Introduction
The subject of consciousness, long shunned by mainstream psychology and the scientific community, has over the last two decades become a legitimate topic of scientific research. One of the most thorough attempts to formulate a theory of consciousness has come from Bernard Baars, a psychologist working at the Wright Institute. Baars proposes that consciousness is the result of a Global Workspace in the brain that distributes information to the huge number of parallel unconscious processors that form the rest of the brain. This paper critiques the central hypothesis of Baars' theory of consciousness.
What is Conscious?
Baars uses an operational definition of consciousness that considers people to be conscious of an event if (1) they can say immediately afterwards that they were conscious of it and (2) we can independently verify the accuracy of their report (Baars, 1988, p 15). In other words, if immediately after being presented with X, a person can say that they have experienced X, but after presented with Y, they can say that they have not experienced X, they are deemed to have been conscious of X. As Baars readily acknowledges, his definition is less than perfect in that it presupposes the abilities to use language (in telling of the conscious experience), to have volition (the ability and inclination to do the telling), and to use metacognition (to become conscious of the conscious experience of X, in order to tell of the experience), and thus doesn't isolate consciousness in its essence as would an ideal definition. However, the primary purpose of the definition is to facilitate and make precise the process of scientific study, and a definition which is pragmatic is more valuable than one which isolates "consciousness" but does not enable the phenomenon to be studied. It is difficult to see how to study consciousness in a way that would completely separate it from other mental phenomena with which it so frequently occurs, and thus Baars' definition is acceptable in spite of not semantically isolating consciousness.
Furthermore, this definition is reasonable in that it seems to correspond to our intuitive notions of what something conscious would be like. (Though not necessarily what it would be like to be conscious!) Every attempt to talk about consciousness is inevitably rooted in the observations people make about their own minds or in the observations people make about other people's minds. In the first case, it is reasonable to say that what a person can observe about their own mind corresponds, in an epistemological sense, to what they are able to express in some manner or another (i.e. if a person observes that they think something, they can communicate that which they observe themselves to think, and vice versa). In the second case, people only observe the aspects of other people's minds which those other people express. Thus the equation of conscious mental content with that which can be consistently communicated (a better term might be "signaled") makes sense to a large degree. It should not be ignored, however, that this definition and the theory founded on it will have to be modified in order to truly describe consciousness and extend the phenomenon to organisms or machines not possessing one of the three faculties implicit in Baars' definition.
One example of people with dissociations of consciousness and the faculties implicit in Baars' operational definition of consciousness can be found in subjects with frontal lobe brain lesions. Such patients are frequently described as lacking "volition," in that they do not voluntarily engage in action, and are characterized by an inability to carry out non-automatic, complex tasks. They are, however, by all appearances "aware" of the world around them, in that they orient to stimuli and seem generally cognizant of their environment, and few people would deny them consciousness. This dissociation between consciousness and volition deserves further study and concern, but Baars' definition remains effective for an initial study of consciousness. Indeed, without volition, the data relevant to studies of consciousness is all but eliminated.
As a final aside, it should be noted that Baars attempts to explain only the cognitive aspects of consciousness. He seeks to explain what occurs cognitively when we are conscious of something, and makes no attempt to explain, and indeed does not even mention, the possibly more profound problem of qualia, the qualitative experience we have when we consciously experience something. Why is it that when we look at something red we "see" a redness that seems completely unrelated to the frequency of light that is reflected off the red object onto our retina and the processing that the information recorded by the retina undergoes as it makes its way through the visual pathway? A cognitive model such as the one Baars promotes does not offer any explanation as to how or why qualia arise from the neural computation at work in our brains. Global workspace theory offers what we can term a cognitive and neural correlate of consciousness - a structure in the brain which supports certain cognitive processes that happen to be conscious. Baars does not offer explanations of what it means for something to be conscious or why the system he proposes is necessary for conscious experience. As we shall see, he attempts to provide a cognitive framework for consciousness, which is itself no mean task.
