A Connectionist Model of Impaired Reading and Reading Interventions Michael W. Harm Center for the Neural Basis of Cognition Carnegie Mellon University 115 MI, 4400 Fifth Avenue, Pittsburgh, PA 15213. (412) 268-3922 mharm@cnbc.cmu.edu Bruce McCandliss Sackler Institute for Developmental Psychobiology Weill Medical College of Cornell University Mark S. Seidenberg Neuroscience Program University of Southern California ABSTRACT: We investigated the impact of remediation techniques within the Harm and Seidenberg (1999) model of developmental phonological dyslexia. A phonologically impaired model demonstrates patterns of impairment found in reading disabled children. This model was then subjected to an remediation scheme targeting spelling to sound correspondences, using materials from Isabel Beck's Word Building intervention (McCandliss et al., 1999). This produced significant improvements in the model's nonword reading. Other interventions targeted at improvements in phonological processing had little effect. The research shows that the effectiveness of different interventions can be explained by looking at the internal structure of the computational model. SUMMARY: We used the Harm and Seidenberg (1999) model of word recognition to explore remediation techniques in developmental dyslexia. The model maps orthographic onto phonological representations. The phonological component is represented as an attractor network which can be damaged. Previously reported simulations (Harm and Seidenberg 1999, Psychological Review 106(3)) demonstrated that damage to the phonological component of the model gives rise to a pattern of reading impairments quite similar to those seen in developmental phonological dyslexia. This is because impairments in phonological representations cause the intermediate representations that map orthography to phonology to become less componential, and more wholistic. For the current study, we attempted to remediate the phonologically impaired model using materials from a successful remediation scheme reported by McCandliss et al (1999). In this scheme, based on Isabel Beck's Word Building technique, lessons are constructed in which there is strong overlap between subsequent training items (e.g., BAT-CAT-CAB). When errors occur, the word is broken into its subsequent parts to force the child's attention to the subword units. This procedure is thought to increases pressure on mapping grapheme-phoneme correspondences by changing only a single grapheme at each successive trial. The model was subjected to remediation training using the same items and same training paradigm. The model exhibited improvements in nonword reading levels comparable to those observed in the McCandliss et al. (1999) study. The model was also subjected to interventions targeting solely the phonological representation. These took the form of auditory training on subword units, and a more dramatic test involving the removal of the phonological damage which caused the reading impairment. It was found that such interventions were only effective when introduced extremely early in training; once the model had been exposed to even a small amount of reading instruction such remediations had little or no effect. This is because the model had already begun to form wholistic representations mediating orthography and phonology; repairing phonology exerted no pressure on the network to undo these poor reading-specific representations. Finally, an attempt to remediate the model by training it on nonwords was attempted. This produced no benefit, because the nonwords were still represented by the model in a wholistic manner, rather than as combinations of regular parts. The model thus demonstrates why remediations targeted solely at phonology typically produce inferior results to interventions that target both phonology and letter-sounds correspondences, and why even though phonological impairments can cause poor reading, remediation schemes need to target the spelling to sound correspondences to overcome this impairment. It provides further insights into why attention to the subword components of visual words is crucial to overcome the poor representations that result from phonological impairments.