A Bilingual Language Production Model

Naoko Tomioka

McGill University


The attempt to adapt a monolingual version of a language production model to bilingual language processing is not new. De Bot (1992) was the first to postulate a bilingual language production model based on Levelt’s (1989) model for monolinguals. A bilingual language production model should not qualitatively differ from the monolingual model, and yet has to be able to account for the phenomena observed in second language production. This paper presents a new model of language production based on Levelt’s (1989, 1995) model. By adapting an integrated speaking model like Levelt’s, the generalized relation between two language systems can be explored.

The subsystems hypothesis assumes that each language system constitutes a subsystem of the language system, as opposed to an extended system in which both languages would be stored together (Paradis, 1981). De Bot (1992) adheres to the subsystems hypothesis framework for the lexicon and the formulator, but assumes what amounts to an extended system for the articulator. In this paper, de Bot’s rationale for not using the subsystems hypothesis in the case of the articulator is examined and it will be shown that the phenomena invoked do not, in fact, favor the extended system hypothesis over the subsystems hypothesis. Given that both models can account for the same phenomena, the model with a generalized relation between the two language systems is better than a model with different relations holding at different levels of language processing.

Levelt’s model of language production

            Levelt (1989, 1995) proposes a model of language production. His model is unique in that it attempts to integrate independent, automatic modules into a complete ‘speaking’ system. His theory differs from most of the processing theories of bilinguals which only involve the lexicon (e.g., Kroll and de Groot, 1997; Green, 1998). In Levelt’s model, five modules involved in language production (conceptualizer, lexicon, formulator, monitor system and articulator) are put together. (Fig. 1)

            The intention to speak activates the conceptualizer, where the message is refined: The speaker takes into consideration discourse constraints, pragmatic factors, and the portion of the message to be realized para-linguistically (e.g., gestures) and linguistically.

            The preverbal message activates the formulator. Lexical items (lemmas) are activated on the basis of their semantic features. Once they are active, their syntactic and morphological properties are available and the formulator encodes them into sentences. Once the lemmas are put in correct order (grammatical encoding), their phonological forms (lexemes) are retrieved and encoded (phonological encoding).

            The output of the formulator is phonological form. It is acoustic in nature. The phonological form is sent to two parts of the system: The monitor and the articulator. The acoustic signals need to be converted into articulatory movements in order for them to come out as speech. The articulator is responsible for this conversion.

            The monitor system provides a direct link from the production system to the comprehension system and thus allows the speaker to evaluate the output of the formulator (inner speech). This is the only feedback system in his model. The extended use of the monitoring system in the L2 production is discussed below.

de Bot’s model of language production

            De Bot (1992, 2000) considers the bilingual version of the model as a default system, with the possibility of using it for only one language when a person remains monolingual. Thus, it is crucial that the bilingual version of the model should not qualitatively differ from the monolingual version.

            In conformity with the three-store hypothesis that separates the conceptual from the linguistic system (Paradis, 1980), de Bot postulates one common conceptualizer, responsible for the choice of language as well as for the formulation of the message to be verbalized in either language. (Fig. 2) The independence of the conceptualizer of the language systems is supported by the theory of bilingual lexicons (Kroll and de Groot, 1997). Following the assumptions of theories of bilinguals’ lexical systems, de Bot adopts the subsystems hypothesis for the lexicon. There is one lexicon which consists of two language specific sub-lexica. He postulates two formulators to accommodate the two lexica, each of which contains language-specific morphosyntactic information. Then he goes on to propose an extended system for the articulator. He argues that bilinguals normally have a ‘foreign accent’ in their second language because they possess only one articulator without a systematic division for the two languages.

The aim of the revised model of bilingual language production

In de Bot’s model, the type of organization between the two language systems differs from one level to another. At the levels of the lexicon and the formulator, their organization is what is characterized in the subsystems hypothesis. At the level of the articulator, however, the two languages are part of an extended linguistic system and there is no systematic distinction between the two. De Bot assumes that this level-based organization between the two language systems is necessary to account for the observed phenomenon of foreign accent. He conceives a foreign accent as representing the functional influence of the L1 on L2 production, thus an extended system is required to account for this usual interference.

            If a model can employ one type of organization between the two language systems regardless of the functioning levels, the model is simpler and thus should be preferred over a model with different types of organization holding at each level. In the following sections, I will present arguments for a simpler model. As for the general relation between the two language systems, I will adopt the subsystems hypothesis. This is the framework generally adopted for bilingual lexicon theories (e.g. Green, 1998) and is also the only theory (among extended, tripartite and dual system frameworks of bilingual language systems) that is compatible with all the well-known bilingual aphasics’ recovery patterns (Paradis, 1981, 2001). I will then show that the phenomenon of foreign accent can be accounted for within the subsystems hypothesis framework.

The Subsystems Hypothesis

The subsystems hypothesis (Paradis, 2001) proposes that the linguistic system is a collection of modules (e.g., phonology, morphosyntax, lexicon), and that each language system (e.g. English, and Japanese) constitutes a subsystem of the larger linguistic system. Hence each module can be selectively activated or inhibited. (Figs. 3-10) Speaking in the monolingual mode involves activation of one language system and inhibiting the other (cf. Grosjean, 2001; Green, 1998). Brain damage can bring about inhibition of specific modules (syntax, morphology, phonology, or lexicon of a specific language) or all of the language systems.

