duction, which is both in line with current theories of speech processing and accommodates the research findings that I described in the previous chapters. The model uses Levelt’s (1999a) blueprint for the speaker as a starting point, but some of its theoretical underpinnings are modified in order to account for the results of recent studies that suggest the possibility of cascading of activa- tion. In this bilingual speech production model, I not only incorporate L2 knowledge stores and processing systems but also make an attempt to explain how formulae are encoded and how speech production mechanisms are ac- quired. I first outline the general theoretical considerations that underlie the model, which are then followed by the presentation of the model. Finally, I also describe how transfer, code-switching, communication strategies, and the development of proficiency can be accommodated in this new bilingual speech production framework.
THE GENERAL CHARACTERISTICS OF THE BILINGUAL SPEECH PRODUCTION MODEL
The bilingual speech production model I propose is based on Levelt’s (1989, 1999a) theory of speech production because, as I argued in the Sum- mary section of chapter 2, this model is the empirically best supported theory of monolingual speech processing. Consequently, I assume that bilingual speech production is modular in the sense that it consists of separate encoding modules: the conceptualizer, the formulator, and the articulator, which work with their own characteristic input. It is postulated that similarly to L1 speech processing, L2 speech production can also work incrementally; that is, a frag- ment of a module’s characteristic input can trigger encoding procedures in this module. For example, once the first syllable of a word is phonologically en- coded, its articulation can start in the articulator. This also entails that for learners above a certain level of proficiency, parallel processing is theoreti- cally possible. However, as long as an encoding process requires conscious attentional control, encoding can only work serially. Nonetheless, this bilin-
gual speech production model is not a strictly serial model in the sense that the cascading of activation is allowed from the lexical to the phonological level. In other words, activated but not selected word nodes can pass on activation to lower level phonological nodes. On the other hand, the model does not permit the backward flow of activation between levels, and monitoring is done with the help of the speech comprehension system.
In Levelt’s (1999a) model, there are three knowledge stores: the store for the knowledge of external and internal world, the mental lexicon, and the syl- labary. Based on major theories of memory research (e.g., Tulving, 1972), I propose that the new model contains one large memory store, called long-term memory, which consists of several subcomponents: episodic memory, seman- tic memory including the mental lexicon, the syllabary, and a store for declara- tive knowledge of L2 rules (see Fig. 9.1). Semantic memory contains linguistic and nonlinguistic concepts as well as meaning-related memory traces associated with these concepts, whereas episodic memory is the store of temporally organized events or episodes experienced in one’s life. In order to account for findings of speech production research, semantic memory is as- sumed to have a hierarchical structure and consists of three levels: conceptual, lemma, and lexeme level. The lemma level contains syntactic information and the lexeme level morpho-phonological information related to lexical items. The syllabary stores the automatized gestural scores used to produce syllables. Based on empirical findings discussed in the earlier chapters of this book, it is hypothesized that all the knowledge stores described so far are shared between L1 and L2; in other words, there is a common episodic and semantic memory for L1 and L2, a shared store for L1 and L2 lemmas and lexemes, and for L1 and L2 articulatory scores. In L2 production, however, we need to postulate the existence of a fourth and L2 specific knowledge store: a declarative mem- ory of syntactic and phonological rules in L2. In L1 production, rules are as- sumed to be automatized and to be part of the encoding systems (Levelt, 1989). On the other hand, for bilingual speakers many of the phrase- and clause-build- ing as well as lexical and postlexical phonological rules are not automatic and are stored in the form of declarative knowledge. T. Ullman (2001) cited several pieces of evidence from neuroimaging research (for details, see the Theories of Automaticity and the Development of L2 Fluency section in chap. 8) that declarative knowledge concerning grammar is stored in a brain region distinct from the area that is responsible for the processing of automatized rules of grammar. Therefore, it seems to be justified that for L2 speakers a fourth knowledge store for not yet automatized syntactic and phonological rules is in- cluded in the model (see Fig. 9.1).
