERP responses to assumed to index syntactic processing (see p. 163) (the final words of) correct sentences on the one and in a more recent Japanese–German study hand and to each of the three types of incorrect Hahne and Friederici (2001) also focused on all sentences on the other hand (these differences three components. The results of both studies being the ERP effects). The brain responses of the suggest that syntactic analysis may be qualita- German L2 speakers were compared to those of tively different in bilinguals and monolinguals. native German speakers who had been presented Furthermore, and contrary to the above conclu- with exactly the same materials in an earlier study. sion, the latter study suggests that also semantic The pattern of results obtained for the L2 processing may, in one respect, differ qualitatively listeners differed from that shown by the native between monolinguals and bilinguals. To get listeners in a number of respects. In the ensuing some idea of how semantic and syntactic process- discussion, I will first deal with the N400 effect, ing was manipulated in these studies, consider the marker of semantic processing, and will then sentences 5a–5d below, which exemplify the four continue with a discussion of ELAN and the types of German passive sentences that Hahne P600, the components that are thought to reflect and Friederici presented to their participants syntactic processing. The focus will be on how a (English translations are given in parentheses). semantic and/or syntactic violation affects these The participants in this study were late Japanese components. learners of German who were clearly dominant in Japanese and the sentences were presented aurally In the bilingual participants the incorrect (instead of visually, as in Weber-Fox & Neville, sentences that only involved a semantic violation 1996). (5b) significantly modulated the N400 over centro-parietal electrodes as compared to the 5a syntactically correct Das Brot wurde correct sentences and this N400 effect did not semantically correct gegessen (The differ statistically from the analogous effect in the (no violation) bread was eaten) native-speaker group (see pp. 408–409 for infor- mation about the locus of these electrodes on the skull). This finding converges with the results

4. COMPREHENSION PROCESSES 215 obtained by Ardal et al. (1990) and Weber-Fox to reflect an automatic first-pass parsing process and Neville (1996). Unlike in these two earlier stage in which the parser “assigns the initial syn- studies, and even though the participants were tactic structure on the basis of word category late bilinguals, this effect was not delayed as com- information only” (Hahne & Friederici, 1999, pared to the corresponding effect in native p. 195). Therefore, the absence of the ELAN speakers. It lasted longer in the bilingual group, effect in bilinguals suggests this process has not however, suggesting that bilinguals’ semantic been automated yet in L2 learners. The second integration processes take relatively long. From marker of syntactic processing, the P600, also these results one might once again be inclined to showed a difference between native and L2 conclude that semantic processing is qualitatively speakers: Whereas in native speakers a syntactic the same in native speakers and L2 speakers violation modulated this component, in the (albeit slower). Yet this conclusion would be pre- L2 speakers the P600 did not differ between cor- mature, because both conditions that involved a rect and incorrect sentences. Because this com- semantic anomaly (5b and 5d) showed an effect in ponent is thought to reflect controlled syntactic bilinguals that was not observed for native repair and re-analysis processes, Hahne and speakers: a late right anterior-central negativity Friederici (2001) suggested this result indicates around 700–1100 ms following the critical that L2 speakers do not engage in these repair sentence-final word in incorrect sentences as processes. compared with correct sentences. To summarize, in both L1 and L2 sentence To account for this pattern of results, the processing, bilinguals show N400 effects similar authors suggested that two types of semantic pro- to those observed for monolinguals, suggesting cessing must be distinguished: the processing of that semantic integration processes are qualita- lexical semantic knowledge and of more general tively (largely; see below) the same in mono- conceptual semantic knowledge that is also linguals and bilinguals. There is some evidence implicated in non-verbal semantic tasks. The data that semantic integration is delayed in bilinguals can then be accounted for by assuming that as compared to monolinguals and that the delay lexical-semantic processing, tapped by the N400 is relatively large when bilinguals process their component, is the same in native and L2 speakers weaker language. Other evidence suggests the of a language, whereas only bilinguals exploit delay only occurs in late bilinguals. Yet a further conceptual-semantic information while process- set of experimental results suggests that the N400 ing verbal materials in their L2. The assumed lasts longer in bilinguals than in monolinguals, distinction between lexical semantics and indicating that bilinguals’ semantic integration language-independent conceptual knowledge processes take relatively long to complete. Finally, that underlies this analysis has been made by one of the above studies showed a second negativ- other researchers as well (e.g., Paradis 2004; see ity, occurring later than the N400, in bilinguals pp. 234–236 for a discussion). but not in monolinguals. This may indicate that in one respect semantic integration processes are Further differences between native and L2 different in bilinguals than in monolinguals after speakers of German were observed when the all. The authors hypothesized this ERP com- processing of correct sentences was compared ponent indexes the use of conceptual-semantic with the processing of sentences that included information by bilinguals but not monolinguals syntactic violations (5c and 5d). The pattern of during semantic integration. In contrast, the results resembled those of the analogous con- N400, common to monolinguals and bilinguals, ditions in the reading study by Weber-Fox may reflect the use of lexical-semantic infor- and Neville (1996): Whereas in native speakers mation. Finally, the data suggest that syntactic syntactic violations modulated the early left analysis differs qualitatively between native and anterior negativity (ELAN) as compared to L2 speakers. This showed, first, from the presence correct sentences, this effect was not evident in of a differential ELAN to syntactically correct L2 speakers. The ELAN component is thought

