The paper opens with an evaluation of the BIA model of bilingual word recognition in the light of recent empirical evidence. After pointing out problems and omissions, a new model, called the BIA+, is proposed. Structurally, this new model extends the old one by adding phonological and semantic lexical representations to the available orthographic ones, and assigns a different role to the so-called language nodes. Furthermore, it makes a distinction between the effects of non-linguistic context (such as instruction and stimulus list composition) and linguistic context (such as the semantic and syntactic effects of sentence context), based on a distinction between the word identification system itself and a task/decision system that regulates control. At the end of the paper, the generalizability of the BIA+ model to different tasks and modalities is discussed.

It is not easy to comment on Dijkstra and Van Heuven's model because there are many more aspects we agree with than aspects we feel uncomfortable about. Indeed, the BIA model has played an enormous role in showing us how bilingual visual word recognition can be achieved without recurrence to the intuitively appealing – but wrong – ideas of separate, language-specific lexicons and language-selective access. As in many other research areas, a working computational model has been much more influential in convincing critical readers (and researchers) than any series of empirical findings. The BIA+ model inherits this strength and, hopefully, in the coming years will be implemented in enough detail to exceed its predecessor. In the rest of this comment, we would like to put a cautionary note behind the temporal delay assumption introduced in the target article and provide some additional corroborating evidence for the lack of non-linguistic effects on early processes in the identification system.

Dijkstra and van Heuven have made an admirable attempt to develop a new model of bilingual memory, the BIA+. Their article presents a clear and well-reasoned theoretical justification of their model, followed by a description of their model. The BIA+ is, as the name implies, an extension of the Bilingual Interactive Activation (BIA) model (Dijkstra and van Heuven, 1998; Van Heuven, Dijkstra and Grainger, 1998; etc), which was itself an adaptation to bilingual memory of McClelland and Rumelhart's (1981) Interactive Activation model of monolingual memory.

Dijkstra and van Heuven lucidly summarize the important research generated by the BIA model and provide an excellent case for the BIA+ model with its critical separation of the identification system from the task/decision system. A keynote article necessarily offers a selective exposition of the authors' thinking and so my remarks are an invitation to expand. My first question concerns the scope of the BIA+ model. My remaining questions broadly address a key feature of the BIA+ model – its ability to explain performance changes as a function of task demands

Central questions in psycholinguistic studies on bilingualism are how bilinguals access words in their two languages, and how they control their language systems and solve the problem of cross-language competition. In their excellent paper “The architecture of the bilingual word recognition system: From identification to decision”, Dijkstra and Van Heuven expound their BIA+ model on bilingual word recognition. BIA+ builds on its predecessor BIA, one of the first connectionist models on bilingual word recognition. BIA+ preserves one of BIA's crucial assumptions, namely that the bilingual lexicon is integrated across languages and is accessed in a language non-selective way, an assumption that is supported in many empirical studies and that is now widely accepted in the bilingual literature. Compared to the original BIA model, the BIA+ architecture is further developed (in fact, much more so than the subtle ‘plus’ denotes). BIA+ now includes orthographic, as well as phonological and semantic representations in the word identification system, and a distinction is made between a word identification system and a task/decision system. This latter extension resembles the language task schemas in Green's (1998) Inhibitory Control model. Dijkstra and Van Heuven also distinguish between effects of linguistic and non-linguistic context on performance: linguistic context effects, that arise from lexical, syntactic and semantic sources, are assumed to affect the activity in the word identification system, whereas non-linguistic effects, that can arise from instruction, task demands or participant expectancies, are assumed to affect the task/decision system.

Each year the Cognitive Science Society honors David Rumelhart by awarding the Rumelhart Prize to an outstanding cognitive scientist whose research makes a significant contribution to the formal analysis of human cognition. Formal models of language, including those of Rumelhart and his associates, are well known to psycholinguists in the monolingual context. The formalism of language in the bilingual context, however, is lamentable: to this date, there are only a handful of models (connectionist or otherwise) that are designed specifically to account for bilingual language processing and acquisition (cf. Li and Farkas, 2002). BIA is one of them. BIA+ is now another.

Dijkstra and Van Heuven sketch the BIA+ model for visual word processing in bilinguals. BIA+ differs in a number of respects from its predecessor, BIA, the leading implemented model of bilingual visual word recognition. Notably, BIA+ contains a new processing component that deals with task demands. BIA+ has not been computationally implemented yet and design decisions still need to be taken. In this commentary, I outline a proposal for modeling the control of tasks in BIA+.

The target article represents a significant advance in the level of sophistication applied to models of bilingual word recognition, and Dijkstra and van Heuven are to be congratulated on this endeavour. Bearing in mind the success of the (computational) BIA model in capturing detailed patterns of experimental data, I look forward to future simulation results from the BIA+ when the proposals of this new framework are implemented. It is an essential step to draw a distinction between recognition systems and the decision mechanisms that drive responses, and the authors have provided a novel way of apportioning empirical evidence of context effects in bilingual word recognition across this divide. Given the explanatory weight now being placed on decision mechanisms rather than the word recognition system itself, perhaps indeed it is now time to make some simplifying assumptions about the recognition system and start building detailed computational models of the decision component of the system. Implementation will provide the clarity of theorisation and evaluation of theory viability that have been the hallmark of the BIA model thus far.

The detailed comments of our colleagues on our keynote article suggest that they all value the integration of various sorts of empirical data into one formalized model of bilingual word recognition. Taking this position, our peers propose further specifications or adaptations of the proposed BIA+ model with respect to its representations, processes or control structure; and they indicate what, in their view, should be the next step in model development or empirical testing. We are grateful for their contributions, which we will carefully think through in the coming time and use as a basis for future implementations.

This study investigated whether bilinguals simultaneously activate phonological representations from both of their languages when reading words in just one. The critical stimuli were interlingual homographs (e.g., PAIN) that were low in frequency in the target language of the study (English) and high in frequency in the nontarget language (French). Both English-French and French-English bilinguals were tested. In each experiment, participants named a block of English experimental words, a block of French filler words, and then a second block of English experimental words. In the first block of English trials, the English-French bilinguals had similar naming latencies for homographs and English-only control words, although they made more errors on homographs. In contrast, the French-English bilinguals showed a homograph disadvantage in both the latency and error data. In the second block of English trials, both the English-French bilinguals and the French-English bilinguals showed homograph interference on latency and error measures. We interpret these results as indicating that the activation of phonological representations can appear to be both language-specific and nonspecific, depending on the characteristics of the bilingual and whether they have recently named words in the nontarget language.

German-English bilinguals decided whether a visually presented word, either German or English, referred to an animate or to an inanimate entity. Bilinguals were slower to respond on a language switch trial than on language non-switch trials but only if they had to make the same response as on the prior trial (e.g., two successive “animate” responses). If they had to change their response from the prior trial, a switch of language speeded responses. This interaction between language switching and response type was abolished for the translations of words previously presented. We consider the implications for theories of language control.