Dual-Route 1 Dual-Route Model Versus Parallel-Distribution Processing Model in Reading Gideon Bromand Queens University Dual-Route 3 Introducation It has been often predicted that some facts about reading aloud can be explained only by the dual-route model (lexical and nonlexical routes from print to speech). This view was challenged by Seidenberg & McClelland (1989, 1990). This model consists of sets of orthographic, phonological, and hidden units that use back propagation learning algorithms to modify the connection's weights between units. These two models will be viewed by 5 major aspects of reading. The single-route model proposed by Seidenberg & McClelland can account for the first one but not for the remaining 4. On the other hand, the dual-route model can explain all the 5 facts of reading. This may suggest that the dual-route model is a more reasonable model for skilled reading and learning to read. The recognition and pronunciation of words have been one of the central topics in research of reading and have been studied into depth in the last several years. In the dual-route models of reading, the readers can choose between two different procedures while converting print to speech. In this model there are a dictionary search procedure and a letter-to-sound rule procedure (Coltheart, Curtis, Atkins and Haller, 1993). All the learned words by the reader are represented as an entry in a mental dictionary or internal lexicon. Each word is read aloud by entering the word's lexical entry from its printed form and retrieving from that entry the pronunciation of the word. Readers can pronounce some nonwords regardless of whether the readers have seen them before. The dual-route theorists suggest that readers have a nonlexical route for reading aloud, thus there is a system of rules that specify the relationship between letters and sounds in English. The nonlexical route allows the pronunciation of non-words and the pronunciation of words that obey the rules of spelling to sound in English. This route will fail to pronounce words that disobey the rules of spelling to sound such as colonel. In summary, the lexical route will have an output when the string is a word but will not provide an output when the string is a nonword. Whereas, the nonlexical route will pass on the correct output when the input is nonword or a regular word and will result with an incorrect output when the input string is an exception word. Therefore, the term dual-route model means a model that has one route for reading word but cannot read nonwords, and another route that can read nonwords and regular words but misreads exception words by regularizing them. The dual-route model has been popular because it has succeeded in accounting for several facts about normal reading (children & skilled readers) and abnormal reading (acquired and developmental dyslexics). The contribution of abnormal reading in the dual-route approach was work on patterns of reading breakdown caused by brain damage in a reader. Marshall and Newcombe (1973) were the first to do work of this kind. They discussed data from 6 patients with reading impairments and considered how the patterns of reading abilities exhibited by these patients might be interpreted in relation to a dual-route model of reading. The dual-route model has been challenged by the development of the Parallel Distribution Processing (PDP) connectionist model (single-route) of Seidenberg and McClelland(1989) which was a computer program, accepting letter input strings and producing a form of a phonological output. They have argued that a single-route model can also explain different facts about reading. It has been suggested in their article that there is a single, uniform procedure for computing a phonological representation from an orthographic representation suitable for the regular words, exception words and nonwords (Seidenberg and McClelland, 1989). The present paper attempts to determine the differences between the two models, dual-route and single-route (PDP) in 5 important aspects of reading. Furthermore, it has been predicted by dual-route modelers that human reading systems do not contain a processing procedure that can correctly translate both exception words and nonwords from print to speech. However, Seidenberg and McClelland (1989) have suggested that human can translate exception words and nonwords from print to speech. The main theme of this paper is the conflict between the dual-route view and single-route view (PDP). Aspects of the Seidenberg and McClelland (1989) Model The Seidenberg-McClelland model consists of three layers, orthographic input units, hidden units and phonological output units which are feed-forward network. There are 400 orthographic units, 200 hidden units and 460 phonological units. In addition, there are 80000 connections between orthographic units and the hidden units and 92000 connections between the hidden units and the