The MOGUL framework is a research framework aiming to provide a theoretical perspective on the nature of language. MOGUL (Modular On-line Growth and Use of Language) draws on the common ground underlying various related areas of cognitive science including psycholinguistics, theoretical linguistics, first- and second-language acquisition, neurolinguistics and cognitive psychology; it is designed to be applicable to all these fields of research.
The MOGUL framework's background assumption is that the mind is composed of expert systems which have evolved over time, one of which is responsible for human linguistic ability. Historically, scientific studies of language have been divided between many sub-disciplines; theoretical linguists focus on the abstract properties of language[1] and researchers in other fields investigate how language is used and processed in real time: either in psychological terms or (in the case of neurolinguistics) through a study of the physical systems of neurons in the brain. Each field of study has developed its own research traditions and technical vocabulary, making it difficult to integrate insights across disciplines. The MOGUL Framework represents an attempt to identify common themes and compatible approaches in different (but related fields), and hence to facilitate integration.
Contents |
MOGUL is designed as a platform for the generation of new hypotheses, with a terminology and a set of interrelated concepts that can be used across more than one discipline. Although not a full-fledged theory in its own right, it sets out various theoretical claims based on the current literature in a range of research domains. As suggested below it is not connectionist in the most widely-used sense of that term, although it employs terms and concepts familiar to connectionists such as competition, activation, activation levels and so forth. The same processing perspective also requires the long-established Chomskyan distinction between competence and performance to be viewed in a somewhat different way. Linguistic competence (knowledge) is not compartmentalised and placed in a separate box in the model, which then forms the object for a processor (or processors) to act on. Rather, it is instantiated and implicit within the processing system. The distinction between mental representation and processing is still valid, but is expressed in a way that ought to allow the abstract properties of linguistic systems in the mind to be more easily investigated in conjunction with real-time linguistic phenomena. The inspiration for this approach is the work of Ray Jackendoff. Here the architecture of the language faculty is formulated in such a way as to facilitate statements about competence-related, representational issues and their relationship to real-time language processing within the same model.
Empirical evidence for the claims and assumptions set out in MOGUL rests essentially on the experimental findings of the various fields of research on which it is based. The MOGUL framework queries current solutions and raises new questions based upon this body of evidence, in the hope that answers to such questions will be sought in the related areas by appropriate researchers.
More specifically, MOGUL represents an attempt to provide ways of investigating how first- or second-language processing relates to linguistic development and how language in the monolingual and multilingual language user is situated within human cognitive architecture in general. What exactly happens, even in milliseconds, when someone is exposed to the sights and sounds of a new language? How do we explain what gets noticed and what gets regularly ignored? What about the fixed stages that people seem to pass though? How is it that we have so little conscious control over what we are attempting to acquire? Why do children appear to be worse language learners in the short run and much better in the long run? How can two or more language systems cohabit in one mind? Answers to these questions require an interdisciplinary approach. Two major sources for inspiration for the approach described here are the ideas of generative linguist Ray Jackendoff on language faculty and modularity and Global Workspace Theory, as advanced by Bernard Baars.[2] The first MOGUL publication was released in 2004 and was a keynote article in Bilingualism: Language and Cognition (Truscott and Sharwood Smith 2004).[3]
In order to situate languages within a wider cognitive context, the MOGUL framework is designed to provide a basic architecture that purports to explain how (in general terms) the mind functions. This means that it is not an account of brain function, although it is set out in such a way that might facilitate such accounts.
Briefly, MOGUL architecture follows Jackendoff in positing a generic cognitive module (or, in MOGUL, a processing unit) which consists of an integrative processor operating on structural elements (structures) in a memory store. Specifically, it works with activated elements that are raised into working memory from that memory store.
Modules are only somewhat generic in structure; the rest is very specific to each module, since each module operates with its own code. Visual memory contains elements that only the visual processor can access; syntactic memory contains elements that only the syntactic processor can deal with, and so on. While Jackendoff prefers to think of working memory as a separate blackboard onto which elements of long-term memory are written, in MOGUL the alternative view of the working-memory blackboard is adopted: it is part of the unified memory system, and elements are raised to the working memory "surface" within the same store. In both versions, elements must cross an activation threshold to be visible to the integrative processor.[4]) Note that memory is modularised, so visual memory (including visual working memory) is separate from – and independent of – for example, phonological memory. Modules, useless if entirely autonomous and cut off from other parts of the system as a whole, require some way whereby they can link up and cooperate. This is done by means of interface processors matching structural elements in the memory stores of adjacent modules. The cognitive system as a whole copes with a multitude of different tasks like tying a knot, listening to a speech, driving through a busy city and salsa-dancing by forming appropriate coalitions of processors producing assemblies of structures (linked together via appropriate interfaces) from different modules. The generic structure of a MOGUL module (based on Jackendoff's model) is portrayed in the figure on the right, and also reflected in various places in the third figure in this article (see "MOGUL architecture in a nutshell" below).