216 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS and incorrect sentences in native speakers but the but included a proficiency manipulation: Two absence of this ELAN effect in L2 speakers. It groups of late Japanese learners of L2 English furthermore showed from the fact that a syntactic were included, one at an intermediate level of violation modulated the P600 in native speakers English proficiency, the second at a high level of but not in L2 speakers. The first of these findings English proficiency. A control group of English indicates that initial first-pass parsing differs native speakers was also included. At variance between native speakers and L2 speakers; the with the results of the studies discussed above, second may mean that, unlike native speakers, this study demonstrated that not only semantic L2 speakers are not involved in second-pass processing but also syntactic processing in L2 re-analysis yet. speakers may resemble that of native speakers: The ERPs to semantic anomalies in English sen- One factor that affects the ERP patterns in tences read by both the intermediate- and high- L2 learners while they process anomalous and level learners of English resembled those of the normal sentences was underexposed in the above native English speakers. In contrast, syntactic discussion; namely, the level of proficiency the anomalies evoked a similar pattern of ERP bilinguals have attained in their L2. Yet, a responses in the high-proficiency learner group growing number of studies show that L2 pro- and the English native speakers but a different ficiency is a more important determinant of the one in the learner group of intermediate pro- ERP patterns to be observed in L2 speakers ficiency. Plausibly then, the fact that the above processing their L2 than L2 acquisition age and studies had shown qualitatively different syntactic that the same holds for the patterns of brain analysis processes in L2 learners and native activation as observed in fMRI research (see speakers was due to the L2 learners’ relatively low e.g., Abutalebi et al., 2005, and Kotz, 2009, for level of L2 proficiency. Ojima and his colleagues reviews; see also Chapter 2, pp. 73–74). To illus- concluded that a late, post-childhood, start of L2 trate the important role that level of L2 pro- learning per se is no impediment to attaining ficiency plays in the way sentence anomalies are nativelike performance in both grammar and processed (as reflected in the ERP patterns) let us semantics, underscoring the current view that L2 look at a study by Ojima et al. (2005) that proficiency level is what matters most. resembled Hahne and Friederici’s (2001) study SUMMARY • Unilingual studies on the resolution of lexical ambiguities using behavioral measures have produced mixed results, suggesting that word recognition is neither fully autonomous nor fully interactive under all circumstances but that the specifics of task and context determine how the ambiguities are resolved. Instead, recent studies employing the ERP methodology, with its high temporal resolution, suggest that word recognition is a highly interactive process, thus providing evidence against a modular view of word recognition. • In many but not all studies that examined the processing of interlexical homographs and homophones in isolation a homograph/homophone effect was obtained. This indicates that even when words are processed in isolation bilingual word recognition is not always language- nonselective. Specifically, word recognition in the stronger language appears to be language- selective if the activation in the weaker language is not boosted by some experimental manipulation. • A strong test of language-nonselective bilingual word recognition requires the presentation of exclusively unilingual language materials. Apparent evidence of language-nonselective word recognition in experiments that also present words from the non-target language is inconclusive because these words, however few, will boost the activation level of all representations in the lexicon of the non-target language, thus increasing their availability.

4. COMPREHENSION PROCESSES 217 • Language-nonselective lexical access is constrained by the relative dominance of a bilingual’s two languages and by context: (1) If the non-target language is the stronger of a bilingual’s two languages, lexical access tends to be language-nonselective; if the non-target language is the weaker language of the two, lexical access tends to be language-selective; (2) Immersing the bilingual participants in the non-target language prior to having them perform a word recognition test increases the degree of activation in the non-target language’s memory system and, consequently, of language-nonselective lexical access. Similarly, immersing them in the target language prior to having them perform a word recognition test increases the degree of activation in the target language’s memory system and, consequently, of language-selective lexical access. • The bilingual interactive activation (BIA) model of bilingual lexical access contains four levels of representation units that represent visual letter features, letters, orthographic word forms, and language information, respectively. Representations at one level can activate and inhibit representations at adjacent levels via excitatory and inhibitory connections. The model assumes language-nonselective lexical access: Activated letter nodes activate word nodes in both of a bilingual’s languages. • BIA explains the interlexical homograph effects by assuming two word node representations for interlexical homographs, one for each language, but just one for non-homographic control words. When a homograph is presented to the system both of its word nodes will become activated but when a control word is presented, only its one word node will be activated. The homograph effects are attributed to this difference in the activation state of the recognition system following the presentation of the two types of words. • In lexical decision experiments, when interlexical homographs are presented in isolation the size of the homograph effect is especially large when the homograph is more frequent in the non-target language than in the target language. BIA explains this effect by assuming that the memory representations of frequent words have higher baseline levels of activation than those of infrequent words. As a consequence, the memory representations of frequent words have a head start in the recognition process. • A word stimulus activates both its within-language neighbors and its cross-language neighbors. This result follows directly from BIA’s assumption that lexical access is language-nonselective. • In BIA+ a task/decision system is added onto the word identification system. This system can explain why the interlexical homograph effect varies across different tasks and with different compositions of the stimulus materials. Whereas the word identification system is only affected by linguistic sources of information, the task/decision system is sensitive to extra-linguistic influences such as participant expectancies. • Studies that examined phonological activation in same-alphabet bilingualism have shown that during visual recognition of L2 words bilinguals assemble the phonological forms of these words just as native speakers do, that this process comes about automatically and unconsciously, and that under certain circumstances the grapheme–phoneme conversion rules of both languages are activated in parallel. The few studies that tested different-alphabet bilinguals and bi-dialectal language users suggest these conclusions also apply to these forms of bilingualism. • Studies that examined the recognition of spoken words using the eye-tracking methodology have shown that the words of weaker L2, but not those of stronger L1, give rise to language- nonselective phonological activation in bilingual lexical memory. Though not totally univocal, the results of gating studies are compatible with these conclusions. • Generally, the non-native (L2) lexicon contains fewer lexical elements than the native lexicon. Nevertheless, a spoken non-native word input causes more spurious activation in a bilingual’s non-native lexicon than the spurious lexical activation caused by this same word if presented to a native speaker.

218 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS • Two models of word recognition, SOPHIA and BIMOLA, may ultimately be suitable to explain language-nonselective phonological activation. One salient difference between the two models is that in SOPHIA the lateral inhibition exerted by a word affects words from the same and the other language, whereas BIMOLA restricts lateral inhibition to units (phonemes, words) of the other language. Another difference is that SOPHIA but not BIMOLA assumes two language nodes, one for each language. • Whereas in many tasks cognates are processed faster than matched non-cognates, in word naming they are often responded to more slowly. The likely reason cognates are processed relatively slowly in word naming is that in this task the stimulus must be named aloud. The activated phonological representation of the cognate’s translation in the non-target language triggers a response that mismatches the correct response. This pending response will act as a nuisance competitor in the naming process. The consequence is a delayed naming response. • There is some evidence that in both high-constraint and low-constraint sentence contexts interlexical homograph effects disappear. This suggests that both types of sentence context block co-activation of an interlexical homograph’s representation in the non-target language. • In high-constraint sentence contexts but generally not in low-constraint sentence contexts cognate effects disappear. This suggests that only highly constraining sentence context blocks co-activation of a cognate’s representation in the non-target language. • In general, the modulating effects of sentence context on interlexical homograph effects and cognate effects indicate that context constrains language-nonselective processing in bilinguals. • The source of cognate effects on word recognition is equivocal: They may either be due to co-activation of a cognate’s translation equivalent in the non-target language, to some representational difference between cognates and non-cognates, or to both. • The most natural interpretation of the observed null effects of the non-target language in the context studies is that when words are presented in a larger linguistic context the non-target language is deactivated. An alternative interpretation is that words presented in a larger linguistic context also activate lexical representations in both of a bilingual’s linguistic subsystems but that the activation in the non-target language’s subsystem is ignored by an attention system that supervises performance. This idea is consistent with the recent addition of a control system to models of word recognition. • When parsing syntactically ambiguous sentences both native speakers and L2 speakers make use of lexical knowledge to resolve the ambiguity. Furthermore, the language the bilingual has recently been exposed to most determines which parsing strategy he or she adopts in resolving a syntactic ambiguity. Both findings challenge the modular syntax-first model of parsing. • If an L2 learner’s L1 and L2 favor different parsing solutions for a given type of grammatical structure, with increasing L2 proficiency the strategy the learner employs to parse L2 sentences may change from the strategy he or she prefers to use while parsing sentences in L1 to one used preferably by native speakers of L2. • When L1 sentences are parsed in an L2 immersion setting, L2-specific parsing strategies may dominate L1 parsing. When bilinguals parse L2 sentences, L1 lexical knowledge influences the parsing solution that is chosen. Both findings suggest the occurrence of language-nonselective activation of parsing procedures. • Studies that examined how native speakers and L2 speakers process semantically anomalous sentences indicate that semantic integration processes are largely similar in native speakers and L2 speakers although they may be delayed and last longer in L2 speakers.

