Learning to assign lexical stress during reading aloud: Corpus, behavioral, and computational investigations

摘要

Models of reading aloud have tended to focus on the mapping between graphemes and phonemes in monosyllables. Critical adaptations of these models are required when considering the reading of polysyllables, which constitute over 90% of word types in English. In this paper, we examined one such adaptation - the process of stress assignment in learning to read. We used a triangulation of corpus, behavioral, and computational modeling techniques. A corpus analysis of age-appropriate reading materials for children aged 5-12 years revealed that the beginnings and endings of English bisyllabic words are highly predictive of stress position, but that endings are more reliable cues in texts for older children. Children aged 5-12 years revealed sensitivity to both the beginnings and endings when reading nonwords, but older children relied more on endings for determining stress assignment. A computational model that learned to map orthography onto stress showed the same age-related trajectory as the children when assigning stress to nonwords. These results reflect the gradual process of learning the statistical properties of written input and provide key constraints for adequate models of reading aloud.