Figure 119

The architecture of the connectionist approach to word reading put forward by Plaut et al. (1996). Copyright © 1986 by the American Psychological Association. Reprinted with permission.

1. The pronunciation of a word or non-word is influenced strongly by consistency based on the pronunciations of all those words similar to it.

2. High-frequency or common words have more influence on pronunciation than do low-frequency or rare words: high-frequency words are encountered more often, and so contribute more to changes in the network.

A successful simulation was based on the architecture shown in Figure 11.9 (hidden units are discussed in Chapter 1). The network learns to pronounce words accurately as connections develop between the visual forms of letters and combinations of letters (grapheme units) and their corresponding phonemes (phoneme units). The network based on this architecture learned by the use of back-propagation, in which the actual outputs or responses of the system are compared against the correct ones. The network received prolonged training with a set of 2998 words. At the end of training, the performance of the network closely resembled that of adult readers:

1. Inconsistent words took longer to name than did consistent words.

2. Rare words took longer to name than common ones.

3. There was an interaction between word frequency and consistency, with the effects of consistency being much greater for rare words than for common ones.

4. The network pronounced over 90% of non-words "correctly", which is comparable with the performance of adult readers; this finding is especially impressive, because the network received no direct training on non-words.

The simulation did not take semantic information into account. However, Plaut et al. (1996, p. 95) argued that, "to the extent that the semantic pathway has learned to derive the meaning and pronunciation of a word, it affords additional input to the phoneme units, pushing them toward their correct activations." They expanded their network model to include semantic information, assuming that such information has more impact on high-frequency words. A network based on this assumption learned to read regular and exception words much faster than a network lacking semantic information.

Surface dyslexia and phonological dyslexia

Plaut et al. (1996, p. 95) used the notion that semantic information contributes to reading performance to advance the following theory of surface dyslexia: "Partial semantic support for word pronunciations alleviates the need for the phonological pathway to master all words such that, when the support is eliminated by brain damage, the surface dyslexic reading pattern emerges."

Plaut et al. (1996) tested this theory by making "lesions" to the network to reduce or eliminate the contribution from semantics. As has been found with surface dyslexics, the network's reading performance was very good on regular high- and low-frequency words and on non-words, worse on irregular high-frequency words, and worst on irregular low-frequency words.

Further support for the theory comes from the study of patients suffering with Alzheimer's disease, which involves progressive dementia or loss of mental powers. Such patients typically have similar reading performance to surface dyslexics, and the severity of the reading impairment is correlated with the extent of semantic deterioration (Patterson, Graham, & Hodges, 1994).

What about phonological dyslexia? Plaut et al. (1996, p. 99) only considered this disorder in general terms: "In the limit of a complete lesion between orthography and phonology, nonword reading would be impossible. Thus, a lesion to the network that severely impaired the phonological pathway while leaving the contribution of semantics to phonology (relatively) intact would replicate the basic characteristics of phonological dyslexia."

Deep dyslexia

Plaut and Shallice (1993) proposed a network similar to the ones later put forward by Plaut et al. (1996) in order to understand deep dyslexia. This network has four key properties:

• Similar patterns of activation represent similar words in the orthographic and semantic domains.

• Learning alters the strengths of the connections between word spellings and meanings.

• The initial pattern of semantic activity produced by a visually presented word moves towards (or is attracted by) the pattern of the nearest known meaning; this is known as the operation of attractors.

• The semantic representations of most high-imageability words contain many more features than those of low-imageability words.

Plaut and Shallice (1993) studied the consequences of damage to the network. They found that virtually all the main symptoms of deep dyslexia could be simulated. The only important symptom that did not emerge from damage to the connectionist network was impaired writing performance.

Plaut and Shallice's (1993) theory is successful in two main ways. First, it predicts about a dozen symptoms of deep dyslexia from only a few theoretical assumptions. Second, the theory is explicit, in that the processes involved were specified in detail before the simulations proceeded.


The connectionist approach of Plaut et al. (1996) and Plaut and Shallice (1993) has various advantages over the traditional dual-route approach. First, the apparent non-independence between the two routes of the dual-route approach poses no problems for the connectionist approach, which assumes that the processing system is interactive. Second, the connectionist approach (unlike the dual-route approach) does not draw a sharp distinction between regular and irregular words. This is an advantage, because the evidence does not support the notion of a rigid distinction, and there is no agreement on the rules determining which words belong to each category. Third, the evidence suggests that speed and accuracy of word and non-word pronunciation depend more on consistency than on regularity. Fourth, Plaut et al. (1996) predicted correctly that damage to the semantic system can affect reading performance in surface dyslexics. In contrast, "dualroute theories that include a lexical, nonsemantic pathway (e.g.,. Coltheart et al., 1993) predict that selective semantic damage should never affect [word] naming accuracy" (Plaut et al, 1996, p. 102).