Episodic and semantic memory are closely related, which is indicated by adjacent circles in the figure. This means that episodic memories can activate concepts and vice versa. The hierarchical nature of the semantic memory en- tails that in speech production activation flows from the conceptual to the lemma and finally to the lexeme level, whereas in speech comprehension acti- vation flows in the opposite direction.
The model aims to follow the principle of ecology and simplicity, which is prevalent in human cognition. Therefore, except for the addition of a new knowledge store for the declarative knowledge of production rules and the in- corporation of L2 concepts, lemmas, lexemes, and syllable programs (gestural scores), the bilingual production model proposed here is not significantly differ- ent from a model constructed for monolingual speakers. Abutelebi et al.’s (2001, 2005) reviews of neuroimaging studies of L2 production seem to provide sup- port for the essentially similar nature of L1 and L2 speech processing. Abutelabi et al.’s meta-analyses of existing research in this field suggest that neither the ex- tent of brain activation nor the regions involved in processing in L1 and L2 are different for bilinguals who learned the L2 early in their lives and for highly pro- ficient speakers with extensive L2 exposure. However, late bilinguals, espe-
cially those who are not proficient in the L2 and have had low exposure to the target language, were found to activate larger and slightly different cerebral ar- eas when speaking in L2 than in L1. The model accounts for this finding because proficient bilinguals do not rely on the separate knowledge store of declarative rules, whereas for learners at lower stages of proficiency grammatical and pho- nological rules are stored in a separate region of the brain.
ENCODING MECHANISMS AND THE STRUCTURE OF KNOWLEDGE STORES IN L2 SPEECH PRODUCTION The processing of L2 speech starts with conceptualizing the message, which involves the activation of the relevant concepts to be encoded and deciding on the language in which the message will be spoken. As already mentioned in the preceding section, L1 and L2 concepts are assumed to be stored together in the semantic memory (see also Francis, 2005). In this model, a concept is seen to be a conglomerate of interrelated memory traces consisting of information concerning word meaning (see de Groot, 2000; Hintzman, 1986). When a con- cept is called on, not all the memory traces are activated; only the contextually relevant pieces of information become active (Hintzman, 1986). The assump- tion that concepts consist of a network of memory traces allows that L1 and L2 concepts can be identical, shared, or rarely completely separate. The extent to which L1 and L2 concepts are shared depends on the concept (e.g., concepts expressed by concrete nouns tend to be shared, whereas those expressed by ab- stract nouns show partial overlap), the situation in which the L2 was acquired (e.g., if the two languages were learned and are used in different environments, concepts might be separate), and the speaker’s level of proficiency (e.g., at the beginning level L2 concepts are completely mapped on L1 concepts, whereas at advanced levels the L2 conceptual representation is greatly enriched) (see de Groot’s, 1992, conceptual feature model in the Models of the Organization of the Bilingual Lexicon section in chap. 4).
The language of the message also needs to be set in the conceptualization phase. Language choice is largely dependent on sociolinguistic factors such as the nature of the communicative situation, relationship of the interlocutors, prestige of the languages involved, and so on. In this model, we opt for the most simple and economical solution, namely that language choice is indi- cated in the form of a language cue, which is added to the activated conceptual information. We have to emphasize that the language cue is added to each con- cept separately, and therefore it is possible that in the case of encoding a sen- tence, a preverbal plan consists of a string of activated concepts to which different language cues are added. For example, upon encoding the sentence,
“The policeman fined the motorist,” a German-English bilingual speaker might add a language cue + English to the concepts of POLICEMAN and MOTORIST, whereas the concept of FINE might receive a tag + German.
In line with Levelt’s (1989, 1999a) and Levelt et al.’s (1999) theory as well as with the spreading activation accounts of speech production (Dell, 1986; Dell & Juliano, 1996), this model also assumes that in semantic memory not only the concept that the speaker wants to encode is activated, but semantically related concepts also receive activation. For example, in the case of the con- cept of MOTHER, related concepts such as FATHER, CHILD, LOVE, and so on, also become active. This also entails that if separate concepts exist for the same notion in L1 and L2, which is a rare case, when the bilingual speaker uses one of his languages, the concept in the other language will also be activated. If conceptual representations are identical or partly overlap, the cohort of the conceptual features will be activated. The bilingual speech production model assumes that only the intended concept in the selected language is chosen for further processing (e.g., Bloem et al., 2004; Levelt, 1989). The selected con- cept activates not only the matching lexical item but also semantically related lemmas including lemmas in the nonselected language.