4. COMPREHENSION PROCESSES 219 • Studies that examined how native speakers and L2 speakers process syntactically anomalous sentences indicate that syntactic analysis differs qualitatively between native speakers and L2 speakers. Specifically these studies suggest that initial first-pass parsing is automated in native speakers but not in L2 speakers and that only native speakers execute a repair or re-analysis on a second pass. • Recent studies on the processing of semantically and syntactically anomalous sentences by native and L2 speakers indicate that the above conclusions apply in particular to L2 speakers who have not attained a high level of L2 proficiency. In contrast, proficient L2 speakers may process these sentences in the same way as native speakers do.



5 Word Production and Speech Accents INTRODUCTION AND PREVIEW whether, when, and to what extent bilinguals can deactivate or switch off the contextually While learning about research on bilingual lan- inappropriate (“non-selected”, “non-target”) lan- guage comprehension in Chapter 4 you have guage during language comprehension. A central probably noticed that much of that work was question examined in work on bilingual language built on the study of monolingual language com- production is the analogous one: In the inter- prehension: Similar questions were posed and mediate stages between conceptualizing and they were tackled using the same tasks and delivering a message in the selected language, methods, adapted to the bilingual case. This does co-activation occur of memory nodes in the situation also holds for studies on bilingual lan- non-selected language or is the non-selected lan- guage production, which have gratefully taken guage totally deactivated? Reflecting the broad advantage of the large body of knowledge interest in this topic, studies that have addressed acquired in earlier research on monolingual lan- this question will be covered extensively in this guage production: They pose similar questions chapter, providing relevant evidence from both about the content and structure of the knowledge picture naming and Stroop studies. One specific base underlying language production, the pro- set of picture-naming studies, those that cessing mechanisms involved, the nature of the employed the so-called “language-switching” processes through which stored knowledge is methodology, will not be covered in this chapter accessed and exploited during language pro- but in Chapter 6 instead, which takes a different duction and the order in which these processes are point of view: After having established, in executed. Furthermore, they often use the same Chapters 4 and 5, that in both bilingual language research methods, thereby endorsing their comprehension and production the bilingual’s rationale. two language systems are often activated in para- llel, how then can it be ensured that (most of the Chapter 4 largely dealt with the question of time) the targeted element is ultimately selected? 221

222 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS Before presenting the bilingual studies, the ceivably the most complex of all forms of lin- models and methods developed in the research guistic behavior. field of monolingual speech production that pro- vided the groundwork for the bilingual work will Dealing with a task that involves both com- be introduced, paying special attention to two prehension and production spontaneously calls central stages of the speech production process: up the question of whether these two types of lemma selection and phonological encoding. The processing exploit the same or different under- way these two stages are sequenced—strictly lying representations and cognitive machinery. serial or partly in parallel—concerns one of two This issue is briefly addressed in a further section. core questions in a lively debate about the The chapter concludes with a discussion of one architecture of the speech production system. The particular characteristic of the output that second core question in this debate is whether emerges from the bilingual speech production these two stages proceed strictly unidirectionally process: The speech accents that are noticeable in or interact with one another. In addition to pre- the L2 speech of the majority of late L2 learners senting the monolingual base work, these two and that can often be detected in early learners as sections will also present the proposals of bi- well. (In Chapter 7 bilinguals’ less noticeable, yet lingual researchers on how to adapt the mono- detectable, accents in their L1 speech will be dis- lingual models to bilingualism. cussed.) The final part of this section will review studies that examined the comprehensibility of A further section reviews studies that accented L2 speech. But first of all I will present employed the word translation task. A salient the most popular research methods employed feature of this task is that it involves both a in speech production research. I will do so with a comprehension component and a production bird’s eye view in the Methods and Tasks section, component. Another salient feature is that, by detailing the picture–word naming and Stroop definition, it always implicates the use of both tasks in the course of the later sections. languages. Unlike picture naming and the Stroop task, which at least in theory can be performed METHODS AND TASKS with the non-target (or “non-response”) language completely at rest, word translation is only Many studies on speech production, both mono- possible when both languages are activated, lingual and bilingual, have focused on one specific either simultaneously or in rapid alternation: Its part of the full speech production process—word comprehension component requires one language production. Word production consists of several (the “source” language) to be active and its processing stages from (and including) the con- production component requires that the other ceptualization of the lexical concept the speaker language (the “target” language) is on the alert. intends to verbalize to articulating the associated The word translation task has already been word. To get the word production process introduced in Chapter 3 in a discussion of how going, the concept to be verbalized must first be bilinguals map word form onto word meaning established in the experimental participant’s (pp. 136–137). In the present chapter I will focus mind, preferably by means of a stimulus that is on its resemblance with other word production non-verbal because in veridical speech produc- tasks and the experimental results obtained using tion the process also starts off with a non-verbal the word translation task will be related to mental representation. The most popular way do those gathered by means of these other tasks. In this is by presenting the participant with a picture Chapter 6 the translation task will reappear depicting the targeted word, with the instruction (pp. 314–336). But whereas the present chapter to name it as rapidly as possible. While generating treats the translation of separate words, there I the naming response the participant passes will primarily deal with the much more complex through various mental stages, starting with the task of translating larger linguistic units and, especially, with simultaneous interpreting, con-

