Models of retrival and the Mental Lexicon, with specific consideration of dyslexia
Created | Updated Jan 28, 2002
On a basic level it can be presumed that there is a part of the brain that contains some kind of knowledge about every word we have the ability to produce and every word we have the ability to comprehend. It therefore further follows that this dictionary or "lexicon" has the space for new words to be added and in some way has links from the words within it to other parts of the mind where the actual information or template of what the word means is stored. It is possible that each word also has links to other words within the lexicon linking them in terms of appearance in a sentence, commonality, colour or description but this is of less importance as yet in explaining how the words inside this Mental Lexicon are accessed for use in speech, reading and writing.
Word recognition has undergone much experimental research and therefore there are some well-established findings of things that can effect the time that words are recognised. Any model which attempts to describe the way in which words are accesed from the mental lexicon must therefore account for these findings. These are that common words are recognised quicker than uncommon ones (the frequency effect); letter strings that couldn't possibly be words in the person’s language are quickly discounted as words (the word/non-word effect); words seen within context are more quickly recognised than words out of context (the context effect); and if a letter string or phoneme string is difficut to hear or see then it is harder to recognise (the degradation effect).
There have been two main types of model that have been proposed to show how words are accessed from the mental lexicon. These are known as either 'direct access'models or 'search' models according to their features.
John Morton's logogen model (1969,1979) is a typical direct access model. The model states that every word that we know has a feature counter or 'logogen' corresponding to it. It suggests that the visual stimuli feed directly into a 'logogen' system, a set of counters that activates words as features become apparent. Hence if the first feature is that a word starts with a straight line then all logogens of words starting with I, L, K, N, M, B, etc. will gain a point. Each logogen has a threshold level, and when this level is reached the corresponding word is given up as the response and all logogens return to zero-count and then the system prepares for the next input. Lexical access is direct (hence its classification as a direct access model) and occurs simultaneously for all words. It is also passive as instead of seeking rejection the logogens wait until they are accepted.
Research into this model has been very complex. Based around the loggogen model's threshold level principle, which attempts to account for both frequency effects (the more frequently a word is used the more quickly it is recognised) and context effects (the words around the target word aid recognition if they are within a sentence for example) experimental research has concentrated on looking for the ways in which these two varibles interact with others. If the threshold level is an accurate description then they should interact in similar, if not indentical ways, with other variables such as stimulus quality.
Norris (1984) showed that frequency and stimulus quality can interact, but that stimulus quality and context interact more, however if stimulus quality actually affects the orignal visual encoding processes then this has no value when lexical access is considered. The problems with the logggen model when conducting emprical experiments is that there are too many units in which errors can occur to give very similar results, and although it can generally explain all experimental data it does so by adding more and more units into the system to cope with the explanations and as such it could be considered vauge.
Dispite the fact that it can explain most of the basic findings of word recognition, it fails to take into account what happens when a person is presented with non-words or words that they haven't seen before, although it can be presumed that each logogen is morpheme specific, not word specific. It does not allow for the speed at which we write or read, and it also suggests that each word has another logogen for the auditory features of the word, but if this is the case then the issue of space within the mind must be questioned, based upon the universal presumption that the mind works as efficiently as possible in terms of storage, the brain has only a limited amount of storage space and so if two areas of the mind have to contain very similar or identical information, then this model must be considered less of an adequete description than one which has muti-function units. As somewhere else must therefore contain "logogens" in reverse: word is "horse" then I need to do a downward stroke then a curve half way up then a downward stroke etc., when this information is already stored in the logogen unit of the reading system this seems simply too inefficient for a process as immediate as reading and writing are for the average human adult.
Forster's (1976, 1979) model demonstrates the other main type of access model, one which utilises the concept of search. A complete perceptual representation of the input is constructed and then compared to a set of access files, one of which is orthographic. The entries are the searched in order and when a match is found information from the mental lexicon is accessed, which contains all the information about the word, duplicating a lot of the information from the access files. Each access file is devided into bins arranged according to frequency and the high-frequency bins are searched first. Context effects are handled by the presumption that as soon as words are accessed related words are used to create a new list which is searched at the same time as the bin. As soon as a matching word is found in either list the mental lexicon is accessed. Degradation is presumed to delay the start of the search as the perceptual representation takes longer to complete.
