Abstract:
The study of word reading and recognition has been strongly influenced by
computational cognitive modeling. These models facilitate theorizing about the
mechanisms that underlie word reading and recognition (e.g., Morton, 1970; McClelland
& Rumelhart, 1981; Seidenberg & McClelland, 1989; Plaut, McClelland, Seidenberg,
and Patterson, 1996; Harm & Seidenberg, 1999; Coltheart, Rastle, Perry, Langdon, &
Ziegler, 2001; Perry, Ziegler, & Zorzi, 2007). However, the preeminent models in this
field only process monosyllabic words. This results from difficulties inherent in
representing the orthography and phonology of multisyllabic words. To address this
issue Sibley, Kello, Plaut, & Elman (in Press) created a connectionist architecture named
the Sequence Encoder. The present work utilizes representations from a Sequence
Encoder to build models that address an order of magnitude more data than previous
models.
A second goal of this work is to explore the possibility of hypothesizing fewer
mechanisms in models of the reading system. The three preeminent models of reading all
implement two distinct pathways from orthography to phonology. A sublexical route
encodes statistical relationships between letters and phonemes, while a lexical route
encodes whole word information. This dissertation explores whether each of these
pathways are necessary for word reading and recognition. We present three models
trained on 60,000 mono- and multisyllabic English words. Simulation 1 maps from
orthography to phonology using a single sublexical route. It demonstrates substantial
naming capacities, but is incapable of addressing lexical decision data. Simulation 2
utilizes only a lexical route, where reading is achieved by an inductive process that
utilizes whole word information stored in a lexicon. This model addresses naming and
lexical decision data on an unprecedented scale. Simulation 3 integrates sublexical and
lexical routes from the previous models, but exhibits negligible capacities beyond
Simulation 2. Finally, we examine each simulation’s sensitivity to stimuli characteristics
that impact behavioral latencies. Our simulations mimicked the effects of all examined
variables on participants’ latencies. These simulations demonstrate that models can be
scaled up without incorporating new mechanisms specifically to address phenomena of
multisyllabic word reading, such as stress assignment. We conclude that a single lexical
pathway from orthography to phonology is sufficient to simulate word reading and
recognition.