5. WORD PRODUCTION AND SPEECH ACCENTS 223 visual analysis of the picture. This first stage is of the colors of the words, not the words themselves. no particular interest to the speech production Alternatively, they are presented with color researcher but it is a necessary prerequisite for patches, each patch being accompanied by a color triggering the stages of interest because it loads word. In this version of the task the participants the concept to name in the production system. have to name the color of the patch and to ignore Then follow the stages of interest. the word. Both versions of the Stroop task are conceptually very similar to the picture–word Different versions of this picture-naming task interference version of the picture-naming task. have been used. In addition to the “simple” (but, Whereas in picture naming the production system in cognitive terms, still rather complex) picture- is loaded with a concept by means of the presen- naming task, in which the participants simply tation of a picture, in the Stroop task a (color) name the pictures in the specified language and concept is established in the production system no potentially distracting information is pre- by means of the color patch or the color sented with it, the picture–word interference task word’s color. The word component of the Stroop has been used. In this version of the task the pic- stimulus serves as distracter in the naming ture is accompanied by a word and the relation process. The original unilingual version of the between picture and word is systematically varied. task (with word input and response in the same This word can have an interfering or a facilitative language) has been adapted to the bilingual case effect on picture naming and the exact conditions and the task’s bilingual version provides a means under which such effects occur and the direction to study the involvement of the non-target lan- of the effects (facilitative or inhibitory) tell us guage during task performance. In the bilingual something about the word production process version the colors have to be named in the par- (see pp. 237–246). In yet another version of the ticipants’ one language whereas the distracters picture-naming task the picture does not have to are words in their other language. A problematic be named explicitly but a response must be based aspect of both the picture–word interference task on its internally generated name. The instruction and the Stroop color-naming task is that in both may, for instance, be to determine whether the cases the process of interest, word production, depicted object’s name contains a specific phon- is executed in parallel to a word recognition pro- eme. Alternatively, a particular grammatical cess that interacts with the production process. judgment must be based on the picture’s name, Because the precise nature of this interaction is generated tacitly. The exact task to be performed not fully known, the fact that it occurs severely by the participants might, for instance, be a go/ complicates the interpretation of the results. no-go task in which the participants have to Furthermore, a constraint of both tasks (and respond overtly on only a subset of trials. For the picture-naming task in general) is that they example, they may be asked to respond if the can only be applied to the study of a limited picture’s name starts with a consonant but to set of words, concrete words and color words, refrain from responding if it starts with a vowel. respectively. In addition to presenting a picture or The primary question of interest in bilingual a color, a third way to establish a lexical concept picture-naming studies is whether a particular in the speech production system is by presenting component of the depicted object’s lexical repre- the participants with a definition. An advantage sentation in the non-target language affects of this procedure is that it can be used for con- responding. If it does, this would suggest that crete as well as abstract words but a disadvantage bilingual word production is language- is that the word production process is set in nonselective. motion by verbal instead of non-verbal information. A further popular speech production task is the Stroop color-naming task, which we have As we shall see, the picture-naming task can be already briefly encountered in Chapter 4 (p. 190): used to study the cross-language activation of all The participants are presented with color words sorts of lexical information: meaning, phonology, printed in different colors and are asked to name

224 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS and grammar. The question of whether the acti- hesitations and pauses, and errors in spontaneous vation of grammatical knowledge is language- speech (see Bock, 1996, for a review). Both selective or language-nonselective has also been methods have been applied to monolingual and studied with the “sentence completion (or sen- bilingual speech production alike. The reason for tence generation) paradigm”, in which the par- confining myself here to the above experimental ticipants are presented with sentence fragments methods is that these are the ones that were used to complete into full sentences. As with the trans- in the selection of studies to be discussed below. lation task, it involves both a comprehension and a production component. In monolingual MODELS OF MONOLINGUAL AND research on language production the paradigm BILINGUAL SPEECH PRODUCTION has been used to elicit a particular class of gram- matical mistakes—errors in subject–verb number Levelt’s model of monolingual speech agreement (e.g., Bock & Miller, 1991). In experi- production ments using this methodology the participants are given complex noun phrases such as The A major part of the research on monolingual key to the cabinets. Their task is to complete speech production has been inspired by Levelt’s the phrases into full sentences (by adding, for (1989) theory of the speech process as illustrated instance, is lost). The two nouns in the complex in Figure 5.1. It covers the whole trajectory from noun phrase may differ in number (as in the the conceptualization of a message to be example) or may have the same number (The key expressed to actual articulation, and describes the to the cabinet). The former situation leads to various types of knowledge structures that are more agreement errors during completion (as in accessed and retrieved en route and their tem- The key to the cabinets . . . are lost) than the latter. poral order. Boxes in the figure represent process- This finding is attributed to the fact that the ing components; the circle and ellipse represent two nouns prime completions with differential knowledge stores. numbers. This creates a response conflict, which may lead to an error. This paradigm has been The first steps in the production process adapted to the study of bilingual speech pro- involve the conception of the intention to express duction by exploiting the fact that noun number a particular thought in words, the selection of the does not always converge across a bilingual’s two required information from the relevant memory languages. For instance, scissors and trousers are stores, and the ordering of this information for plural in English but their Greek equivalents are expression. This aggregate of activities is called singular whereas, conversely, hair and money conceptualizing, and the system involved is called are singular in English but plural in Greek the “conceptualizer”. The output of the concep- (Hatzidaki, 2007). If in a unilingual condition tualizer is a conceptual structure called the bilinguals are affected by this incongruence while preverbal message. This output possibly takes performing the sentence completion task, as the form of a collection of activated memory compared to a condition where noun number is nodes, each representing a particular conceptual the same in the two languages, this would suggest component, an element of meaning. Speakers are language-nonselective grammatical processing. known to monitor their own speech. This activity A similar technique, both in a monolingual (e.g., is also taken care of by the conceptualizer. The Bock, 1995) and a bilingual version (see p. 252), ellipse that is connected with the conceptualizer has been used to study pronoun–antecedent in the figure and from which the conceptualizer number agreement. draws the information required for expressing its intention, represents a database of declarative This overview of ways to study speech produc- knowledge. It includes all the factual (“encyclo- tion is by no means exhaustive. Most notably pedic”) knowledge stored in the speaker’s missing are the common observational methods used to analyze speech dysfluencies, such as

5. WORD PRODUCTION AND SPEECH ACCENTS 225 Levelt’s (1989) model of the speech process. Boxes represent processing components. Circle and ellipse represent know- ledge stores. From Levelt (1989), © 1989 Massachusetts Institute of Technology, by permission of The MIT Press. long-term memory as well as knowledge on the encoder”, respectively. The grammatical encoder present discourse situation (e.g., who the inter- consists of procedures to access entries in the locutors are, what their goals are, what everybody mental lexicon (the knowledge store represented has said so far). by the circle in Figure 5.1) and to build syntactic constructions. Lexical items (or “entries”) are The preverbal message constitutes the input composed of two parts that, together, contain for the second processing component, the formu- four types of information: the word’s meaning, its lator, which translates the preverbal message into syntactical characteristics, its morphological a linguistic structure. It does so in two steps: characteristics, and its phonological form grammatical encoding and phonological encoding. (because the model deals with speaking, the fact These two processes are taken care of by the that, in literate language users, lexical entries also “grammatical encoder” and the “phonological