The main criteria for selection of a model that can be used to helpfully consider dyslexia is that the model is as efficient as possible and that it therefore presents obvious areas that the speed required for the actions of reading and writing hinders the process. Like the loggogen model, Forster’s model makes no account for the speed in which we read and write and in fact presumes that the process takes longer than the loggogen model does. It also suffers from overkill, like the loggogen model it includes extra units in order to deal with the basic findings and suffers from a lack of refinment. The search model utilises serial processing, which would suggest that a dyslexic suffers from a wide-spread complex problem involving the whole system and as the fundamental presumption here is that there is a simple problem that causes all these effects it can be discounted.
Although Forster's model has a phonological file that means that words can be accessed simply by their phonological features which could explain why the majority of errors a dyslexic makes are phonlogical rather than simple spelling slips (Doyle 1996) (eg. the recogniton of 'fizzicks'as 'physics' without acknowledgement of the mistake) the way in which it does this is over-complicated, as it presumes that the dyslexic seraches the orphographic file and then the phonological file without taking twice the amount of time to do so in comparison to recognising a word spelt correctly.
Spelling slips in general are not considered at all in either model, although this is not a major oversight as these models are both concered primarily with recognition and not production. However it is logical to suggest that recognition and production are dealt with by the same area, due again to restraints on space.
Instead therefore it has been proposed that there is a learned set of rules, known as the grapheme-phoneme correspondence rules that exist and the mental lexicon can only be accessed phonologically. As spoken language comes before written language, both species-wise and in child development, this assumption seems logical. One of the theories presented with this process in mind is Rumelhart and McClelland's (1981-2) Interactive Activation model which belongs to a whole subset of theories within psycholinguists known as Connectionism. These theories present processing that occurs through the actions of many simple inter-connected units. Some of these units are dedicated to the recognition of letters and presenting to the units next along the line the phonemes these correspond to.
Rumelhart and McClelland's (1981-2) Interactive Activation model was originally presented to account for word context effects on letter identification, i.e. to account for the fact that it is easier to tell which letter is which if it occurs within a word, and as it is reading difficulties that are of paramount importance to this discussion this model proves to be of further relevance.
The model is made of lots of processing units arranged in three levels. An input level where the units look for visual features, a level where units look at the correspondence between these features and gives the next level a set of letters and finally an output level which has units which correspond to each word. Within a level each unit is connected to every unit in the level above it and if it is the middle level the units are also connected to the units in the level below. These connections fall into two types, a connection is either excitatory or inhibitory. Excitatory or positive connections make the units at the end of them (the ones in the higher level) more active, whereas inhibitory or negative connections make the end units less active. As well as these connections between levels there are also intra-level connections, which are inhibitory and connect each unit within a level to each other. This has the effect of quickly causing this pathway to have preferential access to the next level and causing all the other units to defer. If this were not the case it would be possible that there would be too many access points into the next level and thus into the mental lexicon, causing there to be too much abiguity in the choice of word meaning.
When a unit becomes activated it sends a signal down each of its connections simultaneously. If the connection is inhibitory then it will decrease the level of activation of the unit it is connected to, or if it is excitatory it will increase its activation level. Eventually, although at the start there is a lot of energy flowing through the system, the system relaxes until systematically only one word is left.
In comparison to other two models discussed, this model still needs a mental lexicon, it is this that is accessed by the final level, however it becomes more interactive with the process of reading, and this is also the only type of model which seriously proposes parallel processing as an integral part of reading, which due to speed at which this process occurs must be the case. It bears a passing similarity to Morton's logogen model, and another direct access model, the cohort model, however they suffer in analysis because they contains a serious amount of overkill to explain processes such as word context which the Connectionist model does efficiently. The search model states that full features profile is created and then the lexicon is checked systematically for the word which contains these features. This model bears similarity here too, the model does in way build up a features list, however this list is produced in parallel to the comprehension of the word. If the input was phonological then there is another first level that deals with it, sending messages on to the same second level for access of the mental lexicon. In a way this second level acts as a kind of grapheme to phoneme conversion processor.