phonological units. There are also 80000 feedback loops from the hidden units to phonological units. The connection weights initially compute a random pronunciation for the orthographic inputs. The network learns by using back propagation which produces adjustment to weights accuracy of the activation's patterns across the phonological units in response to orthographic inputs (Seidenberg and McClelland, 1989). The orthographic unit has a full set of letter detectors for every position in the word. Each of the 400 input units consisted a list of 10 possible first characters, a list of 10 possible second characters and a list of 10 possible third characters. The character's vocabulary includes all the alphabet letters and a boundary letter symbol. For example, if the input word is "MADE" and it has "@" as a boundary letter, then the word "MADE" will turn on any unit that consists any of the triples @MA, MAD, ADE or DE@. The average of the three consecutive characters in the input activates approximately 20 orthographic units. Thus, the probability will be zero to activate the same set of orthographic units and all the input string should be discriminable at this input level. The phonological units deal with the problems of repetitions and use an output system similar to their input system. Each of the 460 output units are presented only with a single triples which were chosen systematically ( not randomly). The output method they adapted was originally used in Rumelhart and McClelland (1986) in a model obtaining past tense verbs in English. This output derives from the concept of Wickelphone which is a sequence of three consecutive phonemes($$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$). Aspects of the Dual-Route Model The Dual-Route model of reading ( Morton and Patterson, 1980; Paap and Noel, 1991) is also a hypothesis of procedures skilled readers use to read aloud. According to this hypothesis, such readers simultaneously apply lexical and nonlexical letter-to-sound conversion procedures when attempting to read aloud. The lexical procedure is a dictionary look up, and the nonlexical procedure is application of letter-to-sound rules. In other words, the first stage in the process of reading in the dual-route hypothesis is letter identification, and this stage is common in both procedures. After the letters of letter string have been identified, a representation of the letter-string is used to access an orthographic input lexicon, a "sight vocabulary", which is a system of word-specific orthographic representations. If the letter string is indeed one of the words the reader knows by eye, then the word's representation in the orthographic input lexicon is activated. If the letter string is not such a word (e.i. pronounceable pseudo word) it will not have an entry in the orthographic input lexicon, and so this lexical procedure for reading aloud will not succeed in generating a correct response. If the stimulus does activate an orthographic representation in the orthographic input lexicon, this leads in turn to activate the word's phonological representation in a phonological output lexicon. The last stage of the lexical reading procedure is the phoneme activation stage. Any word's entry in the phonological output lexicon has connections to the appropriate phoneme units in the phoneme activation level. So the phonemes representing that word's pronunciation are all activated when the word's entry in the phonological output lexicon is activated. This stage is also used by the nonlexical procedure. Thus both procedures have a common initial stage (the letter identification stage) and a common final stage (the phoneme activation stage). The nonlexical procedure, according to the dual-route concept works as follows. Once the letters of letter string have been identified, a representation of the letter-string is used as input to a rule system that consists of a set of rules specifying how subword-sized letter strings are pronounced. There are different ideas about the sizes of the orthographic and phonological units represented in these nonlexical rules. The units cannot be just letters and phonemes, because there are many examples in English where a phoneme corresponds to a letter sequence ( not just a single letter) and a rule system that consisted solely of single letter to phoneme rules would therefore be grossly inadequate. The orthographic units in the rule system must be graphemes, where "grapheme" means the written representation of a phoneme (in the word "thigh" there are two graphemes, th' and igh'), and here the rules would be grapheme-phoneme rules. It has been proposed that orthographic units larger than the grapheme and phonological units larger than the phoneme are also represented in this nonlexical rule system. Five Questions about Reading There are several differences between the dual-route model and the single-route model (PDP) in accounting for various facts