226 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS contain orthographical information is ignored). encoding and phonological encoding are An entry’s part that specifies its meaning and mediated by the information that becomes avail- syntax is called its lemma, whereas the part that able when lexical entries are activated by the specifies its morphological and phonological preverbal message. Figure 5.1 does not show a properties is called lexeme (Kempen & Huijbers, direct connection from the preverbal message to 1983). It is the lemma component of the lexical the lexicon. Yet it is the preverbal message entries that is contacted by the procedures that “triggers lexical items into activity” (Levelt, executed by the grammatical encoder. The output 1989, p. 181). More specifically, the meaning of the grammatical encoder is a “surface struc- component of a lexical item’s lemma becomes ture”: an ordered string of lemmas grouped in activated automatically when it matches the phrases or sub-phrases whose sound is not yet information in the preverbal message. The specified. moment this happens, the associated syntactical information also becomes available (e.g., that the During the next processing step, phonological item is a female noun, which is pluralized with encoding, a phonetic plan is built for each lemma an -s or that it is an intransitive verb), and this and the utterance as a whole. During this process, information drives the grammatical encoding the lexical entries’ lexemes are accessed and the process. elements in the grammatical encoder’s output are morphologically and phonologically adapted A final hypothesis reflected in the model is that to their linguistic environment. The output of the information flow is top-down, with no feed- phonological encoding is an internal representa- back from processors that start operating at a tion of how the planned utterance should be later point in time (e.g., the formulator) back to articulated. It is variously called the “phonetic those that start earlier (e.g., the conceptualizer). plan” or “internal speech”. But even when first presenting his model, Levelt (1989) considered the possibility there might be The next step is articulating; that is, the one exception; namely, that the outcome of execution of the phonetic plan. This process is phonological encoding might feed back to the taken care of by the “articulator”. As already process of grammatical encoding and, especially, mentioned, speakers monitor their speech, a to the lemma selection part of that process. The process that exploits the speech comprehension reason for this qualification was that several system. The model assumes monitoring to take sets of experimental results appeared hard to place on both internal speech—that is, the output reconcile with a strict top-down serial view of of the formulator—and on overt speech, the processing. One such source of counterevidence is output of the articulator. the so-called “lexical bias” effect; that is, the observation that phonological speech errors Levelt (1989) hypothesized the different pro- constitute real words more often than would be cessing components to be dedicated specialists: expected on the basis of chance (Baars, Motley, & He assumed that each of them serves one (and MacKay, 1975; Dell & Reich, 1981; and see just one) function, for instance to translate a Costa, Roelstraete, & Hartsuiker, 2006, for a preverbal message into a surface structure, that bilingual lexical bias effect across languages). If it can, in principle, only operate on one type of grammatical and phonological encoding take input, and that it works autonomously. This latter place in two serial and independent steps, errors characteristic means that it is not affected by pro- that occur during the second stage should result cessing activity elsewhere in the system. In terms in words and nonwords equally often. A second of Fodor’s (1983) modularity of mind theory source of counterevidence is the occurrence of presented earlier (p. 163), the processing com- “mixed” errors, errors that resemble the targeted ponents in Levelt’s model are “informationally lexical item both in form and meaning (such as encapsulated” modules. oyster for lobster). These errors were found to occur more often than would be expected on the A further processing assumption of the model is that the operations of the formulator are “lexically driven”. This means that grammatical

5. WORD PRODUCTION AND SPEECH ACCENTS 227 basis of a chance estimate (e.g., Dell & Reich, “unidirectional cascaded-processing models” 1981). These findings have led several authors (e.g., Humphreys, Riddock, & Quinlan, 1988; (most notoriously Dell, 1986) to assume feedback Peterson & Savoy, 1998), and “interactive acti- from the phonological encoding stage back to the vation models” (e.g., Dell, 1986). These models all grammatical encoding stage (see, e.g., Levelt & distinguish between three levels of representation Maassen, 1981; Levelt et al., 1991; Schriefers, in the speech production system: a prelexical Meyer, & Levelt, 1990, for more complete discus- conceptual level and two lexical levels, one of the sions, including suggestions of how to reconcile latter storing the lexical items’ lemmas and this apparent counterevidence with serial produc- the second storing their lexemes; that is, their tion models). phonological forms. However, they differ in their assumptions regarding the way activation The primary goal of much of the monolingual flows between these representational levels and, speech production research to date has been to especially, between the two lexical levels. answer this very question, whether grammatical encoding and phonological encoding occur Discrete two-stage models assume that acti- strictly top-down (“forward”) or whether there is vation can only flow in a forward direction from feedback from the phonological level back to the lemma level to the lexeme (or phonological) the level of grammatical encoding. However, level. In addition they assume that activation the terminology has changed somewhat over the is not transmitted from the lemma level to the years, presumably because of the fact that the lexeme level until after processing at the lemma lexicalization component of the complete pro- level has been completed; that is, after one lemma duction model has become the main focus of has been selected from the initially activated set attention in studies on speech production (see of lemmas (the semantic cohort). Unidirectional pp. 229–234). Current terminology distinguishes cascade models also hold that activation only between the stages of “lemma selection” and flows forward, but contrary to the discrete two- phonological encoding, and both processes are stage models they assume that this activation flow conceived of as the spreading of activation in a does not await the completion of processing at lexical network. In addition to a distinction the lemma stage. The moment a lemma becomes between forward-only processing and forward- activated by the preverbal message it starts and-backward (“interactive”) processing, a sending activation down to the lexeme level. The second distinction is made in current theories of consequence of this combination of processing speech production. It is commonly assumed that assumptions is that all elements in the semantic a particular conceptual content (the preverbal cohort are temporarily phonologically encoded. message) activates not one lemma but a set of Finally, as with the unidirectional cascade lemmas, each of them sharing a part of its seman- models, the interactive activation models assume tic content with the conceptualized content (this that all initially activated lemmas immediately set of activated lemmas is known as the “semantic pass on their activation to the corresponding cohort”). Ultimately only one of the activated lexemes. But unlike the unidirectional cascade lemmas must be selected for production. The models, these models assume that activation may second distinction referred to here concerns the also flow back from activated lexemes to lemma view that only the selected lemma is phonol- representations. ogically encoded, versus the view that shortly after their activation, and prior to one of them As mentioned, a conceptual structure activates being selected, the phonology of all activated all lemmas that share content with the con- lemmas becomes temporarily activated. ceptualized structure. Although translation “equivalent” words typically do not share mean- In practice, the above two theoretical distinc- ing completely between the two languages in tions have led to three types of speech production question, their meanings generally overlap to a models: “discrete two-stage processing models” large extent, maybe even more so than the mean- (e.g., Levelt et al., 1991; Schriefers et al., 1990), ings of within-language synonyms (Hermans,