In short it seems that the Interactive Activation model, although at first looking like the most complicated model does the most thorough job when attempting to explain the process of written language comprehension and writing. The dyslexic could suffer from errors at any point along this system, although perhaps the most obvious place would be the second level as it is here that the most complicated processes occur, there is definitive observation to state that there is an error somewhere to do with this level, and therefore due to the interconnections between the levels there is either an error with the access of the mental lexicon, the units within the second level themselves or with the information coming from the first level.
Regardless of which model, if any, is actually an accurate approximation to the way in which we read and write, each model states that there is a mental lexicon.
Therfore, putting these models aside for a moment, in summary it is a logical presumption that the goal of reading or listening is to divine meaning from the marks seen or the waves heard. From this deduction it is further logical that there must therefore be a store within the brain that holds not only the meanings of the individual words but also their sound patterns and symbol patterns, in similarity to a good dictionary and hence the title of this mental area: The Mental Lexicon.
Although the analogy between the dictionary and the lexicon is often made the comparison is really only a base one. It is unlikely that the lexicon actually contains real definitions of any word but rather pointers to areas of long term memory that contain these definitions, due to the universal assumption that the brain works as efficiently as possible in terms of space and it would be inefficient to store a definition within an already huge mental area if it is only based on previous perceptions of objects anyway. However like a dictionary it certainly would contain the phonological patterns of words, the template of the written form of the word, its part of speech and possibly links to other words which are regularly found with it in sentences (Meyer and Schvaneveldt, 1971 in Garnham, 1985 : 45). If there is an error within the lexicon then it is with one or more of these factors.
The fact that a dyslexic can produce speech as well as any other of similar age does not necessarily have an effect on this lexicon. Child language development may suggest that there are two lexicons, one for production and one for comprehension, the forms that a child can produce and comprehend are different and that does suggest that there are two different lexicons (Menn, 1983). This idea is redundant when it is considered that most dyslexics have difficulties with both production and comprehension of the written form.
With similar reasoning the links to the semantics of the lexical items can also be concluded to be without error and it is tempting to concluded that the phonological storage of the words is also without fault and (as logic would also suggest) it is simply the written form's storage that is at fault. However as the written form comes after the spoken in both the terms of evolution and child development, and as most languages have at least some phonetic input to their alphabets it is not unreasonable to assume that the way in which phonological storage occurs within the mental lexicon has some affect, either directly (there is some kind of grapheme to phoneme conversion process) or indirectly (the way in which graphemes are ordered is in accordance to phoneme storage rules, whatever they may be) on how the graphical representation of words is stored.
Word recognition has undergone much experimental research and therefore there are some well-established findings of things that can effect the time that words are recognised. Any model which attempts to describe the way in which words are accesed from the mental lexicon must therefore account for these findings. These are that common words are recognised quicker than uncommon ones (the frequency effect); letter strings that couldn't possibly be words in the person’s language are quickly discounted as words (the word/non-word effect); words seen within context are more quickly recognised than words out of context (the context effect); and if a letter string or phoneme string is difficut to hear or see then it is harder to recognise (the degradation effect).
There have been two main types of model that have been proposed to show how words are accessed from the mental lexicon. These are known as either 'direct access'models or 'search' models according to their features.
John Morton's logogen model (1969,1979) is a typical direct access model. The model states that every word that we know has a feature counter or 'logogen' corresponding to it. It suggests that the visual stimuli feed directly into a 'logogen' system, a set of counters that activates words as features become apparent. Hence if the first feature is that a word starts with a straight line then all logogens of words starting with I, L, K, N, M, B, etc. will gain a point. Each logogen has a threshold level, and when this level is reached the corresponding word is given up as the response and all logogens return to zero-count and then the system prepares for the next input. Lexical access is direct (hence its classification as a direct access model) and occurs simultaneously for all words. It is also passive as instead of seeking rejection the logogens wait until they are accepted.
Research into this model has been very complex. Based around the loggogen model's threshold level principle, which attempts to account for both frequency effects (the more frequently a word is used the more quickly it is recognised) and context effects (the words around the target word aid recognition if they are within a sentence for example) experimental research has concentrated on looking for the ways in which these two varibles interact with others. If the threshold level is an accurate description then they should interact in similar, if not indentical ways, with other variables such as stimulus quality.