This functional architecture may be neurally and anatomically instantiated in different ways: the design of systems such as that proposed by Jackendoff and others (including MOGUL) involves no specific claims about the precise nature and location of the neural system. These are two different levels of description. The auditory module (translated into cortical terms) involves several brain regions distributed across the supratemporal plane; activation in MOGUL terms (when translated into neurological terms) may involve both excitatory and inhibitory neurotransmitters.[5] By the same token, functional auditory structure (as described within the MOGUL framework) must be distinguished from anatomical and physiological auditory structure.
An important portion of the MOGUL account is devoted to the claim that the acquired (and atrophied, or lost) linguistic structures are natural byproducts of online processing. This means that acquisitional mechanisms do not exist as such, but are embodied in the operations of the parser. This claim is made explicit in Acquisition by Processing Theory (APT). Processing concepts (activation levels, competition and so on), as exploited by other non-modular, domain-specific approaches, are realigned in MOGUL to explain how new cognitive systems emerge. APT is a hypothesis put forward by Sharwood Smith and Truscott as an integral part of MOGUL and is applicable to all kinds of cognitive development, not only language acquisition and language attrition (for a similar approach acquisition which is framed within an emergentist perspective, see O'Grady 2005).[6]
Essentially, the claim is that all development is the lingering effect of processing. As the mind attempts to build mental representations online, various structures at its disposal are activated. Structures compete with one another to be selected for the current representation. As a simple example, on hearing the word "ship" an English speaker's processing system will activate various candidate structures so (for instance) the phonological structures underlying "sheep" and "shape" will compete for selection (and other candidates as well, including phonological structures belonging to other languages known to the listener).[7] The "ship" structure is normally selected as the best fit, and thereby its likelihood of being selected in the future is correspondingly strengthened by a small amount. In this way, a basic "use it or lose it" principle is invoked; this development principle works throughout the cognitive system as a whole.
As we match various types of cognitive structure available to us in order to find the best fit for unfamiliar input from the environment new connections are developed, initially with the relevant structures possessing a low resting level of activation. This means they will have a relatively poor chance of selection for future instances of the same input. However, the more they are selected the more they will show up in the observable behaviour of the individual concerned.
Although the frequency in which we experience given phenomena influences our development, cognitive growth is not an automatic consequence of experience; our mind is modular, and each module is controlled by its own unique processing principles. This limits, for instance, what we can learn to see or hear or say; seeing, hearing and speaking each involve dedicated processing units, which control their own internal operations. In this way, high-frequency events in the environment may still not impact development because the relevant parts of our mind are incapable of processing them. This may be because a) the relevant modules (processing units) are simply not designed to do so, or b) they are not yet ready to process them, because their internal principles require some prerequisite state of affairs before the potential new input can be integrated. By the same token, an internal operation may be halted because although one processing unit has processed input, an adjacent processing unit (module) with which it is connected is not in a ready state to cooperate (find matching structures within its own memory store). Hence frequent events in the external environment may indeed get processed by some part(s) of the cognitive system but still not by every relevant part. MOGUL espouses both modularity of mind and a constrained form of connectionism with a complex web of expert systems, each with its own particular organising principles. In this way, MOGUL makes claims not only about language abilities but about the mind in general.
The core modules involved in language comprehension work in two directions. Phonological, syntactic and conceptual processing work in one direction in response to visual and auditory events in the environment (typically in listening to speech or reading a text), but they also work in the other direction to produce speech or writing (or messages in sign language). Production involves a physical response to internal events, the creation of a message to be conveyed. This requires articulation of different parts of the body, following the commands of motor structures. Meanings in the conceptual processor are matched with syntactic structures which in turn are matched with phonological structures; this structural chain continues to be built following different routes according to the selected mode of articulation. The required motor structures that drive the articulation of speech will be different from those involved in writing or signing.
When the basic direction of processing is, say, from conceptual structure to speech, the assembling of an appropriate structural chain is not carried out in a rigid unidirectional sequence; at different levels, different options will be available and compete for selection. In production, some options may be partially formed and then be dropped in favour of a better-fitting rival structure. In this way the ultimate structural chain is built up incrementally, out of the more accessible candidate structures that happen to provide the best overall fit at the time. Comprehension works according to the same principles, with parallel processing and split-second editing and revising until a best-fit interpretation is arrived at. It may be the wrong interpretation but it will be the best fit, given the person's current processing resources at that particular moment in time.
