Cognitive psychology is the study of cognitive skills as the flow of information through the brain as stimuli are transformed into responses. It explains reading aloud as being divisible into a number of stages, between the presentation of a word on the page and the facial muscles moving to speak.
The problem of explaining a complex phenomenon like reading aloud can be seen on different levels of �as if� explanation. Here, the emphasis is on understanding in functional terms what is happening.
The hope is that by describing the behaviour in sufficient detail, we can develop a theory about the internal workings of the brain at a systems-based level. As with describing the behaviour of our perceptual mechanisms, it is by contrasting success with illusions and the problems that we encounter in various ways that we can learn most about which mechanisms must be at work, since they are identifiable by where they fall down. One of the most important clues to brain mechanism in reading aloud is the reaction time (RT), i.e. the time taken to read one word as opposed to another (all other things, such as the length of word, being equal). As I will show below, this time is dependent on the amount of processing required for one word as opposed to another, and that this processing time can be speeded up by all sorts of factors, such as the processing of a related word or increased familiarity.
When the theorists has outlined what they consider to be a functional working abstract model, computational modelling is perhaps the closest an experimenter can come to directly testing the implementation of a theory. By comparing observed with model data, especially notable or striking features, the model�s closeness and so hopefully, accuracy of replication of real-life mechanisms, can be ascertained. Building a computational model demonstrates that the explanation adequately explains the phenomenon. In the model outlined below, the Interactive Activation model, the behaviour is reproduced with nodes representing words and activation passing between them, which ultimately speeds processing.
Having established that a computational model does produce similar results to observed data, the next step is usually to identify corresponding neural mechanisms which actually implement the behaviour, e.g. voluntary action being a product of dopamine flow in the basal ganglia. There is some debate as to the worth of computational models which are not grounded in the hard evidence of neurophysiology, since they are fundamentally just hypothetical reverse-engineered models.
The theoretical stages involved in reading aloud were defined as follows:
word on page
pattern falls on retina
letter recognition
word recognition
production of a motor plan
muscles move
word produced
Obviously this would be more complicated and would require further semantic analysis in order to read a sentence with the proper intonation.
In
order to make any computational model intended to be based on the brain as
realistic and close as possible, connectionist models like the
Interactive Activation model are neuronally-inspired. A connectionist model�s
basic elements conceptually mimic some aspects of the way the brain works.
Neurons function by integrating information, through myriad excitatory and
inhibitory connections of varying weights. More sophisticated and developed
versions of the interactive activationist model exist, but they are based on
the same system of activation flowing from one level to another or within a
level. The aim is always the same: to get from the outside world to the model�s
knowledge of a domain.
The Interactive Activation model (Rumelhart & McClelland, 1981) uses the presence of features at different levels to excite or inhibit the next level, thus increasing the probability of exceeding the threshold for a particular word. In this model, letter-level and word-level information are different.
To give an example of its functioning, the model would recognise the letter �T' because it would only fire if both the mid-vertical bar and the top-horizontal bar are present at the feature-detector level, and if no other features which would not be found in �T' are present to inhibit the pathway, then the threshold is exceeded. Information processing is thus a very large number of feature detectors operating in parallel, trying to provide evidence about the state of the world to a system which works out on the basis of how the outside world�s been in the past.
This particular model took all the four-letter words in the English language with a frequency greater than 2 per million (just over a thousand) as its domain. Though an obvious simplification, the model is sufficiently realistic in scope to be convincing when it works. It can be tested by matching its reaction times of processing to normals�.
The
last step in the explanation of the model is the mutual inhibition suppressing
alternatives. If two paths seem almost identical, then the slightly more
excited node�s inhibition outweighs the other path, which consequently inhibits
the first path slightly less, leading to a positive feedback loop, ensuring
that decisions are come to quickly. The system
operates by constraint satisfaction, whereby the node which has most things
consistent with it (in the form of excitatory inputs) is the most active. Thus
the decisions are the best bet about the input given all available sources of
information. This appears to be how the brain works � crucial for an organism
adapted to survive in a complex environment, the model does not work on the
basis of certainty, but simply choosing the most likely path. Otherwise, how
would we understand another person�s voice in a crowded or handwriting on messy
paper.
There is the possibility that noise would be magnified, which would lead to inherent instability in a complicated noisy environment. This was tested on the model by placing ink blots strategically over letters in different positions, simulating a degraded stimulus. In one example, the stimulus matched both of the known words �work� and �word�, while the obscured letter matched either a �K� or an �R�. In this case, it is necessary to integrate information back down the levels; the activations for the different words at the word level pass back down to the letter level to further activate the �K� detector, and since it could not be a �D� and there is no such word as �WORR�, the model eventually came to the right conclusion.
Computational models are a vital tool to combine with data from both normal subjects and patients with brain damage or some form of cognitive dysfunction. Further complications in reaction time emerge during priming experiments. In these, the subject is presented with a mask, followed by a very brief �prime� word, then the �target� word, for slightly longer, then a mask again. The timings are such that the prime word cannot be recognised by subjects, but a proportion of subjects can recognise the target word. The chance of recognising the target word is greater if the subject is primed in some way, by being shown similar letters (letter level) or a word in a similar semantic category (meaning level).
Reicher�s
1969 word superiority effect experiment presented subjects with either (i) a
word (�work�) or (ii) a letter (�***k�), followed by a mask, so that the target
identification was about 50%. He asked subjects after whether the last letter
was a �D� or a �K�? Surprisingly, subjects do better in (i) than (ii), from
which it can be inferred that it is easier to detect a letter in a word than in
isolation. This begs the question, �how can you know about a word before you
know about the letters of which it is composed?� Here, the model is especially
instructive, since it shows that knowledge about a word influences knowledge
about the letters which make it up, showing that it
makes sense to see the perception of words as involving independent stages of
letter and word identification, confirming the evidence from both patients and
normals.
Experiments by Oldfield & Wingfield (1965) and later Scarborough et al (1977) demonstrate a frequency effect and a repetition effect, as well as an interaction between them. Patients shown pictures of words with differing commonness in the English language took longer (in terms of reaction time) to name the uncommon pictures than the common ones. When asked to read words aloud, again the more common words took less time, since the nervous system processes things more quickly if it has seen them before. If words had already been seen recently, the reaction times were also lower, since the nervous system finds it easier to access information which it has just accessed. It can be seen then, and demonstrated by the model that the mind does not work like a dictionary, taking the same average length of time to find any single word, but rather is adaptive in their ease of access. Moroever, the reaction time for uncommon words was most drastically decreased by the recency effect (an interaction effect).
All of these effects can be demonstrated with the Interactive Activation model. By varying the strength of the connection, the frequency effect can be seen, just as synaptic weights change in learning in the brain. Frequent use builds strong connections and activation in nodes with strong connections builds up faster because it is nearer the threshold already. Temporal summation (also seen in synapses) shows the recency effect.
Subtle
differences in the speed of processing can lead to these quite elaborate
appreciations of the likely brain mechanisms involved. As to whether having a computational model which gives the correct answer
means that we know any more about how the brain works, even though our
understanding is hypothetical, it is a far better educated guess than without
computational modelling.