Norris (1984) showed that frequency and stimulus quality can interact, but that stimulus quality and context interact more, however if stimulus quality actually affects the orignal visual encoding processes then this has no value when lexical access is considered. The problems with the logggen model when conducting emprical experiments is that there are too many units in which errors can occur to give very similar results, and although it can generally explain all experimental data it does so by adding more and more units into the system to cope with the explanations and as such it could be considered vauge.
Dispite the fact that it can explain most of the basic findings of word recognition, it fails to take into account what happens when a person is presented with non-words or words that they haven't seen before, although it can be presumed that each logogen is morpheme specific, not word specific. It does not allow for the speed at which we write or read, and it also suggests that each word has another logogen for the auditory features of the word, but if this is the case then the issue of space within the mind must be questioned, based upon the universal presumption that the mind works as efficiently as possible in terms of storage, the brain has only a limited amount of storage space and so if two areas of the mind have to contain very similar or identical information, then this model must be considered less of an adequete description than one which has muti-function units. As somewhere else must therefore contain "logogens" in reverse: word is "horse" then I need to do a downward stroke then a curve half way up then a downward stroke etc., when this information is already stored in the logogen unit of the reading system this seems simply too inefficient for a process as immediate as reading and writing are for the average human adult.
Forster's (1976, 1979) model demonstrates the other main type of access model, one which utilises the concept of search. A complete perceptual representation of the input is constructed and then compared to a set of access files, one of which is orthographic. The entries are the searched in order and when a match is found information from the mental lexicon is accessed, which contains all the information about the word, duplicating a lot of the information from the access files. Each access file is devided into bins arranged according to frequency and the high-frequency bins are searched first. Context effects are handled by the presumption that as soon as words are accessed related words are used to create a new list which is searched at the same time as the bin. As soon as a matching word is found in either list the mental lexicon is accessed. Degradation is presumed to delay the start of the search as the perceptual representation takes longer to complete.
The main criteria for selection of a model that can be used to helpfully consider dyslexia is that the model is as efficient as possible and that it therefore presents obvious areas that the speed required for the actions of reading and writing hinders the process. Like the loggogen model, Forster’s model makes no account for the speed in which we read and write and in fact presumes that the process takes longer than the loggogen model does. It also suffers from overkill, like the loggogen model it includes extra units in order to deal with the basic findings and suffers from a lack of refinment. The search model utilises serial processing, which would suggest that a dyslexic suffers from a wide-spread complex problem involving the whole system and as the fundamental presumption here is that there is a simple problem that causes all these effects it can be discounted.
Although Forster's model has a phonological file that means that words can be accessed simply by their phonological features which could explain why the majority of errors a dyslexic makes are phonlogical rather than simple spelling slips (Doyle 1996) (eg. the recogniton of 'fizzicks'as 'physics' without acknowledgement of the mistake) the way in which it does this is over-complicated, as it presumes that the dyslexic seraches the orphographic file and then the phonological file without taking twice the amount of time to do so in comparison to recognising a word spelt correctly.
Spelling slips in general are not considered at all in either model, although this is not a major oversight as these models are both concered primarily with recognition and not production. However it is logical to suggest that recognition and production are dealt with by the same area, due again to restraints on space.
Instead therefore it has been proposed that there is a learned set of rules, known as the grapheme-phoneme correspondence rules that exist and the mental lexicon can only be accessed phonologically. As spoken language comes before written language, both species-wise and in child development, this assumption seems logical. One of the theories presented with this process in mind is Rumelhart and McClelland's (1981-2) Interactive Activation model which belongs to a whole subset of theories within psycholinguists known as Connectionism. These theories present processing that occurs through the actions of many simple inter-connected units. Some of these units are dedicated to the recognition of letters and presenting to the units next along the line the phonemes these correspond to.
Rumelhart and McClelland's (1981-2) Interactive Activation model was originally presented to account for word context effects on letter identification, i.e. to account for the fact that it is easier to tell which letter is which if it occurs within a word, and as it is reading difficulties that are of paramount importance to this discussion this model proves to be of further relevance.