Conscious versus Unconscious
We now turn to the basis of Baars' theory. To establish his theory Baars examines the cognitive properties of consciousness by contrasting conscious with unconscious processes. These contrasts serve as the basis for his theory. I begin by reproducing a table of them:
Conscious Processes versus Unconscious Processes
serial, with limited capacity parallel, with great capacity
internally consistent diverse
broad range of contents limited range for a given processor
relational ability autonomous and isolated
context sensitive largely ignorant of context
inefficient efficient
error-prone few errors
Baars' theory is founded on the view that the brain is composed of many different parallel processors, synonymously referred to as modules, each capable of performing some task on the symbolic representations that it receives as input. The modules are flexible in that they can combine to form new processors capable of performing novel tasks, and can decompose into smaller component processors. This model of the brain is fairly well-supported by evidence from anatomical studies and studies of patients with brain damage. Most brain theorists have been hesitant to label mental processes "unconscious" or "conscious," but there is growing evidence from a variety of sources that many unconscious processes underlie ordinary perception and cognition. (By an unconscious process, I mean a process that takes place in the brain of which we are unaware. I am not referring to the term unconscious as used to describe people who have received hard blows to the head etc.) "Ordinary" perception and cognition are not normally described as "conscious," and most psychological experiments conducted in this century have not mentioned consciousness as having anything to do with the processes that they study. However, it does not seem a huge leap of faith to term such processes "conscious." When a person "sees" something, it is clear that they are conscious of what they see. Recognizing that processes that can be tested explicitly are consciously experienced in some manner, and that those mental processes that subjects can describe as having taken place are similarly "conscious" is a major step, for it allows Baars and other researchers to draw on the masses of perceptual and cognitive psychology experiments that have been performed over the last century, instead of having to start independently of prior research not directly aimed at consciousness. Many of the contrasts between conscious and unconscious processes listed above are rooted in such experiments that Baars interprets in the new light of conscious and unconscious experience. Assimilating the mass of experimental results into a framework of conscious and unconscious processes is not without potential pitfalls, for it is not always clear what aspects of a process are consciously experienced or whether or not a process is conscious at all. However, because Baars' conclusions are not very controversial, I will not dwell on this and proceed to looking at the individual contrasts that Baars describes.
Baars treats the brain as a large group of separable, very specialized systems that are unconscious much of the time that they operate. At least some of these processes can, one by one, become conscious, and the successive outputs of these processes constitute conscious experience. Significant, though, is the idea that only one process can be conscious at one instant of time. In other words, consciousness is a serial phenomenon. This is a fairly standard observation, but one that is deceptively simple. Any respectable theory of consciousness will have to explain how it is that a serial stream of thought emerges from the masses of unconscious mental computations that are proceeding simultaneously in the brain, what it is about this stream of thought that makes it conscious, and what the evolutionary advantages of such a system are.
Baars' second claim about consciousness is that it has internal consistency, a property not shared by the collection of unconscious processes in the brain. Baars cites as an example of this property the experience of viewing a Necker cube, an optical illusion which we can consciously see in one of two different orientations. The two views of the cube can "flip" back and forth, but we cannot entertain both of them simultaneously. In other words, our conscious experience of the cube is consistent. A similar situation is found with ambiguous words. People seem to be capable of having but one meaning of a given word in mind at one time. There is evidence, though, that the alternative meanings are represented unconsciously in the brain at the same time as the conscious meaning, in that the other meanings of such words often show priming effects on sentence comprehension (Tanenhaus, Carlson, and Seidenberg, 1985). This indicates that, while conscious processes are consistent, the collection of unconscious processes are not. I will note in passing that a criterion of "consistency" should not be taken completely at face value, simply because consistency is a vague term. It isn't clear what or who the arbiter of consistency is. What Baars means by this criterion is that conscious experience is coherent to the processes of thought, with local ambiguities resolved to define one "consistent" experience. While this is circular and in some sense redundant, it is nonetheless important to notice. Within the context of Baars' model, there turn out to be functional reasons for the internal consistency of conscious experience. For the time being, though, I will put this aside.
The third contrast should be clear, namely, that a huge variety of things can be experienced consciously, but that by definition, an unconscious specialized processor can perform but a limited range of tasks. The visual pathway is designed to process information from the retina, and not much more. This is the case with seemingly every such part of the brain. It is remarkable, though, that consciousness can consist of nearly any sort of experience. We can consciously see things, hear things, do math problems, and learn a complex task like driving a car, to mention a few of the myriad things we are conscious of over the course of our lives.
