Incorporating learning mechanisms into the Dual-Route Cascaded (DRC) model of reading aloud and word recognition
thesisposted on 28.03.2022, 10:41 by Stephen C. Pritchard
The dual-route cascaded (DRC) model of reading aloud and word recognition has achieved considerable success. Despite this, it has faced ongoing criticism for being a static model of skilled reading that does not describe reading acquisition. This PhD research focused on incorporating learning mechanisms into DRC. Work was divided into two broad areas: orthographic learning within DRC's lexical route, and grapheme-phoneme correspondence learning in DRC's sublexical route. To model orthographic learning, a “learning DRC” (L-DRC) was created. L-DRC provides a computational account of the self-teaching hypothesis, and in accordance with this, models orthographic learning as being self-driven via phonological recoding, with context supporting irregular word learning. L-DRC effectively modelled self-teaching and orthographic learning, and suggested mechanisms for the difficulties children may face when self-teaching difficult words like potentiophones or heterophonic homographs. To model sublexical learning, a grapheme-phoneme correspondence (GPC) Learning Model was created and tested. This model effectively demonstrated GPC learning, especially when trained on an input corpus limited to mono-morphemic words presented once each. However, it experienced difficulties when trained on more realistic input corpuses. The model's performance suggests that sublexical route learning is sensitive to morpheme structure, and to type-based rather than token-based features in written material. The investigation of sublexical-route learning was preceded by a comparison of the sublexical routes of two competing dual-route models, the DRC and connectionist dualprocess (CDP+) models. These were assessed against new empirical data on how people pronounce nonwords. While neither model provided a good match to the human data, DRC performed significantly better than CDP+, or its successor CDP++.