As an example of what processing online means, let us examine speech production in a fluent speaker. The construction of a message will be initiated in the conceptual processor. Conceptual structures will be chosen, which then activate the interface between the conceptual and syntactic system. The conceptual structures are matched up with particular syntactic structures forming the first stage – in other words, a CS+SS (conceptual structure plus syntactic structure) chain. A semantic argument structure in CS code which specifies an action with an agent (the doer) and a patient (what is acted upon), as in "a boy hit the ball", is matched up with a syntactic argument structure with the requisite verb and noun phrases (determiner phrases), each in the appropriate case: one in nominative case and the other in objective case. The interface between SS and PS kicks in, causing various appropriate phonological structures to be activated; an SS/PS match is made, the outcome now being a CS+SS+PS chain. As is generally the case, more than one option may be selected in parallel before one particular option is settled on. Structural chains are formed incrementally; as more CS is built so more SS and PS are constructed with more context, earlier options that were provisionally selected are dropped as the representation develops and becomes more complex. The PS is matched up with motor structures responsible for the articulation of speech, and the utterance is produced. Each type of structure (AS, PS, SS, CS and so on) is constructed in its own particular module and by its own unique integrative processor, following its own particular set of principles. Comprehension is a similar process, with the general direction going in reverse; auditory structures are formed in response to acoustic stimuli in the environment. These auditory structures match PS and SS, finally culminating in the interpretation of the message (its conceptual structure).
One feature of the MOGUL approach is an attempt to spell out in coherent terms the role of perception and affect in issues of language development in the individual. In the earlier example of "a boy hit the ball" (where a simple example of a how a CS+SS+PS chain was built up), we assume that the conceptual structures evoked also have interfaces with various perceptual and affective structures to account for the associations individuals have with, for example, the concept "boy". In other words, the activation of a chain effectively becomes the activation of a whole network of associations contributed by different modular systems in the mind. The co-activation of structural networks of various kinds is important in explaining facts about attention and noticing. MOGUL follows the view that high levels of activation are strongly implicated in the phenomenon of conscious awareness. This is made explicit in the role of perceptual output structures (POpS).
POpS is a generic term covering the output of various perceptual systems corresponding to the five senses – or, more properly, the sensory systems currently believed to exist (which number more than the traditional five). Of greatest interest in MOGUL are visual structures (VS) and auditory structures (AS), given the fact that language is usually perceived in visual and auditory terms. Structures reside in the memory stores of the appropriate processing system (processing unit, or module). Exposed to the sound of the word "dog" or the sound of a creaking door, the auditory system activates a particular auditory structure (or set of auditory structures) in response to this sensory input. In this way, a given structure can be thought of as an auditory memory that may be re-activated even when there is no external sensory input (in hallucinations and dreams, for example) and grow stronger or weaker (less accessible) depending on the frequency with which it is activated successfully.
Perceptual output structures are strongly interconnected. This enables a coordinated responses to events in the environment and assists the organism's chances of survival. This coordinating function of the POpS system may be related to the role of the global workspace in Baars' theory.[8] At the same time, since POpS mediate between the external environment and the internal operations of the mind that are inaccessible to awareness, POpS conform closely with Jackendoff's Intermediate Theory of Consciousness.[9] The rich interconnections between POpS permit very high levels of activation to be achieved, a hypothetical prerequisite for awareness to occur. It is also reflected in the condition known as synaesthesia: here the connection between two POpS reaches levels that result in an awareness whereby the two senses appear to merge; for example, a particular sound takes on the quality of a particular taste. Another example of connection is provided by the experimentally-induced McGurk effect. Here, two sensory signals are generated to produce a conflict; the subject's awareness is the result of an attempt to resolve the conflict, so that hearing one sound and simultaneously seeing the speaker pronounce a different sound will create the illusion of hearing a sound halfway between the actual sound and the sound suggested by the speaker's facial gestures (especially the movement of the lips).
The main point is that awareness (whether it can be classified as an illusion or not) is generated indirectly via POpS; we do not have direct access to the contents of any module or processing unit. While we can never become aware of the fine phonological and syntactic properties of the word "dog" - or its semantic and pragmatic properties, its conceptual structure(CS) - we can certainly become aware of its sound. In the same way, we can become aware of the sound of the creaking door. In both cases, it is auditory structure (one of the POpS) that gives rise to the conscious experience. As implied above, it is in the nature of linguistic structures (PS and SS) and its conceptual structures (CS) that they do not have the rich interconnectivity described above, and consequently do not (and cannot) achieve the appropriate levels of activation for consciousness to occur. What we become conscious of arises directly from POpS activity. In other words, consciousness is always perceptual in nature.