Another property of consciousness is its ability to relate what seem to be any two conscious experiences to each other. The best example of this is classical conditioning, where virtually any conscious stimulus may serve as a signal for virtually any other event. This relating cannot occur if the experiences are unconscious. Baars cites a study showing that Pavlovian association cannot occur if the signal stimulus has been repeated to the point of habituation (when the stimulus ceases to be consciously experienced). (Razran, 1961) The findings suggest that in order for the relating of stimuli to occur, both must be consciously experienced, and the fact that they are related must be conscious as well. This supports the fourth contrast of consciousness having relational capacities, while unconscious processes are unable to communicate in any sense with other processors other than the extent to which they do by receiving input representations and producing output representations in their established pathways of operation.
A fifth contrast is that conscious experiences are what Baars terms "context-sensitive," while representations processed unconsciously are not. Context-sensitivity is defined by Baars as "the way in which conscious events are shaped by unconscious factors." (Baars, 1988, p 79) Our conscious experiences are constantly affected by unconscious assumptions. Unconscious events are, in contrast, not influenced by such contextual assumptions, as can be seen in the frequent occurrences of actions that are performed unconsciously in completely inappropriate circumstances, as is the case when a person walks into their bedroom, combs their hair, and walks out, forgetting that they had gone in their to get a book they were reading. In such instances, actions are performed in accordance with habit, as opposed to current intention. In order to insure that we actually retrieve the book in our bedroom, we need to be conscious of our action so that the context of our intention can guide what we do.
Finally, there are the contrasts of inefficient, error-prone conscious processes with efficient, relatively error-free unconscious processes. These can be illustrated with any task that a person learns. While unlearned, a task has to be performed consciously, at which point it is done slowly and with frequent errors. Once learned , the task is unconscious, and is performed with comparative speed and accuracy. This is one example of a classic domain of psychological study (namely, learning) whose results may be used and reinterpreted, albeit with care, to study consciousness.
These last two contrasts are accurate, but must be cautiously applied to consciousness, as they are concerned with the voluntary learning of tasks. Consciousness, in spite of the controversy surrounding the question of "is consciousness necessary for learning?", is unquestionably central to learning, and thus it would be foolish to ignore the connection. It is unclear, though, as to whether anything about the nature of consciousness itself can be inferred from these last two properties, since inefficiency and the tendency to make mistakes obviously have little to do with a purely phenomenological experience, like the conscious experience of seeing a tree. It turns out that these two properties are not central to Baars' theory, though, and they make sense given an interpretation of learning in light of his ideas about consciousness. The other contrasts appear to be general enough to encompass all forms of conscious experience, and they suffice to give initial support to Baars' basic claims. Thus the lack of clarity about whether the final contrasts truly apply to consciousness is not fatal to Baars' theory. That having been said, we will briefly describe the principal structures of Baars' theory, and then proceed to discuss them.
The Global Workspace
To explain the differences between conscious and unconscious processes, Baars turns to a popular model in artificial intelligence distributed-processing systems. Many artificial intelligence systems rely on a globally accessible block of working memory to mediate communication and novel interaction between the various individual processors. Baars proposes that a similar structure exists in the human brain, and that it supports conscious experience. The structure, which he terms the global workspace, is accessible by most processors, meaning that most processors potentially can have their contents occupy the working memory. The global workspace can also "broadcast" its contents globally in such a way that every processor receives or has access to the conscious content. It is serial and of limited capacity in that only one processor's contents may be "broadcast" over the global workspace to all the other processors at any given time. Baars acknowledges that an important part of a complete global workspace theory of consciousness would be an explanation of how a given processor gains access to the global workspace. Presumably there are a multitude of processors that could have their contents broadcast by the global workspace. However, only one processor's representations are broadcast at any one moment (consciousness is serial), and in order for the global workspace to be functionally advantageous, as consciousness is assumed to be, there must be an explanation for why one processor gains access to the global workspace and another does not. Baars describes two possible systems for controlling access to consciousness, and they in themselves are large enough in scope for a research paper. I will not discuss these to maintain a reasonable focus, and we will proceed under the assumption that an access system exists in conjunction with the global workspace.
In Baars' theory, conscious experience consists of the series of representations produced by certain processors that are broadcast by the global workspace to the brain's subsystems. Due to its proposed accessibility and presumed capability to broadcast the output of nearly any unconscious processor, the global workspace satisfies one stated property of consciousness - that it has a great range of potential contents.