In spite of its successes, Plaut et al.'s (1996) connectionist approach to the reading of words has various limitations. First, as Plaut et al. (1996, p. 108) admitted, "the nature of processing within the semantic pathway has been characterised in only the coarsest way." Second, the approach has only been tested with one-syllabled words, and clearly needs additional testing with multisyllabled words. Third, the approach provides only a sketchy account of some key issues, such as the nature of the impairment in phonological dyslexia. Fourth, as Ellis and Humphreys (1999, p. 340) pointed out, "evidence suggests that.short words can be named without recourse to sequential letter processing., so the validity of the sequential read-out mechanism might be questioned." Fifth, as Ellis and Humphreys (1999, p. 537) argued:

Plaut et al. proposed that their second route was semantic in nature; however, they did not attempt to represent semantic knowledge in any plausible manner.Hence it may equally be argued that this route was lexical rather than semantic and that, in lesioning the model, they simulate operation of a non-lexical route in isolation from lexical reading processes—much as argued in dual-route accounts of reading.

Phonological theory of reading

According to the dual-route model, the reading performance of normal individuals is generally little affected by phonological coding. This is because reading via the indirect route (grapheme-phoneme conversion) tends to be much slower than reading via the direct route. Frost (1998) argued that phonological coding is much more important in reading than is implied by the dual-route model. Frost (1998, p. 76) put forward a phonological model of reading based on the following assumption:

A phonological representation is a necessary product of processing printed words, even though the explicit pronunciation of their phonological structure is not required. Thus, the strong phonological model would predict that phonological processing will be mandatory [obligatory], perhaps automatic.

Two predictions that follow from the phonological model, and that seem inconsistent with the dual-route model, are as follows:

1. Phonological coding will occur even when it impairs performance.

2. Some phonological coding occurs rapidly when a word is presented visually.

A study supporting prediction (1) was reported by Tzelgov et al. (1996). It was based on the Stroop effect (described earlier), in which naming the colours in which words are printed is slowed when the words themselves are different colour names. The participants in the study were English-Hebrew bilinguals, and in the key condition they had to name the colours of non-words in one of the two languages. Each non-word had an unfamiliar printed form, but its phonological translation was a colour name in the other language. Tzelgov et al. (1996) obtained a strong Stroop effect with these non-words. Thus, the participants engaged in phonological coding of the non-words even though it was disadvantageous to do so.

A study supporting prediction (2) was reported by Berent and Perfetti (1995). They used a backward masking technique involving the following stages: (1) a target word was presented very briefly; (2) the target word was masked by a pseudo-word that was presented very briefly; (3) a pattern mask was presented; and (4) the participants wrote down what they had seen. The main measure was the proportion of trials on which the target word was detected. The key finding was that target detection was higher when the non-words were phonemically similar to the target words than when they were graphemically similar. The implication is that basic phonological coding can occur within about 60 milliseconds of the presentation of a word.


There is reasonable support for the central assumption of the phonological model that phonological coding typically occurs during the processing of printed words. There is evidence for phonological coding even when it disrupts performance. Overall, it seems that phonological coding occurs more often and more rapidly than is assumed by the rival dual-route model.

On the negative side, the phonological model does not provide an explicit and detailed account of the processes involved in reading. In addition, it seems more applicable to reading in normals than in dyslexics. For example, phonological dyslexics have great difficulties with phonological coding, but are reasonably good at reading familiar words. This is puzzling if one assumes that phonological coding is of major importance in reading. As Frost (1998, p. 93) admitted, "Evidence that is.damaging to the strong phonological model comes from phonological dyslexia."

Section summary

There is still some theoretical controversy about the processes involved in reading individual words, in large part because of the inconclusiveness of much of the evidence. There are two main reasons for this inconclusiveness. First, the findings that are obtained depend on the methods that are employed. For example, as Harley (1995, p. 82) pointed out, "Lexical decision and naming do not always give the same results.. .the differences probably arise because while naming times are a relatively pure measure of the time it takes for automatic access to the lexicon, lexical decision times may include a substantial amount of attentional processing." Second, methods such as lexical decision and naming provide measures of the total time taken to complete a series of processes, but are fairly uninformative about those underlying processes. As Frost (1998, p. 95) suggested, "Instead of setting one's experimental camera at the finish line of the cognitive events, one should aim at filming their on-line, step-by-step development."

Stop Anxiety Attacks

Stop Anxiety Attacks

Here's How You Could End Anxiety and Panic Attacks For Good Prevent Anxiety in Your Golden Years Without Harmful Prescription Drugs. If You Give Me 15 minutes, I Will Show You a Breakthrough That Will Change The Way You Think About Anxiety and Panic Attacks Forever! If you are still suffering because your doctor can't help you, here's some great news...!

Get My Free Ebook

Post a comment