Remaining still at the phase of conceptualization, we also need to consider the problem that not every instance of language is creatively constructed. In fact, the majority of our utterances are combinations of memorized phrases, clauses, and sentences, which together are called formulaic language (Pawley & Syder, 1983). In order to account for the use of formulas, we need to assume that chunking, that is, the creation of larger production units, is done at the level of the conceptualizer. In other words, it is postulated that for expressing various communicative functions such as requesting, apologizing, expressing surprise, and so forth, native speakers have conceptual chunks consisting of a group of concepts, which they activate as one unit when routinely expressing certain communicative intentions. These conceptual chunks spread activation to the corresponding linguistic chunks, which are also stored and retrieved as one unit (i.e., one lemma).
In this bilingual speech production model, lexical encoding means the matching of the conceptual specifications and the language cue with the ap- propriate lexical entry in the mental lexicon. Based on empirical evidence summarized in the Lexical Activation and Selection in L2 section in chapter 4, the conceptual specifications send activation to both L1 and L2 lemmas (Costa et al., 2000; Hermans et al., 1998), and they both compete for selection (Costa, Colomé, et al., 2003; Hermans et al., 1998; Lee & Williams, 2001). The winner of the competition is the lemma whose features match all the conceptual speci-
fications including the language cue (La Heij, 2005; Poulisse, 1999; Poulisse & Bongaerts, 1994). Neither inhibitory nor additional checking mechanisms are believed to be necessary to control bilingual lexical encoding.
The mental lexicon contains L1 and L2 lemmas and lexemes; in other words, it is a depository of a speaker’s knowledge of word forms (lexemes) and their syntactic and morphological features (lemmas). The bilingual lexicon is assumed to consist of single L1 and L2 words as well as longer word sequences in L1 and L2 that correspond to conceptual chunks. These longer sequences can be idioms, conventionalized expressions, and phrases, which form a single entry and have their own syntactic information. Like the conceptual system, the lexicon is conceived of as a network in which entries have connections with each other. Connections might exist between L1 and L2 lemmas and lexemes, and between items within languages. Well-known and frequently used L2 en- tries occupy a central position in the network and have a high number of links with other items, whereas words not known very well by L2 speakers can be found at the periphery of the network (Wilks & Meara, 2002; Wolter, 2001). The strength of connections might also vary; at the beginning of the learning process links between L1 and L2 items might be stronger than links among L2 lexical entries (Kroll & Stewart, 1990, 1994). Moreover, connections might also be asymmetrical, which means that in certain cases it is possible that there is only a one-way link pointing for example from an L2 entry to an L1 item (passive vocabulary that one is able to recognize) (Meara, 1997).
Syntactic encoding in L1 production entails two important procedures: the activation of syntactic information related to a lexical item such as gender, countability status, and optional and obligatory complements, and the use of syntactic encoding mechanisms to assemble phrases and clauses using the ac- tivated words and their syntactic features. In the first phase, the L1 speaker re- lies on declarative knowledge, whereas the second stage involves applying procedural knowledge. We have rather limited knowledge of syntactic encod- ing in L2; therefore, some of the claims made in the bilingual speech produc- tion model are going to be speculative. In this model, I assume that as regards the general process of syntactic encoding there is no fundamental difference between L1 and L2 production, and that syntactic processing follows the steps of Kempen and Hoenkamp’s (1987) Incremental Procedural Grammar (see the Syntactic Processing section in chap. 2 and the General Overview section in chap. 5). This means that syntactic encoding is lexically driven and consists of distinct stages that follow each other.
The first major stage of the process is the activation of the syntactic proper- ties of the lemma that corresponds to the first conceptual chunk of the mes-