A major benefit of computational models is their ability to demonstrate which theoretical assumptions are truly necessary to explain a pattern of data. Dijkstra, Wahl, Buytenhuijs, van Halem, Al-jibouri, de Korte, and Rekké (Reference Dijkstra, Wahl, Buytenhuijs, van Halem, Al-jibouri, de Korte and Rekkéin press) have impressively shown with Multilink that it is possible to account for a range of findings from bilingual lexical decision, word naming, and forward and backward translation tasks with an integrated lexicon, without lateral connections between translation equivalents, and without inhibition. In this commentary, we consider the applicability of the current model to other multilingual language production tasks, and note where the model's assumptions might need revision as its scope is expanded.
Language mixing effects in picture naming
Much evidence suggests that when bilinguals use both languages (e.g., in picture naming or verbal fluency), usage of the non-dominant language hinders dominant language production more than vice-versa; as recently demonstrated, this effect is cumulative (Kleinman & Gollan, Reference Kleinman and Gollan2018). To account for it, adjustments to Multilink seem necessary – likely, the existence of at least some lateral connections, though it is a matter of debate whether these connections exist only between translation equivalents, between synonyms as well, or between all words in a speaker's lexicon. We also note that a similar mechanism may be required to simulate within-language production, as naming a picture with a subordinate-level name (dalmatian) slows its subsequent naming with a basic-level name (dog; Ivanova, Reference Ivanovain preparation).
On the other hand, inhibition of any kind seems problematic for the facilitatory effect of translation equivalent distractors in picture-word interference studies (Costa, Miozzo & Caramazza, Reference Costa, Miozzo and Caramazza1999). Currently, Multilink nicely accounts for this finding by avoiding lateral connections between translation equivalents but instead connecting them only through their shared concepts (although the model might need adjustments to account for other picture-word interference effects).
Semantic effects in word translation
In Kroll and Stewart's (Reference Kroll and Stewart1994) study, proficient bilinguals’ L1-to-L2 word translation latencies were slower when items were blocked by semantic category versus presented in a randomized order, while L2-to-L1 translation latencies were unaffected. A comparable asymmetry emerged when monolinguals “translated” between low- and high-frequency words (e.g., purchase and buy): only high-frequency-to-low-frequency latencies, but not the reverse, were slowed by semantic blocking (Ibrahim, Cowell & Vorley, Reference Ibrahim, Cowell and Varley2017). Aspects of these data fit nicely with Multilink's simplifying assumption that word-strength differences can be represented in terms of word frequency both within and across languages. The semantic blocking effect could arise from the mechanism checking the activation of target semantic representations during translation, as increased activation of semantically related non-targets would make this checking process more difficult. However, the current model seems unable to account for the asymmetries. If the semantic checking mechanism were more engaged during translation when the potential output word were lower-frequency – which would account for these data – it would be more likely to miss phonologically activated high-frequency competitors during L2-to-L1 translation. At present, it is not clear to us how Multilink could best account for these results.
Language interference in bilingual and trilingual phoneme monitoring
The assumption of non-selective access seems to account well for evidence showing simultaneous activation of both a target name and a translation equivalent. In phoneme monitoring of target picture names, bilinguals took longer to reject phonemes from the picture name in the non-target language relative to phonemes not present in the name in any of their languages (Colomé, Reference Colomé2001).
An interesting extension of Multilink (although, we recognize, one that may need to await further evidence) would be to expand the integrated lexicon to three languages. For example, Tomoschuk, Duyck, Ferreira, Hartsuiker, and Gollan (Reference Tomoschuk, Duyck, Ferreira, Hartsuiker and Gollan2017) showed in a phoneme monitoring task that (1) cross-language interference (measured as proportion of incorrect acceptances of phonemes from non-target languages) was most pronounced when the phoneme came from trilinguals’ second language and when the target language was their least dominant language; and that (2) for bilinguals who produced novel-language words, the phoneme interference effect was smaller for the language with which they had associated the novel words than for the language with which no associations were formed. As translation equivalents are connected in Multilink only indirectly, there is currently no mechanism that would permit L3 production to receive more interference from L2 than from L1. To account for these asymmetries, inhibitory connections might be required, with stronger connections between representations paired during the learning process.
Conclusions
Overall, we think there is much to admire about Multilink. In its present form, it seems able to account for a range of bilingual production findings that fall beyond its current scope (translation facilitation effects in picture-word interference, bilingual phoneme monitoring, and (not discussed here) some cognate effects and the pattern of larger frequency effects in L2 than in L1). We have noted that changes to some of the model's assumptions might be necessary to account for asymmetrical effects of language dominance and frequency, and that many between-language effects would likely need to be paralleled within-language.