As Baars points out, it is also conceivable that the relational ability attributed to consciousness could accompany the global nature of the global workspace. By broadcasting the output of any given module (for example, the representations of the sound of a bell from the auditory system) across the brain, and then broadcasting a second representation from a second module (sensations of food coming from the olfactory system), the two could be related via a "relating module," and conditioning could take place. This is a non-trivial claim, in that the brain has no way of "knowing" what stimuli will be paired together in a conditioning situation, and so could not simply send them off to our hypothetical "relating module" to be conditioned. It would also be of questionable feasibility to have all perceptual input automatically sent to such a module (with consciousness playing no role in the process), for it then would have to distinguish between a huge amount of data. Global broadcast over a limited capacity system offers a possible explanation for the occurrence of conditioning, in that by making conscious experiences available to all processors, any processor could respond, and in this case after a number of repetitions of the two stimuli, the module responsible for establishing a link between two experiences would create the conditioning effect. Furthermore, the information bottleneck that would occur as a result of the serial nature of the global workspace could have the effect of only making available through global broadcast information deemed significant by the hypothetical access system, so that information overload is prevented. Note that in actual conditioning experiments, as cited above, if one of the stimuli is habituated and not conscious, then the conditioning does not take place. This corresponds to the event not being globally broadcast, and thus in Baars' model the failure of conditioning is accurately predicted.
Another observed property of consciousness, which we discussed in the preceding section, is that of internal consistency. The construct of the global workspace does not explain the property of internal consistency, but it is possible to see that internal consistency is to some degree necessary for anything meaningful to occur as a result of a global broadcast. If an ambiguous message were broadcast to a host of different unconscious processors, different interpretations could conceivably result in conflicting responses, which is not desirable. The point of the global workspace is to facilitate the integration of diverse groups of processors, and conflicting responses would be counterproductive in this regard.
The fifth property of consciousness that Baars cites and that we have discussed is that of context-sensitivity. This property has little to do with the global workspace itself, and we will not discuss it, concentrating instead on the main hypothesis of the global workspace.
Global Broadcasting
The claim that consciousness is supported by a working memory of serial, limited capacity accessible to most of the processors in the brain, while necessary, is not terribly eye-opening or explanatory as it cites only properties of consciousness that seem rather intuitive and essential. While they are properties that any theory of consciousness will have to explain, they cannot be used to any real extent to differentiate between theories of consciousness because of their intrinsic role in conscious thought. The definitive claim made with the postulation of the global workspace is the idea that conscious contents are globally broadcast to all unconscious processors, and it is this hypothesis that I will discuss in detail. As noted above, global broadcasting could be responsible for the remarkable ability of disparate conscious experiences to be related to each other. Furthermore, global broadcasting could conceivably allow the brain to learn to solve new tasks, by submitting a single goal to every processor, and allowing them to coordinate themselves into a coalition capable of completing the task. This is the chief purpose of global workspaces in artificial intelligence systems, and their utility is at first glance equally plausible in the context of the human brain.
Another advantage of the global workspace theory of consciousness is that it resolves what is known as the "filter paradox." The filter paradox is the result of the common claim that the functional advantage of attention is that it singles out information in the environment that is in some sense "important" to the brain, thus saving computation by preventing the brain from processing all information. This explanation is paradoxical in that in order to single out information, the attentional systems would have to process all available information to determine which is important. Thus, there really isn't any savings in processing at all. Baars' theory resolves this paradox by hypothesizing that attention singles out the information which is broadcast via the global workspace to all processors. Clearly, it would be ridiculous to flood all parts of the brain with all available information. Attention filters through information, and only certain things become conscious and are globally broadcast. According to Baars, the "savings" come from only broadcasting certain information.
