The Science of Reading. Группа авторов
Читать онлайн книгу.
however. Instead of using a set of conversion rules, CDP+ includes a neural network to translate graphemes into phonemes (Figure 4.3). A neural network has two or more layers of nodes connected to each other. In CDP+, the input layer consists of graphemes and the output layer of phonemes. The model is trained in such a way that it learns to activate the most likely phonological output when presented with the graphemic input of a monosyllabic English word.
This neural network solves the problems with the assembled phonology route in the DRC model. First, there is no need to define default conversions (set of rules). Instead, the network learns to activate the best fitting output on the basis of the input units activated. It does so by changing the weights of the connections between the units. When there is a consistent correspondence between input patterns and output patterns, the model rapidly catches the correspondence and returns the correct output when given the input. Thus, the neural network rapidly learns to activate the output pronunciation /‐id/ for input words ending in –eed, as all monosyllabic words ending in these letters have the same pronunciation.
Second, when there are inconsistencies in the stimulus set (like the pronunciation of the end letters –ead), the network learns to activate the different pronunciations in line with the input. So, both the phonological forms /‐id/ and /‐εd/ are activated when the model is presented with a visual English word ending–ead. The degree of activation depends on the frequencies of the words pronounced as /‐id/ (read, bead, lead) relative to the frequencies of the words pronounced as /‐εd/ (head, lead, bread, spread).
Figure 4.3 The CDP+ model of visual word naming
(Perry et al., 2007/With permission of American Psychological Association).
The route for addressed phonology is the same as in DRC, but a neural network replaces the rule system in the assembled route.
The CDP+ model has two more advantages over the DRC model. First, it more naturally accounts for the fact that not all people name pseudowords the same (Pritchard et al., 2012). Indeed, not everyone pronounces the pseudowords nead as /nid/, which they should if they follow strict grapheme‐phoneme correspondence rules as in the DRC. Some people pronounce nead as /nεd/, or give different pronunciations on different occasions. Such differences can be understood as individual differences in the learning of the neural network, or in the activation dynamics of the network. Second, the CDP+ model has more scope for interactions between the graphemes and phonemes in a word. Because the conversion occurs in parallel across the entire word, the model may learn that the pronunciation of the vowel in subtle ways depends on the consonants before and/or after the vowel. So, the model may end up naming the pseudoword glive as rhyming with give and the pseudoword brive as rhyming with drive because of the larger overlap between glive and give and brive and drive. Or the model may pronounce glive as rhyming with drive if it has shortly before been presented with the word drive. Similar dynamics are seen in people (Pritchard et al., 2012).
Triangle model
Both the DRC and CDP+ models stay close to verbal descriptions of the processes involved in word reading from which they were derived. For instance, they assume the existence of orthographic and phonological word nodes, representing written and spoken words, respectively. Activation flows between meaningful units, which become more or less likely to be selected for conscious output.
Neural networks allow lexical processing to be simulated in more radical ways. Rather than dedicated word nodes that include disambiguating information, information about words can be captured by patterns of activation distributed across network units and shared with other words. The new approach was called the distributed approach, in contrast with the localist approach of DRC and DCP+. The first model presented along these lines was by Seidenberg and McClelland (1989; see Seidenberg et al., this volume). Here, we discuss a more elaborate version proposed by Harm and Seidenberg (2004), shown in Figure 4.4.
The starting point of the Triangle model is that visual input can be described by a limited number of features (graphemes) and phonological output by a limited number of features (phonemes), similar to DRC and CDP+. Importantly, however, the Triangle model also includes word meaning (semantics). This allows print‐to‐meaning activation to be simulated, not just reading aloud; both DRC and CDP+ deal only with word naming. Including meaning helps decrease ambiguity at the level of grapheme‐phoneme mappings. Take the words head and knead. A model only consisting of graphemes and phonemes will not be able to derive the correct phonology of both words and distinguish them from their homophones heed and need. An additional route with addressed phonology is needed, as we have seen for DRC and CDP+. However, the situation changes when meaning is added. The meaning of the written word head is very different from the meaning of heed. Similarly, the meaning of the spoken word /hεd/ is different from the meaning of /hid/. In addition, the meanings of head and /hεd/ are very much the same, as are the meaning of heed and /hid/. So, the written word head can activate the phonology /hεd/ not only by direct orthography‐phonology connections, but also indirectly, via semantic information.
Figure 4.4 The Triangle model
(Harm & Seidenberg, 2004/With permission of American Psychological Association)
showing full interactions between three types of information: orthography, phonology, and semantics.
The orthographic layer in Harm and Seidenberg’s (2004) model required 111 units to represent the various graphemes that can occur at different positions in monosyllabic English words. The phonological layer included 200 units to represent the phoneme features at various positions in words. Finally, the network included 1,989 semantic features (e.g., is a person, is a piece of furniture, involves movement). If the meaning of a word contained the feature, the unit was set to 1, otherwise it remained at 0.
The model was taught in two steps. First, training was limited to the correspondences between phonemes and semantic features, akin to native language acquisition. In a second stage, orthography was added, so that the connections from print to phonology and print to meaning could be learned, just as children learn to read after a number of years of spoken language experience.
Following training, Harm and Seidenberg (2004) noticed that their model produced the correct pronunciation of 99.2% of the words without requiring a route with word nodes (addressed phonology). The model further simulated all effects in visual word recognition captured by DRC (and later CDP+). The model was recently used successfully by Chang et al. (2020) to investigate the effects of spoken word knowledge and different reading instructions on word reading.
The Triangle model makes another interesting prediction. Because the three types of representation (orthography, phonology, semantics) fully interact, orthography activates phonology, but phonology also activates orthography. Learning to read should therefore bring about changes in how phonology is represented. This is consistent with evidence from illiterate populations. For example, Morais et al. (1987) reported that illiterate people perform less well than peers on phonological awareness tasks such as taking away the first sound of a spoken word. This finding and other data indicate that knowledge of the form of spoken words is less detailed and less stable in illiterate people, and people with reduced reading practice (Huettig et al., 2018).