228 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS Bongaerts, De Bot, & Schreuder, 1998). Unless extension of Levelt’s model with an extra com- languages can be fully sealed off from one ponent, the “verbalizer”, which receives input another (which we have seen not to be the case), from the preverbal message (the “conceptual this prompts the question of whether, when structure”) and carves it up in such a way that it bilinguals intend to utter a word, the associated matches the semantic information in the targeted conceptual content will activate not only lemmas lemmas. If the information in the conceptual in the target language but also the lemma of the structure is indeed lexicalized differently in a targeted word’s translation in the non-target bilingual’s two languages, this process results in language (and maybe other lemmas in the non- different sets of information chunks, depending target language as well). Such an outcome would on the language currently spoken. Subsequently, constitute evidence of language-nonselective the constructed chunks are output by the verbal- speech production. This is one of the main izer and trigger the lemmas into action. (Note the questions addressed in the bilingual speech pro- difference with Levelt’s model, in which lemmas duction research to be reviewed below. If the are accessed directly by the information in the semantic cohort indeed contains lemmas from preverbal message.) both languages, a further question is whether all lemmas in the semantic cohort, irrespective of As acknowledged by the authors, this is all language, are phonologically encoded or just easier said than done, because chunking the con- one; namely the one that emerges as the winner in ceptual information such that it matches the the stage where lemmas compete for selection. lemmas’ meaning components implies that the Accordingly, a second main question posed in the verbalizer knows in advance what semantic bilingual speech production research is whether information it can expect to find in the lemmas. the lemmas of both terms in a translation pair On what other grounds could it carve up the con- transmit their activation to the corresponding ceptual structure into chunks that the lexicon phonological forms, the lexemes, or whether can handle? In this respect the original model in phonological form activation is restricted to the which lemma triggering requires no fore- target-language lemma. Before presenting the knowledge of knowledge structures that are only relevant evidence I will first discuss the earliest accessed later in the production process, seems attempts to apply Levelt’s (1989) model to preferable. But the point made by De Bot and bilingual speech production, and next zoom in on Schreuder—that languages differ in the way they the lexicalization process in some more detail. carve up conceptual space in their vocabulary—is an important one and will have to be addressed Levelt’s model applied to bilingualism in any model of bilingual language production that strives for completeness. De Bot (1992) and De Bot and Schreuder (1993) were the first to extend the insights obtained A second issue these authors addressed was in monolingual speech production research to how language selection comes about: How do bilingualism, building on Levelt’s (1989) model. bilinguals manage to produce speech according One of the facts about language that a bilingual to their choice—that is, to produce (relatively) production model should deal with is that lan- pure monolingual speech in either of their two guages differ in the way they lexicalize the con- languages—or to code-switch freely between the ceptual information in the preverbal message. A two, depending on their current intentions? What particular subset of the preverbal message’s con- adjustment to the monolingual production ceptual components may be expressed in a single system could account for this flexible behavior of word in one language but may require a whole the bilingual speaker? The authors’ solution to phrase to be expressed in another language. This guarantee monolingual output was to assume is one of the issues De Bot and Schreuder (1993) that the decision about what language to speak is attempted to solve. Their solution involved an determined by the conceptualizer, because only this component of the production system has access to the knowledge relevant for making a

5. WORD PRODUCTION AND SPEECH ACCENTS 229 language choice (e.g., knowledge regarding the seminal work—the language cue as part of the present discourse and what languages the inter- conceptual structure (the preverbal message) and locutors master). The information representing some piece of information specifying the lan- this choice is one component of the conceptual- guage membership of a lexical item (the language izer’s output—that is, of the (language- tag)—have since become more generally accepted nonspecific) preverbal message—it is called the in models of bilingual speech production. A language cue. In addition the authors assumed number of the studies that imply these constructs that the semantic information within each lemma will be discussed further on in this chapter. includes the knowledge of what language the Chapter 6, which is exclusively dedicated to the lemma belongs to (this piece of information is way bilinguals select and control their two often called the “language tag”). With this set-up, languages, covers studies that incorporate these the match between the preverbal message and the concepts more explicitly. In addition to the semantic information in the lemma of the target present view of language selection in terms of a language will generally be larger than between language cue, other views on how language the former and the semantic information in the selection comes about will be reviewed there as translation-equivalent lemma in the non-target well. language. As a result, the targeted lemma will generally become more highly activated than the THE CORE OF SPEECH PRODUCTION lemma of its translation equivalent so that MODELS: LEXICALIZATION the words that exit the production system will generally be words of the selected language. As mentioned, much research on speech pro- duction has focused on the lexicalization com- But what about a situation in which language ponent of the complete production process and mixing does not impede communication or is specifically on lemma selection and phonological even considered natural and desirable, for encoding. Because of this focus, visual illustra- instance because the addressee is fluent in the tions of the speech production process usually same two languages as the speaker and both share zoom in on the lexicon, ignoring the broader con- the same bilingual social background? De Bot text of the complete process. An illustration of (1992) adopted the earlier idea (Green, 1986) how this is done in reports of bilingual speech that bilinguals generate two speech plans simul- production studies is presented in Figure 5.2, taneously, one for the selected language and one which I have borrowed, in adapted form, from for the other language, which is also active. Kroll, Sumutka, and Schwartz (2005). The figure Because this way the language not currently in depicts one possible pattern of activation that use is always present “in the background”, fore- may be established in bilingual memory when grounding this language with a language switch Dutch–English bilingual participants are asked can occur rapidly and effortlessly. De Bot and to name pictures in one of their languages, here Schreuder (1993) proposed a different solution, English. In addition to the picture to be named, it namely that the language cue in the preverbal shows the three levels of representations com- message may be assigned a different weight in dif- monly assumed in speech production models: the ferent situations: If switching languages would prelexical conceptual level and the lexical lemma incur the risk of a communicative breakdown, the and phonological form (lexeme) levels. language cue may be assigned a high value, thus ensuring it will increase the activation level of Picture naming is the task used most fre- the targeted lemma. In a context where language quently in word production research because switching is quite common, would not hinder the outcome of the picture analysis process is communication, and would be experienced as regarded similar to the output of the mental quite natural, the language cue may receive a low conceptualization process in natural speech value, thus allowing mixed output to occur. Two of the theoretical constructs in this