However, I question the true efficacy of a global broadcast in a human brain, and also whether conscious experience even suggests that this occurs. In terms of its functional feasibility, a global broadcast would presumably interrupt the goings-on of all affected processors with whatever its message would be. While there are certainly some instances when this would be effective, it would seem that most of the time, conscious contents are not terribly relevant to the responsibilities of most of the modules in the brain, and that whatever there is to be gained by broadcasting a thought or percept to all the various modules would surely be more than offset by the disruption of the myriad of other tasks that the brain is unconsciously occupied with. Much of this is difficult to talk about in any realistic and useful sense, because so little is known about what modules exist in the brain, how they are constructed from neural tissue, and how a symbolic representation is manifested in neuron interactions. It is possible that representations come in different classes, for instance, and that modules of roughly similar function understand the same classes of representation. In this scenario, modules to whom the conscious content was irrelevant might not be impaired by receiving the globally broadcast message, because it would not be of usable form. Thus a global broadcast would not necessarily be uniformly disruptive. Admittedly, though, this is a rather arbitrary claim given what is known about the workings of the brain, and it would be foolish to define the language of the brain in order to accommodate an unconfirmed theory of consciousness. It is not clear that it would even be practical for modules to be divided along functional lines. Given any two functional modules in the brain, I would argue that one can conceive of a task which requires the integration of the two. Granted, this is a weak argument, but it demonstrates that a compartmentalization of the brain postulated to accommodate global workspace theory is itself possibly unrealistic. Furthermore, weakening the claim of global broadcast in this manner, by specifying that only certain modules receive certain conscious representations, also weakens Baars' claims for its effectiveness, in that the range of the relational ability of consciousness would be impaired.
How Global Is Global?
We are left with the question of, as Baars puts it, "how global is global?" Baars himself makes very strong claims about the breadth of the broadcasting done by the global workspace, suggesting that conscious events have "very wide" distribution in the nervous system. The strength of this claim comes from the fact that it is implicit that all conscious events have the same distribution. Baars cites several lines of evidence to support this claim. His first source of evidence is the fact that as conditioning studies have shown, virtually any stimulus can serve as a signal to perform virtually any voluntary act. The distinction here is that all signals are by their nature perceptual in quality, and properties of conditioning are not necessarily characteristic of consciousness across all forms of thought.
Baars also cites evidence from biofeedback studies which show that a degree of voluntary control can be gained over virtually any neural event. Again, this is just as likely to be an isolated property of volitional acts, which, though they depend on conscious awareness of the biofeedback, are certainly not identical to conscious experience. It is perfectly conceivable that voluntary intentions would be subject to some manner of global broadcasting or that they by some other means are able to access any relevant part of the brain. But although they are consciously experienced, the broadcasting which we are hypothetically associating with volition could be completely distinct from consciousness. That is, it could very well be the case that conscious thought in its most general form (when we are not trying to perform an action or attend to anything specific) is not in any way broadcast, that such broadcasting is restricted to acts of volition or attention. Furthermore, there is evidence that attentional systems can affect any part of the brain involved in the task that is being attended to, which may account for the biofeedback data. It is still unclear as to how the attentional systems would initiate the action, since causing those single neurons to fire would be a novel task without established pathways. However, there is no reason to suggest consciousness as an explanation for this feature of attention over anything else that might be involved. Additionally, gaining control over neuron firing via biofeedback is a form of learning, and while learning appears to require consciousness, we cannot be sure how consciousness is involved and if consciousness is connected in any way with global broadcasting that may be taking place.
Baars' third line of evidence is the existence of the well-known Orienting Response in humans consisting of the automatic widespread bodily changes that accompany the onset of a novel stimulus. It is fairly obvious from personal experience that novel stimuli are likely to be conscious, and experimental results show that such stimuli cause reflexive responses in the brain including changes in EEG patterns and cortical activity. The widespread nature of these effects gives possible support for the hypothesis of conscious experience being globally broadcast.
Another source of evidence comes from event-related potential studies performed on awake cats. Recording electrodes implanted throughout the cat's brain display uniformly high levels of activity at the initiation of a series of bright flashes or loud clicks. Interestingly, neural activity is not limited to the respective visual or auditory pathways at the outset of the stimulation. As long as the stimuli are presented, brain activity as measured by the event related potentials becomes increasingly localized until it is limited to the expected particular sensory pathway. Baars assumes that the cat is conscious of the stimulus at the beginning of the presentation, as would be a human to a new stimulus, and cites this study as evidence for the global breadth of conscious experience. There are several objections to this claim. Ignoring for the moment the fact that the cat is nonhuman, the main problem is that it is again questionable as to whether what is measured has anything to do with consciousness. Presumably the stimuli are such that they cause an orienting response to take place, which, in accordance with studies in humans, might well cause widespread initial activity across the brain. However, it is clear that not all conscious events cause such a response; otherwise it would be impossible to detect the Orienting Response or the cat's event-related potentials with electrodes in the first place.