The model is made of lots of processing units arranged in three levels. An input level where the units look for visual features, a level where units look at the correspondence between these features and gives the next level a set of letters and finally an output level which has units which correspond to each word. Within a level each unit is connected to every unit in the level above it and if it is the middle level the units are also connected to the units in the level below. These connections fall into two types, a connection is either excitatory or inhibitory. Excitatory or positive connections make the units at the end of them (the ones in the higher level) more active, whereas inhibitory or negative connections make the end units less active. As well as these connections between levels there are also intra-level connections, which are inhibitory and connect each unit within a level to each other. This has the effect of quickly causing this pathway to have preferential access to the next level and causing all the other units to defer. If this were not the case it would be possible that there would be too many access points into the next level and thus into the mental lexicon, causing there to be too much abiguity in the choice of word meaning.
When a unit becomes activated it sends a signal down each of its connections simultaneously. If the connection is inhibitory then it will decrease the level of activation of the unit it is connected to, or if it is excitatory it will increase its activation level. Eventually, although at the start there is a lot of energy flowing through the system, the system relaxes until systematically only one word is left.
In comparison to other two models discussed, this model still needs a mental lexicon, it is this that is accessed by the final level, however it becomes more interactive with the process of reading, and this is also the only type of model which seriously proposes parallel processing as an integral part of reading, which due to speed at which this process occurs must be the case. It bears a passing similarity to Morton's logogen model, and another direct access model, the cohort model, however they suffer in analysis because they contains a serious amount of overkill to explain processes such as word context which the Connectionist model does efficiently. The search model states that full features profile is created and then the lexicon is checked systematically for the word which contains these features. This model bears similarity here too, the model does in way build up a features list, however this list is produced in parallel to the comprehension of the word. If the input was phonological then there is another first level that deals with it, sending messages on to the same second level for access of the mental lexicon. In a way this second level acts as a kind of grapheme to phoneme conversion processor.
In short it seems that the Interactive Activation model, although at first looking like the most complicated model does the most thorough job when attempting to explain the process of written language comprehension and writing. The dyslexic could suffer from errors at any point along this system, although perhaps the most obvious place would be the second level as it is here that the most complicated processes occur, there is definitive observation to state that there is an error somewhere to do with this level, and therefore due to the interconnections between the levels there is either an error with the access of the mental lexicon, the units within the second level themselves or with the information coming from the first level.
Regardless of which model, if any, is actually an accurate approximation to the way in which we read and write, each model states that there is a mental lexicon.
Therfore, putting these models aside for a moment, in summary it is a logical presumption that the goal of reading or listening is to divine meaning from the marks seen or the waves heard. From this deduction it is further logical that there must therefore be a store within the brain that holds not only the meanings of the individual words but also their sound patterns and symbol patterns, in similarity to a good dictionary and hence the title of this mental area: The Mental Lexicon.
Although the analogy between the dictionary and the lexicon is often made the comparison is really only a base one. It is unlikely that the lexicon actually contains real definitions of any word but rather pointers to areas of long term memory that contain these definitions, due to the universal assumption that the brain works as efficiently as possible in terms of space and it would be inefficient to store a definition within an already huge mental area if it is only based on previous perceptions of objects anyway. However like a dictionary it certainly would contain the phonological patterns of words, the template of the written form of the word, its part of speech and possibly links to other words which are regularly found with it in sentences (Meyer and Schvaneveldt, 1971 in Garnham, 1985 : 45). If there is an error within the lexicon then it is with one or more of these factors.
The fact that a dyslexic can produce speech as well as any other of similar age does not necessarily have an effect on this lexicon. Child language development may suggest that there are two lexicons, one for production and one for comprehension, the forms that a child can produce and comprehend are different and that does suggest that there are two different lexicons (Menn, 1983). This idea is redundant when it is considered that most dyslexics have difficulties with both production and comprehension of the written form.
With similar reasoning the links to the semantics of the lexical items can also be concluded to be without error and it is tempting to concluded that the phonological storage of the words is also without fault and (as logic would also suggest) it is simply the written form's storage that is at fault. However as the written form comes after the spoken in both the terms of evolution and child development, and as most languages have at least some phonetic input to their alphabets it is not unreasonable to assume that the way in which phonological storage occurs within the mental lexicon has some affect, either directly (there is some kind of grapheme to phoneme conversion process) or indirectly (the way in which graphemes are ordered is in accordance to phoneme storage rules, whatever they may be) on how the graphical representation of words is stored.