230 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS A model of picture naming in aspect (or “conceptual component” or “con- bilinguals. The (Dutch– ceptual feature”) of the targeted word (e.g., English) participants’ task is –animate or +artifact). In terms of Levelt’s to name the picture in English (1989) model, together with the language cue (stoel = chair; fiets = bike; hypothesized by De Bot (1992), De Bot and brommer = motorcycle). In Schreuder (1993) and others (e.g., Poulisse, 1997; this example model lemma Poulisse & Bongaerts, 1994), these activated activation is language- conceptual components constitute the preverbal nonselective and message. This type of representation, where a phonological activation is word’s meaning is spread out over a number of language-selective. Circles more elementary meaning units, is called “dis- represent memory nodes. tributed” or “componential”, or “decomposed” Black nodes are highly (see p. 133 and Figures 3.8 and 3.11b, Chapter 3, activated; grey nodes are for a similar model), weakly activated; unfilled nodes are totally deactivated. The activated components at the conceptual Adapted from Kroll et al. level feed their activation down to the lemma (2005). level, along unidirectional links. In this specific model, at the lemma level each word is repre- production, and this output is thought to set off sented in a single, “non-decomposed” unit (see, the remainder of the production process in the e.g., the monolingual models of Jescheniak & same way in picture naming and common speech Levelt, 1994, and Roelofs, 1992, 1997, and the production. In brief, the full picture-naming bilingual model of Poulisse, 1997, and Poulisse & process is assumed to consist of the following chain of operations: The computation of the visual percept, the activation of an appropriate lexical concept, the selection of the target word from the mental lexicon, phonological encoding, phonetic encoding, and the initiation of articula- tion (Levelt, Praamsma, Meyer, Helenius, & Salmelin, 1998). In the sections to follow, these steps, and especially those in between visual per- cept computation and phonetic encoding, will be fleshed out in more detail. Just as conceptualizing a word to speak out loud, the first stage of the picture-naming process as sketched in Figure 5.2, the perceptual analysis of the picture, results in the activation of a subset of conceptual memory units usually called “nodes” (the filled circles are the activated nodes). Each of the activated nodes specifies one meaning

5. WORD PRODUCTION AND SPEECH ACCENTS 231 Bongaerts, 1994). Such type of representation, for consists of a set of semantic components, each of which a one-to-one mapping holds between a them representing one aspect of the word’s mean- linguistic entity to be represented (here, a lemma) ing. Recall that Levelt’s (1989) original model and the representing memory structure, is often assumed that a preverbal message, consisting of a called a “localist” representation (see also p. 132). set of activated conceptual elements, triggered Figure 5.2 depicts a situation in which lemmas for lemmas whose meaning specification matched words semantically related to the targeted word this set into activity. The componential view of are also activated, but to a lesser degree than the lexical semantics provides a natural explanation target word’s lemma. Furthermore, activation is of this triggering process, because the activated not restricted to the targeted language (the one (prelexical) conceptual elements may map directly the participants are asked to use while naming the onto the lexical-semantic components. It also pictures) but lemmas in the non-target language provides a ready explanation of the fact that not are activated as well. In other words, lemma one lemma but a set of semantically related activation is hypothesized to be language- lemmas are triggered, because the (componential) nonselective. From the lemma level activation meaning part of semantically related lemmas will spreads further down, again along unidirectional all share a subset of their meaning components links, to the phonological level, where a word’s with the information content of the preverbal phonological word form is represented in a message. Consequently, they will all be directly pattern of activation over a number of nodes each triggered by the preverbal message. The degree to representing a part of its sound, for instance, one which each of them is activated depends on the phoneme. Figure 5.2 depicts the situation where degree of overlap between preverbal message and this only happens for the most highly activated the lexical-semantic content of the lemma. lemma in the target language and for none of the activated lemmas in the non-target language. In In models that assume the lemmas’ meaning other words, in the depicted model phonological representations to be non-decomposed, multiple encoding is language-selective and it exemplifies activation of semantically related lemmas is more the discrete two-stage models introduced earlier. readily explained in terms of spreading activation in a lexical-semantic network in which the The mixture of representation and processing lemmas of semantically related words are con- assumptions assumed in Figure 5.2 is just one of nected to one another: The activated conceptual a larger number of possible sets of assumptions. elements in the preverbal message trigger one For instance, as we have seen, instead of acti- lemma into activity and from this lemma, acti- vation only spreading forward along uni- vation spreads through the network to those of directional links, some monolingual models semantically related lemmas. According to this assume forward as well as backward spreading of account, the lemmas of words semantically activation, a state of affairs that presupposes related to the target word are thus not directly bidirectional links between the units in adjacent activated by the preverbal message, but indirectly, representation levels (Dell, 1986). Roelofs’ following prior activation of the target word’s (1997) WEAVER model assumes a mixture of lemma (see Levelt, 1999, for a more complete bidirectional and unidirectional connections, and discussion of the contrast between decomposed similar proposals have been made for bilingual and non-decomposed lexical semantics). production models (e.g., Costa, Santesteban, & Caño, 2005). However, assuming both the existence of prelexical conceptual elements as well as lexical- Furthermore, instead of non-decomposed semantic conceptual elements, the latter being representations at the lemma level, some authors part of a lexical entry’s lemma, seems to involve (e.g., Bierwisch & Schreuder, 1992; Butterworth, an unattractive doubling of memory components. 1989; Dell, 1986) assumed that the meaning part This may be the reason why in more recent work, of a lexical entry’s lemma is “componential” or instead of containing both the lexical item’s “decomposed”. As mentioned, this means that it semantic and syntactic information (as in Levelt,

232 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS 1989), the lemma consists of the associated models is that the access of a word’s phonological word’s syntactic information only (see e.g., Bock form is mediated by the access of the word’s & Levelt, 1994; Levelt, 1999, Levelt, Roelofs, & lemma and the associated syntactic features: Meyer, 1999; Roelofs, 1992, 1997). In these Activation from the conceptual level spreads to models the upper layer in the visualized models the lemma level where syntax is specified, and (e.g., the conceptual level in Figure 5.2) no longer from there to the lexeme level (the phonological depicts a prelexical representation level but level), where the phonological form is specified. lexical semantics instead, and all processing Caramazza (1997) called this assumption the between the perceptual processing of the picture “syntactic mediation hypothesis”. and lexical-semantic activation remains unspeci- fied (if further intermediate processing occurs Caramazza (1997) proposed a model, the at all). Figure 5.3 shows a (monolingual English) independent network model, which is radically version of this type of model, with non- different from those reviewed so far. In this decomposed meaning representations organized model, not only do lemmas no longer contain in a semantic network and spreading activation semantic information, but they are dismissed between them. Because the concepts semantically altogether (see Starreveld & La Heij, 1995, for related to the targeted concept are only indirectly a similar proposal). His main reason to get rid activated, through spreading activation, and, of lemmas is that the results of a diverse set therefore, relatively weakly activated, the corre- of cognitive-neuropsychological studies point sponding lemmas are also activated less than the towards the conclusion that access to a word’s targeted lemma. A core assumption in these phonology is independent from access to syn- tactic information, thus falsifying the syntactic A model of picture naming in monolinguals. The (English) participants’ task is to name the picture in English. The lexical conceptual representations are non- componential and organized in a semantic network. Concepts semantically related to the targeted concept are indirectly activated through spreading activation in the semantic network.