You have
Access
A major benefit of computational models is their ability to demonstrate which theoretical assumptions are truly necessary to explain a pattern of data. Dijkstra, Wahl, Buytenhuijs, van Halem, Al-jibouri, de Korte, and Rekké (Reference Dijkstra, Wahl, Buytenhuijs, van Halem, Al-jibouri, de Korte and Rekkéin press) have impressively shown with Multilink that it is possible to account for a range of findings from bilingual lexical decision, word naming, and forward and backward translation tasks with an integrated lexicon, without lateral connections between translation equivalents, and without inhibition. In this commentary, we consider the applicability of the current model to other multilingual language production tasks, and note where the model's assumptions might need revision as its scope is expanded.
Language mixing effects in picture naming
Much evidence suggests that when bilinguals use both languages (e.g., in picture naming or verbal fluency), usage of the non-dominant language hinders dominant language production more than vice-versa; as recently demonstrated, this effect is cumulative (Kleinman & Gollan, Reference Kleinman and Gollan2018). To account for it, adjustments to Multilink seem necessary – likely, the existence of at least some lateral connections, though it is a matter of debate whether these connections exist only between translation equivalents, between synonyms as well, or between all words in a speaker's lexicon. We also note that a similar mechanism may be required to simulate within-language production, as naming a picture with a subordinate-level name (dalmatian) slows its subsequent naming with a basic-level name (dog; Ivanova, Reference Ivanovain preparation).
On the other hand, inhibition of any kind seems problematic for the facilitatory effect of translation equivalent distractors in picture-word interference studies (Costa, Miozzo & Caramazza, Reference Costa, Miozzo and Caramazza1999). Currently, Multilink nicely accounts for this finding by avoiding lateral connections between translation equivalents but instead connecting them only through their shared concepts (although the model might need adjustments to account for other picture-word interference effects).
Semantic effects in word translation
In Kroll and Stewart's (Reference Kroll and Stewart1994) study, proficient bilinguals’ L1-to-L2 word translation latencies were slower when items were blocked by semantic category versus presented in a randomized order, while L2-to-L1 translation latencies were unaffected. A comparable asymmetry emerged when monolinguals “translated” between low- and high-frequency words (e.g., purchase and buy): only high-frequency-to-low-frequency latencies, but not the reverse, were slowed by semantic blocking (Ibrahim, Cowell & Vorley, Reference Ibrahim, Cowell and Varley2017). Aspects of these data fit nicely with Multilink's simplifying assumption that word-strength differences can be represented in terms of word frequency both within and across languages. The semantic blocking effect could arise from the mechanism checking the activation of target semantic representations during translation, as increased activation of semantically related non-targets would make this checking process more difficult. However, the current model seems unable to account for the asymmetries. If the semantic checking mechanism were more engaged during translation when the potential output word were lower-frequency – which would account for these data – it would be more likely to miss phonologically activated high-frequency competitors during L2-to-L1 translation. At present, it is not clear to us how Multilink could best account for these results.
Language interference in bilingual and trilingual phoneme monitoring
The assumption of non-selective access seems to account well for evidence showing simultaneous activation of both a target name and a translation equivalent. In phoneme monitoring of target picture names, bilinguals took longer to reject phonemes from the picture name in the non-target language relative to phonemes not present in the name in any of their languages (Colomé, Reference Colomé2001).
An interesting extension of Multilink (although, we recognize, one that may need to await further evidence) would be to expand the integrated lexicon to three languages. For example, Tomoschuk, Duyck, Ferreira, Hartsuiker, and Gollan (Reference Tomoschuk, Duyck, Ferreira, Hartsuiker and Gollan2017) showed in a phoneme monitoring task that (1) cross-language interference (measured as proportion of incorrect acceptances of phonemes from non-target languages) was most pronounced when the phoneme came from trilinguals’ second language and when the target language was their least dominant language; and that (2) for bilinguals who produced novel-language words, the phoneme interference effect was smaller for the language with which they had associated the novel words than for the language with which no associations were formed. As translation equivalents are connected in Multilink only indirectly, there is currently no mechanism that would permit L3 production to receive more interference from L2 than from L1. To account for these asymmetries, inhibitory connections might be required, with stronger connections between representations paired during the learning process.
Conclusions
Overall, we think there is much to admire about Multilink. In its present form, it seems able to account for a range of bilingual production findings that fall beyond its current scope (translation facilitation effects in picture-word interference, bilingual phoneme monitoring, and (not discussed here) some cognate effects and the pattern of larger frequency effects in L2 than in L1). We have noted that changes to some of the model's assumptions might be necessary to account for asymmetrical effects of language dominance and frequency, and that many between-language effects would likely need to be paralleled within-language.