about normal reading and abnormal reading (dyslexia). These models are viewed by 5 major aspects of reading such as exception words, nonwords, lexical decision task, surface dyslexia and phonological dyslexia. Exception Words In the dual-route model, skilled readers read exception words aloud by the lexical route that contains word representations. This model suggests that the general rules of print-to-speech conversion are not helpful in this process. However, many exception words contain unique mapping from letters to sounds. For example, the vowel phoneme in "pint" or the first vowel phoneme in "colonel". In the Seiderberg and McClelland(1989) model they also succeeded in simulating exception-word reading. Non-words In the task of reading some nonwords in the dual-route model, the reader is required a letter sound rule system that has more than one letter rule (e.g. there are different rules for "eigh" and "igh"). This rule system is called the Graphem-Phoneme Correspondence (GPC) system. On the other hand, the PDP model had scored results with two set of nonwords that were 59% and 50% and the results of human subjects were 94% and 84% ( Colheart, Curtis, Atkins, Haller, 1993). Thus, the PDP model cannot read non-words aloud as well as it can read words and the performance of this model is far below a skilled reader performance. Seidenberg and McClelland (1990) have argued that poor performance of nonword in the PDP was due to a major limitation in the implementation of the model. The model had a limited size of the training corpus and was only exposed to 3000 words whereas a skilled reader was exposed to approximately 10 times that number. Lexical Decision Tast The presentation of the stimulus in the lexicon can determine the performance of the visual lexical decision task in the dual-route model. All the words are represented in the internal lexicon and the nonwords are not in the lexicon. Hence, if the word is in the lexicon then the respond is "yes", otherwise the respond is "no". In the PDP model, as the input word enters the orthographic level, and the activation of hidden units are computed, the feedback connections compute a representation back to the orthographic level. The difference between the original input patterns and the feedback loop yielding an orthographic error score. This error indicates the familiarity of the input pattern and the subject assigns a criterion to make a lexical decision. If the error score is less than the criterion score then the respond is "yes" otherwise its "no". It is shown that PDP model achieves an error rate of about 80% and would yield an error rate of over 6% in an average reader . The PDP model cannot explain why people are so accurate at lexical decision. Therefore, the PDP model results strongly indicate that this model is much poorer than people are at the lexical decision task. Surface Dyslexia After brain damage, the reading of some people is affected in the following way: nonword reading is still normal, but many exception words, even quite common ones, are wrongly read. Moreover, the erroneous responses are the ones that would be predicted from applying spelling-sound rules (e.g. reading PINT as if it rhymed with "mint"). This is called surface dyslexia. Two of the clearest cases are patients MP (Bub, Cancelliere and Kertesz, 1985) and KT (McCarthy and Warrington, 1986) whose reading of nonwords was normal in accuracy but misread in exception word with regularization errors. The dual-route model attempts to explain the surface dyslexia by damage to the lexical route and by complete performance of the nonlexical route. In addition, the exception words are often read as a GPC rule specifies. For example, "flood" may be pronounced as if it rhymes with "mood". This error is referred as a regularization error. Thus, the dual-route explanation is that the lexical route for reading is damaged but the nonlexical (rule-based) route intact. Attempts have been made to simulate this by damaging ("lesion") the trained PDP model (e.g., by deleting hidden units). These attempts have not succeeded. It seems highly unlikely that they ever will succeed because the scores of damaged patients (KT & MP) for reading nonwords was 95% correct and between 51- 65% correct in the unlesioned Seidenberg and McClelland model . Phonological Dyslexia After brain damage in some people, reading is affected in the following way: word reading is still good, but nonword reading is very bad. This is called phonological dyslexia. One case was a patient of Funnell (1983). Her patient (WB) could not read any nonwords at all, but achieved scores of around 90% correct in tests of word reading. The dual-route explanation would be that there was abolition of the nonlexical route and sparing of the lexical route. Seidenberg & McClelland appeal to a way (not implemented in their model) of reading from orthography through meaning to phonology. This would fail for a