The affective system works with affective structures (AfS) that are constructed using primitives such as "fear" and "disgust", which are discussed in affective neuroscience in the work of António Damásio.[10] AfS are connected to POpS and contribute directly to appropriate responses of the organism to events in the immediate environment, to its chances of survival and to the experience of conscious awareness. Making these aspects of mental life more explicit helps researchers to understand and devise hypotheses about how consciousness and emotion affect online language performance and language acquisition. By the same token, it should help to show how certain aspects of language behaviour are (or might be) immune to such influence.
Language is a vague term and is its widest sense involves a host of different processing units, only some of which are what makes human language unique to this species, that is, the phonological and syntactic expert systems that make up what, in generative linguistics, is called the language module and whose special properties are often subsumed under the name Universal Grammar. Closely associated with this core system, but not part of the language module per se, are a rich set of conceptual structures developed over the lifespan, as also b) auditory structures that have been matched up with phonological structures and c) visual structures created for interpreting writing and sign language. In addition, there are the various motor structures involved in the production of language in its different modes. One might lump all of this together and call it knowledge of language. None of it, however, constitutes 'knowledge of language' if, by this term, we rather mean language that we can think and talk about and analyse, in other words metalinguistic knowledge, which, as with all forms of metacognition, is constructed out of conceptual structure and various perceptual output structures. To take a simple example, the word "horse" can be discussed or pondered; all that is needed for this is an auditory structure (the sound of the word) and its visual structure (representing its orthographic, written form), both of which are matched up with its meaning as illustrated in the figure on the right. Not represented in the figure is the additional conceptual structure (over above the meaning of horse itself), consisting of metalinguistic concepts such as word, syllable, noun, definition and the like. These concepts are required for any analytic thinking about language and may vary widely in degree and complexity, depending on an individual's metalinguistic sophistication. In any case, the language module is not directly implicated in any explicit discussion (or explicit thinking) about what is actually a linguistic form. In the MOGUL version of metalinguistic processing, the essential chain of connections that are assembled completely bypasses the PS and SS systems. Processing within the language module will always kick in automatically on exposure of language stimuli, but metalinguistic (explicit) activity is processed elsewhere. Following the standard arguments from learnability theory as applied in generative linguistics, the language module is nevertheless crucial for all other types of linguistic activity and is vital for explaining the acquisition of language.[11] No amount of activation of auditory and conceptual structure alone can explain the growth of language in the mind following repeated exposure to speech, most obviously in the case of little children and (arguably) adults as well.
Because of these two modes of knowing, we can appear to be very knowledgeable about the grammar of a particular language or grammar in general and yet be very poor users of that language. Or, like many native speakers, we can make metalinguistic assertions about the rules of our language which are not at all borne out by the way we actually speak and comprehend our mother tongue. Knowledge of a language, in this metalinguistic sense of the word, can also be "right" or "wrong". We can have misconceptions about language, or we can have a view of grammar that accords with the facts. By way of contrast, the system operated subconsciously within our language module can never be right or wrong: it is just the way it is, the way it as has developed in us over time. Metalinguistic knowledge is not useless, however. An essential part of education in many cultures is acquiring such consciously-accessible knowledge about language, and especially about the mother tongue. In MOGUL, however, this should not be confused with the implicit knowledge of language that drives language performance.
In this figure, the MOGUL version of the cognitive system is sketched in simplified form. The circles or ovals represent specialised processors, and the rectangular boxes each represent the particular memory store of structures which they are designed to handle. Affective structures (AfS) are not differentiated in this figure. Also, for simplicity's sake, only five "senses" are portrayed in the POpS system. All languages are handled by the same syntactic and phonological processors, but different languages will involve both shared and unique syntactic structures (SS) and phonological structures (PS) residing in (respectively) their syntactic and phonological memory stores. Mental activity is characterised by the formation of chains or networks of structures across different modules. This is achieved by means of interface processors (the dotted arrows) following Jackendoff (1987). These are bimodular, in that they match elements in two adjacent modules; they can only access and link certain structural elements in one or the other module. Moreover, the matching is not a "translation" of one element into terms (the code) of another system; there is no exchange of information, merely a chaining of elements that can otherwise only be processed within its own particular processing unit or module.