An empirical and computational investigation of generalisation in nonword reading

Reading aloud new words requires an ability to generalise linguistic knowledge acquired via experience in reading. Yet, the exact cognitive mechanisms by which this happens are still unknown. In this PhD project, I investigated generalisation in reading aloud in English, focusing on pronunciations assigned to nonwords by skilled readers. This work consisted of computational, empirical and methodological investigations.

Firstly, I developed a new, symbolic model of reading aloud – the Weighted Segments Pronunciation (WSP) model. This model converts letter strings into speech sounds based on different statistical properties of the writing system, across varying sized print-to-sound correspondences. The WSP model simulated central tendencies in human nonword reading responses comparably to prominent computational models of reading (the DRC and the CDP++ models). Furthermore, the WSP model showed some promise in simulating variability in nonword reading, and the present work illustrated some ways to evaluate models that produce variable output. Issues in the performance of the WSP model were identified and several avenues for improving the model were discussed.

Secondly, I conducted two empirical studies, aiming to clarify which statistical properties of the writing system skilled readers are sensitive to. Both type and token frequency measures of print-to-sound correspondences were shown to be influential in nonword processing, with likely larger influence of type frequency.

Thirdly, I compared two methods of collecting information about how skilled readers process nonwords: the traditional nonword naming method (where participants read aloud nonwords) and a relatively new nonword rating method (where participants give acceptability ratings to pronunciations assigned to nonwords). These comparisons revealed that the rating method is a feasible alternative to the naming method, and it may reveal aspects about skilled readers' knowledge of print-to-sound correspondences that the nonword naming method cannot. These findings bear relevance to future empirical investigations and theory development of reading aloud.