Constraints for computational models of reading: Evidence from learning lexical stress

Models of reading have typically focused on monosyllabic words, consequently lexical stress assignment has been neglected in such models. Yet, determining how and when stress is applied in naming is an important consideration in providing a comprehensive account of reading processing. A corpus analysis of English age-appropriate reading materials between ages 5-12 indicated two properties: a general tendency that bisyllabic words have initial stress; and the orthographic endings of words increase in importance as indicators of stress position. A behavioural study on children aged 5 to 12 reading nonwords showed that children are sensitive to these properties in their responses. Finally, a developmental computational model mapping orthography onto stress showed the same developmental trend, suggesting that probabilistic orthographic information is used for stress assignment, and that the child’s biases may be a consequence of learning the statistical properties of written words. Models of Reading and Stress Assignment Models of reading have typically focused on processing the mapping between the orthography and the phonology of the word. The debate over the architecture of the reading system has converged to the question of whether the mapping between written and spoken forms of words is mediated by a single route or a dual route. The single route perspective assumes that the statistical regularities of the properties of letters and sets of letters mapping onto phonemes and sets of phonemes are learned by a mechanism where all these sources of information are available in parallel (Seidenberg & McClelland, 1989). The alternative dual-route view is that the mapping is served by a system that maps the regularities between graphemes and phonemes, and another, separate system that consults a lexicon for the pronunciation of the whole word (Coltheart et al., 2001; Perry, Ziegler, & Zorzi, 2007). The role of stress assignment provides a test case for constraining the architecture of such models of reading, but it has generally been ignored, presumably due to the limitation of models of reading to monosyllabic words, where stress assignment is trivial. Yet, considering how stress assignment is applied to the word is critical in understanding reading processes. This is because stress must be assigned before the word can be correctly pronounced. In the DRC dual-route model of reading, for regularly pronounced words, the lexical route is accessed after the grapheme-phoneme mapping system is completed. Thus, stress assignment for regular words cannot be stored in the lexicon for this model. Rastle and Coltheart (2000) attempted a solution to this difficulty by deriving a rule-based system for stress assignment that performed a sublexical search through the word for morphemes, and then consulted a database for whether the morpheme carried stress or not. This model was able to classify a large proportion of the lexicon of English, and could generalize to nonwords when they contained identifiable morphemes. Yet, the model performed poorly when no morphemes were present in the nonwords (Kelly, 2004; Seva, Monaghan, & Arciuli, submitted). In contrast, the single-route model of reading predicts that stress position can be determined by the learning of regularities in the mapping from orthography to stress position. Zevin and Joanisse (2000) undertook some pilot modeling that examined the regularities potentially available in the orthography, morphology, and phonology of words for determining stress assignment, and found that such cues were highly predictive of stress position (see also Daelemans, Gillis, & Durieux, 1994, for a word-analogy approach to stress assignment). It is one aim of the current study to determine the extent to which a single-route computational model with no pre-determined structure can effectively learn stress assignment from the regularities in the orthography alone for reading in English. Models of Reading Development A critical aspect of the single-route, statistical regularity approach to modeling of reading is that the mapping between written and spoken forms of words is learned. So, additional constraints on the architecture of the reading system can be provided by developmental studies. A second principal aim of this paper is to determine the patterns of stress assignment during reading, and determine whether a computational model can produce a similar performance trajectory when trained on a realistic reading corpus. Previous studies of the development of stress assignment in children have indicated that morpheme frequency has an influence on stress assignment. Jarmulowicz (2002) found that 9 year old children were more likely to stress multimorphemic words correctly than 7 year olds, but the advantage increased for more frequent morphemes. Protopapas, Gerakaki, and Alexandri (2006) report that Greek children make more errors in stress assignment for nonwords than for words, suggesting that stress may be lexically stored in Greek. The same study also found a default stress assignment on the penultimate syllable in early reading development. Similarly, Gutierrez-Palma and Reyes (2007) found that, for Spanish children aged 7 to 8 years, sensitivity to stress predicted reading ability, and that children with lower reading ability showed a bias toward penultimate stress, the most common stress position in Spanish. However, there is no even distribution of information across words regarding stress position. The distributions of phonemes and letters at the very beginning and ending of words have been found to be particularly reliable indicators of word boundaries (Hockema, 2006) and grammatical category (Arciuli & Monaghan, 2006), as well as stress position (Arciuli & Cupples, 2006, 2007). In the current study, we extend these previous analyses by combining corpus analysis, experimental, and computational modeling to investigate the role of stress in the reading system. In particular, we: 1. Investigate the extent to which the start and the end of words’ orthography can predict stress assignment in the lexicon of English; 2. Determine whether these cues are available in the words that children are exposed to at early stages of reading; 3. Determine the extent to which beginnings and endings of words are utilized by children at different stages of reading; 4. Test whether a computational model instantiating a single route that learns the statistical regularities mapping orthography onto stress position, and without in-built structure, is sensitive to words’ beginnings and endings in the same way as children learning to read. Study 1: Corpus Analyses of Stress