In an extended system, there is no systematic division between each language system. Items of L2 are stored among items of L1. Thus, as described in de Bot (1992), extensive functional interference occurs. In this framework, however, it is not clear how one can ever manage to use a language without interference from another (or others).

In a tripartite system, items of L1 and L2 that are identical (cognates) are stored as one item, but different items are stored separately in each system. The language system can be considered as a set of items. The items that are in the intersection of the L1 and L2 systems are stored only once. This differs from the subsystems hypothesis in which language-membership is considered to be exclusive. In the subsystems hypothesis framework, if  an item is stored as part of L1, it cannot be part of the L2 system. Thus, identical items in L1 and L2 are stored twice. The tripartite framework cannot explain why an item in a language cannot be accessed while its cognate in another language can be accessed in bilingual aphasia.

The dual systems hypothesis postulates that there are separate systems for each language, just as music and language are stored separately. Unlike other hypotheses of language organization, code-switching, and parallel recovery pattern of bilingual aphasics cannot be easily explained in this system.

All recovery patterns of bilingual aphasics are compatible with the subsystems hypothesis, but not with the extended, dual, or tripartite system. Thus, the bilingual language model should adopt the subsystems hypothesis.

A bilingual language model that states that the relationship between the two language systems is that of a subsystem is superior to a theory that states that, at level X (lexicon, syntax, or phonology), the organization of the two language systems is that of an extended system, at level Y, tripartite, and at level Z, subsystems. Given the general view that the subsystems hypothesis is supported at the level of lexicon, then we should, by null hypothesis, assume that the subsystems hypothesis holds at every level of the linguistic system.

            In accordance with the three-store hypothesis (Paradis, 1980) that recognizes one common conceptual store external to the linguistic system, the conceptualizer is not part of the linguistic system. Thus, while the two languages of a bilingual speaker are assumed to form subsystems in the linguistic system, a speaker nevertheless has only one conceptualizer.

the revised model

            As stated above, as far as the subsystems hypothesis can account for all the observed facts of bilingual production, we should choose the subsystems hypothesis over the extended or tripartite systems hypotheses. (Fig. 11)

            The fact that bilingual speakers often have a foreign accent in one of their languages (L2) led de Bot (1992) to favor what amounts to an extended system for the articulator. He argues that a system that assumes functional independence of the two language systems (such as the subsystems hypothesis) cannot account for the phenomenon of L1 interference on L2 such as foreign accent.

The fact that the dominant language (generally L1) can influence the L2 may be conceived of in two ways: Functional (dynamic) interference and representational (static) interference (Paradis, 1993). The former assumes that the use of L2 is disrupted by the activation of L1 items. The latter occurs at the point of acquisition. Bilinguals’ representation of L2 contains some (illegitimate) traits of their L1. Thus, these traits of L1 may be observed in L2 production even when the L1 system is functionally inhibited.

An argument that may support the functional interference account comes from the theory of performance error. The awareness of one’s own errors is a characteristic of a performance error (Baars, 1992). Thus, the fact that bilinguals are often aware of their errors during L2 production may be taken as evidence for the functional interference account.  One may assume that L2 users’ awareness of their error reflects their competence. However, the awareness of one’s own error does not necessarily reflect one’s competence. As shown in Levelt’s model (1989, 1995), there is a system that allows the speaker to check the output of the formulator (the monitor). Levelt’s monitor may be interpreted as a system that allows speakers to compare their production and what would be correct according to their meta-linguistic knowledge (cf. Krashen (1977) for this view of monitoring). That is, the awareness of one’s own errors may reflect one’s meta-linguistic knowledge rather than one’s competence. This greater reliance on meta-linguistic knowledge receives support form  neuroimaging studies, in which activation of hippocampal structures (Perani et al., 1996, 1998) was observed in the L2 processing task, but not in the L1. Hippocampal structures are known to subserve the use of explicit knowledge, such as meta-linguistic knowledge. Thus, L2 users would recognize their own errors when they check their production against their meta-linguistic knowledge.

            De Bot (1992) presupposes functional interference to account for bilinguals’ production errors and thus, chooses to adopt a system which assumes an extended system for the articulator. However, if L1 has influenced the L2 at the level of representation, the two systems can be functionally independent and L2 output will still contain L1 traits. This is a more likely explanation for the foreign accent phenomena as it would account for such interference at any other level of linguistic structure (e.g. lexical, or syntactic). Functional interference is an error, just like speech errors observed in the native speakers of any given language and its occurrence should not be systematic. Systematic production of the deviant forms should be considered as the reflection of the nature of the representation.

            De Bot (1992) considered the phenomenon of foreign accent as evidence against the subsystems hypothesis account of the articulator. However, as shown above, a foreign accent can actually be accounted for within the subsystems hypothesis framework.


            Levelt’s integrated language production model can be adapted to bilingual language production. This integrated production model provides the possibility to develop a theory of generalized interaction of the two language systems in bilinguals, instead of postulating a different organization of the two languages at each level. Two observed facts that were taken as evidence against the subsystems hypothesis at the articulator level have been shown to be, in fact, compatible with the subsystems hypothesis. A bilingual language production model with consistent relations between the two language systems across levels is thus developed.