5. WORD PRODUCTION AND SPEECH ACCENTS 233 mediation hypothesis. In other words, a word’s exist between the phonological (and ortho- syntactic make-up is not prepared prior to, but graphic) lexemes and the syntactic features net- independent from, specifying its phonology. work. This set-up, illustrated in Figure 5.4 is Caramazza then argues that the original moti- called “independent” because the activation of vation to postulate a separate lemma level was the lexemes is no longer mediated by syntax. the assumed dependence of lexeme activation on (Note that the orthographic component of the the representations that specify syntax; that is, the model and its connections with other parts of the lemmas. If this dependence does not hold and, at system are not shown.) As can be seen, the model the same time, lemmas are devoid of meaning, assumes componential instead of localist word why then would lemma representations exist in meanings. How these lexical meaning com- the first place? The only role a lemma would ponents are activated as a result of prior con- continue to play is to serve as a “way station to ceptual processing remains unspecified, as was syntactic and phonological representations” and also the case for the “lemma-dependent” models “the lemma node would have been rendered that only assign syntax, not meaning, to the superfluous” (Caramazza, 1997, p. 188). lemmas. The alternative proposed by Caramazza is a Considering these more recent views, there production process in which activation from appears to be a second possible reading of Figure a lexical-semantic network flows on directly to 5.2, one that is crucially different from the one both phonological (and orthographic) lexemes advanced above: Also here, the conceptual level and to a network of syntactic features, and may represent lexical semantics instead of, as from the phonological lexemes on to segmental assumed before, a set of activated elements in a phonological information (e.g., units that repre- (non-linguistic) conceptual system which acti- sent phonemes). In addition, direct connections vates the lemma’s meaning part (the latter no The independent network model of lexical access illustrated for French. Access of a word’s phonology is independent from syntactic access and no lemma representations are assumed. Adapted from Caramazza (1997).

234 LANGUAGE AND COGNITION IN BILINGUALS AND MULTILINGUALS longer being stored in the lemma). In other words, sion of close similarity to another model thus the units at the conceptual level may either repre- carries a serious risk of biasing a reader towards a sent prelexical conceptual knowledge or lexical- completely wrong interpretation of a model and semantic knowledge. Phrased in yet a different the data that led to it. manner, the three visualized representation levels may represent lexical knowledge solely (meaning, On the other hand, the opposite may occur as syntax, and phonology) or a mixture of prelexical well: Visual illustrations of models that look and lexical knowledge. rather different at first sight do not necessarily imply radically different assumptions regarding In reports of monolingual and bilingual representation and processing. For instance, the speech production research the exact assumptions number of representation layers shown is often regarding representation and processing are often not an indicator of the number of processing not explicated but taken for granted. This does steps assumed by the modeler, but simply a way to not just involve the risk of uncertainty about focus the reader’s attention on the steps that are how to interpret a particular illustration of the the object of study in a particular investigation. production process (as in Figure 5.2). A further To give one example, because in many bilingual risk is that two similar-looking illustrations, in studies the way word meaning maps onto word fact representing two different views on represen- form has been the main object of study, visual tation and processing in word production, may renderings of the underlying representation and mistakenly be interpreted as representing one and processing assumptions often only display two the same type of model. To illustrate this point, representational levels, one representing word Figure 5.5 depicts the representation and pro- meaning and a second representing the word’s cessing assumptions that Costa, Caramazza, and complete form. A number of these models were Sebastián-Gallés (2000) hypothesized to hold for presented in Chapter 3, when discussing the picture naming by bilinguals, in this specific case revised hierarchical model of bilingual memory Spanish–Catalan bilinguals. This figure illustrates and its precursors (see pp. 129–135). It is only the production process hypothesized for words because these models do not pretend to have any- that have a cognate- or a non-cognate translation thing to say about the representation of, say, equivalent in the other language (upper and lower syntax or sublexical units such as phonemes or panels, respectively). letters, that no further levels are visualized. The research designed to test the hypotheses in Conceptual knowledge and word question will be discussed later (p. 246). The point meaning to make here is that, at first sight, these figures look very similar to Figure 5.2 and therefore To conclude this section, a brief discussion is in suggest they represent the same levels of represen- order on a research dispute that arguably under- tation but labeled differently. For instance, the lies much of this ambiguity and especially the two lemma level in Figure 5.2 appears to be the same possible readings of the conceptual level of repre- as the lexical nodes level in Figure 5.5. But, on the sentation in Figure 5.2: the one in terms of lexical contrary, the models depicted in the latter illustra- semantics (word meaning) versus the one in terms tion represent bilingual versions of Caramazza’s of the preverbal conceptual structure. The dis- (1997) independent network model (Figure 5.4): pute in question is whether or not word meanings The semantic nodes represent Caramazza’s exist separately from more general conceptual lexical-semantic network and the lexical and knowledge. Levelt’s original model assumes sublexical nodes represent the phonological such to be the case, and other researchers have lexemes and sublexical phonological nodes in advanced this view as well. For instance, the dis- Caramazza’s model, respectively. As we have tinction between lexical semantics and general, seen, this model (which dismisses the existence of non-linguistic, conceptual knowledge forms the lemmas) is radically different from the majority basis of the three-store hypothesis of Paradis of the models discussed above. The first impres-

5. WORD PRODUCTION AND SPEECH ACCENTS 235 A model of picture naming in bilinguals. The (Catalan– Spanish) participants’ task is to name the pictures in Spanish. The upper and lower panels illustrate the memory representations of the Catalan–Spanish cognate pair gat–gato (“cat”) and non-cognate pair taula– mesa (“table”), respectively. Lexical and sublexical activation are both language- nonselective. Lexical nodes represent phonological lexemes. From Costa et al. (2000). Copyright © 2000 American Psychological Society. (e.g., Paradis, 2004), and, as we have seen, De Bot of the distinction between lexical semantics and and Schreuder (1993) also distinguished between non-linguistic conceptual knowledge is that con- the two (and see also Pavlenko, 1999). In terms of cepts exist that are not lexicalized. the above discussion, all theorists who assume (1) that lemmas contain meaning and (2) that Francis (1999, 2005) advanced the following lemmas are triggered by some match between parsimonious view of the relationship between prelexical conceptual information and the general conceptual knowledge and lexical- lemmas’ meaning components appear to hold semantic concepts: All humans possess a (non- this view. An argument in favor of the validity linguistic) conceptual system consisting of a large set of conceptual elements. Semantics comes into