meaningless letter string, so anyone reading solely by such a route can read words but not nonwords. The explanation fails, however, because for phonological dyslexia referred to above (Funnell, 1983), the patient also had a semantic impairment and would have shown semantic confusions in reading aloud if he had been reading semantically. He did not make such confusions. Therefore, Seidenberg and McClelland's reconciliation of phonological dyslexia with their model cannot be correct. Conclusions It has been shown that many important facts about reading discussed by Seidenberg and McClelland (1989) cannot be explained by the PDP model but can be explained by the dual-route model. Some reachers have claimed that the Seidenberg and McClelland (1989) demonstration has not extended our understanding of word reading because the operation of the model is too complex to understand. The SM89 model errors were due to deficiencies in its phonological representation that caused it to produce small deviations from the intended targets (Seidenberg and McClelland, 1990). Plaut el al. (1996) have suggested that an improvement in the phonological representation yield accuracy similar to human accuracy. However, it should be noticed that the SM89 model did exhibit a nonword effect (words that sound like a word like "joak" or "brane" ) although it has no lexical representations. The model performance was better on the nonword than on the random word (such as "fgeh" or "dhes") regarding accuracy and orthographic and phonological error scores. These results occurred because the model picked up on the systematic orthographic and phonological differences between two types of nonwords. In summary, the PDP model presented by SM89 does not offer a tenable explanation of what is known about the aspect of human language. A recent article by Plaut, McClelland, Seidenberg and Patterson (1996) has presented a new architecture for the PDP network. This single-route network consists 105 grapheme units, 100 hidden units and 61 phoneme units. They have found an increase of accuracy in nonword performance to 94-96 %. Furthermore, they have also suggested that the performance of the network directly refutes the claims of the dual-route theorists that the skilled reader requires separation of lexical and nonlexical procedures for conversion of print to sound. The dual-route and single-route models introduced above are still under investigation and the mystery of reading will soon be discovered. Dual-Route 13 References Bub, D., Cancelliere, A. (1985). Whole word and analytic translation of spelling to sound in a nonsemantic reader. In K. E. Patterson, J. C. Marshall, & M. Coltheart (Eds.), Surface dyslexia: Neuropsychological and cognitive studies of phonological reading (pp.15-40). Hillsdale, NJ: Erlbaum. Coltheart, M., Curtis, B., Atkins, P., & Haller, M. (1993). Model of reading aloud: dual- route and parallel- distributed-processing approaches. Psychology Review, 100, 589-608. Funnell, E. (1983). Phonological processes in reading: New evidence from acquired dyslexia. British Journal of Psychology, 74, 159-180. Marshall, J. C., & Newcombe, F. (1973). Patterns of paralexia: A psychological approach. Journal of Psychololinguistic Research, 2, 175-199. McCarthy, R., & Warrington, E. K. (1986). Phonological reading: Phenomena and paradoxes. Cortex, 22, 359-380. McClelland, J. L., & Rumelhart, D. E. (1981). An interactive activation model of context effects in letter perception: Part 1. An account of basic findings. Psychology Review, 88, 375-407. Morton, J., & Patterson, K. (1980). A new attempt at an interpretation, or an attempt at a new interpretation. In M. Colthart, K. Patterson, & J. C. Marshall (Eds.), Deep dyslexia (pp. 91-118). London: Routledge & Kegan Paul. Paap, K. R., & Noel, R. W. (1991). Dual route models of print to sound: Still a good horse race. Psychology Research, 53, 13-24. Plaut, D. C., McCclelland, J. L., Seidenberg, M. S., & Patterson, K.(1996). Understanding Dual-Route 14 normal and impaired word reading: Computational principles in quasi-regular domains. Psychology Review, 103, 56-115. Rumelhart, D. E., & McClelland, J. L. (1982). An interactive activation model of the context effects in letter perception: Part 2. The contextual enhancement effect and some tests and extensions of the model. Psychology Review, 89, 60-94. Rumelhart, D. E., & McClelland, J. L. (1986). On learning the past tenses of English verbs. In D. E. Rumelhart & J. L. McClelland (Eds.), Parallel distribution processing: Vol. 2. Psychological and biological models (pp. 216-271). Cambridge, MA: MIT Press. Seidenberg, M. S., & McClelland, J. L. (1989). A distributed, developmental model of word recognition and naming. Psychology Review, 96, 523-568. Seidenberg, M. S., & McClelland, J. L. (1990).More words but still no lexicon: Reply to Besner el al(1990). Psychology Review, 97, 447-452. Dual-Route 15 ANY COMMENTS OR THOUGHT ?