If similar Event-Related Potential studies were performed on humans, it might be possible to determine through protocol analysis and dual task measures how the subjective experience of a new stimulus corresponds to the brain activity observed in cats, assuming that the brain patterns observed in cats occur in humans as well. Such an opportunity has come about through the use of non-intrusive brain-imaging technology currently available with PET and fMRI. Although no studies have as of yet directly addressed the issue of conscious experience, there are studies that may have relevance to the question of global broadcasting in the brain (Haier, 1992).
Baars' final source of evidence for the global workspace is error-detection. He cites the phenomenon of effortlessly being able to catch all types of errors in sentences simply by consciously attending to them. A person fluent in a language can detect errors in a huge variety of domains, including pronunciation, syntax, semantics, and pragmatic intentions. This ability would seem to involve many different processing components, and thus might support the claim that conscious contents are globally broadcast, in that some mechanism or other must be responsible for transmitting the speech signal to all of the various modules involved. Significantly, the ability to detect errors is drastically impaired when a person is not conscious of a spoken sentence, as can be seen from the fact that people often do not perceive errors in their own speech if preoccupied with something else at the time.
Error detection in speech is perhaps the best piece of evidence for global broadcasting that Baars cites. There do, however, remain several questions. There remains the possibility that language has its own error detection modules, and that error detection capabilities are not an indication of a general property of consciousness. This is particularly conceivable in light of the evolutionary importance of language. Again, it also isn't clear that consciousness is really at work here. Certainly, people have to attend to a spoken sentence in order to catch any errors, but the errors could be detected prior to the consciousness of the part of the sentence containing the error. There is limited evidence that there is a delay involved in some types of conscious events, and that much unconscious preprocessing goes on prior to the conscious experience of something (Libet, 1978). Thus it is conceivable that error detection has nothing at all to do with consciousness.
Also significant is that the error detection mechanisms involved can all easily be construed as related to the task at hand of understanding spoken speech. Thus even if error detection does have something to do with consciousness, it does little to support the hypothesis of a global workspace, in that no processors outside of the normal domain of speech respond to the conscious experience. Under Baars theory such processors would simply have not responded to the global broadcast of the speech representation. However, error detection is supposed to be evidence for global broadcasting. Error detection in speech does not give direct support to Baars hypothesis because no modules respond outside of the normal domain of language, as would be necessary to directly support some sort of global broadcast system.
This observation rings true for many conscious experiences, in that it is difficult to see how it would functionally make sense for the entire brain to be apprised of every mundane conscious experience. Any given type of conscious experience would seem to have modules that it would be functionally desirable to broadcast to, but rarely do these consist of the entire brain, or even a significant portion of the brain.
In spite of the absence of clear experimental evidence in support of the global workspace theory of consciousness, it offers explanations of several properties of conscious experience that otherwise are unexplained. Baars should be credited with tossing aside the obscuring effects of both the mystic and the behaviorist attitudes toward consciousness. He has adopted an approach that will prove useful to studying consciousness, regardless of whether the global workspace turns out to exist or not. In addition, many of the problems associated with verifying a global workspace will remain problems for all future attempts to discover neural and cognitive correlates of consciousness. It is difficult to verify the consciousness of something, and even more difficult to experimentally isolate consciousness from other mental or brain phenomena.
Bibliography
Baars, Bernard J. (1988). A cognitive theory of consciousness. Cambridge University
Press, New York
Haier, Richard J. (1992) Intelligence and changes in regional cerebral glucose
metabolic rate following learning. Intelligence. 16, n3-4, 415-426.
Libet, Benjamin. (1978) Neuronal vs. subjective timing for a conscious sensory
experience. In P. A. Buser & A. Rougeul-Buser (Eds.), Cerebral correlates of
conscious experience (pp. 69-82). INSERM Symposium No. 6. Amsterdam:
North Holland/Elsevier.
Razran, G. (1961).The observable unconscious and inferable conscious in current
Soviet psychophysiology. Psychological Review, 68, 81-147.
Tanenhaus, M K, Carlson, G N, Seidenberg, M S. (1985). Do listeners compute
linguistic representations? In D R Dowty, L Karttunen, & A Zwicky (Eds.) Natural
Language Parsing: Psychological, computational,and theoretical perspectives
(pp 359-408). London: Cambridge University Press.